WO2026013205A2 - Clustering parameter data to identify behavior pattern of industrial equipment - Google Patents

Clustering parameter data to identify behavior pattern of industrial equipment

Info

Publication number
WO2026013205A2
WO2026013205A2 PCT/EP2025/069785 EP2025069785W WO2026013205A2 WO 2026013205 A2 WO2026013205 A2 WO 2026013205A2 EP 2025069785 W EP2025069785 W EP 2025069785W WO 2026013205 A2 WO2026013205 A2 WO 2026013205A2
Authority
WO
WIPO (PCT)
Prior art keywords
parameter data
components
data
computer
cluster
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
PCT/EP2025/069785
Other languages
French (fr)
Other versions
WO2026013205A3 (en
Inventor
Cédric SCHOCKAERT
Henri HOYEZ
Marcos TEIXEIRA
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Paul Wurth SA
Original Assignee
Paul Wurth SA
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Paul Wurth SA filed Critical Paul Wurth SA
Publication of WO2026013205A2 publication Critical patent/WO2026013205A2/en
Publication of WO2026013205A3 publication Critical patent/WO2026013205A3/en
Pending legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
    • G05B23/0221—Preprocessing measurements, e.g. data collection rate adjustment; Standardization of measurements; Time series or signal analysis, e.g. frequency analysis or wavelets; Trustworthiness of measurements; Indexes therefor; Measurements using easily measured parameters to estimate parameters difficult to measure; Virtual sensor creation; De-noising; Sensor fusion; Unconventional preprocessing inherently present in specific fault detection methods like PCA-based methods
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B23/00—Testing or monitoring of control systems or parts thereof
    • G05B23/02—Electric testing or monitoring
    • G05B23/0205—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults
    • G05B23/0218—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterised by the fault detection method dealing with either existing or incipient faults
    • G05B23/0224—Process history based detection method, e.g. whereby history implies the availability of large amounts of data
    • G05B23/024—Quantitative history assessment, e.g. mathematical relationships between available data; Functions therefor; Principal component analysis [PCA]; Partial least square [PLS]; Statistical classifiers, e.g. Bayesian networks, linear regression or correlation analysis; Neural networks

Definitions

  • the disclosure relates to industrial production processes, and more in particularly, the disclosure relates to computer systems, methods and computerprogram products to identify behavior patterns of industrial equipment that performs the industrial production process.
  • Industrial equipment such as industrial machines perform processes.
  • the equipment has components in a particular arrangement so such that the process can be performed.
  • the components would be arranged in a chain to sequentially forward intermediate products from a predecessor component to a successor component.
  • Parameter data can be obtained from sensors that measure phenomena that are relevant to the process, such as temperature, pressure, material properties and others. Parameter data can also be obtained from actuators. For example, a heater has a state (i.e., "on” or “off"), a rated power (e.g., a kW value number given by the manufacturer), an actual power consumption (e.g., a kW value that is measured); a valve may be "open” or “closed” (to influence the pressure in a vessel), and so on. Parameter data can also be obtained from shop-floor systems (e.g., regarding the amounts of materials etc.).
  • Parameter data can usually be related to particular components, although a sub-set of parameter data can be related to multiple components. For example, precipitations usually increase the air humidity in a factory as a whole, but when the Sun returns, the humidity decreases again.
  • the skilled person is also able to aggregate a sub-plurality of parameter data to one or more particular indicators.
  • an indicator can be a quality indicator.
  • Quality indicators can be related, for example, (a) to the result of the process (such as to the product that results from a production process), (b) to the performance of the equipment (such as energy consumption, use of auxiliary materials such as oil), the time it took to perform the process, and (d) others.
  • the time-series have durations that can correspond to process instances, and the durations of these instances change from instance to instance.
  • the skilled person who evaluates data can harmonize instance duration with techniques such as dynamic time warping (DTW).
  • DTW dynamic time warping
  • machine-learning (ML) tools allow the identification of parameter deviations, the prediction of the mentioned quality indicators, and the like. ML tools may also indicate the absence or presence of equipment anomaly at present, or may predict anomaly events for the future.
  • the definition of an "anomaly" is usually adapted to particular processes. For example, an unexpected stop of a production line would certainly qualify as anomaly, but for some processes a relatively minor deviation in the product quality would be regarded as anomaly as well.
  • the ML tools are trained with training data that is usually a collection of historical data from historical process instances (i.e., past usage data). During operation of the equipment, the tools process data of the current process.
  • the ML tools perform algorithms to process the parameter data at the tool input and to derive a function at the tool output.
  • the function can be an estimation of equipment failure in the future, of a particular property of the product at the end of the process, the detection or anticipation of an anomaly, or others.
  • the function is a (mathematical) function of a plurality of parameter values.
  • this (mathematical) function is usually unknown by explicit formulas, the ML tools are implemented by neural networks or the like.
  • a component may be relatively new, in the sense to have been manufactured recently, to have been installed to the equipment recently, etc. As a consequence, historical parameter data to train the ML tool may not (yet) be available, or may not be available in sufficient amount.
  • Parameter data in the multi-variate time-series may become incomplete, due to sensor failure or similar events. If a sensor - that is a sub-component to a component - stops providing data, the process at the equipment may continue. However, the ML tool might not be prepared for such situations, and it might derive an incorrect conclusion.
  • Parameter data may be disturbed, because of effects that are colloquially attributed to be noise.
  • a component may vibrate and a sensor might provide not only the sensor-specific signal but also a vibration specific signal.
  • the ML tool might process noisy signals and clean signals differently.
  • the ML tool has already been trained by noisy data, because the skilled person occasionally augments training data by noise.
  • ML tools that rely on data only are strong in identifying correlations, such as between two or more variates in a multivariate time-series. Identifying a correlation is that same as identifying a cause-effect relation.
  • US 2020/0226490 Al discusses a machine tool by illustrating infrastructure components that provide real-time machine metrics, by explaining computer modules that process the metrics and that identify anomalies and corresponding remedial actions. The actions can be applied to the infrastructure components.
  • Industrial equipment performs an industrial process by components (i.e., equipment components), and a computer executes a method to identify a behavior pattern of the equipment, for the equipment as a whole or for individual components.
  • components i.e., equipment components
  • a computer executes a method to identify a behavior pattern of the equipment, for the equipment as a whole or for individual components.
  • the computer-implemented method also has further aspects or implies further activities such as clustering the variates of parameter data according to static or dynamic relations between the components or process stages.
  • further activities such as clustering the variates of parameter data according to static or dynamic relations between the components or process stages.
  • static and dynamic are convenient labels that indicate knowledge of the relations in view of the operation phase.
  • Clustering can involve representing the components (or stages) by so-called virtual components (or “virtualizing” in short). In principle, clustering can be performed during all three phases.
  • the computer executes the method to identify the behavior pattern in the operation phase.
  • the computer receives parameter data as multi-variate timeseries from data providers - such as sensors - of the industrial equipment.
  • the computer processes the parameter data by two computing modules, and provides behavior data.
  • the first computing module can receive at least a sub-set of the parameter data as original parameter data and turns that sub-set into adapted parameter data.
  • the first module is also called the "adaptor module”. Clustering plays a role for this adaptor module.
  • the parameter data is adapted by applying cluster values (as weights to original parameter data).
  • the second computing module can receive the adapted parameter data and performs behavior prediction to provide the behavior data.
  • the second computing module can also receive parameter data that by-passes the first computing module.
  • the second module is also called the "predictor module”.
  • the predictor module is implemented as a pre-trained machine-learning (ML) tool.
  • the noise pattern can be determined from one component and that noise pattern can be used to cancel the noise from the data of the other components in that cluster.
  • components may influence each other only temporarily. If such a relation is known, the adaptor can apply the appropriate weights. For example, to have less weights to parameter data during certain operations.
  • clustering the parameter data is performed by representing the components by component-representations and representing process stages by stage-representations.
  • the steps clustering and assigning are performed in repetitions, when a dynamic relation between the components is changed due to changed parameter data that are being received.
  • a computer-implemented method is provided to identify a behavior pattern of industrial equipment that has components, and that performs an industrial process.
  • the components of the industrial equipment are associated with data providers that correspond to variates in multi-variate original parameter data.
  • Relation between the components belong to four relation groups that are differentiated according to the following criteria: the presence of interactions between the components, with the interactions being defined by the industrial process, or the absence of such interactions that are process-defined; and the presence of changes in these interactions after the industrial equipment has started to perform the industrial process, or the absence of such changes (in these interactions).
  • the computer assigns cluster values to the data clusters.
  • the computer applies the cluster values as weights to the original parameter data to obtain adapted parameter data.
  • a computer program product that - when loaded into a memory of a computer and being executed by at least one processor of the computer causes the computer to perform the steps of the methods.
  • an arrangement comprising industrial equipment and a computer.
  • the equipment can be an industrial machine and the equipment is adapted to provide parameter data.
  • the computer with an adaptor module and with a predictor module is adapted to process parameter data by performing the computer-implemented method.
  • FIG. 1 illustrates a writing convention for parameter data that are being clustered
  • FIG. 3 illustrates a component, a component-representation, and a componentspecific predictor
  • FIG. 4 illustrates clustering by way of examples
  • FIG. 5 illustrates component-representations with connecting lines that symbolize clusters
  • FIG. 6 illustrates a flow-chart diagram of a computer-implemented method to identify a behavior pattern of industrial equipment that performs an industrial process by components
  • FIG. 10 illustrates a diagram with two clustering dimensions, each with two alternatives.
  • Clustering is mainly driven by rules, not by data. In that sense, there is a balance between data- driven computation (cf. step 450) and rule-driven computation (cf. steps 420, 430, 440). Correlations may arise from confounding data or from biasing data, but for clustering, such correlations would be ignored.
  • This pre-processing involves certain activities at a still earlier points in time (pre-pre- processing, cf. FIG. 7 at item 700) and/or during processing.
  • FIG. 1 illustrates a writing convention for original parameter data [X] that are being clustered by a computer (as well as for adapted parameter data [Y] that result from clustering).
  • [X] [XI, X2, ..., Xv, ..., XV], and
  • [Y] [Yl, Y2, ..., Yv, ..., YV]
  • Multi-variate parameter data [X] and [Y] can be input to processing tools, such as to neural networks (cf. predictor module 500 in FIG. 2).
  • processing tools such as to neural networks (cf. predictor module 500 in FIG. 2).
  • original parameter data [X] shown on the left side
  • adapted parameter data [Y] shown on the right side
  • Pre-processing (or “adapting") comprises
  • the example in FIG. 1 assumes to have two clusters, with values Cl and C2.
  • the network receives original Xv and Wv, wherein Wv is an indicator that the neural network has to process Xv with its original value.
  • the network receives original parameter data [X] with a weight set [W], [W] can indicate to use (at least a particular part of) the neural network that has been trained with certain training (for example to accommodate missing historical data).
  • the weights are cluster-specific, and in the example of FIG. 1, there should be two clusters and hence two cluster values Cl and C2.
  • XI, X3 and X4 would have been clustered to a first cluster with the cluster value Cl
  • X2 and X5 would have been clustered to a second cluster with the cluster value C2.
  • Cluster vector [C] that expresses the separation only would, for example, be a vector with a cluster identifier to indicate cluster membership of a variate, and with "0" otherwise, for example [1, 2, 1, 1, 2, 0, 0],
  • variates XI, X3, X4 would be clustered to cluster 1
  • variates X2 and X5 would be clustered to cluster 2.
  • Variates X6 and X7 would not be belong to any cluster.
  • Cluster vector [C] that expresses the cluster values would comprises the values (such as Cl and C2) and would be "zero” or “neural” otherwise, for example [Cl, C2, Cl, Cl, C2, neutral, neutral].
  • the skilled person can use different notations.
  • clustering (as computation activity) depends on relations that are static or dynamic.
  • cluster vector [C] stands for a cluster that is static (i.e., that remains while the equipment performs the process)
  • time-series cluster vector ⁇ [C] ⁇ stands more for a cluster that is more dynamic (i.e., that may change during the execution of the process).
  • Such and other changes in [C] can trigger the computer to execute the method again (cf. FIG. 7), or to repeat at least the steps from clustering (step 420 in FIG. 7).
  • parameter data change over time in many situations, some parameter data is relatively static. Using parameter data and cluster data in time-series is therefore not mandatory.
  • Two or more clusters may have some variates in common.
  • stage differentiates processing time intervals for industrial processes (first stage, second stage and so on).
  • module stands for the implementations of a computing function by one ore more computers.
  • component and “equipment” stand for hardware.
  • representation stands for a data-structure by that a computer represents a component, a stage, or the equipment (as a whole).
  • clustering stands for a method step that is performed by a computer when data has been received.
  • Cluster preparation can be computer-implemented (cf. FIG. 7) and can be human-implemented.
  • the computer can process a computer-readable equipment-descriptor document that has pointers to clusters, at least implicitly.
  • the computer can parse this document (cf. step 710 in FIG. 7).
  • descriptor documents can be, for example,
  • Such diagrams, lists or plans can include verbal descriptors, such as short phrases. It can be assumed that the computer understands their semantics. For example, the phrase "process connected to” implicitly induces a relation that two or more components that participate in the industrial process are "mechanically connected to (each other)". The reverse is however not implicit (mechanical connection not necessarily being a process connection).
  • clusters and rules for clustering can be prepared by human experts.
  • Such an expert is familiar with the industrial domain (e.g., the particular equipment performing a particular process, as in FIG. 2).
  • the domain expert could establish a database entry with a list of components and could annotated them to be "main" or "aux" components.
  • the computer can then receive input from a domain expert (cf. step 720 in FIG. 7).
  • FIG. 2 illustrates a block diagram of industrial equipment 100 with components 110- n, 120-k, and also illustrates computer 200/500 with adaptor module 200 and with predictor module 500.
  • the industrial process can be, for example, a material processing process by that raw material (left side of the figure) is being processed via intermediate material to a final product (right side).
  • auxiliary component 120-k should provide oil or other lubricants to main components 110-2 and 110-n.
  • Index n is the "main index” that identifies the main components
  • index k is the "auxiliary index” that identifies the auxiliary components.
  • the industrial process can be differentiated into a main process (bold arrows, main components) and one or more auxiliary processes (thin arrows, auxiliary components).
  • auxiliary component 120-k may run out of oil, this is not problematic for the main process, at least for some time. On the long run, oil must be provided, otherwise a component would be damaged.
  • Main and auxiliary processes can be differentiated to have different process stages (or process steps, or sub-processes).
  • One and the same component can perform stages sequentially or in parallel (e.g., heating material and subsequently mixing it with a further material, or heating and mixing simultaneously).
  • the data providers for original parameter data [X] can be implemented, for example, by sensors, by data outputs of actuators, by shop-floor systems etc.
  • the data providers can be regarded as sub-components.
  • the skilled person is familiar with devices (that comprise on-board processors and that are software/firmware controlled).
  • original parameter data XI (from data provider 1) should be a sensor reading regarding the operation of component 110-1 (e.g., a temperature value);
  • original parameter data X2 should be a sensor reading regarding intermediate material that leaves component 110-1 and that arrives at component 110-2;
  • original parameter data X3 should be a sensor reading regarding the operation of component 110-2 (e.g., a temperature value as well);
  • original parameter data X4 should be a sensor reading regarding the operation of auxiliary component 110-k (e.g., oil delivery data), the sensor for X4 would be a sub-component to component 110-k;
  • auxiliary component 110-k e.g., oil delivery data
  • original parameter data X5 should be a sensor reading regarding the operation of component 110-n (e.g., a pressure value);
  • • original parameter data X6 should be a sensor reading regarding intermediate material that leaves component 110-n and that arrives at component 110-N; and • original parameter data X7 should be a sensor reading regarding the operation of component 110-n (e.g., its energy consumption).
  • the notation Xv with a single variate index v is a simplification, it is also possible to use double-indices that identify the component (like indices n and k) and identify a data provider that belongs to a particular component.
  • the time by that the original parameter data become available should be the present time (i.e., during the operation of equipment 100).
  • the computer can obtain the associations from the above-mentioned equipmentdescriptor document (that has pointers to clusters). Such associations can be used as a basis for clustering.
  • Predictor module 500 receives [Y] and performs the behavior prediction for equipment 100 (and/or for particular components), cf. step 450 in FIG. 6.
  • Computer 200/500 performs the computer-implemented method (400 in FIG. 6) to identify behavior pattern Z of industrial equipment 100.
  • the small icon at the right side of predictor module 500 merely symbolizes that computer 200/500 would inform operator 190 about the result (i.e., the identified behavior pattern Z).
  • the result would be converted to a control signal to equipment 100 (or to an individual component).
  • the skilled person is familiar with control engineering and the like.
  • the behavior pattern can comprise, for example, the following:
  • the behavior pattern can be represented by a more complex data structure.
  • the parameter deviation could be provided with meta-data, such as the identification of the parameters (i.e., v), a quantity of the deviation (e.g., from a reference), with a time-frame of actual or expected occurrence, and so on.
  • behavior pattern Z is that of industrial equipment that performs an industrial process by components
  • operator 190 can also see a pattern at the level of the components (i.e., deviations at component granularity).
  • adapting comprises clustering and applying cluster-specific weights.
  • adaptor module 200 adapts [X] to [Y]
  • Operator 190 may also be interested in understanding the causes for particular patterns. If, for example, equipment 100 fails as a whole, operator 190 might inspect the last component in the chain first.
  • the "trouble making" component may however be a component at an earlier processing stage (i.e., a component with a lower component index n), or may be an auxiliary component.
  • clustering may highlight a broken oil supply component.
  • Adaptor module 200 is illustrated by bold lines, simply because it performs method steps that are to be explained with more details below.
  • clustering plays a role, it should be noted that - from a high-level perspective - the parameter data from the main components form a first cluster, and the parameter data from the auxiliary components form a second cluster.
  • Such simply main/aux clusters can be obtained, for example, by the domain expert writing a database table with the components in a description document, and by the computer processing that document (e.g., by parsing, cf. step 710 in FIG. 7).
  • Component 110-n receives intermediate material from component llO-(n-l), and component 110-n provides (different) intermediate material to component 110-(n+l). This is applicable during the process performance.
  • Component 110-k provides oil or the like components 110-2 and 110-n, etc.
  • the inter-component transfer of intermediate material may depend on a quality of the intermediate materials, cf. X2 and X7. That quality is not constant all of the time.
  • Component 110-k provides oil, but that oil supply could be interrupted. Or, an oil receiving component may suddenly demand for more, due to a leakage or the like.
  • the computer is able to obtain the (static and dynamic) relations from the computer-readable equipment-descriptor document.
  • Changes can be normal or can point to an undesired situation (such as failure events).
  • Individual variates Xv in original parameter data [X] can be clustered as well, for example, for parameters at the first processing stage (e.g., XI, X2), at intermediate processing stages (e.g., X3, X4, X5), and at the last processing stage (e.g., X6, X7).
  • first processing stage e.g., XI, X2
  • intermediate processing stages e.g., X3, X4, X5
  • the last processing stage e.g., X6, X7
  • variates that are specific to particular components e.g., first cluster XI, X2, X3, X5, X6, X7
  • second cluster X4, 110-2, 110-n and 120-k e.g., second cluster X4, 110-2, 110-n and 120-k.
  • FIG. 3 illustrates component 110-n, component-representation 210-n, and component-specific predictor 310-n.
  • Using component-representation 210-n and component-specific predictor 310-n is optional. While the index "n" refers to the main components (cf. FIG. 2), the approach can also be applied to the auxiliary components (i.e., to components 210-k).
  • the computer represents component 110-n in component-representation 210-n that can comprise for example:
  • cluster information that indicates the membership of the component in a component cluster and that indicates a type of the cluster (e.g., cluster with main component, cluster with auxiliary components, etc. as mentioned with FIG. 1), hence the parameters data can be clustered accordingly ,
  • That data can be used in the step clustering 420 (method 400, FIG. 6) as well as in cluster preparation activities (cf. 700 in FIG. 7).
  • Component-representations 210-n can be visually presented to a computer user, that is however not required.
  • FIG. 3 symbolizes the data by rectangles.
  • Component-specific predictor 310-n for that component can identify componentspecific behavior at present (or in the future, by prediction), such as deviations, etc.
  • predictor 310-n provides predicted parameter data that is parameter data that is expected to be applicable in the future.
  • predictor 310-n can calculate a probability Pn that the component may fail (within a certain time-interval into the future).
  • component-representation 210-n may have a first state variable that is set to the present state ON (i.e., the component is operating), and may have a second state variable that is set for the future (to a certain time-interval) to OFF.
  • the first state variable may be derived from parameter data
  • the second state variable may be derived from parameter data that is being processed by component-specific predictor 310-n.
  • Predicted parameter data can be used as original parameter data [X], In other words, predictor 310-n can have the same function as a data provider introduced above in FIG. 2.
  • predictor 310-n by a ML tool. This is similar to the implementation of predictor module 500 (cf. FIG. 1), but as a different granularity (parameter data for the components vs. parameter data for the equipment as a whole).
  • stage-representations can be used for process stages, and the predictors can be stage-predictors.
  • FIG. 4 illustrates clustering by way of examples, with CLUSTER_1 to CLUSTER_9.
  • the figure merely illustrates possible ways to cluster data.
  • the figure thereby also symbolizes how pre-defined rules can be used as clustering criteria. Clustering will be applied to the parameter data, but clustering is best explained by looking at the components that provide the parameter data. Writing "clustering components” therefore also comprises that the variates of these components are being clustered.
  • the computer can process the representations of the components (e.g. item 210-n in FIG. 3), and the (original) parameter data [X] that are associated with the components).
  • clustering should not be data-driven alone (i.e., performed by a machine learning tool only) because evaluating parameter data may introduce random correlations that are not necessarily causal relations.
  • CLUSTER_1 shows a linear cluster (or a "component chain cluster") with component-representations 210-1, 210-2 and 210-3 that are visualized along a graph.
  • the representations are shown as nodes and the relations between them (i.e., the relations between the components that are represented) are shown by direct edges (arrows).
  • event probabilities can be combined, such as the mentioned failure probabilities Pn. Assuming that failures propagate from component to component, such failure probability would be the higher at the end (i.e., at representation 210-3 for the corresponding component. For example, the above- mentioned failure probability P3 would be higher than Pl.
  • Parameter data for these components can be clustered accordingly.
  • the cluster vector [C] would, for example, identify the variates that belong to these components.
  • the static relation between the components that is the basis for clustering
  • the static relation between the components is the arrangement of the components in a processing chain (to process materials, or the like).
  • CLUSTER_2 shows a location cluster with componentrepresentations 210-2, 210-3 and 210-5 for components in proximate physical location.
  • some common parameters such as humidity or environmental temperature
  • the parameter data would be clustered, and the cluster vector [C] would, for example, identify the variates that belong to co-located components.
  • the static relation between the components is the arrangement of the components at a shared physical location for that like environmental conditions apply to the components.
  • CLUSTER_3 shows a physical relation cluster with componentrepresentations 210-2, 210-3 and 210-5 for that the components are in physical connection (210-2 and 210-3, 210-3 and 210-5).
  • a failure in a first component (such as over-heating) may trigger a failure in a physically connected second component (domino effect).
  • the vibrations could propagate to the other components in the cluster.
  • the description takes CLUSTER_3 as an example for a cluster that can be derived from documentation (either by parsing, cf. step 710 in FIG. 7) or manually (to be available to the computer in step 720).
  • the static relation between the components is the arrangement of the components in physical connection.
  • CLUSTER_4 and CLUSTER_5 show an active-component cluster (ON-cluster) and a non-active component cluster (OFF-cluster). This is symbolized by the acronym ON and OFF (for the states).
  • ON and OFF for the states.
  • the probability Pn that an OFF- component fails is lower than the probability Pn that an ON-component fails.
  • the cluster values would be provided accordingly, for example, to let predictor 500 ignore data from OFF-components.
  • the dynamic relation between the components is the equality of ON/OFF states that the components can assume.
  • CLUSTER_6 indicates that a component may be relatively new, in the sense to have been manufactured recently, to have been installed to the equipment recently, etc.
  • historical parameter data to train the ML tool may not (yet) be available, or may not be available in sufficient amount. Therefore, the cluster values may indicate that predictor module 500 may not perform certain operations.
  • CLUSTER_7 symbolizes by a line trough [X] that parameter data may become incomplete, due to sensor failure or similar events. If a sensor - that is a sub-component to a component - stops providing data, the process at the equipment may continue. As already mentioned above, predictor 500 (cf. FIG. 1) might not be prepared for such situations, and it might derive an incorrect conclusion. To mitigate that, parameter data [X] are separated, at least into reliable data (cluster [X]
  • CLUSTER_8 symbolizes that parameter data may be disturbed.
  • noise or other effects.
  • a component may vibrate and a sensor might provide not only the sensor-specific signal but also a vibration specific signal.
  • the ML tool might process noisy signals and clean signals differently.
  • the ML tool has already been trained by noisy data, because the skilled person occasionally augments training data by noise. In this case, the presences of noise turns [X] to [X]
  • CLUSTER_9 symbolizes the identification of correlations between two or more variates (in original parameter data [X]).
  • first variate Xa would should a U-shaped drop in its value
  • second variate would show a similar U-shaped drop in its value as well, even with some delay, the skilled person would see a correlation.
  • Automation is possible, and there are many mathematical approaches to identify correlation, the calculation of the Pearson correlation coefficient (PCC) is just an example.
  • a further example would be clustering by data processing, such as by using the well- known k-nearest neighbors algorithm, using hierarchical clustering and other approaches.
  • a further example is a cluster of variates that do not change frequently, but for that changes occur.
  • the parameter for oil supply (from component 120-k) may suddenly change, and with that a follow-up parameter will change as well.
  • a further example would be the identification of so-called process signatures. As the process (or a sub-process, or a stage) would not show its characteristic signature, the process could be considered as deviating (even as abnormal), with the consequence that process data can be clustered to be treated separately.
  • the temperature readings for Alpha and Beta correlate to energy consumption for Alpha and Beta, respectively, but there can also be a correlation between the temperature of Alpha to the consumption of Beta, the temperature of Beta to the consumption of Alpha.
  • V 4 variates, with two correlations that indicate causal relations, and with two correlations that are rather random correlations, without causal relation. In such a simplified situation, the operator can easily see the sort out any misleading correlation.
  • Clustering by the computer steps in here If a correlation across cluster borders is detected, it can be marked as such. Or, the cluster value (here for cluster Alpha and for cluster Beta) may prevent the computer from evaluating parameter data in a mixed fashion. If the cluster values are applied as weights (step 440), predictor module 500 would, for example, process the parameter data from both clusters separately. This approach limits the changes for mis-interpreting behavior.
  • FIG. 5 illustrates component-representations 210-1, 210-2, 201-3 and 210-4 for the corresponding components (omitted here for simplicity).
  • FIG. 5 also illustrates clusters: symbolically by connection lines.
  • Representation 210-1 shows that the corresponding component is located at "location A" and that its current state is ON.
  • Representation 210-2 shows that the corresponding component is located at "location A", that it is component that participates in performing a sub-process (1), and that its current state is OFF.
  • Representation 210-3 shows that the component participates in performing subprocess (1) (or sub-stage), and that its current state is ON. The location is not specified.
  • Representation 210-4 shows that the component is located at B, and that its state is ON.
  • 210-1 and 210-4 there is a dynamic relation between 210-1 and 210-4, according the state, both ON, state cluster ON.
  • the state can change, of 210-1 (from ON to OFF) or of 210-4 (from ON to OFF). This is therefore a dynamic cluster.
  • the parameter data from a component in state OFF can be weighted with W-values (that let predictor module 500 ignore them). Or, for example, the parameter data from that component in state OFF would have to be considered in the further cluster (the parameter data of the components of the cluster for sub-process (1)).
  • FIG. 5 can be regarded as a partial copy of an equipment-descriptor document (as mentioned, e.g., an instrumentation diagram). The copy is not complete but limited to the information that the computer needs for clustering. Advanced clustering with the process of time
  • FIG. 5 takes the component states as an example for a dynamic change, and as an example, for a change that can be predicted to occur in the future.
  • the asterisk * next to "ON” or “OFF” symbolizes that the state can be predicted for a time-interval in the future by a predictor, such as by predictor 310-n (FIG. 3).
  • [X] can be initially by clustered into parameter data belonging to an initial ON- cluster (i.e., the variates from the component that correspond to representations 210-1, -3- and 4) and to an initial OFF-cluster (component that corresponds to representation 210-2).
  • [C] [ON, OFF, ON, ON] is simplified, there are multiple variates for each component at vector positions 1, 3 and 4 that would be clustered to the ON-cluster, multiple variates for position 2 as well.
  • Predefined rules to assign and to apply weights can be applied once re-clustering is completed, cf. FIG. 7 with a repetition of method 400 for such situations.
  • screen icons or other symbols
  • the components in OFF state may be illustrates in different style than the components in ON state.
  • FIG. 5 illustrates dashed arrows 250-A, 250-B and 250-C that are directed from an icon that indicates the state of component #3 (that is represented by item 210-3), indicates the stage or sub-process of that component #3, indicates that components #3 and #2 (cf. item 210-2) are statically related (both participate in the same stage), and also indicate by the last arrow 250-C that component #2 is OFF.
  • This is not yet an indication of a behavior that is potentially a failure (of equipment 100), because the behavior indication is performed by predictor 500. But the arrows could give the operator a hint to have a closer look at component #2 that is OFF.
  • predictor 500 Based on clustering (in step 420) and based on the weights (steps 430 and 440), predictor 500 indicates the behavior. If the behavior would be "failure” or the like, the operators can follow the arrows to see more details. The last arrow may point to an approximation of the root cause for the failure.
  • FIG. 6 illustrates a flow-chart diagram of computer-implemented method 400 to identify behavior pattern Z of industrial equipment 100 that performs an industrial process.
  • industrial equipment 100 has components 110-n, 120-k and performs an industrial process.
  • pre-trained machinelearning tool cf. predictor module 500 in FIG. 2
  • the computer has two modules: adaptor module 200 and predictor module 500, cf. FIG. 2.
  • the computer i.e., adaptor module 200 receives original parameter data [X] from data providers that belong to the components of the industrial equipment.
  • clustering step 420 the computer (i.e., adaptor module 200) clusters original parameter data [X] to data clusters - such as [X]@[C] in FIG. 1 - according to relations between the components.
  • Step 420 is illustrated with an arrow that is labelled [C], This indicates that a cluster vector [C] can be received prior to step execution.
  • Static relations between the components are pre-defined by the industrial process, and dynamic relations between the components are defined by parameter data that are being received in step 410.
  • step 430 the computer (i.e., adaptor module 200) assigns cluster values (cf. Cl, C2 in FIG. 1) to the data clusters.
  • step 440 the computer applies the cluster values as weights [W] to the parameter data [X] to obtain adapted parameter data [Y]).
  • step 450 the computer (i.e., predictor module 500) performs the behavior prediction with adapted parameter data [Y], This step results in behavior pattern Z.
  • FIG. 7 illustrates a (flow-chart) diagram of activities, including cluster preparation activities 700 and computer-implemented method 400.
  • the box 400 on the right side stands for method 400, but details are omitted for simplicity.
  • the result of cluster preparation activities 700 can be identification of original parameter data [X] that would have to be clustered in a particular way.
  • the identification is symbolized here by cluster vector [C], cf. FIG. 1 and FIG. 6.
  • [C] symbolizes the result of activities 700, but the vector format is just an example.
  • Cluster vector [C] can be available even before receiving parameter data [X] in step 410. This in particular relevant for [C] (or part of it) that is based on static relations. Method 400 can be repeated for new parameter data arriving (i.e., receiving 410 in FIG. 6) but also when [C] changes.
  • Cluster preparation activities 700 are presented as separate steps that are computer-implemented, not as a mandatory flow.
  • Activity 710 stands for the computer parsing a document, such as the above- mentioned description.
  • Activity 720 stands for the computer receiving input from the human domain expert.
  • Activity 730 stands for the computer parsing parameter data according to the availability while the equipment performs the industrial process.
  • a computer can divide parameter data (i.e., particular variates) into groups.
  • the parameter of a first group may change during first interval but may remain unchanged during a second, subsequent time interval.
  • the time intervals may be identical to particular process stages.
  • FIG. 7 shows activities 710, 720 and 730 separately, but they could be combined.
  • industrial equipment cf. FIG. 1 could be a furnace. In a first stage, it is being charged with materials, in the second stage it will be heated up, and so on.
  • FIG. 8 illustrates a block diagram for a computer that performs behavior prediction with adapted parameter data by separate predictor sub-networks 510 and 520 that belong to predictor module 500.
  • multi-variate parameter data [X] and [Y] or subsets thereof can be input to processing tools, such as neural networks in the function of predictor module 500.
  • processing tools such as neural networks in the function of predictor module 500.
  • [X] going to predictor module 500 is omitted for simplicity, by [Y] is presented in the two clusters of FIG. 1.
  • Yl, Y3 and Y4 should belong to cluster Cl, and Y2 and Y5 should belong to cluster C2.
  • the weights W1 to W5 can indicate cluster membership.
  • Sub-network 510 processes data from the first cluster (i.e., Yl, Y3 and Y4) and subnetwork 520 processes data from the second cluster (i.e., Y2 and Y5).
  • first cluster i.e., Yl, Y3 and Y4
  • subnetwork 520 processes data from the second cluster (i.e., Y2 and Y5).
  • the same principle can be applied to further clusters and to further sub-networks.
  • Sub-networks 510 and 520 perform (cf. step 450) the behavior prediction with adapted parameter data, but the result is not (yet) the identified behavior pattern of industrial equipment 100. But there are preliminary cluster-specific patterns at the outputs, 511 and 521.
  • Combiner module 550 receives these patterns and obtains the behavior pattern of industrial equipment 100 as a whole.
  • Combiner module 550 may provide the behavior pattern based on logical relations (e.g., NORMAL operation of equipment if all components show NORMAL; or ABNORMAL if at least one component shows ABNORMAL)
  • Combiner module 550 may provide the behavior pattern based on logical relations as well, according to the same principle (all, at least one)
  • the clusters and the sub-networks can be associated with parameter data that are redundant and that are available in parallel.
  • Combiner module 550 would compare the preliminary patterns.
  • Combiner module 550 may provide the behavior pattern based on logical relations as well.
  • FIG. 8 can be considered as a cascade with two steps, but further steps can be applied as well.
  • a first sub-network can provide input to a second subnetwork
  • the second sub-network can provide input to a third sub-network and so on.
  • FIG. 8 symbolizes an optimal second sub-network by a dashed box 515.
  • training for serially connected sub-networks (e.g., 510 and 515), training can be performed serially (the parameter data receiving subnetwork first), and parallel sub-networks can be trained individually.
  • predictor module 500 that is a machine-learning tool
  • predictor module 500 can be trained according to usual procedures, with training sets that comprise historical parameter data and that comprise historical behavior patterns Z.
  • steps clustering 420, assigning 430, and applying 430 of method 400 can be performed for historical data as well.
  • the training set would therefore also comprise historical parameter data with cluster values as weights.
  • FIG. 9 illustrates a flow-chart diagram of computer-implemented method 600 to enable a machine-learning tool (cf. predictor module 500) to identify behavior pattern Z of industrial equipment 100 (that has components 110-n, 120-k and that performs an industrial process), such that machine-learning tool 500 is being enabled to perform behavior prediction during process performance.
  • a machine-learning tool cf. predictor module 500
  • the method-performing computer receives historical parameter data from data providers that belong to the components 110-n, 120-k of industrial equipment 100.
  • Step 610 has similar to step 410, but performed with historical data that are original data from historical process performances, from equipment 100, or from reference equipment.
  • step 620 the computer clusters the historical parameter data to data clusters according to relations between the components 110-n, 120-k.
  • static relations between the components are pre-defined by the industrial process, and dynamic relations between the components are defined by the parameter data that are being received (but for historical parameter data).
  • step classifying 625 - that is the additional step just mentioned - the computer classifies the dynamic relations into relation groups.
  • FIG. 9 illustrates the step to be performed together with clustering, in parallel. This is convenient, but not required.
  • Classifying 625 the dynamic relations into relation groups can be performed by an auxiliary machine learning tool that is a classifier. As the skilled person can easily implement that, such an auxiliary ML tool is omitted from illustration.
  • the method continues with assigning 630 cluster values to the data clusters and with applying 640 the cluster values as weights to the historic parameter data to obtain adapted historical parameter data. That is similar to steps 430 and 440, but with historical data.
  • step using 650 the adapted historical parameter data as training set to train the machine-learning tool 500
  • the computer differentiates training sets for the machine-learning tool according to the classified relation groups.
  • FIG. 9 shows that differentiation by a dashed line from step 625 to step 650.
  • step using 650 the adapted historical parameter data as training set can be performed for different sub-networks (cf. 510, 520 in FIG. 8, or others) of machine-learning tool 500 that are being trained for different relation groups separately.
  • Training sub-networks separately is also applicable for configurations in that the networks are cascaded. Training data would have to be differentiated as well.
  • the method-executing computer performs behavior prediction with adapted parameter data [Y]
  • Adaptor module 200 performs clustering 420 the variates Xv of the original parameter data [X] to data clusters [X]@[C] according to relations between the components 110-n, 120-k.
  • the computer can use cluster vector [C] as a data structure to implement clustering.
  • the cluster granularity i.e., what parameter data belong to what cluster influences the determination of the behavior pattern Z.
  • the description has already explained aspects (cf. FIG. 4, Advanced Clustering) and has explained that the operation of ML tool 500 can be cluster-specific (cf. FIG. 8, sub-networks 510 and 520). It is noted that clustering (in step 420) can require that the computer "knows" how the components are related.
  • Timing This aspect could be summarized in the question whether the pattern Z becomes available early enough to take measures, if needed.
  • the computer may use electrical energy to obtain a pattern that is of no interest, or that would not represent the state of the industrial machine. It is therefore desired to identify a pattern Z that would indicate failures (at present, or in the future), that would indicate inefficiencies etc.
  • data from a single variate Xv may influence accuracy, timing and computation (influence to one aspect only, or to multiple aspects).
  • data from a particular parameter may be disturbing to correctly identify Z, or a particular parameter might be essential to the cluster, otherwise Z could not be identified.
  • clustering (step 420) can be performed by
  • ML tools that are trained with or without supervision
  • human experts who prepare rules to be applied during clustering 420.
  • the technical efforts to cluster with ML tools may not be justified for some approaches, but some approaches may by most efficient with ML tools only.
  • the description outlines a further methodology to select the suitable clustering approach.
  • the methodology can be applied to provide criteria for clustering 420, that means at a point in time when clustering 420 is not yet performed, or even earlier (prior to receiving original parameter data [X] in step 410. From a different perspective, the methodology is for preparing the clustering step (i.e., by "pre-clustering").
  • the first criterium looks at the presence or absence of interactions (between the components) that are defined by the industrial process (i.e., interactions with each other in order to perform the industrial process, process-inherent relations).
  • the second criterium looks at the presence or absence of changes in these interactions during the performance of the process (i.e., after the industrial equipment has started to perform the industrial process).
  • Clustering (cf. step 420) at least some of the variates Xv of the multi-variate original parameter data [X] to one or more data clusters [X]@[C] is performed such that in at least one data cluster [X]@[C], the majority of variates Xv belong to components in the same relation group (cf. FIG. 10, groups i, ii, iii, iv).
  • Each cluster has a number of variates.
  • Each cluster can be regarded as a set, and the number would be the cardinality of that set.
  • the simplified example of FIG. 1 shows a first cluster with 3 variates XI, X3 and X4, and shows a second cluster with 2 variates X2 and X5.
  • a cluster with 2 variates can't have a majority of variates.
  • two variates form the majority if the variates belong to components of the same relation group.
  • the majority rule can be applied likewise.
  • the components have data providers (cf. 1, 2, 3, 4, 5, 6, and 7 in the example of FIG. 2) that correspond to variates Xv in multi-variate original parameter data, cf. FIG. 1.
  • Some variates of the parameter data [X] can therefore belong to components of the same relation group.
  • a particular cluster may comprise parameter variates from relations in more than one group (cf. "the majority of”), and that the parameter data from a particular relation group is not yet a cluster. Parameter data from the components that for a particular relation group is rather the starting point to obtain a cluster.
  • FIG. 10 illustrates a diagram with two dimensions ORIGIN and DEVELOPMENT (labelled in uppercase letters). Each dimension is given with two alternatives (labelled in quotation marks "". Therefore, the diagram has four quadrants (Roman figures i, ii, iii, iv), one for each relation group. The dimensions with their alternatives can be taken as criteria to select an approach to perform clustering.
  • step 420 As clustering in step 420 is performed prior to performing the behavior prediction in step 450, the data within the adapted parameter data [Y] can change.
  • the dimension ORIGIN can be applicable to relations, in two alternatives (presence, absence): • (ii, iii) Relations in "process" are defined by the interactions between the components as they are defined with each other in order to perform the industrial process. In other words, the relations are process-intrinsic and defined by the industrial process.
  • a provider-consumer relation can belong to the industrial process because component 110-1 provides material to component 110-2, and component 110-2 consumes that provided material (cf. for example CLUSTER_1 in FIG. 4). Otherwise, the industrial process can't be performed. A provider-consumer relation remains even in case of failure or during maintenance.
  • an air-conditioning component is in relation to the components in the same building (FIG. 4, CLUSTER_2).
  • the process could be performed in a non-air-conditioned building. But as temperature and/or humidity may influence the performance of the process, such parameters would have to be taken into account.
  • original parameter data [X] for clustering can be available: in the alternative "process”, for example, the process definition allows to obtain clustering rules to be applied to parameter data. For example, clustering according to "not by process” would have to cluster variates from parameter data that the computer receives. For example, if the provider components for a given point in time provides material (or not) is a variate that the computer receives.
  • the ORIGIN is related to the physical space by that the components are arranged, the ORIGIN dimension (in general) has the space dimension (in particular as a sub-dimension, cf. CLUSTER_2).
  • the components are arranged in a certain topology (e.g., material flow in a particular direction between components) and the topology can be taken into account.
  • the dimension DEVELOPMENT has an aspect in repetitions of method steps.
  • Predictor module 500 would perform behavior prediction, to pattern Z (step 450) substantially all of the time while the industrial process is ongoing (the industrial equipment has already started to perform the industrial process). Previous steps, such as clustering 420 can be repeated as well, but rather for relations in the "change” alternative and not for relations in the "constant” alternative. As a sidenote, not repeating a step (such as clustering) may save computation (in terms of processor and memory consumption).
  • component 110-1 may provide the material to component 110-2 via a pipeline or via a vehicle.
  • the relation does not change (iii, iv), but for the vehicle there are intervals during that providing is interrupted.
  • component 110-2 should have a storage so that consuming should be possible at any time. Not providing does not have to be a failure).
  • original parameter data ([X]) for clustering can be available as well: in alternative “change” (ii, i) for example from measurement data (e.g., "Is the vehicle in motion or not?") and "constant” from a process definition, material from 110-1 to 110-2 but not vice versa.
  • DEVELOPMENT "change"
  • the parameter data [X] can indicate that relations change.
  • the relations in the first quadrants can be attributed to be "dynamic". For example, a failure event may turn the provider component out of service so that the provider-consumer relation stops to exists. But non-failure events can change the relations as well.
  • clustering can be time-based clustering.
  • the relations can be considered to be pre-defined so that clustering rules can be "hardcoded". Changes over time (i.e., during the performance of the process) of the relations are not to be expected.
  • clustering parameter data [X] from the provider and from the consumer component
  • the parameter data may change, and - to stay with the example - the provider may stop providing or the consumer may step consuming.
  • Predictor module 500 can identify a pattern Z that stands for a failure.
  • predictor module 500 continues to process parameter data ([W] to [X] applied, step 440) that had been clustered for third quadrant relations. If the parameter data suddenly show first quadrant behavior, predictor module 500 may qualify the pattern Z as failure.
  • energy use can be distributed to certain hours of the day, such as for high-power components (such as heaters or electrical furnaces).
  • high-power components such as heaters or electrical furnaces.
  • a parameter that stands for energy availability would be part of a cluster in a second quadrant relation.
  • CLUSTER_2 shows a location cluster with components in proximate physical location, and the interactions between the components are process- defined and constant, relation group iii.
  • CLUSTER_3 shows a physical relation cluster with the assumption that events may be propagated.
  • a failure in a first component (such as overheating) may trigger a failure in a physically connected second component (domino effect).
  • Monitoring data (i.e., variates) in such a cluster would be in relation group iii as well.
  • CLUSTER_4 and CLUSTER_5 show an active-component cluster (ON-cluster) and a non-active component cluster (OFF-cluster).
  • the cluster values would be provided accordingly, for example, to let predictor 500 ignore data from OFF-components.
  • the relations are dynamic, group i.
  • CLUSTER_6 indicates that a component may be relatively new, in the sense to have been manufactured recently. The relation to other components is therefore static, group iii, because the component is part of the equipment.
  • CLUSTER_7 symbolizes by a line trough [X] that parameter data may become incomplete, due to sensor failure or similar events. This is an example for a highly dynamic cluster, group i, with the relatively high probability that re-clustering may be required (during process performance), group i. Using computer- implemented clustering appears an option to take.
  • CLUSTER_9 symbolizes the identification of correlations between two or more variates (in original parameter data [X]).
  • the identification of correlations etc. can be regarded as steps to perform clustering automatically.
  • the variates can be related to components in groups i or ii.
  • FIG. 11 illustrates an example of a generic computer device which may be used with the techniques described here.
  • FIG. 11 is a diagram that shows an example of a generic computer device 900 and a generic mobile computer device 950, which may be used with the techniques described here.
  • Computing device 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers.
  • Generic computer device may 900 correspond to the computer system 200/500 of FIG. 2.
  • Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smart phones, driving assistance systems or board computers of vehicles and other similar computing devices.
  • computing device 950 may be used as a frontend by a user (e.g., an operator of a blast furnace) to interact with the computing device 900.
  • a user e.g., an operator of a blast furnace
  • the components shown here, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the inventions described and/or claimed in this document.
  • Computing device 900 includes a processor 902, memory 904, a storage device 906, a high-speed interface 908 connecting to memory 904 and high-speed expansion ports 910, and a low speed interface 912 connecting to low speed bus 914 and storage device 906.
  • Each of the components 902, 904, 906, 908, 910, and 912, are interconnected using various busses, and may be mounted on a common motherboard or in other manners as appropriate.
  • the processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or on the storage device 906 to display graphical information for a GUI on an external input/output device, such as display 916 coupled to high speed interface 908.
  • multiple processors and/or multiple buses may be used, as appropriate, along with multiple memories and types of memory.
  • multiple computing devices 900 may be connected, with each device providing portions of the necessary operations (e.g., as a server bank, a group of blade servers, or a multi-processor system).
  • the memory 904 stores information within the computing device 900.
  • the memory 904 is a volatile memory unit or units.
  • the memory 904 is a non-volatile memory unit or units.
  • the memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.
  • the storage device 906 is capable of providing mass storage for the computing device 900.
  • the storage device 906 may be or contain a computer-readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid state memory device, or an array of devices, including devices in a storage area network or other configurations.
  • a computer program product can be tangibly embodied in an information carrier.
  • the computer program product may also contain instructions that, when executed, perform one or more methods, such as those described above.
  • the information carrier is a computer- or machine-readable medium, such as the memory 904, the storage device 906, or memory on processor 902.
  • the high speed controller 908 manages bandwidth-intensive operations for the computing device 900, while the low speed controller 912 manages lower bandwidth-intensive operations.
  • the high-speed controller 908 is coupled to memory 904, display 916 (e.g., through a graphics processor or accelerator), and to high-speed expansion ports 910, which may accept various expansion cards (not shown).
  • low-speed controller 912 is coupled to storage device 906 and low- speed expansion port 914.
  • the low-speed expansion port which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet) may be coupled to one or more input/output devices, such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.
  • input/output devices such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.
  • the computing device 900 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a standard server 920, or multiple times in a group of such servers. It may also be implemented as part of a rack server system 924. In addition, it may be implemented in a personal computer such as a laptop computer 922. Alternatively, components from computing device 900 may be combined with other components in a mobile device (not shown), such as device 950. Each of such devices may contain one or more of computing device 900, 950, and an entire system may be made up of multiple computing devices 900, 950 communicating with each other.
  • Computing device 950 includes a processor 952, memory 964, an input/output device such as a display 954, a communication interface 966, and a transceiver 968, among other components.
  • the device 950 may also be provided with a storage device, such as a microdrive or other device, to provide additional storage.
  • a storage device such as a microdrive or other device, to provide additional storage.
  • Each of the components 950, 952, 964, 954, 966, and 968 are interconnected using various buses, and several of the components may be mounted on a common motherboard or in other manners as appropriate.
  • the processor 952 can execute instructions within the computing device 950, including instructions stored in the memory 964.
  • the processor may be implemented as a chipset of chips that include separate and multiple analog and digital processors.
  • the processor may provide, for example, for coordination of the other components of the device 950, such as control of user interfaces, applications run by device 950, and wireless communication by device 950.
  • Processor 952 may communicate with a user through control interface 958 and display interface 956 coupled to a display 954.
  • the display 954 may be, for example, a TFT LCD (Thin-Film-Transistor Liquid Crystal Display) or an OLED (Organic Light Emitting Diode) display, or other appropriate display technology.
  • the display interface 956 may comprise appropriate circuitry for driving the display 954 to present graphical and other information to a user.
  • the control interface 958 may receive commands from a user and convert them for submission to the processor 952.
  • an external interface 962 may be provide in communication with processor 952, so as to enable near area communication of device 950 with other devices.
  • External interface 962 may provide, for example, for wired communication in some implementations, or for wireless communication in other implementations, and multiple interfaces may also be used.
  • the memory 964 stores information within the computing device 950.
  • the memory 964 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or units, or a non-volatile memory unit or units.
  • Expansion memory 984 may also be provided and connected to device 950 through expansion interface 982, which may include, for example, a SIMM (Single In Line Memory Module) card interface.
  • SIMM Single In Line Memory Module
  • expansion memory 984 may provide extra storage space for device 950, or may also store applications or other information for device 950.
  • expansion memory 984 may include instructions to carry out or supplement the processes described above, and may include secure information also.
  • expansion memory 984 may act as a security module for device 950, and may be programmed with instructions that permit secure use of device 950.
  • secure applications may be provided via the SIMM cards, along with additional information, such as placing the identifying information on the SIMM card in a non-hackable manner.
  • the memory may include, for example, flash memory and/or NVRAM memory, as discussed below.
  • a computer program product is tangibly embodied in an information carrier.
  • the computer program product contains instructions that, when executed, perform one or more methods, such as those described above.
  • the information carrier is a computer- or machine-readable medium, such as the memory 964, expansion memory 984, or memory on processor 952 that may be received, for example, over transceiver 968 or external interface 962.
  • Device 950 may communicate wirelessly through communication interface 966, which may include digital signal processing circuitry where necessary. Communication interface 966 may provide for communications under various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. Such communication may occur, for example, through radio-frequency transceiver 968. In addition, short-range communication may occur, such as using a Bluetooth, WiFi, or other such transceiver (not shown). In addition, GPS (Global Positioning System) receiver module 980 may provide additional navigation- and location-related wireless data to device 950, which may be used as appropriate by applications running on device 950.
  • GPS Global Positioning System
  • Device 950 may also communicate audibly using audio codec 960, which may receive spoken information from a user and convert it to usable digital information. Audio codec 960 may likewise generate audible sound for a user, such as through a speaker, e.g., in a handset of device 950. Such sound may include sound from voice telephone calls, may include recorded sound (e.g., voice messages, music files, etc.) and may also include sound generated by applications operating on device 950. [00296] The computing device 950 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a cellular telephone 980. It may also be implemented as part of a smart phone 982, personal digital assistant, or other similar mobile device.
  • audio codec 960 may receive spoken information from a user and convert it to usable digital information. Audio codec 960 may likewise generate audible sound for a user, such as through a speaker, e.g., in a handset of device 950. Such sound may include sound
  • Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof.
  • ASICs application specific integrated circuits
  • These various implementations can include implementation in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.
  • the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer.
  • a display device e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor
  • a keyboard and a pointing device e.g., a mouse or a trackball
  • Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.
  • the systems and techniques described here can be implemented in a computing device that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components.
  • the components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN”), a wide area network (“WAN”), and the Internet.
  • LAN local area network
  • WAN wide area network
  • the Internet the global information network
  • the computing device can include clients and servers.
  • a client and server are generally remote from each other and typically interact through a communication network.
  • the relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
  • Method (400) according to any of clauses 1 to 4, wherein during the process performance, the step clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) - to data clusters ([X]@[C]) - is performed once for data clusters that remain unchanged, and wherein the step assigning (430) cluster values (Cl, C2) to the data clusters is repeated for changing cluster values.

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Mathematical Physics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Stored Programmes (AREA)

Abstract

To identify a behavior pattern (Z) of industrial equipment that performs an industrial process, a computer uses an adaptor module (200) and a predictor module (500) to perform a computer-implemented method (400). The adaptor module (200) receives (410) original parameter data [X] from data providers that belong to components of the industrial equipment, the adaptor module (200) clusters (420) original parameter data [X] to data clusters according to relations between the components. Static relations between the components are pre-defined by the industrial process, and dynamic relations between the components are defined by parameter data that are being received. The adaptor module (200) assigns (430) cluster values to the data clusters, and applies (440) the cluster values as weights [W] to the parameter data [X] to obtain adapted parameter data [Y]. The predictor module (500) performs (450) the behavior prediction with adapted parameter data [Y] results in the behavior pattern (Z).

Description

CLUSTERING PARAMETER DATA TO IDENTIFY BEHAVIOR PATTERN OF INDUSTRIAL
EQUIPMENT
Technical Field
[001] In general, the disclosure relates to industrial production processes, and more in particularly, the disclosure relates to computer systems, methods and computerprogram products to identify behavior patterns of industrial equipment that performs the industrial production process.
Background
[002] Industrial equipment such as industrial machines perform processes. The equipment has components in a particular arrangement so such that the process can be performed. In a simplified example, the components would be arranged in a chain to sequentially forward intermediate products from a predecessor component to a successor component.
[003] Before, during and after process performance, the equipment provides parameter data, mostly in the form of multi-variate time-series. Parameter data can be obtained from sensors that measure phenomena that are relevant to the process, such as temperature, pressure, material properties and others. Parameter data can also be obtained from actuators. For example, a heater has a state (i.e., "on" or "off"), a rated power (e.g., a kW value number given by the manufacturer), an actual power consumption (e.g., a kW value that is measured); a valve may be "open" or "closed" (to influence the pressure in a vessel), and so on. Parameter data can also be obtained from shop-floor systems (e.g., regarding the amounts of materials etc.).
[004] Parameter data can usually be related to particular components, although a sub-set of parameter data can be related to multiple components. For example, precipitations usually increase the air humidity in a factory as a whole, but when the Sun returns, the humidity decreases again. [005] The skilled person is also able to aggregate a sub-plurality of parameter data to one or more particular indicators. For example, such an indicator can be a quality indicator. Quality indicators can be related, for example, (a) to the result of the process (such as to the product that results from a production process), (b) to the performance of the equipment (such as energy consumption, use of auxiliary materials such as oil), the time it took to perform the process, and (d) others.
[006] The time-series have durations that can correspond to process instances, and the durations of these instances change from instance to instance. The skilled person who evaluates data can harmonize instance duration with techniques such as dynamic time warping (DTW).
[007] As equipment failure need to be prevented, machine-learning (ML) tools allow the identification of parameter deviations, the prediction of the mentioned quality indicators, and the like. ML tools may also indicate the absence or presence of equipment anomaly at present, or may predict anomaly events for the future. The definition of an "anomaly" is usually adapted to particular processes. For example, an unexpected stop of a production line would certainly qualify as anomaly, but for some processes a relatively minor deviation in the product quality would be regarded as anomaly as well.
[008] The skilled person would define the occurrence of an anomaly by applying domain knowledge, and terms such as "anomaly", "equipment anomaly", "equipment or component failure" is a convenient simplification of a more complex behavior pattern of the equipment.
[009] The ML tools are trained with training data that is usually a collection of historical data from historical process instances (i.e., past usage data). During operation of the equipment, the tools process data of the current process.
[0010] Simplified, the ML tools perform algorithms to process the parameter data at the tool input and to derive a function at the tool output. By way of example, the function can be an estimation of equipment failure in the future, of a particular property of the product at the end of the process, the detection or anticipation of an anomaly, or others.
[0011] Formally, the function is a (mathematical) function of a plurality of parameter values. As this (mathematical) function is usually unknown by explicit formulas, the ML tools are implemented by neural networks or the like.
[0012] From a high-level perspective, this concept to use ML tools requires data in a proper data quantity. The accuracy of the prediction suffers if there is not enough data available, but also suffers if there is too much data available. More in detail, the deficiencies can be explained by the following:
• A component may be relatively new, in the sense to have been manufactured recently, to have been installed to the equipment recently, etc. As a consequence, historical parameter data to train the ML tool may not (yet) be available, or may not be available in sufficient amount.
• Parameter data (in the multi-variate time-series) may become incomplete, due to sensor failure or similar events. If a sensor - that is a sub-component to a component - stops providing data, the process at the equipment may continue. However, the ML tool might not be prepared for such situations, and it might derive an incorrect conclusion.
• Parameter data may be disturbed, because of effects that are colloquially attributed to be noise. For example, during operation a component may vibrate and a sensor might provide not only the sensor-specific signal but also a vibration specific signal. The ML tool might process noisy signals and clean signals differently. At best, the ML tool has already been trained by noisy data, because the skilled person occasionally augments training data by noise.
• ML tools that rely on data only (i.e., data driven approaches) are strong in identifying correlations, such as between two or more variates in a multivariate time-series. Identifying a correlation is that same as identifying a cause-effect relation.
[0013] These and other deficiencies must be taken into account, especially when the result of the ML tool is used to control the equipment. For example, if equipment failure is predicted (for the future), the equipment operator may stop the operation of the equipment or of a component (for exceptional maintenance, for trouble-shooting or the like). However, such a stop would interrupt process performance, with undesired consequences (such as wasting materials, additional energy to re-start, etc.).
[0014] There is a need to make the operation of ML tools more robust to some or all of the mentioned deficiencies.
[0015] US 2020/0226490 Al discusses a machine tool by illustrating infrastructure components that provide real-time machine metrics, by explaining computer modules that process the metrics and that identify anomalies and corresponding remedial actions. The actions can be applied to the infrastructure components.
[0016] Industrial equipment performs an industrial process by components (i.e., equipment components), and a computer executes a method to identify a behavior pattern of the equipment, for the equipment as a whole or for individual components.
[0017] As this computer-implemented method involves the use of machine learning (ML) tools, it is convenient to differentiate
• a preparation phase (to collect historical data),
• a training phase (to apply historical data to train the tool), and
• an operation phase (while the industrial equipment performs this industrial process, and - optionally - to use the identified behavior pattern to control the equipment).
[0018] But the computer-implemented method also has further aspects or implies further activities such as clustering the variates of parameter data according to static or dynamic relations between the components or process stages. The terms "static" and "dynamic" are convenient labels that indicate knowledge of the relations in view of the operation phase.
[0019] Clustering can involve representing the components (or stages) by so-called virtual components (or "virtualizing" in short). In principle, clustering can be performed during all three phases.
[0020] The computer executes the method to identify the behavior pattern in the operation phase. The computer receives parameter data as multi-variate timeseries from data providers - such as sensors - of the industrial equipment. The computer processes the parameter data by two computing modules, and provides behavior data.
[0021] In implementations, the first computing module can receive at least a sub-set of the parameter data as original parameter data and turns that sub-set into adapted parameter data. The first module is also called the "adaptor module". Clustering plays a role for this adaptor module. The parameter data is adapted by applying cluster values (as weights to original parameter data).
[0022] In implementations, the second computing module can receive the adapted parameter data and performs behavior prediction to provide the behavior data. The second computing module can also receive parameter data that by-passes the first computing module. The second module is also called the "predictor module". The predictor module is implemented as a pre-trained machine-learning (ML) tool.
[0023] By coupling the adaptor module into the data path from the data provider to the predictor module, the operation of the computer is made more robust (to deficiencies), not by changing the predictor module itself, but by pre-processing (i.e., adapting) the data that reaches the predictor module. This aspect is applicable for the training phase and for the production phase).
[0024] As disclosed herein, the predictor module receives (at least a subset of) the parameter data that are adapted according to clusters. More in detail, cluster values are applied as weights to the original parameter data. Clustering components (and sub-processes) according to static and dynamic relations is expected to accommodate deficiencies.
[0025] Clustering applies to the training phase when the predictor is being trained with historical data, and also applies to the operation phase when the predictor actually serves as a predictor (by performing the method).
[0026] In view of the above-mentioned deficiencies, the following might be advantageous:
• A relatively a new component might not have provided enough historical data yet. As the result of clustering, parameter data can be differentiated as data from a "newcomer" component and data from a "veteran" component. Differentiating according to a finer granularity is also possible, but the main criteria is the amount of historical data to perform training. Adapted parameter data have less weight for the predictor module than original parameter data.
• Sensor failure - if detected - would classify the sensor to a sensor cluster, and eventually would indicate the predictor to ignore that sensor (i.e., to ignore the variates from that sensor). In implementations, the adaptor can switch data (such as by changing to an alternative sensor, changing to alternative variates).
• If a noise-causing effects originates from components of a "noise cluster", the noise pattern can be determined from one component and that noise pattern can be used to cancel the noise from the data of the other components in that cluster. Alternatively, in a dynamic relation, components may influence each other only temporarily. If such a relation is known, the adaptor can apply the appropriate weights. For example, to have less weights to parameter data during certain operations.
[0027] A computer-implemented method is disclosed to identify a behavior pattern of industrial equipment that has components and that performs an industrial process. During process performance, a pre-trained machine-learning tool is performing a behavior prediction.
[0028] By using an adaptor module, the computer is receiving original parameter data from data providers that belong to the components of the industrial equipment. The computer is then clustering the original parameter data - the variates to be more accurate - to data clusters according to relations between the components. Static relations between the components are pre-defined by the industrial process, with the definition being completed before the industrial equipment starts to perform the industrial process. Dynamic relations between the components are defined by parameter data that are being received, after the industrial equipment has started to perform the industrial process (i.e. during the process performance).
[0029] As the label suggests, static relations between components are relations that are therefore (i) known to the method-executing computer when it starts receiving data and (ii) that do not change during the process (the definition of the relations has been completed). As the computer executes the method while the industrial equipment performs the industrial process (the industrial equipment has started performing the industrial process, (i) receiving data can coincide with starting the process, and the relations remain unchanged during the process.
[0030] On the other hand, dynamic relations between components are relations that may (i) not be known to the method-executing computer when it starts receiving data and (ii) that may change during the process. As the computer executes the method while the industrial equipment performs the industrial process, (i) receiving data and identifying the relations coincides with starting the process, and the relations change during the process.
[0031] The computer is assigning cluster values to the data clusters, and applying the cluster values as weights to the parameter data to obtain adapted parameter data.
[0032] By using a predictor module, the computer is performing the behavior prediction with adapted parameter data.
[0033] Optionally, upon detecting a change in the relations between the components, the computer can repeat the method step assigning, clustering, applying and performing behavior prediction. In other words, a change in the component relations - especially in the relations that are dynamic relations - can lead to a repetition of the clustering steps and of the steps that follow. When during the process performance, a change in the data that the computer receives can lead the computer to detect a change in a dynamic relation between the components and subsequently to perform clustering again.
[0034] Optionally, receiving original parameter data from data providers comprises predicting parameter data by one or more component-specific predictor modules. In the following, clustering the variates of the original parameter data can be expressed in the short version: "clustering the parameter data".
[0035] Optionally, clustering the parameter data is performed by representing the components by component-representations and representing process stages by stage-representations.
[0036] Optionally, clustering the parameter data is performed by representing the components by component-representations, wherein the static relations between the components are identified by parsing a computer-readable equipmentdescriptor document that has pointers to clusters.
[0037] Optionally, during the process performance, the step clustering the parameter data to data clusters is performed once for data clusters that remain unchanged, and wherein the step assigning cluster values to the data clusters is repeated for changing cluster values.
[0038] Optionally, during the process performance, the steps clustering and assigning are performed in repetitions, when a dynamic relation between the components is changed due to changed parameter data that are being received.
[0039] Optionally, clustering the original parameter data to data clusters according to relations is performed for static relations between the components by parsing computer-readable equipment-descriptor documents that are selected from the following: documents that differentiate main and auxiliary components, piping and instrumentation diagrams for the equipment, mechanical flow diagrams of the equipment, part or material lists, process specifications, recipes for the product, batch control documents, and floor plans.
[0040] Optionally, clustering the original parameter data to data clusters according to relations is performed for static relations between components that are arranged in a processing chain, that experience common environmental conditions, that are arranged in physical connection, or components that have been installed to the equipment within a common time-frame.
[0041] Optionally, clustering the original parameter data to data clusters according to relations is performed based on parameter data that indicate, selected from the following: quality and quantity of materials that are exchanged between components, equality of activity states or ON/OFF states that the components can assume, data reliability of the parameter data, distortions in the parameter data, and correlations within parameter data.
[0042] Optionally, clustering is repeated when the dynamic relations change.
[0043] From a different perspective, a computer-implemented method is provided to identify a behavior pattern of industrial equipment that has components, and that performs an industrial process. The components of the industrial equipment are associated with data providers that correspond to variates in multi-variate original parameter data.
[0044] Relations between the components belong to four relation groups that are differentiated according to the following criteria: the presence of interactions between the components, with the interactions being defined by the industrial process, or the absence of such interactions that are process-defined; and the presence of changes in these interactions after the industrial equipment has started to perform the industrial process, or the absence of such changes (in these interactions).
[0045] When the industrial equipment has started to perform the industrial process, a pretrained machine-learning tool is performing a behavior prediction.
[0046] In a receiving step, the computer receives the multi-variate original parameter data from the data providers. In a clustering step, the computer clusters at least some of the variates of the multi-variate original parameter data to one or more data clusters . The computer is clustering such that - in least one data cluster - the majority of variates belong to components in the same relation group.
[0047] In an assigning step, the computer assigns cluster values to the data clusters.
[0048] In an applying step, the computer applies the cluster values as weights to the original parameter data to obtain adapted parameter data.
[0049] By operating the pre-trained machine-learning tool with the adapted parameter data at the input of the tool, the computer is performing the behavior prediction with adapted parameter data, and is thereby identifying the behavior pattern that become available at the output of the pre-trained machine-learning tool.
[0050] The use of the method is presented to identify the behavior pattern of industrial equipment, in that the components are arranged to sequentially forward intermediate products.
[0051] A computer system (with an adaptor module and with a predictor module) is associated with industrial equipment and that is adapted to execute the computer- implemented method to identify a behavior pattern of the industrial equipment.
[0052] A further computer-implemented method to enable a machine-learning tool to identify a behavior pattern of industrial equipment has similar steps, but there is a further step to classify the dynamic relations into relation group. The groups are then applied to differentiate training sets for the machine-learning tool.
[0053] A computer program product that - when loaded into a memory of a computer and being executed by at least one processor of the computer causes the computer to perform the steps of the methods.
[0054] Further disclosed is an arrangement (or machine-computer system) comprising industrial equipment and a computer. The equipment can be an industrial machine and the equipment is adapted to provide parameter data. The computer with an adaptor module and with a predictor module is adapted to process parameter data by performing the computer-implemented method. Brief Description of the Drawings
[0055] FIG. 1 illustrates a writing convention for parameter data that are being clustered;
[0056] FIG. 2 illustrates a block diagram of industrial equipment with components, and also illustrates a computer with an adaptor module and with a predictor module;
[0057] FIG. 3 illustrates a component, a component-representation, and a componentspecific predictor;
[0058] FIG. 4 illustrates clustering by way of examples;
[0059] FIG. 5 illustrates component-representations with connecting lines that symbolize clusters;
[0060] FIG. 6 illustrates a flow-chart diagram of a computer-implemented method to identify a behavior pattern of industrial equipment that performs an industrial process by components;
[0061] FIG. 7 illustrates a flow-chart diagram of activities, including cluster preparation activities and the computer-implemented method of FIG. 6;
[0062] FIG. 8 illustrates a block diagram for a computer that performs behavior prediction with adapted parameter data by separate predictor sub-networks;
[0063] FIG. 9 illustrates a flow-chart diagram of a computer-implemented method to train the machine-learning tool - the predictor module of FIG. 2 - to identify the behavior pattern;
[0064] FIG. 10 illustrates a diagram with two clustering dimensions, each with two alternatives; and
[0065] FIG. 11 illustrates a generic computer.
Detailed
Overview
[0066] The description explains a computer-implemented method (400 in FIG. 6) to identify a behavior pattern Z of industrial equipment that has components and that performs an industrial process. During process performance, a pre-trained machinelearning tool such as a neural network predicts the behavior (cf. the last method step 450).
[0067] While operating the neural network is mainly driven by data (such as training data to train the network, production data to use the network during process performance, etc.), the above-mentioned deficiencies are addressed by preprocessing data (cf. steps 420, 430 and 440 in FIG. 6). Consequently, the data is provided to the ML tool in a quantity (and quality) that is optimized for the prediction.
[0068] One of the method steps in pre-processing is clustering (cf. step 420). Clustering is mainly driven by rules, not by data. In that sense, there is a balance between data- driven computation (cf. step 450) and rule-driven computation (cf. steps 420, 430, 440). Correlations may arise from confounding data or from biasing data, but for clustering, such correlations would be ignored.
[0069] This pre-processing involves certain activities at a still earlier points in time (pre-pre- processing, cf. FIG. 7 at item 700) and/or during processing.
Notation for clusters
[0070] FIG. 1 illustrates a writing convention for original parameter data [X] that are being clustered by a computer (as well as for adapted parameter data [Y] that result from clustering). As used herein, the description notes multi-variate parameter data by single square brackets [ ], such in the expressions [X] and [Y], with [X] and [Y] being vectors with V single variates Xv from v = 1 to v = V, and single variates Yv from v = 1 to v = V as well. In the example of FIG. 1, each vector has V = 7 variates.
[X] = [XI, X2, ..., Xv, ..., XV], and
[Y] = [Yl, Y2, ..., Yv, ..., YV]
[0071] Multi-variate parameter data [X] and [Y] (or sub-sets thereof) can be input to processing tools, such as to neural networks (cf. predictor module 500 in FIG. 2). [0072] As it will be explained for the computer-implemented method (item 400 in FIG. 6), original parameter data [X] (shown on the left side) are adapted to adapted parameter data [Y] (shown on the right side).
[0073] Pre-processing (or "adapting") comprises
• clustering variates Xv (cf. step 420 in FIG. 6),
• assigning cluster values Cl, C2 (cf. step 430),
• applying cluster values Cl, C2 as weights to the data (cf. step 440).
[0074] In other words, the cluster values Cl, C2 can act as weights Wv (v = 1 to V) to the neural network. The example in FIG. 1 assumes to have two clusters, with values Cl and C2.
[0075] The notation Yv for a single variate in the adapted parameter data can be considered as a combination of Xv and Wv, and that combination can be supplied to the neural network. The figure symbolizes such a combination by the @ symbol. Combinations can be defined, for example as follows:
• The network receives original Xv and Wv, wherein Wv is an indicator that the neural network can ignore Xv.
• The network receives original Xv and Wv, wherein Wv is an indicator that the neural network has to process Xv with its original value.
• The network comprises two sub-networks that have been trained differently. The network receives original Xv, and Wv is an indicator that Xv has to be processed by a particular sub-network. In that sense, the sub-networks can be cluster-specific sub-networks (cf. FIG. 8).
• The network receives original Xv and Wv, wherein Wv is an indicator that the neural network has to process Xv in combination with a further variate (that is not variate v). In that sense, Wv is an indicator to process parameter data together. The identification of the further variate could be obtained by machine-learning as well.
• The network receives original parameter data [X] with a weight set [W] (same notation), wherein [W] comprises indicators that a first sub-set of [X] | Cl variates has to be processed separately from a second sub-set of [X] | C2 variates. Here in the text, the notation with the vertical bar | and with cluster values Cl and C2 symbolizes that parameter data [X] can be clustered into disjunct parameter data. Such a split (with separately processing adapted parameter data [Y] | Cl and [Y] | C2) can be used to differentiate causations from random correlations (details in the section "correlation ..." below).
• The network receives original parameter data [X] with a weight set [W], [W] can indicate to use (at least a particular part of) the neural network that has been trained with certain training (for example to accommodate missing historical data).
• The network can be a calculation function that calculates a particular value, such as a value VALUE that represents a parameter (e.g., a temperature). This calculation function VALUE can be a function of other data, such as VALUE = function(INPUT, INTERMEDIATE_VALUE, ..., OUTPUT, FML). While INPUT, INTERMEDIATE_VALUE, ..., OUTPUT would be available as parameter data, that means as a variate, FML would be a function to by calculated by machine learning (in the production phase, or even earlier in a training phase). In that case, the weight set [W] would serve as additional input, even as training data.
[0076] The skilled person can use modified approaches. For example, if the predictor module would still have to be trained, [W] may indicate to use a particular training set for training.
[0077] In terms of adapted parameter data [Y], the notation can be
[Y] = [X1@W1, X2@W2, ..., Xv@Wv, ..., XV@WV]
[0078] The weights are cluster-specific, and in the example of FIG. 1, there should be two clusters and hence two cluster values Cl and C2. In the example, XI, X3 and X4 would have been clustered to a first cluster with the cluster value Cl, and X2 and X5 would have been clustered to a second cluster with the cluster value C2.
[0079] The cluster values can be expressed by a cluster vector [C], [C] can have V elements that indicate cluster values for each variate Xv. For simplicity of explanation, the description occasionally writes cluster vector [C] merely as a data structure that identifies WHAT parameter data (what variate) should be clustered, but not HOW (i.e., cluster value). As parameter data changes over time, multiple cluster vectors [C] could be considered as a matrix, with different vectors for different points in time. The description will explain such scenarios below when it explains who dynamic relations between components can define the clusters.
[0080] Cluster vector [C] that expresses the separation only would, for example, be a vector with a cluster identifier to indicate cluster membership of a variate, and with "0" otherwise, for example [1, 2, 1, 1, 2, 0, 0], In the example, variates XI, X3, X4 would be clustered to cluster 1, and variates X2 and X5 would be clustered to cluster 2. Variates X6 and X7 would not be belong to any cluster.
[0081] Cluster vector [C] that expresses the cluster values would comprises the values (such as Cl and C2) and would be "zero" or "neural" otherwise, for example [Cl, C2, Cl, Cl, C2, neutral, neutral]. The skilled person can use different notations.
[0082] Not all variates do have to belong to a particular cluster. In the example, nonclustered variates X5, X6, X7, X8 and X9 would not be combined with cluster values (cluster neural variates). For such non-clustered variates, the rules is: Yv = Xv.
[0083] It is also possible, to have particular variates assigned to two or more clusters.
Time-series
[0084] In many situations, parameter data would change over time, so that the vectors change as well. {[X]} and {[Y]} are simplified notations for multi-variate time-series, in that the variate Xv may have different values for consecutive time-slots tm (m = 1 for the first slot, to m = M for the last slot).
[0085] Clustering can change over time, so that a particular variate Xv may changes its membership to a cluster (within tm to tM). In terms of cluster vector [C], it would become a time-series {[C]}.
[0086] The description will explain below that clustering (as computation activity) depends on relations that are static or dynamic. In terms of the notation with cluster vector [C] stands for a cluster that is static (i.e., that remains while the equipment performs the process), and the time-series cluster vector { [C]} stands more for a cluster that is more dynamic (i.e., that may change during the execution of the process).
[0087] Just for example, {[C]} = {[1, 2, 1, 1, 2, 0, 0], ..., [1, 2, 1, 2, 1, 0, 0], ...} would symbolize a dynamically changing cluster setting (variates X4 and X5 would change membership). Such and other changes in [C] can trigger the computer to execute the method again (cf. FIG. 7), or to repeat at least the steps from clustering (step 420 in FIG. 7).
[0088] Although parameter data change over time in many situations, some parameter data is relatively static. Using parameter data and cluster data in time-series is therefore not mandatory.
[0089] It is noted that - while the variate Xv may have different values for consecutive time-slots tm - the different values within a variate Xv are not being clustered. In other words, the computer clusters variates but not values within the variates. In that sense, clustering means selecting variates for being processed together to perform the behavior prediction.
Membership in clusters
[0090] Memberships of particular variates can be disjunct (Xv belongs to a particular cluster, but in one cluster only). In that case only one cluster value Wv is applied.
[0091] A particular variate can be member of more than one cluster. In that case multiple cluster values have to be applied. Multi-cluster membership can be defined by [C] that is expanded to a matrix.
[0092] Two or more clusters may have some variates in common.
Wording convention
[0093] The progress of time has many aspects.
As used herein, the term "stage" differentiates processing time intervals for industrial processes (first stage, second stage and so on).
• The term "phase" is exclusively used to differentiate time intervals to operate neural networks.
• The time can be divided into "past", "present" and "future". Unless indicated otherwise, the figures show the operation of industrial equipment and of computers at present.
[0094] The term "module" stands for the implementations of a computing function by one ore more computers. The terms "component" and "equipment" stand for hardware.
[0095] The term "representation" stands for a data-structure by that a computer represents a component, a stage, or the equipment (as a whole).
[0096] The skilled person is familiar with such data-structures, and can look up for details in the literature with metaphorical terms such as "Digital Twin" or the like.
[0097] The term "clustering" stands for a method step that is performed by a computer when data has been received.
[0098] This has to be differentiated from activities that are performed in advance, "cluster preparation" in the following. Cluster preparation can be computer-implemented (cf. FIG. 7) and can be human-implemented.
[0099] In case of computer-implemented cluster preparation, the computer can process a computer-readable equipment-descriptor document that has pointers to clusters, at least implicitly. The computer can parse this document (cf. step 710 in FIG. 7). Such descriptor documents can be, for example,
• piping and instrumentation diagrams (PID) for the equipment,
• mechanical flow diagrams (MFD) of the equipment,
• part or material lists, process specifications, recipes for the product, batch control documents, or the like, or
• floor plans (for the physical arrangement of equipment components).
[00100] Such diagrams, lists or plans can include verbal descriptors, such as short phrases. It can be assumed that the computer understands their semantics. For example, the phrase "process connected to" implicitly induces a relation that two or more components that participate in the industrial process are "mechanically connected to (each other)". The reverse is however not implicit (mechanical connection not necessarily being a process connection).
[00101] In case of human-implemented cluster preparation, clusters and rules for clustering can be prepared by human experts. Such an expert is familiar with the industrial domain (e.g., the particular equipment performing a particular process, as in FIG. 2). For example, the domain expert could establish a database entry with a list of components and could annotated them to be "main" or "aux" components. The computer can then receive input from a domain expert (cf. step 720 in FIG. 7).
Equipment and data
[00102] FIG. 2 illustrates a block diagram of industrial equipment 100 with components 110- n, 120-k, and also illustrates computer 200/500 with adaptor module 200 and with predictor module 500.
[00103] The components are differentiated as main components 110-n (n = 1 to N) that perform the industrial process and auxiliary components 120-k (k = 1 to K) that support the performance. As illustrated by bold arrows, the industrial process can be, for example, a material processing process by that raw material (left side of the figure) is being processed via intermediate material to a final product (right side). By way of example (thin arrows), auxiliary component 120-k should provide oil or other lubricants to main components 110-2 and 110-n.
[00104] Index n is the "main index" that identifies the main components, and index k is the "auxiliary index" that identifies the auxiliary components.
[00105] The industrial process can be differentiated into a main process (bold arrows, main components) and one or more auxiliary processes (thin arrows, auxiliary components). There should be some interdependency, both main and auxiliary processes are necessary for the success of the industrial process as a whole. But they have different importance. For example, auxiliary component 120-k may run out of oil, this is not problematic for the main process, at least for some time. On the long run, oil must be provided, otherwise a component would be damaged.
[00106] Main and auxiliary processes can be differentiated to have different process stages (or process steps, or sub-processes). One and the same component can perform stages sequentially or in parallel (e.g., heating material and subsequently mixing it with a further material, or heating and mixing simultaneously).
[00107] Circle symbols stand for data providers that are simply numbered from 1 to 7, corresponding to v = 1 to V variates Xv in the multi-variate original parameter data. Having V = 7 is merely a convenient simplification, in real equipment, V is expected to be much higher.
[00108] The data providers for original parameter data [X] can be implemented, for example, by sensors, by data outputs of actuators, by shop-floor systems etc. The data providers can be regarded as sub-components. The skilled person is familiar with devices (that comprise on-board processors and that are software/firmware controlled).
[00109] By way of example (and simplified here to sensor readings),
• original parameter data XI (from data provider 1) should be a sensor reading regarding the operation of component 110-1 (e.g., a temperature value);
• original parameter data X2 should be a sensor reading regarding intermediate material that leaves component 110-1 and that arrives at component 110-2;
• original parameter data X3 should be a sensor reading regarding the operation of component 110-2 (e.g., a temperature value as well);
• original parameter data X4 should be a sensor reading regarding the operation of auxiliary component 110-k (e.g., oil delivery data), the sensor for X4 would be a sub-component to component 110-k;
• original parameter data X5 should be a sensor reading regarding the operation of component 110-n (e.g., a pressure value);
• original parameter data X6 should be a sensor reading regarding intermediate material that leaves component 110-n and that arrives at component 110-N; and • original parameter data X7 should be a sensor reading regarding the operation of component 110-n (e.g., its energy consumption).
[00110] The notation Xv with a single variate index v is a simplification, it is also possible to use double-indices that identify the component (like indices n and k) and identify a data provider that belongs to a particular component.
[00111] As mentioned, the time by that the original parameter data become available should be the present time (i.e., during the operation of equipment 100).
[00112] It is however possible - during operation - to obtain predicted parameter data (in the function of original parameter data) that will be applicable for future. Such prediction will be explained in connection with FIG. 3.
[00113] There are variate-to-component associations, for example in the following:
• data XI associates to component 110-1
• data XI associates to components 110- and 110-2
• data X3 associates to component 110-2
• data X4 to associates to auxiliary component 120-k (i.e., to the auxiliary component with a particular index k)
• data X5 to 110-n
• data X6 to 110-n and 110-N
• data X7 to 110-n
[00114] The computer can obtain the associations from the above-mentioned equipmentdescriptor document (that has pointers to clusters). Such associations can be used as a basis for clustering.
[00115] Adaptor module 200 receives original parameter data [X] and provides adapted parameter data [Y], cf. steps 410, 420, 430 and 440 in FIG. 6. With the variate index v remaining unchanged, [Y] can be considered to be a function [X] --> [Y] and that can be expressed as a [Y] = [X]@[C], cf. FIG. 1.
[00116] Predictor module 500 receives [Y] and performs the behavior prediction for equipment 100 (and/or for particular components), cf. step 450 in FIG. 6.
[00117] Computer 200/500 performs the computer-implemented method (400 in FIG. 6) to identify behavior pattern Z of industrial equipment 100.
[00118] The small icon at the right side of predictor module 500 merely symbolizes that computer 200/500 would inform operator 190 about the result (i.e., the identified behavior pattern Z). In an alternative, the result would be converted to a control signal to equipment 100 (or to an individual component). The skilled person is familiar with control engineering and the like.
[00119] The behavior pattern can comprise, for example, the following:
• a parameter deviation (from a reference component, from reference equipment, at present or for the future)
• the indication of a severity of such parameter deviation (e.g., "not severe" if no equipment failure is to be expected, or with a warning if failure would follow)
• an estimation of a quantity in the output (last component) or quality indicator [00120] Although symbolized by a letter Z only, the behavior pattern can be represented by a more complex data structure. For example, the parameter deviation could be provided with meta-data, such as the identification of the parameters (i.e., v), a quantity of the deviation (e.g., from a reference), with a time-frame of actual or expected occurrence, and so on.
[00121] The skilled person is familiar with such predictor modules, for example from WO 2022/258835 Al.
[00122] As the behavior pattern Z is that of industrial equipment that performs an industrial process by components, operator 190 can also see a pattern at the level of the components (i.e., deviations at component granularity).
[00123] As already mentioned with FIG. 1, adapting comprises clustering and applying cluster-specific weights. In that case, adaptor module 200 adapts [X] to [Y], However, there is a synergistic side-effect that can be used further: Since adapting involves clustering, the components and/or the process stages can be visualized to operator 190 in an adapted representation (cf. FIG. 5 for more details).
Root-cause
[00124] Operator 190 may also be interested in understanding the causes for particular patterns. If, for example, equipment 100 fails as a whole, operator 190 might inspect the last component in the chain first. The "trouble making" component may however be a component at an earlier processing stage (i.e., a component with a lower component index n), or may be an auxiliary component.
[00125] For example, if clustering identifies, for example, missing oil supply as a nonexpected event (cf. FIG. 2, oil from component 120-k), clustering may highlight a broken oil supply component.
[00126] Operators seek to identify the root cause (no oil supply because the power supply to a motor failed). A cluster-adapted representation could also indicate correlations and even could indicates causal relations, up to root causes (cf. a path 250-A, 250-B, 250-C in FIG. 5).
[00127] Adaptor module 200 is illustrated by bold lines, simply because it performs method steps that are to be explained with more details below.
Basic Clustering
[00128] As clustering plays a role, it should be noted that - from a high-level perspective - the parameter data from the main components form a first cluster, and the parameter data from the auxiliary components form a second cluster. Such simply main/aux clusters can be obtained, for example, by the domain expert writing a database table with the components in a description document, and by the computer processing that document (e.g., by parsing, cf. step 710 in FIG. 7).
[00129] The relations between the components are static relations when - during process performance - the relations remain the same. The following examples should explain that:
Component 110-n receives intermediate material from component llO-(n-l), and component 110-n provides (different) intermediate material to component 110-(n+l). This is applicable during the process performance.
• Component 110-k provides oil or the like components 110-2 and 110-n, etc.
[00130] The relations between the components are dynamic because there are changes, for example in the following:
• The inter-component transfer of intermediate material may depend on a quality of the intermediate materials, cf. X2 and X7. That quality is not constant all of the time.
• Component 110-k provides oil, but that oil supply could be interrupted. Or, an oil receiving component may suddenly demand for more, due to a leakage or the like.
[00131] The computer is able to obtain the (static and dynamic) relations from the computer-readable equipment-descriptor document.
[00132] Changes can be normal or can point to an undesired situation (such as failure events).
[00133] Individual variates Xv in original parameter data [X] can be clustered as well, for example, for parameters at the first processing stage (e.g., XI, X2), at intermediate processing stages (e.g., X3, X4, X5), and at the last processing stage (e.g., X6, X7). There are variates that are specific to particular components (e.g., first cluster XI, X2, X3, X5, X6, X7), and that are specific to multiple components (e.g., second cluster X4, 110-2, 110-n and 120-k).
[00134] To evaluate the relations (no matter if the relation is static or dynamic), it is possible to perform the evaluation in the computer. The description continues with that aspect. It is noted that using the computer allows clustering (by the domain expert) but also allows clustering otherwise (such as according to rules, while the domain expert is not available).
Components and representations
[00135] FIG. 3 illustrates component 110-n, component-representation 210-n, and component-specific predictor 310-n. Using component-representation 210-n and component-specific predictor 310-n is optional. While the index "n" refers to the main components (cf. FIG. 2), the approach can also be applied to the auxiliary components (i.e., to components 210-k).
[00136] The computer represents component 110-n in component-representation 210-n that can comprise for example:
• present parameter data (from the component, such as [X]),
• historical parameter data (from the component, such as [X] as well),
• one or more variables that indicate the actual state of the component,
• cluster information that indicates the membership of the component in a component cluster and that indicates a type of the cluster (e.g., cluster with main component, cluster with auxiliary components, etc. as mentioned with FIG. 1), hence the parameters data can be clustered accordingly ,
• data regarding the process (or sub-process) in that the component participates, and
• event probabilities, such as failure probabilities
[00137] That data can be used in the step clustering 420 (method 400, FIG. 6) as well as in cluster preparation activities (cf. 700 in FIG. 7).
[00138] Component-representations 210-n can be visually presented to a computer user, that is however not required. FIG. 3 symbolizes the data by rectangles.
[00139] Component-specific predictor 310-n for that component can identify componentspecific behavior at present (or in the future, by prediction), such as deviations, etc.
[00140] As used herein, predictor 310-n provides predicted parameter data that is parameter data that is expected to be applicable in the future.
[00141] For example, predictor 310-n can calculate a probability Pn that the component may fail (within a certain time-interval into the future). The failure probability Pn is given here simplified. In many situations, Pn would be accompanied by further data (e.g., Pn = 80% within the next hour).
[00142] For example, component-representation 210-n may have a first state variable that is set to the present state ON (i.e., the component is operating), and may have a second state variable that is set for the future (to a certain time-interval) to OFF. The first state variable may be derived from parameter data, and the second state variable may be derived from parameter data that is being processed by component-specific predictor 310-n.
[00143] Predicted parameter data can be used as original parameter data [X], In other words, predictor 310-n can have the same function as a data provider introduced above in FIG. 2.
Implementations of the predictor by ML tools
[00144] The skilled person can implement predictor 310-n by a ML tool. This is similar to the implementation of predictor module 500 (cf. FIG. 1), but as a different granularity (parameter data for the components vs. parameter data for the equipment as a whole).
Same principle for process stages
[00145] The principle to use component-representations 210-n is not limited to components. Likewise, stage-representations can be used for process stages, and the predictors can be stage-predictors.
Advanced Clustering
[00146] FIG. 4 illustrates clustering by way of examples, with CLUSTER_1 to CLUSTER_9.
There is no particular meaning in the arrangement order, the figure merely illustrates possible ways to cluster data. The figure thereby also symbolizes how pre-defined rules can be used as clustering criteria. Clustering will be applied to the parameter data, but clustering is best explained by looking at the components that provide the parameter data. Writing "clustering components" therefore also comprises that the variates of these components are being clustered. (The computer can process the representations of the components (e.g. item 210-n in FIG. 3), and the (original) parameter data [X] that are associated with the components).
[00147] Although the description refers to components, parameter data that relates to processing stages and the like can be clustered as well, with the same principles. [00148] Whenever possible, clustering should not be data-driven alone (i.e., performed by a machine learning tool only) because evaluating parameter data may introduce random correlations that are not necessarily causal relations.
[00149] The example in CLUSTER_1 shows a linear cluster (or a "component chain cluster") with component-representations 210-1, 210-2 and 210-3 that are visualized along a graph. The representations are shown as nodes and the relations between them (i.e., the relations between the components that are represented) are shown by direct edges (arrows).
[00150] For the components, event probabilities can be combined, such as the mentioned failure probabilities Pn. Assuming that failures propagate from component to component, such failure probability would be the higher at the end (i.e., at representation 210-3 for the corresponding component. For example, the above- mentioned failure probability P3 would be higher than Pl.
[00151] Parameter data for these components (in that chain cluster) can be clustered accordingly. The cluster vector [C] would, for example, identify the variates that belong to these components. In other words, the static relation between the components (that is the basis for clustering) is the arrangement of the components in a processing chain (to process materials, or the like).
[00152] The example in CLUSTER_2 shows a location cluster with componentrepresentations 210-2, 210-3 and 210-5 for components in proximate physical location. There is similar influence of some common parameters (such as humidity or environmental temperature) to all of the components in the cluster. This is an example for a cluster in that static relations between the components are predefined by the process. The corresponding components stay at their locations at least relative to each other (and in many cases stay also at their location, e.g., in the same factory hall). Likewise, the parameter data would be clustered, and the cluster vector [C] would, for example, identify the variates that belong to co-located components. In other words, the static relation between the components (that is the basis for clustering) is the arrangement of the components at a shared physical location for that like environmental conditions apply to the components.
[00153] The example in CLUSTER_3 shows a physical relation cluster with componentrepresentations 210-2, 210-3 and 210-5 for that the components are in physical connection (210-2 and 210-3, 210-3 and 210-5). There is an assumption that events may be propagated. A failure in a first component (such as over-heating) may trigger a failure in a physically connected second component (domino effect). Or, if one component starts to vibrate, the vibrations could propagate to the other components in the cluster. The description takes CLUSTER_3 as an example for a cluster that can be derived from documentation (either by parsing, cf. step 710 in FIG. 7) or manually (to be available to the computer in step 720). In other words, the static relation between the components (that is the basis for clustering) is the arrangement of the components in physical connection.
[00154] The examples in CLUSTER_4 and CLUSTER_5 show an active-component cluster (ON-cluster) and a non-active component cluster (OFF-cluster). This is symbolized by the acronym ON and OFF (for the states). The probability Pn that an OFF- component fails is lower than the probability Pn that an ON-component fails. The cluster values would be provided accordingly, for example, to let predictor 500 ignore data from OFF-components.
[00155] Establishing that a component is ON or OFF at present time and relatively easy and can be done by pre-defined rules (no machine-learning required). Establishing that a component will be ON or OFF in the future does not rely on ML either, for example by applying rules that consider manufacturing schedules or the like (shop-floor data). Using ML learning (such as in predictor 310-n of FIG. 3) is possible.
[00156] In other words, the dynamic relation between the components (that is the basis for clustering) is the equality of ON/OFF states that the components can assume.
[00157] The example in CLUSTER_6 indicates that a component may be relatively new, in the sense to have been manufactured recently, to have been installed to the equipment recently, etc. As a consequence, historical parameter data to train the ML tool may not (yet) be available, or may not be available in sufficient amount. Therefore, the cluster values may indicate that predictor module 500 may not perform certain operations.
[00158] In other words, the static relation between the components is a common but relatively short time lapsed from installation. Over time, the historical parameter data should become available.
[00159] The example in CLUSTER_7 symbolizes by a line trough [X] that parameter data may become incomplete, due to sensor failure or similar events. If a sensor - that is a sub-component to a component - stops providing data, the process at the equipment may continue. As already mentioned above, predictor 500 (cf. FIG. 1) might not be prepared for such situations, and it might derive an incorrect conclusion. To mitigate that, parameter data [X] are separated, at least into reliable data (cluster [X] | reliable) and potentially unreliable data (cluster [X] | unreliable).
[00160] In terms of reliability, the relation between the components would be dynamic.
[00161] The example in CLUSTER_8 symbolizes that parameter data may be disturbed. As mentioned above, there could be noise or other effects. For example, during operation a component may vibrate and a sensor might provide not only the sensor-specific signal but also a vibration specific signal. The ML tool might process noisy signals and clean signals differently. At best, the ML tool has already been trained by noisy data, because the skilled person occasionally augments training data by noise. In this case, the presences of noise turns [X] to [X] | noise.
[00162] In other words, the dynamic relation between the components the property that the parameter data of the components are affected by distortions.
[00163] The example in CLUSTER_9 symbolizes the identification of correlations between two or more variates (in original parameter data [X]). For example, of a first variate Xa would should a U-shaped drop in its value, and a second variate would show a similar U-shaped drop in its value as well, even with some delay, the skilled person would see a correlation. Automation is possible, and there are many mathematical approaches to identify correlation, the calculation of the Pearson correlation coefficient (PCC) is just an example.
[00164] In other words, the dynamic relation between the components the property that their parameter data correlate.
Further examples for advanced clustering
[00165] There are many more clustering approaches that are well-known so that they can be omitted from the figures.
[00166] A further example would be clustering by data processing, such as by using the well- known k-nearest neighbors algorithm, using hierarchical clustering and other approaches.
[00167] A further example is a cluster of variates that do not change frequently, but for that changes occur. For example, the parameter for oil supply (from component 120-k) may suddenly change, and with that a follow-up parameter will change as well.
[00168] A further example would be the identification of so-called process signatures. As the process (or a sub-process, or a stage) would not show its characteristic signature, the process could be considered as deviating (even as abnormal), with the consequence that process data can be clustered to be treated separately.
Correlation and causation
[00169] For example, for two components (Alpha and Beta) that both comprises heaters, the temperature readings for Alpha and Beta correlate to energy consumption for Alpha and Beta, respectively, but there can also be a correlation between the temperature of Alpha to the consumption of Beta, the temperature of Beta to the consumption of Alpha. In other words, there would be V = 4 variates, with two correlations that indicate causal relations, and with two correlations that are rather random correlations, without causal relation. In such a simplified situation, the operator can easily see the sort out any misleading correlation.
[00170] In case of clustering (here with two variates in each cluster Alpha and Beta), the correlations across cluster borders (between Alpha and Beta) can easily identified as accidental. Or, the cluster values may indicate that variates for Alpha need to be processed from variates for Beta separately. In the above notation, the weights (cluster values) would indicate a separation into [X] | Alpha and [X] | Beta (i.e., [Y] = [Y] | Alpha and [Y] | Beta).
[00171] However, the quantity of V = 4 would not correspond to reality, V is much higher. Clustering by the computer steps in here. If a correlation across cluster borders is detected, it can be marked as such. Or, the cluster value (here for cluster Alpha and for cluster Beta) may prevent the computer from evaluating parameter data in a mixed fashion. If the cluster values are applied as weights (step 440), predictor module 500 would, for example, process the parameter data from both clusters separately. This approach limits the changes for mis-interpreting behavior.
Cluster types
[00172] The examples (i.e., CLUSTER_1 to CLUSTER_9 in FIG. 4 as wells as the examples without figures) can be regarded as cluster types and some of these types show up again in the description of the next figure.
Clustering and component-interrelations
[00173] FIG. 5 illustrates component-representations 210-1, 210-2, 201-3 and 210-4 for the corresponding components (omitted here for simplicity). FIG. 5 also illustrates clusters: symbolically by connection lines.
[00174] These bold lines indicate current knowledge that can be used for clustering. In other words, the figure shows the representations by nodes (in a graph), and shows clustering by edges (in that graph). The illustration can be regarded as a knowledge graph, that is a dynamic one because during a process instances (i.e., from tl to tM), some of the relations (and hence the clusters) will change.
[00175] Representation 210-1 shows that the corresponding component is located at "location A" and that its current state is ON.
[00176] Representation 210-2 shows that the corresponding component is located at "location A", that it is component that participates in performing a sub-process (1), and that its current state is OFF. [00177] Representation 210-3 shows that the component participates in performing subprocess (1) (or sub-stage), and that its current state is ON. The location is not specified.
[00178] Representation 210-4 shows that the component is located at B, and that its state is ON.
[00179] There is a static relation between 210-1 and 210-2 because the location A is shared. The components should be stationary components, not vehicles. This corresponds to the location cluster A (cf. CLUSTER_2).
[00180] There is a dynamic relation between 210-1 and 210-4, according the state, both ON, state cluster ON. The state can change, of 210-1 (from ON to OFF) or of 210-4 (from ON to OFF). This is therefore a dynamic cluster.
[00181] There is a static relation between 210-1 and 210-3, according to the process stage, a sub-process cluster (for sub-process (1)). In other words, there is a cluster that is defined according to a process interrelation (both components participate in the same sub-process) or sub-stage.
[00182] Having identified clusters does not yet provide cluster values (cf. Cl, C2 etc. in FIG. 1, steps 420 and 430 in FIG. 6).
[00183] But pre-defined rules can help here. For example, the parameter data from a component in state OFF (e.g., the component that is represented by 210-2) can be weighted with W-values (that let predictor module 500 ignore them). Or, for example, the parameter data from that component in state OFF would have to be considered in the further cluster (the parameter data of the components of the cluster for sub-process (1)).
Cluster preparation activities
[00184] FIG. 5 can be regarded as a partial copy of an equipment-descriptor document (as mentioned, e.g., an instrumentation diagram). The copy is not complete but limited to the information that the computer needs for clustering. Advanced clustering with the process of time
[00185] FIG. 5 takes the component states as an example for a dynamic change, and as an example, for a change that can be predicted to occur in the future. The asterisk * next to "ON" or "OFF" symbolizes that the state can be predicted for a time-interval in the future by a predictor, such as by predictor 310-n (FIG. 3).
[00186] In terms of the cluster vector [C], the following is noted:
• [X] it would change from initially [..., ON, ..., OFF, ..., ON, ..., ON] to, for example, [..., OFF, ..., ON, ..., ON, ..., ON], The ellipsis "..." merely indicate the states ON and OFF are only same of many variates and that variates such as for temperature, pressure, energy consumption etc. would have been taken into account (and would have to be processed by predictor 500 as well).
• [X] can be initially by clustered into parameter data belonging to an initial ON- cluster (i.e., the variates from the component that correspond to representations 210-1, -3- and 4) and to an initial OFF-cluster (component that corresponds to representation 210-2). [C] = [ON, OFF, ON, ON] is simplified, there are multiple variates for each component at vector positions 1, 3 and 4 that would be clustered to the ON-cluster, multiple variates for position 2 as well.
• In case of a change [X], the state changes, the cluster vector [C] would automatically change as well, to [C] = [OFF, ON, ON, ON],
[00187] Predefined rules to assign and to apply weights (steps 430 and 440) can be applied once re-clustering is completed, cf. FIG. 7 with a repetition of method 400 for such situations.
User interfaces
[00188] The human operator of equipment 100 (cf. FIG. 2) can of course inspect data values for the variates because they are available (as parameter data [X], also as multivariate time-series {[X]}. Some of the data is also visible directly. For example, pressure or temperature values can be indicated in user interfaces, and - despite implementations by digital sensors - traditional barometers and thermometers should be available for the operator to look at. [00189] The number of variates would not be V = 7 as in FIG. 1 or FIG. 2, but the number could reach V = 1.000 or even exceed that number. The skilled person is familiar with showing representations (together with parameter data) graphically on computer screens. The representations would usually by arranged on a screen in a topology that imitates the physical locations of the components.
[00190] However, it is also possible to arrange screen icons (or other symbols) according to the clusters. For example, the components in OFF state may be illustrates in different style than the components in ON state.
Guiding the operator
[00191] Assuming that predictor (FIG. 2) has received adapted parameter data [Y] and has indicated a failure event for the near future. Clustering (and processing cluster values) as described played a role.
[00192] The skilled person is familiar with designing user interfaces that show information only on request (for example, to indicate details for mouse-over actions by the user who could be operator 190). Eventually, the clustering (with the mentioned clusterspecific highlighting) invites the user to follow a path (in the graph on the screen, the graph is similar to that of FIG. 5) that is cluster-biased.
[00193] Merely by example, FIG. 5 illustrates dashed arrows 250-A, 250-B and 250-C that are directed from an icon that indicates the state of component #3 (that is represented by item 210-3), indicates the stage or sub-process of that component #3, indicates that components #3 and #2 (cf. item 210-2) are statically related (both participate in the same stage), and also indicate by the last arrow 250-C that component #2 is OFF. This is not yet an indication of a behavior that is potentially a failure (of equipment 100), because the behavior indication is performed by predictor 500. But the arrows could give the operator a hint to have a closer look at component #2 that is OFF.
[00194] Based on clustering (in step 420) and based on the weights (steps 430 and 440), predictor 500 indicates the behavior. If the behavior would be "failure" or the like, the operators can follow the arrows to see more details. The last arrow may point to an approximation of the root cause for the failure.
[00195] The skilled person can implement the user interface otherwise, not necessarily with arrows as on the figure.
Method to identify the pattern
[00196] FIG. 6 illustrates a flow-chart diagram of computer-implemented method 400 to identify behavior pattern Z of industrial equipment 100 that performs an industrial process.
[00197] As mentioned above, industrial equipment 100 has components 110-n, 120-k and performs an industrial process. During process performance, pre-trained machinelearning tool (cf. predictor module 500 in FIG. 2) is performing a behavior prediction. The computer has two modules: adaptor module 200 and predictor module 500, cf. FIG. 2.
[00198] In receiving step 410, the computer (i.e., adaptor module 200) receives original parameter data [X] from data providers that belong to the components of the industrial equipment.
[00199] In clustering step 420, the computer (i.e., adaptor module 200) clusters original parameter data [X] to data clusters - such as [X]@[C] in FIG. 1 - according to relations between the components. Step 420 is illustrated with an arrow that is labelled [C], This indicates that a cluster vector [C] can be received prior to step execution.
[00200] Static relations between the components are pre-defined by the industrial process, and dynamic relations between the components are defined by parameter data that are being received in step 410.
[00201] In assigning step 430, the computer (i.e., adaptor module 200) assigns cluster values (cf. Cl, C2 in FIG. 1) to the data clusters. In applying step 440, the computer applies the cluster values as weights [W] to the parameter data [X] to obtain adapted parameter data [Y]). [00202] In performing step 450, the computer (i.e., predictor module 500) performs the behavior prediction with adapted parameter data [Y], This step results in behavior pattern Z.
Activities to prepare method execution
[00203] FIG. 7 illustrates a (flow-chart) diagram of activities, including cluster preparation activities 700 and computer-implemented method 400. The box 400 on the right side stands for method 400, but details are omitted for simplicity. Simplified, the result of cluster preparation activities 700 can be identification of original parameter data [X] that would have to be clustered in a particular way. The identification is symbolized here by cluster vector [C], cf. FIG. 1 and FIG. 6. In other words, [C] symbolizes the result of activities 700, but the vector format is just an example.
[00204] Cluster vector [C] can be available even before receiving parameter data [X] in step 410. This in particular relevant for [C] (or part of it) that is based on static relations. Method 400 can be repeated for new parameter data arriving (i.e., receiving 410 in FIG. 6) but also when [C] changes.
[00205] Cluster preparation activities 700 are presented as separate steps that are computer-implemented, not as a mandatory flow.
[00206] Activity 710 stands for the computer parsing a document, such as the above- mentioned description.
[00207] Activity 720 stands for the computer receiving input from the human domain expert.
[00208] Activity 730 stands for the computer parsing parameter data according to the availability while the equipment performs the industrial process. By simple statistical analysis, a computer can divide parameter data (i.e., particular variates) into groups. For example, the parameter of a first group may change during first interval but may remain unchanged during a second, subsequent time interval. The time intervals may be identical to particular process stages. [00209] Although FIG. 7 shows activities 710, 720 and 730 separately, but they could be combined. For example, in iron and steel industry, industrial equipment (cf. FIG. 1) could be a furnace. In a first stage, it is being charged with materials, in the second stage it will be heated up, and so on. There should be documents that describe each stage in much detail, and the computer can parse them to find parameter data for a "first stage cluster", for a "second stage cluster" and so on, activity 710. The cluster composition would be reviewed by domain experts, so that the computer can process both clusters for some parameter data has been added or removed manually, activity 720. Historical data should be available so that the computer may identify further parameter data (that has not yet been clustered) for inclusion (or exclusion) from the stage clusters, activity 730.
Network separation
[00210] FIG. 8 illustrates a block diagram for a computer that performs behavior prediction with adapted parameter data by separate predictor sub-networks 510 and 520 that belong to predictor module 500.
[00211] As already explained with FIG. 2, multi-variate parameter data [X] and [Y] or subsets thereof can be input to processing tools, such as neural networks in the function of predictor module 500. In the example of FIG. 8, [X] going to predictor module 500 is omitted for simplicity, by [Y] is presented in the two clusters of FIG. 1. Yl, Y3 and Y4 should belong to cluster Cl, and Y2 and Y5 should belong to cluster C2. The weights W1 to W5 can indicate cluster membership.
[00212] Sub-network 510 processes data from the first cluster (i.e., Yl, Y3 and Y4) and subnetwork 520 processes data from the second cluster (i.e., Y2 and Y5). The same principle can be applied to further clusters and to further sub-networks.
[00213] Sub-networks 510 and 520 perform (cf. step 450) the behavior prediction with adapted parameter data, but the result is not (yet) the identified behavior pattern of industrial equipment 100. But there are preliminary cluster-specific patterns at the outputs, 511 and 521.
[00214] Combiner module 550 receives these patterns and obtains the behavior pattern of industrial equipment 100 as a whole.
[00215] Much simplified, the following scenarios are contemplated
• The clusters and the sub-networks can be associated with the components (N clusters, N sub-networks (in FIG. 8, N = 2 for 510 and 520), N components 110- N, cf. FIG. 1). Combiner module 550 may provide the behavior pattern based on logical relations (e.g., NORMAL operation of equipment if all components show NORMAL; or ABNORMAL if at least one component shows ABNORMAL)
• The clusters and the sub-networks can be associated with the process stages (cf. activity 730 in FIG. 7). Combiner module 550 may provide the behavior pattern based on logical relations as well, according to the same principle (all, at least one)
• The clusters and the sub-networks can be associated with a mixture of components and stages. Logical relations would be set up accordingly.
• The clusters and the sub-networks can be associated with parameter data that are redundant and that are available in parallel. Combiner module 550 would compare the preliminary patterns.
• The parallelization (e.g., of components) and serialization (e.g., of stages) in the topology of equipment 100 is mirrored by the sub-networks as well. Combiner module 550 may provide the behavior pattern based on logical relations as well.
[00216] FIG. 8 can be considered as a cascade with two steps, but further steps can be applied as well. For example, a first sub-network can provide input to a second subnetwork, the second sub-network can provide input to a third sub-network and so on. FIG. 8 symbolizes an optimal second sub-network by a dashed box 515.
[00217] Regarding training the sub-networks, for serially connected sub-networks (e.g., 510 and 515), training can be performed serially (the parameter data receiving subnetwork first), and parallel sub-networks can be trained individually.
Training aspects
[00218] To perform the behavior prediction for equipment 100 in step 450 (cf. method 400, in FIG: 6), predictor module 500 (that is a machine-learning tool) can be trained according to usual procedures, with training sets that comprise historical parameter data and that comprise historical behavior patterns Z.
[00219] It is noted that the steps clustering 420, assigning 430, and applying 430 of method 400 can be performed for historical data as well. The training set would therefore also comprise historical parameter data with cluster values as weights.
[00220] In view of the discussion with FIG. 8 to - optionally - implement predictor module 500 with separate predictor sub-networks (e.g., 510, 520 and - if cascaded also with 515 etc.), the dynamic relation would introduce a constraint to the training data. Dynamic relations do change so that the variety in training data would raise with the dynamics.
[00221] This constraint can be addressed by introducing a further step to classify the dynamic relations into groups.
[00222] FIG. 9 illustrates a flow-chart diagram of computer-implemented method 600 to enable a machine-learning tool (cf. predictor module 500) to identify behavior pattern Z of industrial equipment 100 (that has components 110-n, 120-k and that performs an industrial process), such that machine-learning tool 500 is being enabled to perform behavior prediction during process performance.
[00223] In receiving step 610, the method-performing computer receives historical parameter data from data providers that belong to the components 110-n, 120-k of industrial equipment 100. Step 610 has similar to step 410, but performed with historical data that are original data from historical process performances, from equipment 100, or from reference equipment.
[00224] In clustering step 620, the computer clusters the historical parameter data to data clusters according to relations between the components 110-n, 120-k. As in step 620, static relations between the components are pre-defined by the industrial process, and dynamic relations between the components are defined by the parameter data that are being received (but for historical parameter data). [00225] In step classifying 625 - that is the additional step just mentioned - the computer classifies the dynamic relations into relation groups. FIG. 9 illustrates the step to be performed together with clustering, in parallel. This is convenient, but not required.
[00226] Classifying 625 the dynamic relations into relation groups can be performed by an auxiliary machine learning tool that is a classifier. As the skilled person can easily implement that, such an auxiliary ML tool is omitted from illustration.
[00227] The method continues with assigning 630 cluster values to the data clusters and with applying 640 the cluster values as weights to the historic parameter data to obtain adapted historical parameter data. That is similar to steps 430 and 440, but with historical data.
[00228] In step using 650 the adapted historical parameter data as training set to train the machine-learning tool 500, the computer differentiates training sets for the machine-learning tool according to the classified relation groups. FIG. 9 shows that differentiation by a dashed line from step 625 to step 650.
[00229] The different sub-networks that are explained with FIG. 8, can be trained separately. In view of method 600, step using 650 the adapted historical parameter data as training set can be performed for different sub-networks (cf. 510, 520 in FIG. 8, or others) of machine-learning tool 500 that are being trained for different relation groups separately.
[00230] Training sub-networks separately is also applicable for configurations in that the networks are cascaded. Training data would have to be differentiated as well.
Further aspects
[00231] As explained with details above, the method-executing computer performs behavior prediction with adapted parameter data [Y], Adaptor module 200 performs clustering 420 the variates Xv of the original parameter data [X] to data clusters [X]@[C] according to relations between the components 110-n, 120-k. The computer can use cluster vector [C] as a data structure to implement clustering.
[00232] As method 400 has its purpose in identifying the behavior pattern Z (at the output of the pre-trained machine-learning tool 500), the cluster granularity (i.e., what parameter data belong to what cluster) influences the determination of the behavior pattern Z. The description has already explained aspects (cf. FIG. 4, Advanced Clustering) and has explained that the operation of ML tool 500 can be cluster-specific (cf. FIG. 8, sub-networks 510 and 520). It is noted that clustering (in step 420) can require that the computer "knows" how the components are related.
[00233] Some further aspects of the method focus more on the behavior patter Z as the result, such as for example:
• Accuracy. This aspect could be summarized, for example, in the question to what extent the pattern Z corresponds to reality.
• Timing. This aspect could be summarized in the question whether the pattern Z becomes available early enough to take measures, if needed.
• Computation. The computer may use electrical energy to obtain a pattern that is of no interest, or that would not represent the state of the industrial machine. It is therefore desired to identify a pattern Z that would indicate failures (at present, or in the future), that would indicate inefficiencies etc.
[00234] In a particular cluster, data from a single variate Xv may influence accuracy, timing and computation (influence to one aspect only, or to multiple aspects). In other words, data from a particular parameter may be disturbing to correctly identify Z, or a particular parameter might be essential to the cluster, otherwise Z could not be identified.
[00235] So far, the description has differentiated the component relations along the static/dynamic dichotomy. As there are many different approaches to perform clustering, merely differentiating between static relations and dynamic relations might not be sufficient. In view of the mentioned aspects, identifying the pattern Z could be improved.
[00236] To name only a few further approaches, clustering (step 420) can be performed by
ML tools (that are trained with or without supervision), can be performed automatically or semi-automatically (for example, based on technical drawings of the equipment, on correlations, etc.), or can be performed by human experts (who prepare rules to be applied during clustering 420). But the technical efforts to cluster with ML tools (such as training the ML tool and operating the ML tool) may not be justified for some approaches, but some approaches may by most efficient with ML tools only.
[00237] In the following, the description outlines a further methodology to select the suitable clustering approach. The methodology can be applied to provide criteria for clustering 420, that means at a point in time when clustering 420 is not yet performed, or even earlier (prior to receiving original parameter data [X] in step 410. From a different perspective, the methodology is for preparing the clustering step (i.e., by "pre-clustering").
[00238] That methodology establishes general rules that can be applied optionally:
[00239] The relations between the components can be differentiated into four relation groups according to criteria.
[00240] The first criterium looks at the presence or absence of interactions (between the components) that are defined by the industrial process (i.e., interactions with each other in order to perform the industrial process, process-inherent relations).
[00241] The second criterium looks at the presence or absence of changes in these interactions during the performance of the process (i.e., after the industrial equipment has started to perform the industrial process).
[00242] Clustering (cf. step 420) at least some of the variates Xv of the multi-variate original parameter data [X] to one or more data clusters [X]@[C] is performed such that in at least one data cluster [X]@[C], the majority of variates Xv belong to components in the same relation group (cf. FIG. 10, groups i, ii, iii, iv). Each cluster has a number of variates. Each cluster can be regarded as a set, and the number would be the cardinality of that set. The simplified example of FIG. 1 shows a first cluster with 3 variates XI, X3 and X4, and shows a second cluster with 2 variates X2 and X5. (Cl, C2 can be ignored here in this discussion.). A cluster with 2 variates can't have a majority of variates. For a cluster with 3 variates, two variates form the majority if the variates belong to components of the same relation group. For a cluster with more than 3 variates, the majority rule can be applied likewise.
[00243] As explained, the components have data providers (cf. 1, 2, 3, 4, 5, 6, and 7 in the example of FIG. 2) that correspond to variates Xv in multi-variate original parameter data, cf. FIG. 1. Some variates of the parameter data [X] can therefore belong to components of the same relation group.
[00244] By limiting cluster-membership (of parameter data) substantially to components in relation groups, the influence to accuracy, timing and computation are expected to be beneficial.
[00245] It is however noted, that (a) a particular cluster may comprise parameter variates from relations in more than one group (cf. "the majority of"), and that the parameter data from a particular relation group is not yet a cluster. Parameter data from the components that for a particular relation group is rather the starting point to obtain a cluster.
[00246] In the following, the description describes the criteria to differentiate four relation groups ("presence" and "absence" squared) with two dimensions each having two alternatives.
[00247] FIG. 10 illustrates a diagram with two dimensions ORIGIN and DEVELOPMENT (labelled in uppercase letters). Each dimension is given with two alternatives (labelled in quotation marks "..."). Therefore, the diagram has four quadrants (Roman figures i, ii, iii, iv), one for each relation group. The dimensions with their alternatives can be taken as criteria to select an approach to perform clustering.
[00248] As clustering in step 420 is performed prior to performing the behavior prediction in step 450, the data within the adapted parameter data [Y] can change.
[00249] The dimension ORIGIN can be applicable to relations, in two alternatives (presence, absence): • (ii, iii) Relations in "process" are defined by the interactions between the components as they are defined with each other in order to perform the industrial process. In other words, the relations are process-intrinsic and defined by the industrial process.
• (i, iv) Relations in "not by process" are all the others.
[00250] For example, in the alternative "process" (ii, iii), a provider-consumer relation can belong to the industrial process because component 110-1 provides material to component 110-2, and component 110-2 consumes that provided material (cf. for example CLUSTER_1 in FIG. 4). Otherwise, the industrial process can't be performed. A provider-consumer relation remains even in case of failure or during maintenance.
[00251] To name an example for the "not by process" alternative, an air-conditioning component is in relation to the components in the same building (FIG. 4, CLUSTER_2). In principle, the process could be performed in a non-air-conditioned building. But as temperature and/or humidity may influence the performance of the process, such parameters would have to be taken into account.
[00252] For both alternatives, original parameter data [X] for clustering can be available: in the alternative "process", for example, the process definition allows to obtain clustering rules to be applied to parameter data. For example, clustering according to "not by process" would have to cluster variates from parameter data that the computer receives. For example, if the provider components for a given point in time provides material (or not) is a variate that the computer receives.
[00253] As for many industrial processes, the ORIGIN is related to the physical space by that the components are arranged, the ORIGIN dimension (in general) has the space dimension (in particular as a sub-dimension, cf. CLUSTER_2). In other words, the components are arranged in a certain topology (e.g., material flow in a particular direction between components) and the topology can be taken into account.
[00254] The dimension DEVELOPMENT has the two alternatives as well.
(ii, i) In the alternative "change", the relation changes, while the process is ongoing that means after the industrial process has started. Changes are present.
• (iii, iv) In the alternative "constant", the relation does not change while the process is ongoing. In other words, the relation remain constant. Changes are absent.
[00255] The dimension DEVELOPMENT has an aspect in repetitions of method steps.
Predictor module 500 would perform behavior prediction, to pattern Z (step 450) substantially all of the time while the industrial process is ongoing (the industrial equipment has already started to perform the industrial process). Previous steps, such as clustering 420 can be repeated as well, but rather for relations in the "change" alternative and not for relations in the "constant" alternative. As a sidenote, not repeating a step (such as clustering) may save computation (in terms of processor and memory consumption).
[00256] To stay with the example with providing and consuming material, component 110-1 may provide the material to component 110-2 via a pipeline or via a vehicle. For the pipeline, the relation does not change (iii, iv), but for the vehicle there are intervals during that providing is interrupted. (For simplicity, component 110-2 should have a storage so that consuming should be possible at any time. Not providing does not have to be a failure).
[00257] For both alternatives, original parameter data ([X]) for clustering can be available as well: in alternative "change" (ii, i) for example from measurement data (e.g., "Is the vehicle in motion or not?") and "constant" from a process definition, material from 110-1 to 110-2 but not vice versa.
[00258] Besides illustrating both categories with their alternatives in a 2x2 diagram of FIG. 10, but there are observations regarding the availability of data (such as parameter data to be clustered) and regarding the available of information to perform clustering.
[00259] As in the first relation group (quadrant i), ORIGIN = "not by process" &
DEVELOPMENT = "change", the parameter data [X] can indicate that relations change. The relations in the first quadrants can be attributed to be "dynamic". For example, a failure event may turn the provider component out of service so that the provider-consumer relation stops to exists. But non-failure events can change the relations as well.
[00260] Since the relations change, involving human expertise for clustering is potentially inefficient. But the skilled person can apply clustering methods that do not reply on human intervention, such as using k-means clustering. (As a side-note, kNN - k nearest neighbors - might not be applicable here, but knowing non-suitable approaches and ignoring them helps to gain overall efficiency).
[00261] The computational efforts for clustering the relations (in the first quadrant) by computers may be justified. As in quadrant i, clustering can be time-based clustering.
[00262] As in the second relation group (quadrant ii), ORIGIN = "process" & DEVELOPMENT = "change", clustering is possible, but expected to be less likely. Changes in the process would usually be implemented during maintenance, while the process is not yet ongoing. For example, introducing a stand-by provider component (that is redundant to component 110-1) as a new component would be documented in drawings or the like, and parameter data from the new component can easily be added to existing clusters.
[00263] Hence, the computational efforts for clustering the relations (in the second quadrant) by computers may not be justified.
[00264] As in the third relation group (quadrant iii), ORIGIN = "process" & DEVELOPMENT = "constant", the relations between the components (or between the stages) would be known when the process has not yet started. For the example of the providerconsumer relations, the relation would remain the same. The provider component would not turn into the consumer component. These relations can be attributed to be "static".
[00265] The relations can be considered to be pre-defined so that clustering rules can be "hardcoded". Changes over time (i.e., during the performance of the process) of the relations are not to be expected. In the example of the provider-consumer relation, clustering parameter data [X] (from the provider and from the consumer component) would not have to be repeated. However, the parameter data may change, and - to stay with the example - the provider may stop providing or the consumer may step consuming. Predictor module 500 can identify a pattern Z that stands for a failure.
[00266] In that sense, in repetitions of step 450, predictor module 500 continues to process parameter data ([W] to [X] applied, step 440) that had been clustered for third quadrant relations. If the parameter data suddenly show first quadrant behavior, predictor module 500 may qualify the pattern Z as failure.
[00267] The relations in quadrant iii can also be considered to be spatial-based clustering (the location of the component is the decisive factor).
[00268] As in the fourth relation group (quadrant iv), ORIGIN = "not by process" & DEVELOPMENT = "constant", clustering 420 may not be required. (It is even contemplated to avoid clustering.) To give an example, most of the components would be powered by electrical energy, and sufficient supply of energy all of the time is a condition to run the industrial process. For such an assumption, monitoring the consumption with clustered parameters may not be required.
[00269] However, for some processes, energy use can be distributed to certain hours of the day, such as for high-power components (such as heaters or electrical furnaces). For such scenarios, a parameter that stands for energy availability would be part of a cluster in a second quadrant relation.
[00270] Discussing the relations in such relation groups (quadrants in the figure) does not mean that relations remain in the quadrants. If a relation changes its quadrantassignment, re-clustering (in a repetition of clustering) might be the consequence.
[00271] Changes in the parameters (and therefore in the relations) are easily detectable (for example, by using ML-tools to detect failure events), so that, for example, a relation -M - changing from "static" to "dynamic" (third to first quadrant, arrow symbol) can trigger re-clustering.
[00272] Having original parameter data [X] from parameters that belong to components in the same relation group (i.e., i, ii, iii, or iv) is not convenient in view of accuracy, timing and computation. That is - however - not mandatory: a cluster could be identified for relations in more than two relation groups.
[00273] Having explained relation groups between components does not replace the earlier discussions, cf. FIG. 4. Considering relation groups is rather a measure to improve accuracy, timing, and/or computation, but not a requirement.
[00274] The example in CLUSTER_1 shows a linear cluster, the interactions between the components are process-defined and constant, relation group iii.
[00275] The example in CLUSTER_2 shows a location cluster with components in proximate physical location, and the interactions between the components are process- defined and constant, relation group iii.
[00276] The example in CLUSTER_3 shows a physical relation cluster with the assumption that events may be propagated. A failure in a first component (such as overheating) may trigger a failure in a physically connected second component (domino effect). Monitoring data (i.e., variates) in such a cluster would be in relation group iii as well.
[00277] The examples in CLUSTER_4 and CLUSTER_5 show an active-component cluster (ON-cluster) and a non-active component cluster (OFF-cluster). As mentioned above, the cluster values would be provided accordingly, for example, to let predictor 500 ignore data from OFF-components. In view of the relation groups, the relations are dynamic, group i.
[00278] The example in CLUSTER_6 indicates that a component may be relatively new, in the sense to have been manufactured recently. The relation to other components is therefore static, group iii, because the component is part of the equipment. [00279] The example in CLUSTER_7 symbolizes by a line trough [X] that parameter data may become incomplete, due to sensor failure or similar events. This is an example for a highly dynamic cluster, group i, with the relatively high probability that re-clustering may be required (during process performance), group i. Using computer- implemented clustering appears an option to take.
[00280] The example in CLUSTER_8 symbolizes that parameter data may be disturbed. In view of the relation groups, the relations are process-defined (despite the noise), group iii, but some of the noise would certainly not the process-defined, group iv. This is an example, that two relation groups may be applicable.
[00281] The example in CLUSTER_9 symbolizes the identification of correlations between two or more variates (in original parameter data [X]). The identification of correlations etc. can be regarded as steps to perform clustering automatically. The variates can be related to components in groups i or ii.
[00282] Considering the relations between the components in four relation group (not only according to the static/dynamic dichotomy) does not change the approach for the training. The principles to train the ML tool as explained with FIG. 9 and method 600 remain applicable.
Generic computer
[00283] FIG. 11 illustrates an example of a generic computer device which may be used with the techniques described here. FIG. 11 is a diagram that shows an example of a generic computer device 900 and a generic mobile computer device 950, which may be used with the techniques described here. Computing device 900 is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. Generic computer device may 900 correspond to the computer system 200/500 of FIG. 2. Computing device 950 is intended to represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smart phones, driving assistance systems or board computers of vehicles and other similar computing devices. For example, computing device 950 may be used as a frontend by a user (e.g., an operator of a blast furnace) to interact with the computing device 900. The components shown here, their connections and relationships, and their functions, are meant to be exemplary only, and are not meant to limit implementations of the inventions described and/or claimed in this document.
[00284] Computing device 900 includes a processor 902, memory 904, a storage device 906, a high-speed interface 908 connecting to memory 904 and high-speed expansion ports 910, and a low speed interface 912 connecting to low speed bus 914 and storage device 906. Each of the components 902, 904, 906, 908, 910, and 912, are interconnected using various busses, and may be mounted on a common motherboard or in other manners as appropriate. The processor 902 can process instructions for execution within the computing device 900, including instructions stored in the memory 904 or on the storage device 906 to display graphical information for a GUI on an external input/output device, such as display 916 coupled to high speed interface 908. In other implementations, multiple processors and/or multiple buses may be used, as appropriate, along with multiple memories and types of memory. Also, multiple computing devices 900 may be connected, with each device providing portions of the necessary operations (e.g., as a server bank, a group of blade servers, or a multi-processor system).
[00285] The memory 904 stores information within the computing device 900. In one implementation, the memory 904 is a volatile memory unit or units. In another implementation, the memory 904 is a non-volatile memory unit or units. The memory 904 may also be another form of computer-readable medium, such as a magnetic or optical disk.
[00286] The storage device 906 is capable of providing mass storage for the computing device 900. In one implementation, the storage device 906 may be or contain a computer-readable medium, such as a floppy disk device, a hard disk device, an optical disk device, or a tape device, a flash memory or other similar solid state memory device, or an array of devices, including devices in a storage area network or other configurations. A computer program product can be tangibly embodied in an information carrier. The computer program product may also contain instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer- or machine-readable medium, such as the memory 904, the storage device 906, or memory on processor 902.
[00287] The high speed controller 908 manages bandwidth-intensive operations for the computing device 900, while the low speed controller 912 manages lower bandwidth-intensive operations. Such allocation of functions is exemplary only. In one implementation, the high-speed controller 908 is coupled to memory 904, display 916 (e.g., through a graphics processor or accelerator), and to high-speed expansion ports 910, which may accept various expansion cards (not shown). In the implementation, low-speed controller 912 is coupled to storage device 906 and low- speed expansion port 914. The low-speed expansion port, which may include various communication ports (e.g., USB, Bluetooth, Ethernet, wireless Ethernet) may be coupled to one or more input/output devices, such as a keyboard, a pointing device, a scanner, or a networking device such as a switch or router, e.g., through a network adapter.
[00288] The computing device 900 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a standard server 920, or multiple times in a group of such servers. It may also be implemented as part of a rack server system 924. In addition, it may be implemented in a personal computer such as a laptop computer 922. Alternatively, components from computing device 900 may be combined with other components in a mobile device (not shown), such as device 950. Each of such devices may contain one or more of computing device 900, 950, and an entire system may be made up of multiple computing devices 900, 950 communicating with each other.
[00289] Computing device 950 includes a processor 952, memory 964, an input/output device such as a display 954, a communication interface 966, and a transceiver 968, among other components. The device 950 may also be provided with a storage device, such as a microdrive or other device, to provide additional storage. Each of the components 950, 952, 964, 954, 966, and 968, are interconnected using various buses, and several of the components may be mounted on a common motherboard or in other manners as appropriate.
[00290] The processor 952 can execute instructions within the computing device 950, including instructions stored in the memory 964. The processor may be implemented as a chipset of chips that include separate and multiple analog and digital processors. The processor may provide, for example, for coordination of the other components of the device 950, such as control of user interfaces, applications run by device 950, and wireless communication by device 950.
[00291] Processor 952 may communicate with a user through control interface 958 and display interface 956 coupled to a display 954. The display 954 may be, for example, a TFT LCD (Thin-Film-Transistor Liquid Crystal Display) or an OLED (Organic Light Emitting Diode) display, or other appropriate display technology. The display interface 956 may comprise appropriate circuitry for driving the display 954 to present graphical and other information to a user. The control interface 958 may receive commands from a user and convert them for submission to the processor 952. In addition, an external interface 962 may be provide in communication with processor 952, so as to enable near area communication of device 950 with other devices. External interface 962 may provide, for example, for wired communication in some implementations, or for wireless communication in other implementations, and multiple interfaces may also be used.
[00292] The memory 964 stores information within the computing device 950. The memory 964 can be implemented as one or more of a computer-readable medium or media, a volatile memory unit or units, or a non-volatile memory unit or units. Expansion memory 984 may also be provided and connected to device 950 through expansion interface 982, which may include, for example, a SIMM (Single In Line Memory Module) card interface. Such expansion memory 984 may provide extra storage space for device 950, or may also store applications or other information for device 950. Specifically, expansion memory 984 may include instructions to carry out or supplement the processes described above, and may include secure information also. Thus, for example, expansion memory 984 may act as a security module for device 950, and may be programmed with instructions that permit secure use of device 950. In addition, secure applications may be provided via the SIMM cards, along with additional information, such as placing the identifying information on the SIMM card in a non-hackable manner.
[00293] The memory may include, for example, flash memory and/or NVRAM memory, as discussed below. In one implementation, a computer program product is tangibly embodied in an information carrier. The computer program product contains instructions that, when executed, perform one or more methods, such as those described above. The information carrier is a computer- or machine-readable medium, such as the memory 964, expansion memory 984, or memory on processor 952 that may be received, for example, over transceiver 968 or external interface 962.
[00294] Device 950 may communicate wirelessly through communication interface 966, which may include digital signal processing circuitry where necessary. Communication interface 966 may provide for communications under various modes or protocols, such as GSM voice calls, SMS, EMS, or MMS messaging, CDMA, TDMA, PDC, WCDMA, CDMA2000, or GPRS, among others. Such communication may occur, for example, through radio-frequency transceiver 968. In addition, short-range communication may occur, such as using a Bluetooth, WiFi, or other such transceiver (not shown). In addition, GPS (Global Positioning System) receiver module 980 may provide additional navigation- and location-related wireless data to device 950, which may be used as appropriate by applications running on device 950.
[00295] Device 950 may also communicate audibly using audio codec 960, which may receive spoken information from a user and convert it to usable digital information. Audio codec 960 may likewise generate audible sound for a user, such as through a speaker, e.g., in a handset of device 950. Such sound may include sound from voice telephone calls, may include recorded sound (e.g., voice messages, music files, etc.) and may also include sound generated by applications operating on device 950. [00296] The computing device 950 may be implemented in a number of different forms, as shown in the figure. For example, it may be implemented as a cellular telephone 980. It may also be implemented as part of a smart phone 982, personal digital assistant, or other similar mobile device.
[00297] Various implementations of the systems and techniques described here can be realized in digital electronic circuitry, integrated circuitry, specially designed ASICs (application specific integrated circuits), computer hardware, firmware, software, and/or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and/or interpretable on a programmable system including at least one programmable processor, which may be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.
[00298] These computer programs (also known as programs, software, software applications or code) include machine instructions for a programmable processor, and can be implemented in a high-level procedural and/or object-oriented programming language, and/or in assembly/machine language. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus and/or device (e.g., magnetic discs, optical disks, memory, Programmable Logic Devices (PLDs)) used to provide machine instructions and/or data to a programmable processor, including a machine- readable medium that receives machine instructions as a machine-readable signal. The term "machine-readable signal" refers to any signal used to provide machine instructions and/or data to a programmable processor.
[00299] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.
[00300] The systems and techniques described here can be implemented in a computing device that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network ("LAN"), a wide area network ("WAN"), and the Internet.
[00301] The computing device can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other.
[00302] A number of embodiments have been described. Nevertheless, it will be understood that various modifications may be made without departing from the spirit and scope of the invention.
[00303] In addition, the logic flows depicted in the figures do not require the particular order shown, or sequential order, to achieve desirable results. In addition, other steps may be provided, or steps may be eliminated, from the described flows, and other components may be added to, or removed from, the described systems. Accordingly, other embodiments are within the scope of the following claims and/or within the following clauses.
Clauses
[00304] Various aspects of the present invention are now summarized in the following numbered clauses:
[00305] (clause 1) Computer-implemented method (400) to identify a behavior pattern (Z) of industrial equipment (100) that has components (110-n, 120-k), and that performs an industrial process, wherein the components (110-n, 120-k) of the industrial equipment (100) are associated with data providers (1, 2, 3, 4, 5, 6, 7) that correspond to variates (Xv) in multi-variate original parameter data ([X]), wherein relations between the components (110-n, 120-k) belong to four relation groups (i, ii, iii, iv), that are differentiated according to the following criteria: the presence (ii, iii) of interactions between the components (110-n, 120-k), with the interactions being defined by the industrial process, or the absence (i, iv) of such interactions; the presence (ii, i) of changes in these interactions after the industrial equipment has started to perform the industrial process, or the absence (iii, iv) of such changes; wherein - when the industrial equipment (100) has started to perform the industrial process - a pre-trained machine-learning tool (500) is performing a behavior prediction, the method (400) comprising:
• receiving (410) the multi-variate original parameter data ([X]) from the data providers (1, 2, 3, 4, 5, 6, 7);
• clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) to one or more data clusters ([X]@[C]) such that in at least one data cluster ([X]@[C]), the majority of variates (Xv) belong to components in the same relation group (i, ii, iii, iv);
• assigning (430) cluster values (Cl, C2) to the data clusters;
• applying (440) the cluster values (Cl, C2) as weights ([W]) to the original parameter data ([X]) to obtain adapted parameter data ( [Y] ); and
• by the pre-trained machine-learning tool (500) with the adapted parameter data ( [Y]) at the input, performing (450) the behavior prediction with adapted parameter data ( [Y]), thereby identifying the behavior pattern (Z) that become available at the output of the pre-trained machine-learning tool (500).
[00306] (clause 2) Method (400) according to clause 1, wherein receiving (410) original multi-variate parameter data ([X]) from data providers (1, 2, 3, 4, 5, 6, 7) comprises predicting parameter data by one or more component-specific predictor modules (310-n).
[00307] (clause 3) Method (400) according to any of clauses 1 or 2, wherein clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) - to one ore more data clusters ([X]@[C]) - is performed by representing the components (110-n, 120-k) by component-representations (210-n, 210-k) and representing process stages by stage-representations.
[00308] (clause 4) Method (400) according to any of clauses 1 to 2, wherein clustering (420) the at least some of the variates (Xv) of the multi-variate original parameter data ([X]) is performed by representing the components (110-n, 120-k) by componentrepresentations (210-n, 210-k), wherein static relations between the components (110-n, 120-k) are identified by parsing a computer-readable equipment-descriptor document that has pointers to clusters.
[00309] (clause 5) Method (400) according to any of clauses 1 to 4, wherein during the process performance, the step clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) - to data clusters ([X]@[C]) - is performed once for data clusters that remain unchanged, and wherein the step assigning (430) cluster values (Cl, C2) to the data clusters is repeated for changing cluster values.
[00310] (clause 6) Method (400) according to any of clauses 1 to 5, wherein during the process performance, a change in the data that the computer receives (410) leads the computer to detect a change in a dynamic relation between the components (110-n, 120-k) and subsequently to perform clustering (420) again.
[00311] (clause 7) Method (400) according to any of clauses 1 to 6, wherein clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) to data clusters according to relations is performed for static relations between the components by parsing computer-readable equipment-descriptor documents that are selected from the following: documents that differentiate main and auxiliary components, piping and instrumentation diagrams (PID) for the equipment, mechanical flow diagrams (MFD) of the equipment, part or material lists, process specifications, recipes for the product, batch control documents, and floor plans.
[00312] (clause 8) Method (400) according to any of clauses 1 to 7, wherein clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) to data clusters according to relations is performed for static relations between the components that are arranged in a processing chain, experience common environmental conditions, arranged in physical connection, or have been installed to the equipment within a common time-frame.
[00313] (clause 9) Method (400) according to any of clauses 1 to 8, wherein clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) to data clusters according to relations is performed based on parameter data that indicate, selected from the following: quality and quantity of materials that are exchanged between components, equality of activity states or ON/OFF states that the components can assume, data reliability of the parameter data, distortions in the parameter data, and correlations within parameter data.
[00314] (clause 10) Use of the method according to any of clauses 1 to 9 to identify the behavior pattern (Z) of industrial equipment, in that the components (110-1 ... 110- N) are arranged to sequentially forward intermediate products.
[00315] (clause 11) Computer system (200/500) that is associated with industrial equipment (100) and that is adapted to execute a computer-implemented method (400) according to any of clauses 1 to 9, to identify a behavior pattern (Z) of the industrial equipment (100).
[00316] (clause 12) Computer program product that - when loaded into a memory of a computer and being executed by at least one processor of the computer causes the computer to perform the steps of the method according to any of clauses 1 to 9.

Claims

Claims
1. Computer-implemented method (400) to identify a behavior pattern (Z) of industrial equipment (100) that has components (110-n, 120-k), and that performs an industrial process, wherein at least some of the components are statically or dynamically related with each other to perform the industrial process, wherein - during process performance that means while the industrial equipment (100) performs the industrial process - a pre-trained machine-learning tool (500) is performing a behavior prediction, the method (400) comprising receiving (410) original parameter data ([X]) from data providers that belong to the components (110-n, 120-k) of the industrial equipment (100); clustering (420) the variates of the original parameter data ([X]) to data clusters ([X]@[C]) according to relations between the components (110-n, 120-k), wherein static relations between the components (110-n, 120-k) had been predefined by the properties of the industrial process before the equipment starts to perform the industrial process, and wherein dynamic relations between the components (110-n, 120-k) are being defined by the original parameter data that are being received (410) after the industrial process has started; assigning (430) cluster values (Cl, C2) to the data clusters; applying (440) the cluster values (Cl, C2) as weights ([W]) to the original parameter data ([X]) to obtain adapted parameter data ( [Y] ); and by the pre-trained machine-learning tool (500) with the adapted parameter data ( [Y] ) at the input, performing (450) the behavior prediction with adapted parameter data ( [Y]), thereby identifying the behavior pattern (Z) that become available at the output of the pre-trained machine-learning tool (500).
2. Method (400) according to claim 1, wherein receiving (410) original parameter data ([X]) from data providers comprises predicting parameter data by one or more component-specific predictor modules (310-n).
3. Method (400) according to any of claims 1 or 2, wherein clustering (420) the parameter data ([X]) is performed by representing the components (110-n, 120-k) by component-representations (210-n, 210-k) and representing process stages by stagerepresentations.
4. Method (400) according to any of claims 1 to 2, wherein clustering (420) the parameter data ([X]) is performed by representing the components (110-n, 120-k) by component-representations (210-n, 210-k), wherein the static relations between the components (110-n, 120-k) are identified by parsing a computer-readable equipmentdescriptor document that has pointers to clusters.
5. Method (400) according to any of claims 1 to 4, wherein during the process performance, the step clustering (420) the parameter data ([X]) to data clusters ([X]@[C]) is performed once for data clusters that remain unchanged, and wherein the step assigning (430) cluster values (Cl, C2) to the data clusters is repeated for changing cluster values.
6. Method (400) according to any of claims 1 to 5, wherein during the process performance, a change in the data that the computer receives (410) leads the computer to detect a change in a dynamic relation between the components (110-n, 120-k) and subsequently to perform clustering (420) again.
7. Method (400) according to any of claims 1 to 6, wherein clustering (420) the original parameter data to data clusters according to relations is performed for static relations between the components by parsing computer-readable equipment-descriptor documents that are selected from the following: documents that differentiate main and auxiliary components, piping and instrumentation diagrams (PID) for the equipment, mechanical flow diagrams (MFD) of the equipment, part or material lists, process specifications, recipes for the product, batch control documents, and floor plans.
8. Method (400) according to any of claims 1 to 7, wherein clustering (420) the original parameter data to data clusters according to relations is performed for static relations between the components that are arranged in a processing chain, experience common environmental conditions, arranged in physical connection, or have been installed to the equipment within a common time-frame.
9. Method (400) according to any of claims 1 to 8, wherein clustering (420) the original parameter data to data clusters according to relations is performed based on parameter data that indicate, selected from the following: quality and quantity of materials that are exchanged between components, equality of activity states or ON/OFF states that the components can assume, data reliability of the parameter data, distortions in the parameter data, and correlations within parameter data.
10. Computer-implemented method (400) to identify a behavior pattern (Z) of industrial equipment (100) that has components (110-n, 120-k), and that performs an industrial process, wherein the components (110-n, 120-k) of the industrial equipment (100) are associated with data providers (1, 2, 3, 4, 5, 6, 7) that correspond to variates (Xv) in multi-variate original parameter data ([X]), wherein relations between the components (110-n, 120-k) belong to four relation groups (i, ii, iii, iv), that are differentiated according to the following criteria: the presence (ii, iii) of interactions between the components (110-n, 120-k), with the interactions being defined by the industrial process, or the absence (i, iv) of such interactions; the presence (ii, i) of changes in these interactions after the industrial equipment has started to perform the industrial process, or the absence (iii, iv) of such changes; wherein - when the industrial equipment (100) has started to perform the industrial process - a pre-trained machine-learning tool (500) is performing a behavior prediction, the method (400) comprising: receiving (410) the multi-variate original parameter data ([X]) from the data providers (1, 2, 3, 4, 5, 6, 7); clustering (420) at least some of the variates (Xv) of the multi-variate original parameter data ([X]) to one or more data clusters ([X]@[C]) such that in at least one data cluster ([X]@[C]), the majority of variates (Xv) belong to components in the same relation group (i, ii, iii, iv); assigning (430) cluster values (Cl, C2) to the data clusters; applying (440) the cluster values (Cl, C2) as weights ([W]) to the original parameter data ([X]) to obtain adapted parameter data ( [Y]); and by the pre-trained machine-learning tool (500) with the adapted parameter data ( [Y] ) at the input, performing (450) the behavior prediction with adapted parameter data ( [Y]), thereby identifying the behavior pattern (Z) that become available at the output of the pre-trained machine-learning tool (500).
11. Use of the method according to any of claims 1 to 10 to identify the behavior pattern (Z) of industrial equipment, in that the components (110-1 ... 110-N) are arranged to sequentially forward intermediate products.
12. Computer system (200/500) that is associated with industrial equipment (100) and that is adapted to execute a computer-implemented method (400) according to any of claims 1 to 11, to identify a behavior pattern (Z) of the industrial equipment (100).
13. Computer program product that - when loaded into a memory of a computer and being executed by at least one processor of the computer causes the computer to perform the steps of the method according to any of claims 1 to 11.
14. Computer-implemented method (600) to enable a machine-learning tool (500) to identify a behavior pattern (Z) of industrial equipment (100) that has components (110-n, 120-k) and that performs an industrial process, such that the machine-learning tool (500) is being enabled to perform behavior prediction during process performance, the method (600) comprising: receiving (610) historical parameter data from data providers that belong to the components (110-n, 120-k) of the industrial equipment (100); clustering (620) the variates of historical parameter data to data clusters according to relations between the components (110-n, 120-k), wherein static relations between the components (110-n, 120-k) are pre-defined by the properties of the industrial process before the equipment starts to perform the industrial process, and wherein dynamic relations between the components (110-n, 120-k) are being defined by the historical parameter data that are being received (610), classifying (625) the dynamic relations into relation groups; assigning (630) cluster values (Cl, C2) to the data clusters; applying (640) the cluster values (Cl, C2) as weights to the historic parameter data to obtain adapted historical parameter data; and using (650) the adapted historical parameter data as training set to train the machinelearning tool (500), wherein training sets for the machine-learning tool (500) are being differentiated according to the classified relation groups.
15. Method (600) according to claim 14, wherein the step using (650) the adapted historical parameter data as training set is performed for different sub-networks (510, 520) of the machine-learning tool (500) that are being trained for different relation groups separately.
16. Method (600) according to any of claims 14 to 15, wherein the step classifying (625) dynamic relations into relation groups is performed by an auxiliary machine learning tool that is a classifier.
17. Computer program product that - when loaded into a memory of a computer and being executed by at least one processor of the computer causes the computer to perform the steps of the method according to any of claims 14 to 16.
PCT/EP2025/069785 2024-07-11 2025-07-10 Clustering parameter data to identify behavior pattern of industrial equipment Pending WO2026013205A2 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
LU507731A LU507731B1 (en) 2024-07-11 2024-07-11 Clustering parameter data to identify behavior pattern of industrial equipment
LULU507731 2024-07-11

Publications (2)

Publication Number Publication Date
WO2026013205A2 true WO2026013205A2 (en) 2026-01-15
WO2026013205A3 WO2026013205A3 (en) 2026-03-19

Family

ID=92043317

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2025/069785 Pending WO2026013205A2 (en) 2024-07-11 2025-07-10 Clustering parameter data to identify behavior pattern of industrial equipment

Country Status (2)

Country Link
LU (1) LU507731B1 (en)
WO (1) WO2026013205A2 (en)

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200226490A1 (en) 2019-01-15 2020-07-16 Ebay Inc. Machine learning-based infrastructure anomaly and incident detection using multi-dimensional machine metrics

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10410135B2 (en) * 2015-05-21 2019-09-10 Software Ag Usa, Inc. Systems and/or methods for dynamic anomaly detection in machine sensor data
US10826932B2 (en) * 2018-08-22 2020-11-03 General Electric Company Situation awareness and dynamic ensemble forecasting of abnormal behavior in cyber-physical system
LU500272B1 (en) 2021-06-11 2022-12-12 Wurth Paul Sa Predictive maintenance for industrial machines

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200226490A1 (en) 2019-01-15 2020-07-16 Ebay Inc. Machine learning-based infrastructure anomaly and incident detection using multi-dimensional machine metrics

Also Published As

Publication number Publication date
LU507731B1 (en) 2026-01-12
WO2026013205A3 (en) 2026-03-19

Similar Documents

Publication Publication Date Title
Morariu et al. Machine learning for predictive scheduling and resource allocation in large scale manufacturing systems
US9842302B2 (en) Population-based learning with deep belief networks
Wang et al. Fault detection behavior and performance analysis of principal component analysis based process monitoring methods
JP7682168B2 (en) System and method for predicting manufacturing process risks - Patents.com
Colledani et al. Integrated analysis of quality and production logistics performance in manufacturing lines
EP3183622B1 (en) Population-based learning with deep belief networks
Bansal et al. A real-time predictive maintenance system for machine systems
Cupek et al. Determination of the machine energy consumption profiles in the mass-customised manufacturing
JP2011145846A (en) Anomaly detection method, anomaly detection system and anomaly detection program
Reis et al. Incorporation of process-specific structure in statistical process monitoring: A review
CN115769235A (en) Method and system for providing alerts related to the accuracy of a training function
CN117355804A (en) Predictive maintenance for industrial machines
Ye et al. Context-aware manufacturing system design using machine learning
Cattaneo et al. A framework to integrate novelty detection and remaining useful life prediction in Industry 4.0-based manufacturing systems
Khoza et al. Comparing machine learning and statistical process control for predicting manufacturing performance
Jacobs et al. Gas turbine engine condition monitoring using Gaussian mixture and hidden Markov models
Chen et al. Hierarchical RNN-based framework for throughput prediction in automotive production systems
Huang et al. Cascading fault analysis and control strategy for computer numerical control machine tools based on meta action
CN101646982A (en) Method and apparatus for monitoring system health
LU507731B1 (en) Clustering parameter data to identify behavior pattern of industrial equipment
Park et al. Multivariate control charts that combine the Hotelling T 2 and classification algorithms
EP3995920B1 (en) EVENT ANALYSIS IN MODULAR INDUSTRIAL PLANTS
Yang et al. A cyber-physical scheme for predicting tool wear based on a hybrid dynamic neural network
Yang Remote diagnosis and detection technology for electrical control of intelligent manufacturing CNC machine tools
Jin et al. A chart allocation strategy for multistage processes

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 25740044

Country of ref document: EP

Kind code of ref document: A2