CA3173398A1 - Data processing for industrial machine learning - Google Patents
Data processing for industrial machine learning Download PDFInfo
- Publication number
- CA3173398A1 CA3173398A1 CA3173398A CA3173398A CA3173398A1 CA 3173398 A1 CA3173398 A1 CA 3173398A1 CA 3173398 A CA3173398 A CA 3173398A CA 3173398 A CA3173398 A CA 3173398A CA 3173398 A1 CA3173398 A1 CA 3173398A1
- Authority
- CA
- Canada
- Prior art keywords
- data
- machine learning
- time series
- computer
- learning model
- 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
Links
- 238000010801 machine learning Methods 0.000 title claims abstract description 135
- 238000012545 processing Methods 0.000 title claims description 30
- 238000000034 method Methods 0.000 claims abstract description 125
- 238000012549 training Methods 0.000 claims abstract description 66
- 230000008569 process Effects 0.000 claims abstract description 58
- 238000004140 cleaning Methods 0.000 claims abstract description 54
- 238000005065 mining Methods 0.000 claims abstract description 21
- 238000004458 analytical method Methods 0.000 claims abstract description 20
- 238000012423 maintenance Methods 0.000 claims abstract description 19
- 238000012544 monitoring process Methods 0.000 claims abstract description 17
- 238000002372 labelling Methods 0.000 claims description 26
- 238000004590 computer program Methods 0.000 claims description 10
- 238000004519 manufacturing process Methods 0.000 claims description 8
- 238000010923 batch production Methods 0.000 claims description 6
- 238000013473 artificial intelligence Methods 0.000 claims description 3
- 238000004220 aggregation Methods 0.000 claims description 2
- 230000002776 aggregation Effects 0.000 claims description 2
- 238000007726 management method Methods 0.000 claims description 2
- 230000001629 suppression Effects 0.000 claims description 2
- 238000011161 development Methods 0.000 abstract description 6
- 230000018109 developmental process Effects 0.000 abstract description 6
- 230000011218 segmentation Effects 0.000 abstract description 5
- 238000004422 calculation algorithm Methods 0.000 description 11
- 230000009471 action Effects 0.000 description 10
- 238000013479 data entry Methods 0.000 description 10
- 238000007781 pre-processing Methods 0.000 description 7
- 238000013459 approach Methods 0.000 description 5
- 230000001419 dependent effect Effects 0.000 description 5
- 230000002547 anomalous effect Effects 0.000 description 4
- 230000006399 behavior Effects 0.000 description 4
- 238000001514 detection method Methods 0.000 description 4
- 238000012360 testing method Methods 0.000 description 3
- 230000015556 catabolic process Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 238000006731 degradation reaction Methods 0.000 description 2
- 238000009826 distribution Methods 0.000 description 2
- 230000003203 everyday effect Effects 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 238000010561 standard procedure Methods 0.000 description 2
- 238000003860 storage Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 230000010267 cellular communication Effects 0.000 description 1
- 238000012517 data analytics Methods 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000001914 filtration Methods 0.000 description 1
- 238000009434 installation Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000002093 peripheral effect Effects 0.000 description 1
- 230000003449 preventive effect Effects 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 230000002195 synergetic effect Effects 0.000 description 1
- 238000013024 troubleshooting Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/20—Ensemble learning
-
- 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
-
- 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
- G05B23/0259—Electric testing or monitoring by means of a monitoring system capable of detecting and responding to faults characterized by the response to fault detection
- G05B23/0283—Predictive maintenance, e.g. involving the monitoring of a system and, based on the monitoring results, taking decisions on the maintenance schedule of the monitored system; Estimating remaining useful life [RUL]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/40—Data acquisition and logging
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N7/00—Computing arrangements based on specific mathematical models
- G06N7/01—Probabilistic graphical models, e.g. probabilistic networks
-
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/80—Management or planning
Abstract
The invention relates to a computer-implemented method (100) for automating the development of industrial machine learning applications in particular for predictive maintenance, process monitoring, event prediction, or root-cause analysis. The method consists of one or more sub-methods that, depending on the industrial machine learning problem, may be executed iteratively. These sub-methods include at least one of a method to automate the data cleaning in training (S10) and later application (S15) of machine learning models, a method to label (S11) time series (in particular signal data) with help of other timestamp records, feature engineering (S12) with the help of process mining, and automated hyperparameter tuning (S14) for data segmentation and classification.
Description
2 Data Processing for Industrial Machine Learning Field of the invention The invention relates to a computer-implemented method for data preprocessing for industrial machine learning. This method may be utilized for example for predictive maintenance, process monitoring, event prediction, or root-cause analysis. The invention further relates to a data processing system configured to carry out the steps of the computer-implemented method, a computer program comprising instructions to cause the data processing system to execute the method, and a computer-readable medium having stored such computer program.
Technical background Machine learning can be used in industry, amongst others, for predictive maintenance, process monitoring, event prediction, or root-cause analysis. For example, in the case of predictive maintenance, the condition of an industrial asset such as a motor or a robot may be predicted in order to estimate the time when maintenance actions should be performed. Thus, maintenance actions may be scheduled depending on machine learning based predictions of the condition of the industrial asset.
This provides cost savings overtime-based preventive maintenance, because maintenance actions are performed only when required. Furthermore, the probability of an unexpected failure of the industrial asset is reduced, since the condition of the asset is monitored continuously.
However, applying machine learning approaches for predictive maintenance is not a trivial task. In particular, the data from a sensor of an industrial asset or from a control system of an industrial process or plant typically needs to be preprocessed before application of the machine learning model. This preprocessing may comprise, for example, the cleaning of raw sensor data, including for instance the removal of outliers and/or the suppression of noise.
Furthermore, the preprocessing typically involves the derivation of features from a time series of data. These preprocessing algorithms are critical for the performance that can be achieved by the machine learning model. Another critical requirement is the provision of a sufficient number of training samples for the training of the machine learning model.
Machine learning applications for predictive maintenance, but also for other objectives such as process monitoring, event prediction, or root-cause analysis are therefore developed by mixed teams of domain and machine learning experts.
Summary However, machine learning and data science experts are rare and often lack the domain expertise required for industrial machine learning. Moreover, the development of industrial machine learning applications is a time-consuming process. Especially the time required for manual data cleaning, feature engineering, data labeling, and hyperparanneter tuning is long.
There is a lack of automated methods that enable domain experts to develop machine learning applications by themselves.
Existing approaches for supporting domain experts in developing machine learning applications such as automated machine learning (AutoML) leverage the homogenous character of mainstream machine learning applications like machine learning on tabular, textual, or image data. These approaches rely on the availability of labeled data to establish an objective function for model selection and hyperparameter tuning. However, such labeled data is usually not available in industrial machine learning applications.
It may therefore be desirable to provide an improved automation for the development of industrial machine learning applications.
This is achieved by the subject matter of the independent claims, wherein further embodiments are incorporated in the dependent claims and the following description. It should be noted that any step, feature or aspect of the computer-implemented method, as described in the following, equally applies to the data processing system configured to carry out the steps of the method, the computer program, and the computer readable medium, as described in the following, and vice versa.
The method for the automated development of industrial machine learning applications consists of one or more sub-methods that, depending on the industrial machine learning problem, may be executed iteratively. Sub-methods may be (a) a method to automate the data cleaning in training and later application of machine learning models, (b) a method to label a
Technical background Machine learning can be used in industry, amongst others, for predictive maintenance, process monitoring, event prediction, or root-cause analysis. For example, in the case of predictive maintenance, the condition of an industrial asset such as a motor or a robot may be predicted in order to estimate the time when maintenance actions should be performed. Thus, maintenance actions may be scheduled depending on machine learning based predictions of the condition of the industrial asset.
This provides cost savings overtime-based preventive maintenance, because maintenance actions are performed only when required. Furthermore, the probability of an unexpected failure of the industrial asset is reduced, since the condition of the asset is monitored continuously.
However, applying machine learning approaches for predictive maintenance is not a trivial task. In particular, the data from a sensor of an industrial asset or from a control system of an industrial process or plant typically needs to be preprocessed before application of the machine learning model. This preprocessing may comprise, for example, the cleaning of raw sensor data, including for instance the removal of outliers and/or the suppression of noise.
Furthermore, the preprocessing typically involves the derivation of features from a time series of data. These preprocessing algorithms are critical for the performance that can be achieved by the machine learning model. Another critical requirement is the provision of a sufficient number of training samples for the training of the machine learning model.
Machine learning applications for predictive maintenance, but also for other objectives such as process monitoring, event prediction, or root-cause analysis are therefore developed by mixed teams of domain and machine learning experts.
Summary However, machine learning and data science experts are rare and often lack the domain expertise required for industrial machine learning. Moreover, the development of industrial machine learning applications is a time-consuming process. Especially the time required for manual data cleaning, feature engineering, data labeling, and hyperparanneter tuning is long.
There is a lack of automated methods that enable domain experts to develop machine learning applications by themselves.
Existing approaches for supporting domain experts in developing machine learning applications such as automated machine learning (AutoML) leverage the homogenous character of mainstream machine learning applications like machine learning on tabular, textual, or image data. These approaches rely on the availability of labeled data to establish an objective function for model selection and hyperparameter tuning. However, such labeled data is usually not available in industrial machine learning applications.
It may therefore be desirable to provide an improved automation for the development of industrial machine learning applications.
This is achieved by the subject matter of the independent claims, wherein further embodiments are incorporated in the dependent claims and the following description. It should be noted that any step, feature or aspect of the computer-implemented method, as described in the following, equally applies to the data processing system configured to carry out the steps of the method, the computer program, and the computer readable medium, as described in the following, and vice versa.
The method for the automated development of industrial machine learning applications consists of one or more sub-methods that, depending on the industrial machine learning problem, may be executed iteratively. Sub-methods may be (a) a method to automate the data cleaning in training and later application of machine learning models, (b) a method to label a
3 time series of data such as a sensor signal using other tinnestamp records, (c) feature engineering with the help of process mining, and (d) automated hyperparameter tuning for data segmentation and classification.
According to a first aspect of the present disclosure, a computer-implemented method for machine learning is presented. The method comprises acquiring a first time series of data from a sensor of an industrial asset or from a control system for an industrial process or plant.
Furthermore, the method comprises processing the first time series of data to obtain an event log and applying process mining to the event log to provide a conformity analysis and/or bottleneck identification.
The first time series of data may be a discrete-time signal from a sensor of an industrial asset such as a motor or robot, or from a control system for an industrial process or plant such as a computerized distributed or centralized control system. Acquiring the first time series of data may mean, for example, to receive the first time series of data from the sensor or the control system, or to load the first time series from a storage medium. For example, the first time series of data may be loaded from a server such as a remote server. The first time series of data may comprise raw data from a sensor or from a control system, or the first time series of data may be processed data, e.g. a cleaned time series of data.
The steps of acquiring the first time series of data, processing the first time series of data, and applying process mining may be preprocessing steps, that may be executed before training or applying a first machine learning model, wherein the first machine learning model may be utilized, for example, for predictive maintenance or for predicting how a batch process will evolve. In particular, the steps of acquiring the first time series of data, processing the first time series of data, and applying process mining may be used for feature engineering, i.e., for determining the input parameters of the first machine learning model.
In an example, the computer-implemented method further comprises determining a condition indicator of the industrial asset based on the conformity analysis and/or bottleneck identification.
The conformity analysis provided by process mining may be quantified into condition indicators for the industrial asset. For example, different types of conformity and thresholds could be used and/or optimized. By calculating these condition indicators periodically (e.g. every second, every minute, every hour, or every day), these metrics can be compared to discover anomalous behavior.
For example, alarms and/or event data from a control system and/or sensor data from a .. motor, for instance, may be leveraged with the help of process mining to monitor its condition as well as to predict its behavior. This approach is agnostic to the sensor or control system
According to a first aspect of the present disclosure, a computer-implemented method for machine learning is presented. The method comprises acquiring a first time series of data from a sensor of an industrial asset or from a control system for an industrial process or plant.
Furthermore, the method comprises processing the first time series of data to obtain an event log and applying process mining to the event log to provide a conformity analysis and/or bottleneck identification.
The first time series of data may be a discrete-time signal from a sensor of an industrial asset such as a motor or robot, or from a control system for an industrial process or plant such as a computerized distributed or centralized control system. Acquiring the first time series of data may mean, for example, to receive the first time series of data from the sensor or the control system, or to load the first time series from a storage medium. For example, the first time series of data may be loaded from a server such as a remote server. The first time series of data may comprise raw data from a sensor or from a control system, or the first time series of data may be processed data, e.g. a cleaned time series of data.
The steps of acquiring the first time series of data, processing the first time series of data, and applying process mining may be preprocessing steps, that may be executed before training or applying a first machine learning model, wherein the first machine learning model may be utilized, for example, for predictive maintenance or for predicting how a batch process will evolve. In particular, the steps of acquiring the first time series of data, processing the first time series of data, and applying process mining may be used for feature engineering, i.e., for determining the input parameters of the first machine learning model.
In an example, the computer-implemented method further comprises determining a condition indicator of the industrial asset based on the conformity analysis and/or bottleneck identification.
The conformity analysis provided by process mining may be quantified into condition indicators for the industrial asset. For example, different types of conformity and thresholds could be used and/or optimized. By calculating these condition indicators periodically (e.g. every second, every minute, every hour, or every day), these metrics can be compared to discover anomalous behavior.
For example, alarms and/or event data from a control system and/or sensor data from a .. motor, for instance, may be leveraged with the help of process mining to monitor its condition as well as to predict its behavior. This approach is agnostic to the sensor or control system
4 used, i.e., it may be applied separately to other industrial assets and control systems as well, as the normal operation of the asset will be inferred as data is collected overtime. In other words, explicit information or a working model is not required to detect anomalies such as a degradation over time.
In an example, the computer-implemented method further comprises training and/or applying a first machine learning model to determine process deviations, to determine potential improvements, to perform condition-based monitoring, to perform predictive maintenance, and/or to predict how a batch process will evolve, wherein input parameters of the first machine learning model are based on the conformity analysis and/or bottleneck identification.
When the first machine learning model is to be trained, the first time series of data may be a time series of data such as a raw or a cleaned training time series of data.
In particular, the training time series of data may be a historic time series of data. In contrast, when the first machine learning model is to be applied, the first time series of data may be a life data stream from an industrial asset or from a control system such as a computerized distributed or centralized control system.
The first machine learning model may be trained to determine process deviations, to determine potential improvements, to perform condition-based monitoring, to perform predictive maintenance, and/or to predict how a batch process will evolve.
The input parameters of the first machine learning model may be or may be based on the conformity analysis and/or bottleneck identification. In particular, some or all input parameters of the first machine learning model may be or may be based on condition indicators of the industrial asset derived from the conformity analysis and/or bottleneck identification.
In another example, the processing of the first time series of data to obtain the event log comprises encoding the first time series of data by applying the symbolic aggregate approximation or artificial intelligence techniques.
In order to perform process mining on time series data, it needs to be transformed into an event log, i.e., a set of discrete events. Such encoding may be done using the symbolic aggregate approximation (SAX) or Al techniques.
In another example, the processing of the first time series of data to obtain the event log further comprises performing abstractions on the encoded first time series of data.
Since performing process mining on raw low-level event logs may be difficult, these logs may be transformed by performing abstractions. In one example, this may include aggregating raw low-level events or applying a filter below a threshold. For example, raw low-level events below a threshold may be set to zero to remove noise. Other abstractions of the raw low-level events are possible as well.
In another example, the computer-implemented method further comprises acquiring a second
In an example, the computer-implemented method further comprises training and/or applying a first machine learning model to determine process deviations, to determine potential improvements, to perform condition-based monitoring, to perform predictive maintenance, and/or to predict how a batch process will evolve, wherein input parameters of the first machine learning model are based on the conformity analysis and/or bottleneck identification.
When the first machine learning model is to be trained, the first time series of data may be a time series of data such as a raw or a cleaned training time series of data.
In particular, the training time series of data may be a historic time series of data. In contrast, when the first machine learning model is to be applied, the first time series of data may be a life data stream from an industrial asset or from a control system such as a computerized distributed or centralized control system.
The first machine learning model may be trained to determine process deviations, to determine potential improvements, to perform condition-based monitoring, to perform predictive maintenance, and/or to predict how a batch process will evolve.
The input parameters of the first machine learning model may be or may be based on the conformity analysis and/or bottleneck identification. In particular, some or all input parameters of the first machine learning model may be or may be based on condition indicators of the industrial asset derived from the conformity analysis and/or bottleneck identification.
In another example, the processing of the first time series of data to obtain the event log comprises encoding the first time series of data by applying the symbolic aggregate approximation or artificial intelligence techniques.
In order to perform process mining on time series data, it needs to be transformed into an event log, i.e., a set of discrete events. Such encoding may be done using the symbolic aggregate approximation (SAX) or Al techniques.
In another example, the processing of the first time series of data to obtain the event log further comprises performing abstractions on the encoded first time series of data.
Since performing process mining on raw low-level event logs may be difficult, these logs may be transformed by performing abstractions. In one example, this may include aggregating raw low-level events or applying a filter below a threshold. For example, raw low-level events below a threshold may be set to zero to remove noise. Other abstractions of the raw low-level events are possible as well.
In another example, the computer-implemented method further comprises acquiring a second
5 time series of data and cleaning the second time series of data to obtain a third time series of data. Furthermore, a data cleaning machine learning model is trained using a plurality of first training samples, wherein first training samples comprise a clean data point from the third time series of data and a plurality of raw data points from the second time series of data.
Hence, the computer-implemented method may comprise the training of a machine learning model for data cleaning. To train this machine learning model, a set of first training samples may be used, wherein the set of first training samples may be derived from the second and third time series of data.
.. The second time series of data may be a raw time series of data from the sensor of the industrial asset or from the control system for the industrial process or plant.
The third time series of data may be determined manually, for example by a domain expert or a machine learning expert. The cleaning of the second time series of data to obtain the third time series of data may comprise handling missing values, removing noise, and/or removing outliers.
Different first training samples may comprise different clean data points from the third time series of data. Each of the first training samples may further comprise a plurality of raw data points from the second time series of data. Thereby, raw data points of the second time series of data may be contained in several first training samples. In particular, the first training samples may comprise the raw data points of the second time series of data within a time window, which may be centered on the time of the corresponding clean data point. For training the data cleaning machine learning model, the clean data point of a training sample may serve as desired output of the machine learning model, whereas the raw data points of the training sample serve as input parameters to the machine learning model.
After training the machine learning model for data cleaning, this machine learning model may be applied to a raw time series of data from the sensor of the industrial asset or from the control system to provide a clean time series of data. This clean time series of data may be equal to the first time series of data.
In another example, the computer-implemented method further comprises acquiring a fourth time series of data from the sensor or from the control system and applying a data cleaning .. machine learning model to the fourth time series of data to obtain the first time series of data.
Hence, the computer-implemented method may comprise the training of a machine learning model for data cleaning. To train this machine learning model, a set of first training samples may be used, wherein the set of first training samples may be derived from the second and third time series of data.
.. The second time series of data may be a raw time series of data from the sensor of the industrial asset or from the control system for the industrial process or plant.
The third time series of data may be determined manually, for example by a domain expert or a machine learning expert. The cleaning of the second time series of data to obtain the third time series of data may comprise handling missing values, removing noise, and/or removing outliers.
Different first training samples may comprise different clean data points from the third time series of data. Each of the first training samples may further comprise a plurality of raw data points from the second time series of data. Thereby, raw data points of the second time series of data may be contained in several first training samples. In particular, the first training samples may comprise the raw data points of the second time series of data within a time window, which may be centered on the time of the corresponding clean data point. For training the data cleaning machine learning model, the clean data point of a training sample may serve as desired output of the machine learning model, whereas the raw data points of the training sample serve as input parameters to the machine learning model.
After training the machine learning model for data cleaning, this machine learning model may be applied to a raw time series of data from the sensor of the industrial asset or from the control system to provide a clean time series of data. This clean time series of data may be equal to the first time series of data.
In another example, the computer-implemented method further comprises acquiring a fourth time series of data from the sensor or from the control system and applying a data cleaning .. machine learning model to the fourth time series of data to obtain the first time series of data.
6 The data cleaning machine learning model may be trained as described above based on second and third time series of data. This may require the manual determination of the third time series of data, for example by a domain expert.
The fourth time series of data may be different from the second time series of data. In other words, the trained data cleaning machine learning model may be applied to new data, which is not in the training set of first training samples. Thus, the data cleaning machine learning model provides a generalized cleaning logic. In particular, the fourth time series of data may be a live data stream from a sensor or from a control system. The fourth time series of data may comprise thousands of data points per second, which may be cleaned by the data cleaning machine learning model.
It is also possible that the second and third time series of data comprise raw and clean time series of data from other applications, i.e., raw and clean time series of data from other applications may be utilized for training the data cleaning machine learning model. This may reduce or avoid the effort for manually determining clean data points of the third time series of data.
Alternatively, a data cleaning machine learning model from another application may be utilized for cleaning the fourth time series of data.
In another example, a dedicated data cleaning algorithm may be used to clean the fourth time series of data. This dedicated data cleaning algorithm may not be based on a machine learning model. This may be required when the data cleaning machine learning model as determined above does not provide a sufficient data cleaning performance.
In another example, the computer-implemented method further comprises acquiring a first set of labels for training a machine learning model for automatic labelling.
Furthermore, one or more data sources are acquired and a first set of features is extracted from the one or more data sources. The machine learning model for automatic labelling may then be trained using a plurality of second training samples, wherein the second training samples comprise a label from the first set of labels and one or more features from the first set of features.
The labels of the first set of labels may have a tinnestannp. These labels may be used as class labels in a classification process. The labels of the first set of labels may have been determined manually.
The data sources may be unstructured, semi-structured or tabular data sources.
Typical examples are alarm and event data, shift book entries, and entries in the computerized maintenance management system (CMMS).
The fourth time series of data may be different from the second time series of data. In other words, the trained data cleaning machine learning model may be applied to new data, which is not in the training set of first training samples. Thus, the data cleaning machine learning model provides a generalized cleaning logic. In particular, the fourth time series of data may be a live data stream from a sensor or from a control system. The fourth time series of data may comprise thousands of data points per second, which may be cleaned by the data cleaning machine learning model.
It is also possible that the second and third time series of data comprise raw and clean time series of data from other applications, i.e., raw and clean time series of data from other applications may be utilized for training the data cleaning machine learning model. This may reduce or avoid the effort for manually determining clean data points of the third time series of data.
Alternatively, a data cleaning machine learning model from another application may be utilized for cleaning the fourth time series of data.
In another example, a dedicated data cleaning algorithm may be used to clean the fourth time series of data. This dedicated data cleaning algorithm may not be based on a machine learning model. This may be required when the data cleaning machine learning model as determined above does not provide a sufficient data cleaning performance.
In another example, the computer-implemented method further comprises acquiring a first set of labels for training a machine learning model for automatic labelling.
Furthermore, one or more data sources are acquired and a first set of features is extracted from the one or more data sources. The machine learning model for automatic labelling may then be trained using a plurality of second training samples, wherein the second training samples comprise a label from the first set of labels and one or more features from the first set of features.
The labels of the first set of labels may have a tinnestannp. These labels may be used as class labels in a classification process. The labels of the first set of labels may have been determined manually.
The data sources may be unstructured, semi-structured or tabular data sources.
Typical examples are alarm and event data, shift book entries, and entries in the computerized maintenance management system (CMMS).
7 The features extracted from the one or more data sources may comprise typical natural language processing features (e.g. bag-of-words, recognized named entities), but also sentiment analysis or text classifications, statistical figures (alarm rates, # operator actions), quality tests from laboratories, or failure notes on assets in a specific plant area (from CMMS).
Quality tests from laboratories may be Boolean values (e.g. in-spec versus out-of-spec) or numerical or categorical quality indicators.
The entries in the data sources may have an associated tinnestannp, or these entries may comprise time information (e.g. time mentioned in shift book entries). This may be utilized to extract time ranges for labeling process values. One challenge with these data sources is that their tinnestamp may not match precisely with the tinnestannp of the process values. This problem may be resolved by assigning labels with a probability over a time window. Here, process values may be data points of the first time series of data. However, also features of the first machine learning problem such as condition indicators of the industrial asset may be assigned the same label as the process values that they are derived from.
The machine learning model for automatic labelling may be a probabilistic network/model such as a Bayes network. Thus, the features of the first set of features may be used as input into a probabilistic model, which describes a joint probability distribution over the features and the label of interest (e.g. normal vs. anomalous operation).
For each probabilistic model, it may be defined, which documents or entries from the data sources are used to generate the input to the probabilistic model and how a time-window (t_start, t_end) is generated for the output label. For instance, a probabilistic model might generate a label for a four-hour (4h) window from t_start to t_end = t_start +
4h. Thereby, alarms and events between, for example, t_start and t_end may be used.
Additionally or alternatively, shift book entries between, for example, t_start and t_start +
8h (corresponding approximately to one shift) may be used, or shift book entries from t_start until the end of the shift. Additionally or alternatively, CMMS data between, for example, t_start ¨ 12h and t_start +
12h nnay be used.
The notion of the label generated by the machine learning model for automatic labelling may not be that the label is probably present during the entire time-window between t_start and t_end, but that the label is probably present at least for some time between t_start and t_end.
After training the machine learning model for automatic labelling, the model may be used to label so far unlabeled time windows based on the corresponding data in the shift book, the alarm list, the event list, and/or the CMMS.
In another example, the computer-implemented method further comprises extracting a second set of features from the one or more data sources and determining a second set of labels by
Quality tests from laboratories may be Boolean values (e.g. in-spec versus out-of-spec) or numerical or categorical quality indicators.
The entries in the data sources may have an associated tinnestannp, or these entries may comprise time information (e.g. time mentioned in shift book entries). This may be utilized to extract time ranges for labeling process values. One challenge with these data sources is that their tinnestamp may not match precisely with the tinnestannp of the process values. This problem may be resolved by assigning labels with a probability over a time window. Here, process values may be data points of the first time series of data. However, also features of the first machine learning problem such as condition indicators of the industrial asset may be assigned the same label as the process values that they are derived from.
The machine learning model for automatic labelling may be a probabilistic network/model such as a Bayes network. Thus, the features of the first set of features may be used as input into a probabilistic model, which describes a joint probability distribution over the features and the label of interest (e.g. normal vs. anomalous operation).
For each probabilistic model, it may be defined, which documents or entries from the data sources are used to generate the input to the probabilistic model and how a time-window (t_start, t_end) is generated for the output label. For instance, a probabilistic model might generate a label for a four-hour (4h) window from t_start to t_end = t_start +
4h. Thereby, alarms and events between, for example, t_start and t_end may be used.
Additionally or alternatively, shift book entries between, for example, t_start and t_start +
8h (corresponding approximately to one shift) may be used, or shift book entries from t_start until the end of the shift. Additionally or alternatively, CMMS data between, for example, t_start ¨ 12h and t_start +
12h nnay be used.
The notion of the label generated by the machine learning model for automatic labelling may not be that the label is probably present during the entire time-window between t_start and t_end, but that the label is probably present at least for some time between t_start and t_end.
After training the machine learning model for automatic labelling, the model may be used to label so far unlabeled time windows based on the corresponding data in the shift book, the alarm list, the event list, and/or the CMMS.
In another example, the computer-implemented method further comprises extracting a second set of features from the one or more data sources and determining a second set of labels by
8 applying the machine learning model for automatic labelling to features from the second set of features.
The second set of features may be extracted from later entries of the data sources as compared to the first set of features. It is also possible that there is an overlap, so some entries of the data sources may be used for extracting features of both the first and second sets of features.
Given features from the second set of features, the probabilities of the label values may be inferred by means of the machine learning model for automatic labelling.
Hence, a timestarriped label of the second set of labels may be determined by selecting the label value with maximal probability. This may be utilized to label historical processes with labels from the second set of labels.
In another example, multiple labels may be assigned to a process value instead of a single label. Thereto, multiple machine learning models such as multiple probabilistic models may be used. For example, one probabilistic model per data source may be used.
Furthermore, algorithms for the implementation of the actual industrial monitoring and control task may be used, which may be configured to handle inconsistent class labels.
In another example, the first machine learning model is trained using a plurality of third training samples, wherein a third training sample comprises a label from the first or second sets of labels and/or the condition indicator of the industrial asset.
More specifically, for the training of the first machine learning model, labels of the first and/or second sets of labels may be utilized as desired output values of the first machine learning model. Furthermore, condition indicators of the industrial asset may be utilized as input values of the first machine learning model.
According to the present disclosure, also a data processing system is presented. The data processing system is configured to carry out the steps of any of the methods according to the present invention.
The data processing system may comprise a storage medium for storing amongst others, the first, second, third, and/or fourth time series of data. The data processing system may further comprise a processor such as a micro-processor with one or more processor cores. In addition, the data processing system may comprise a graphics processing unit, which may be used for efficiently training the first machine learning model, the machine learning model for data cleaning, and/or the machine learning model for automatic labelling. The data processing system may also comprise communication means such as LAN, VVLAN, or cellular communication modems. The data processing system may be connected to the sensor of the
The second set of features may be extracted from later entries of the data sources as compared to the first set of features. It is also possible that there is an overlap, so some entries of the data sources may be used for extracting features of both the first and second sets of features.
Given features from the second set of features, the probabilities of the label values may be inferred by means of the machine learning model for automatic labelling.
Hence, a timestarriped label of the second set of labels may be determined by selecting the label value with maximal probability. This may be utilized to label historical processes with labels from the second set of labels.
In another example, multiple labels may be assigned to a process value instead of a single label. Thereto, multiple machine learning models such as multiple probabilistic models may be used. For example, one probabilistic model per data source may be used.
Furthermore, algorithms for the implementation of the actual industrial monitoring and control task may be used, which may be configured to handle inconsistent class labels.
In another example, the first machine learning model is trained using a plurality of third training samples, wherein a third training sample comprises a label from the first or second sets of labels and/or the condition indicator of the industrial asset.
More specifically, for the training of the first machine learning model, labels of the first and/or second sets of labels may be utilized as desired output values of the first machine learning model. Furthermore, condition indicators of the industrial asset may be utilized as input values of the first machine learning model.
According to the present disclosure, also a data processing system is presented. The data processing system is configured to carry out the steps of any of the methods according to the present invention.
The data processing system may comprise a storage medium for storing amongst others, the first, second, third, and/or fourth time series of data. The data processing system may further comprise a processor such as a micro-processor with one or more processor cores. In addition, the data processing system may comprise a graphics processing unit, which may be used for efficiently training the first machine learning model, the machine learning model for data cleaning, and/or the machine learning model for automatic labelling. The data processing system may also comprise communication means such as LAN, VVLAN, or cellular communication modems. The data processing system may be connected to the sensor of the
9 industrial asset or to the control system of the industrial process or plant via communication means. The data processing system may further be connected to one or more servers, which may store training samples, or which may execute one or more steps of the computer-implemented method such as the training of the first machine learning model, the machine learning model for data cleaning, and/or the machine learning model for automatic labelling.
Furthermore, the data processing system may comprise peripherals such as screens.
According to the present disclosure, also a computer program is presented, wherein the computer program comprises instructions to cause the data processing system as defined in the independent claims to execute any one of the methods according to the present invention when the computer program is run on the data processing system.
According to the present disclosure, also a computer-readable medium is presented, wherein the computer-readable medium stores the computer program as defined in the independent claims.
It shall be understood that the computer-implemented method for machine learning, the data processing system configured to carry out the steps of the method, the computer program for causing the data processing system to execute the method, and the computer readable medium having stored such computer program have similar and/or identical preferred embodiments, in particular, as defined in the dependent claims. It shall be understood further that a preferred embodiment of the invention can also be any combination of the dependent claims with the respective independent claim.
These and other aspects of the present invention will become apparent from and be elucidated with reference to the embodiments described hereinafter.
Brief description of the drawings Exemplary embodiments of the invention will be described in the following with reference to the accompanying drawings:
Figure 1 illustrates a method for automating the development of industrial machine learning applications.
Figure 2 illustrates a method for training and applying a data cleaning model to achieve an automated data cleaning on raw data received online from an industrial asset.
Figure 3 illustrates a method for automatically determining labels by applying a machine learning model for automatic labelling.
Figure 4 illustrates a method for training a machine learning model for automatic labelling.
Figure 5 illustrates a method for performing process mining on a time series of data.
5 Figure 6 illustrates a workflow from scenario selection to model export.
Figure 7 illustrates a process to generate unsupervised models for anomaly and process phase detection.
Furthermore, the data processing system may comprise peripherals such as screens.
According to the present disclosure, also a computer program is presented, wherein the computer program comprises instructions to cause the data processing system as defined in the independent claims to execute any one of the methods according to the present invention when the computer program is run on the data processing system.
According to the present disclosure, also a computer-readable medium is presented, wherein the computer-readable medium stores the computer program as defined in the independent claims.
It shall be understood that the computer-implemented method for machine learning, the data processing system configured to carry out the steps of the method, the computer program for causing the data processing system to execute the method, and the computer readable medium having stored such computer program have similar and/or identical preferred embodiments, in particular, as defined in the dependent claims. It shall be understood further that a preferred embodiment of the invention can also be any combination of the dependent claims with the respective independent claim.
These and other aspects of the present invention will become apparent from and be elucidated with reference to the embodiments described hereinafter.
Brief description of the drawings Exemplary embodiments of the invention will be described in the following with reference to the accompanying drawings:
Figure 1 illustrates a method for automating the development of industrial machine learning applications.
Figure 2 illustrates a method for training and applying a data cleaning model to achieve an automated data cleaning on raw data received online from an industrial asset.
Figure 3 illustrates a method for automatically determining labels by applying a machine learning model for automatic labelling.
Figure 4 illustrates a method for training a machine learning model for automatic labelling.
Figure 5 illustrates a method for performing process mining on a time series of data.
5 Figure 6 illustrates a workflow from scenario selection to model export.
Figure 7 illustrates a process to generate unsupervised models for anomaly and process phase detection.
10 Detailed description of exemplary embodiments Figure 1 shows a method 100 for automating the development of industrial machine learning applications, in particular for predictive maintenance, process monitoring, event prediction, or root-cause analysis.
In step S10, an automated data cleaning algorithm is applied to historical data. Thereto, a machine learning model for data cleaning may be applied. In step S11, labels are determined, which may be performed by a machine learning model for automatic labelling. In the final pre-processing step, step S12, feature engineering is performed by means of process mining. In step 313, a conventional training of a machine learning model is performed.
This machine learning model may be configured for applications such as predictive maintenance, process monitoring, event prediction, or root-cause analysis. The training data may comprise or may be based on labels as determined in step S11 and features as determined in step S12.
In step S14, an automated machine learning orchestration is performed for steps S10 to S12.
This process is iterative and, depending on the measured performance of the machine learning model obtained from step S13, one or more of the steps S10 to 312 might be revisited. In some embodiments, one or more of the steps S10 to S12 may be performed manually, at least in part, for example the initial data cleaning. The machine learning orchestration may also be performed manually. It is also possible that one or more of the steps S10 to S12 and S14 are skipped, for example the automated data labelling or feature engineering steps.
When the iterations of the machine learning orchestration algorithm end, the final data cleaning algorithm of step S10, the final feature pre-processing algorithm of step S12, and the final machine learning model of step S13 may be provided for the application to new data as illustrated by steps S15 to S17.
In step S15, the final data cleaning algorithm is applied to a live data stream from an industrial installation. In step S16, the final feature determination algorithm is applied to the cleaned data obtained from step S15. In step S17, the trained machine learning model is applied to the
In step S10, an automated data cleaning algorithm is applied to historical data. Thereto, a machine learning model for data cleaning may be applied. In step S11, labels are determined, which may be performed by a machine learning model for automatic labelling. In the final pre-processing step, step S12, feature engineering is performed by means of process mining. In step 313, a conventional training of a machine learning model is performed.
This machine learning model may be configured for applications such as predictive maintenance, process monitoring, event prediction, or root-cause analysis. The training data may comprise or may be based on labels as determined in step S11 and features as determined in step S12.
In step S14, an automated machine learning orchestration is performed for steps S10 to S12.
This process is iterative and, depending on the measured performance of the machine learning model obtained from step S13, one or more of the steps S10 to 312 might be revisited. In some embodiments, one or more of the steps S10 to S12 may be performed manually, at least in part, for example the initial data cleaning. The machine learning orchestration may also be performed manually. It is also possible that one or more of the steps S10 to S12 and S14 are skipped, for example the automated data labelling or feature engineering steps.
When the iterations of the machine learning orchestration algorithm end, the final data cleaning algorithm of step S10, the final feature pre-processing algorithm of step S12, and the final machine learning model of step S13 may be provided for the application to new data as illustrated by steps S15 to S17.
In step S15, the final data cleaning algorithm is applied to a live data stream from an industrial installation. In step S16, the final feature determination algorithm is applied to the cleaned data obtained from step S15. In step S17, the trained machine learning model is applied to the
11 features determined in step S16.
The order of the data cleaning, labelling and feature engineering steps S10, S11, and S12, respectively, may be varied in different embodiments.
Figure 2 shows a method 200 for training and applying a data cleaning model to achieve an automated data cleaning on raw data received online from an industrial asset.
In step S20, raw data from an industrial asset is received and cleaned.
Thereby, raw data points in a received raw time series of data may be mapped onto clean data points in a clean time series of data. The mapping from raw data points onto clean data points may be performed manually, at least in part, for example by a machine learning expert. The cleaning of the received raw data may include handling missing values. For example, missing values may be set to the mean of a preceding and a succeeding data point.
Furthermore, the cleaning of the received raw data may include removing noise. For example, removing noise may be accomplished by setting data points, which are smaller than a threshold, to zero.
Furthermore, the cleaning of the received raw data may include the removal of outliers.
In step S21, the cleaned data points may be used as labels for training a machine learning model for data cleaning. The complete set of raw data is available as regressors. It is also possible that meta-data such as topological connections between measurements or other types of measurements (temperature, level, pressure) is used to select a subset of the complete set of raw data as regressors for a cleaned data point. Thus, a training sample for training the machine learning model for data cleaning may comprise a cleaned data point and a subset of data points of the raw data set. The machine learning model for data cleaning may be trained to predict the value of the cleaned data point from the subset of raw data points in the corresponding training sample. The training of this model may happen in a traditional fashion with manual tuning or automated with concepts like hyperparameter tuning. The output may be a machine learning model or several machine learning models that are capable to produce a cleaned data point based on a plurality of raw data points.
In step S22, the machine learning model for data cleaning obtained from step 521 may be applied to a data stream from an industrial process, i.e. to a time series of data, cleaning the raw online data and making it suitable as input for subsequent monitoring and/or control models. The output of the monitoring and/or control models may be displayed on a human machine interface (HMI). Additionally or alternatively, the output of the monitoring and/or control models may trigger some actions on the technical system, for instance when used as model in a model predictive controller.
When a sufficient number of training samples for data cleaning is already available from other applications, step 520 may be skipped. Then, the training samples from these other
The order of the data cleaning, labelling and feature engineering steps S10, S11, and S12, respectively, may be varied in different embodiments.
Figure 2 shows a method 200 for training and applying a data cleaning model to achieve an automated data cleaning on raw data received online from an industrial asset.
In step S20, raw data from an industrial asset is received and cleaned.
Thereby, raw data points in a received raw time series of data may be mapped onto clean data points in a clean time series of data. The mapping from raw data points onto clean data points may be performed manually, at least in part, for example by a machine learning expert. The cleaning of the received raw data may include handling missing values. For example, missing values may be set to the mean of a preceding and a succeeding data point.
Furthermore, the cleaning of the received raw data may include removing noise. For example, removing noise may be accomplished by setting data points, which are smaller than a threshold, to zero.
Furthermore, the cleaning of the received raw data may include the removal of outliers.
In step S21, the cleaned data points may be used as labels for training a machine learning model for data cleaning. The complete set of raw data is available as regressors. It is also possible that meta-data such as topological connections between measurements or other types of measurements (temperature, level, pressure) is used to select a subset of the complete set of raw data as regressors for a cleaned data point. Thus, a training sample for training the machine learning model for data cleaning may comprise a cleaned data point and a subset of data points of the raw data set. The machine learning model for data cleaning may be trained to predict the value of the cleaned data point from the subset of raw data points in the corresponding training sample. The training of this model may happen in a traditional fashion with manual tuning or automated with concepts like hyperparameter tuning. The output may be a machine learning model or several machine learning models that are capable to produce a cleaned data point based on a plurality of raw data points.
In step S22, the machine learning model for data cleaning obtained from step 521 may be applied to a data stream from an industrial process, i.e. to a time series of data, cleaning the raw online data and making it suitable as input for subsequent monitoring and/or control models. The output of the monitoring and/or control models may be displayed on a human machine interface (HMI). Additionally or alternatively, the output of the monitoring and/or control models may trigger some actions on the technical system, for instance when used as model in a model predictive controller.
When a sufficient number of training samples for data cleaning is already available from other applications, step 520 may be skipped. Then, the training samples from these other
12 applications may be utilized to train the machine learning model for data cleaning. In this case, human effort for determining training data is no longer required.
Alternatively, a machine learning model for data cleaning may be obtained from other applications.
In an embodiment, even though a sufficient number of training samples for data cleaning or a machine learning model for data cleaning may be available from other applications, a training of an improved machine learning model for data cleaning may be performed. This may involve the labelling of additional raw data points (specifying clean data points) in an active learning process. The active learning process may selectively request labels from a machine learning developer or domain expert to provide further information for the training process.
In another embodiment, hyperparameter optimization and other AutoML techniques are used in the training process to find the best possible hyperparanneter setting and machine learning model architecture to learn the data cleaning logic.
Figure 3 shows a method 300 for automatically determining labels using unstructured, semi-structured, or tabular data sources with a tinnestamp. Example data sources are alarm and/or event lists, shift books, or CMMSs.
In step S30, features are extracted from data entries of different data sources. For example, in step S30a, features may be extracted from data entries of a shift book. In step S30b, features may be extracted from data entries of an alarm and/or event list. In step S30c, features may be extracted from data entries in a CMMS. The extracted features may be typical natural language processing features (e.g. bag-of-words, recognized named entities), but also sentiment analysis or text classifications, statistical figures (alarm rates, # operator actions), quality tests from laboratories, or failure notes on assets in a specific plant area (from CMMS).
The entries of the data sources may have an associated tinnestannp or may include time information. From the timestamp associated with the entries in the data sources or time information in the entries itself (e.g. time mentioned in the shift book), time-ranges for labelling the process values may be extracted. One challenge with data sources such as shift books, alarm and/or event lists, and CMMSs is that their tinnestannp cannot be mapped precisely on the timestamp of process values. This issue may be addressed for example by assigning labels with a probability over a time window.
In step S31, the extracted features are used as input into a probabilistic model, e.g. a Bayes network, which may describe a joint probability distribution over the features and the label of interest. For example, the label of interest may indicate an anomaly or normal operation.
Given the features, probabilities of label values may be inferred, and a timestannped label may
Alternatively, a machine learning model for data cleaning may be obtained from other applications.
In an embodiment, even though a sufficient number of training samples for data cleaning or a machine learning model for data cleaning may be available from other applications, a training of an improved machine learning model for data cleaning may be performed. This may involve the labelling of additional raw data points (specifying clean data points) in an active learning process. The active learning process may selectively request labels from a machine learning developer or domain expert to provide further information for the training process.
In another embodiment, hyperparameter optimization and other AutoML techniques are used in the training process to find the best possible hyperparanneter setting and machine learning model architecture to learn the data cleaning logic.
Figure 3 shows a method 300 for automatically determining labels using unstructured, semi-structured, or tabular data sources with a tinnestamp. Example data sources are alarm and/or event lists, shift books, or CMMSs.
In step S30, features are extracted from data entries of different data sources. For example, in step S30a, features may be extracted from data entries of a shift book. In step S30b, features may be extracted from data entries of an alarm and/or event list. In step S30c, features may be extracted from data entries in a CMMS. The extracted features may be typical natural language processing features (e.g. bag-of-words, recognized named entities), but also sentiment analysis or text classifications, statistical figures (alarm rates, # operator actions), quality tests from laboratories, or failure notes on assets in a specific plant area (from CMMS).
The entries of the data sources may have an associated tinnestannp or may include time information. From the timestamp associated with the entries in the data sources or time information in the entries itself (e.g. time mentioned in the shift book), time-ranges for labelling the process values may be extracted. One challenge with data sources such as shift books, alarm and/or event lists, and CMMSs is that their tinnestannp cannot be mapped precisely on the timestamp of process values. This issue may be addressed for example by assigning labels with a probability over a time window.
In step S31, the extracted features are used as input into a probabilistic model, e.g. a Bayes network, which may describe a joint probability distribution over the features and the label of interest. For example, the label of interest may indicate an anomaly or normal operation.
Given the features, probabilities of label values may be inferred, and a timestannped label may
13 be created by selecting the label with maximum probability.
In step S32, the label determined in step S31 is assigned, for example to a process value, i.e., to a data point of a time series of data, or to a quantity derived from one or more process values such as a condition indicator of an industrial asset. Together with features as determined in step 812 of Fig. 1, the determined label may form a training sample for training the machine learning model of step S14 of Fig. 1.
For each probabilistic model, it is defined, which documents or entries from the data sources are used to generate the input to the probabilistic model and how a time-window (t_start, t_end) is generated for the output label.
In one exemplary embodiment, a probabilistic model might generate a label for a four hour window (t_start = t, t_end = t_start + 4 hours), using the alarms and events between t_start and t_end, the shift book entries from t_start to t_start + 8 hours (corresponding approximately to one shift) or from t_start until the end of the shift, and the CMMS entries between t_start ¨
12 hours and t_end + 12 hours.
The notion of the generated label may not be that the label is probably present during the entire time-window between t_start and t_end, but that the generated label is probably present at least for some time between t_start and t_end.
Figure 4 shows a method 400 for training a machine learning model for automatic labelling. In step 640, features are extracted from data entries of different data sources.
For example, in step S40a, features may be extracted from data entries of a shift book. In step S40b, features may be extracted from data entries of an alarm and/or event list. In step S40c, features may be extracted from data entries in the CMMS. The processing of the data entries in the shift book, the alarm/event list, and the CMMS for extracting features may be similar or identical to that of steps S30a to S30c.
In step S41, the machine learning model for automatic labelling is trained.
The machine learning model for automatic labelling may be a probabilistic model such as a Bayes network.
For training the machine learning model for automatic labelling, tinnestannped labels are used as class labels in a classification process.
The trained probabilistic model may be used in steps S11 and S31 to determine labels for so far unlabelled time windows based on data entries in the shift book, the alarm/event list, and/or the CMMS.
In one embodiment, multiple labels may be determined for each time window and/or process value instead of a single label. Thereto, several probabilistic models may be used, even
In step S32, the label determined in step S31 is assigned, for example to a process value, i.e., to a data point of a time series of data, or to a quantity derived from one or more process values such as a condition indicator of an industrial asset. Together with features as determined in step 812 of Fig. 1, the determined label may form a training sample for training the machine learning model of step S14 of Fig. 1.
For each probabilistic model, it is defined, which documents or entries from the data sources are used to generate the input to the probabilistic model and how a time-window (t_start, t_end) is generated for the output label.
In one exemplary embodiment, a probabilistic model might generate a label for a four hour window (t_start = t, t_end = t_start + 4 hours), using the alarms and events between t_start and t_end, the shift book entries from t_start to t_start + 8 hours (corresponding approximately to one shift) or from t_start until the end of the shift, and the CMMS entries between t_start ¨
12 hours and t_end + 12 hours.
The notion of the generated label may not be that the label is probably present during the entire time-window between t_start and t_end, but that the generated label is probably present at least for some time between t_start and t_end.
Figure 4 shows a method 400 for training a machine learning model for automatic labelling. In step 640, features are extracted from data entries of different data sources.
For example, in step S40a, features may be extracted from data entries of a shift book. In step S40b, features may be extracted from data entries of an alarm and/or event list. In step S40c, features may be extracted from data entries in the CMMS. The processing of the data entries in the shift book, the alarm/event list, and the CMMS for extracting features may be similar or identical to that of steps S30a to S30c.
In step S41, the machine learning model for automatic labelling is trained.
The machine learning model for automatic labelling may be a probabilistic model such as a Bayes network.
For training the machine learning model for automatic labelling, tinnestannped labels are used as class labels in a classification process.
The trained probabilistic model may be used in steps S11 and S31 to determine labels for so far unlabelled time windows based on data entries in the shift book, the alarm/event list, and/or the CMMS.
In one embodiment, multiple labels may be determined for each time window and/or process value instead of a single label. Thereto, several probabilistic models may be used, even
14 maybe one probabilistic model per data source, or multiple machine learning models. In this case, algorithms for the implementation of the actual industrial monitoring and/or control task may be used that can handle inconsistent class labels.
Figure 5 shows a method 500 for performing process mining on a time series of data, which may be utilized for feature engineering, in particular for a machine learning model for condition-based monitoring or predictive maintenance for an industrial asset.
Process mining provides the ability to perform conformity analysis. Such conformity reports may be quantified into condition indicators for industrial assets. For example, different types of conformity and thresholds may be used and/or optimized. By calculating these condition indicators periodically (e.g. every second, every minute, every hour, or every day), these metrics can be compared to discover anomalous behavior.
For example, alarms and/or event data from a control system and/or sensor data of an industrial asset such as a motor may be leveraged with the help of process mining to monitor its condition as well as to predict its behavior. This approach is agnostic to the sensor or control system used, i.e., it may be applied separately to other industrial assets and control systems as well (e.g. to robot data), as the normal operation of the asset will be inferred as data is collected overtime. In other words, explicit information or a working model is not required to detect anomalies such as a degradation over time.
On reporting an anomaly to a domain expert, explanations for detecting new data as anomalous may easily be provided as the condition indicators as well as actual historical event logs can all be easily retrieved.
In fact, such a methodology need not be limited to condition based monitoring.
As more data is collected and used for process mining, this collection of historical data can be continuously used to train machine learning models to make predictions of condition indicators and other statistics (e.g. frequency of occurrence of different events) into the future.
For instance, for a batch process, by taking real-time batch data as input, it may be predicted how the process would continue to evolve.
In step S50 of Fig. 5, a time series of data is acquired. This time series may be a raw time series from a sensor of an industrial asset such as a motor or a robot or from a control system such as a distributed or centralized control system for an industrial process or plant.
Alternatively, the time series may be a processed time series from a sensor or from a control system. For example, a cleaned time series from a sensor or from a control system may be acquired.
In step S51, the acquired time series of data is encoded using, for example, the symbolic aggregate approximation (SAX) or artificial intelligence techniques.
Thereby, the time series of data is transformed into a raw low-level event log, i.e., a set of discrete raw low-level events.
5 In the optional step S52, relevant events may be extracted from the raw low-level event log.
Additionally or alternatively, abstractions may be performed on the raw low-level event log.
This may include performing aggregations or filters on the raw low-level event log. For example, a filtering of the raw low-level event log may be performed to remove noise. This may be achieved by setting values below a threshold to zero. Step S52 provides a low-level 10 event log.
In step S53, process mining is applied to the low-level event log to provide conformity analysis and/or bottleneck identification. In particular, bottlenecks in batch processes and/or deviations from standard operating procedures may be discovered.
The process mining in step S53 enables to focus investigations on cases-of-interest. For these cases-of-interest, further data analytics may be performed in step S54. This allows to take contextual information such as the workload of an operator at the time into account, having a closer look at the processes, which deviated from the normal workflow.
Consequently, different actions could be taken to improve process efficiency and safety, for example, by providing training to operators, adapting standard operating procedures, etc.
One simple example for how process mining may be applied is the reaction to an alarm. There may be alarms of different priorities. After the activation of an alarm, an acknowledge of an operator may be expected. Furthermore, depending on the alarm priority, an action of the operator may be expected within a time limit, wherein the time limit may depend on the priority of the alarm. If large deviations are detected, for example, when the reaction to a priority 1 alarm occurs more than 5 minutes after the alarm, this may be used to either reprioritize the alarm or to retrain the operators to act faster. Those action sequences with a fast return to normal should become standard responses for the alarm. In other words, the action sequence may be optimized for shortest time to return to normal.
Figure 6 shows a workflow 600 from scenario selection to model export.
In step S60, the scenario is selected.
In step S61, data is provisioned.
In step S62, a machine learning model is determined with AutoML. This may include the determination of an unsupervised machine learning model with AutoML (step S62a), the determination of a supervised machine learning model with AutoML (step S62b), and the automated machine learning orchestration by a model manager (step S62c).
Starting with raw process / time series data, the method targets two problem classes: Anomaly detection and the segmentation of the time series of data into phases. For both problems, ensembles of unsupervised machine learning models are run to find the best unsupervised machine learning models for both tasks. On top of these results, sequential pattern mining may be applied to derive association rules that may assist with, e.g., root cause analysis.
Association rules may help to identify situations, in which, e.g., specific anomalies tend to occur, or in which productivity of the process suffers (e.g., "in 90% of the cases when phase A
was shorter than 15 minutes, an anomaly occurred in the subsequent phase").
In step S63, a report is generated. A number of results may be presented to the user: a segmentation of the time series into phases, anomalies within the time series of data, and a list of mined rules/patterns. Confidence thresholds for all results may be selected by the user so that only those results are displayed where the machine learning models are highly confident.
The user can then either export (step S64) the machine learning models for productive use, e.g., for monitoring or troubleshooting, or provide feedback (step S65) to the results: true/false (or more detailed labels) for the detected anomalies, higher/lower granularity (and optionally a label) for the detected phases. Based on the feedback, either the unsupervised machine learning model is improved, or a supervised machine learning model is created with AutoML
(step S62b), where the results of the unsupervised machine learning model and the user feedback are used to generate the labels. The process may be repeated until the user accepts a machine learning model for export. This can be either a supervised or unsupervised machine learning model.
Figure 7 illustrates a process 700 to generate unsupervised machine learning models for anomaly and process phase detection. Thus, the process of Fig. 7 may be used for time series segmentation and/or for anomaly detection. In addition, association rules on segments or association rules for anomalies may be derived.
In step S70, a data (pre)processing is performed using for example symbolic aggregate approximation or dynamic time warping.
In step S71, a cluster mining is performed, optionally via ensemble learning.
In step S72, a model and data stability check is performed.
It has to be noted that embodiments of the invention are described with reference to different subject matters. However, a person skilled in the art will gather from the above and the following description that, unless otherwise notified, in addition to any combination of features belonging to one type of subject matter also any combination between features relating to different subject matters is considered to be disclosed with this application.
However, all features can be combined providing synergetic effects that are more than the simple summation of the features.
While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary arid not restrictive. The invention is not limited to the disclosed embodiments.
Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing a claimed invention, from a study of the drawings, the disclosure, and the dependent claims.
In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or other unit may fulfil the functions of several items re-cited in the claims. The mere fact that certain measures are re-cited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope.
Figure 5 shows a method 500 for performing process mining on a time series of data, which may be utilized for feature engineering, in particular for a machine learning model for condition-based monitoring or predictive maintenance for an industrial asset.
Process mining provides the ability to perform conformity analysis. Such conformity reports may be quantified into condition indicators for industrial assets. For example, different types of conformity and thresholds may be used and/or optimized. By calculating these condition indicators periodically (e.g. every second, every minute, every hour, or every day), these metrics can be compared to discover anomalous behavior.
For example, alarms and/or event data from a control system and/or sensor data of an industrial asset such as a motor may be leveraged with the help of process mining to monitor its condition as well as to predict its behavior. This approach is agnostic to the sensor or control system used, i.e., it may be applied separately to other industrial assets and control systems as well (e.g. to robot data), as the normal operation of the asset will be inferred as data is collected overtime. In other words, explicit information or a working model is not required to detect anomalies such as a degradation over time.
On reporting an anomaly to a domain expert, explanations for detecting new data as anomalous may easily be provided as the condition indicators as well as actual historical event logs can all be easily retrieved.
In fact, such a methodology need not be limited to condition based monitoring.
As more data is collected and used for process mining, this collection of historical data can be continuously used to train machine learning models to make predictions of condition indicators and other statistics (e.g. frequency of occurrence of different events) into the future.
For instance, for a batch process, by taking real-time batch data as input, it may be predicted how the process would continue to evolve.
In step S50 of Fig. 5, a time series of data is acquired. This time series may be a raw time series from a sensor of an industrial asset such as a motor or a robot or from a control system such as a distributed or centralized control system for an industrial process or plant.
Alternatively, the time series may be a processed time series from a sensor or from a control system. For example, a cleaned time series from a sensor or from a control system may be acquired.
In step S51, the acquired time series of data is encoded using, for example, the symbolic aggregate approximation (SAX) or artificial intelligence techniques.
Thereby, the time series of data is transformed into a raw low-level event log, i.e., a set of discrete raw low-level events.
5 In the optional step S52, relevant events may be extracted from the raw low-level event log.
Additionally or alternatively, abstractions may be performed on the raw low-level event log.
This may include performing aggregations or filters on the raw low-level event log. For example, a filtering of the raw low-level event log may be performed to remove noise. This may be achieved by setting values below a threshold to zero. Step S52 provides a low-level 10 event log.
In step S53, process mining is applied to the low-level event log to provide conformity analysis and/or bottleneck identification. In particular, bottlenecks in batch processes and/or deviations from standard operating procedures may be discovered.
The process mining in step S53 enables to focus investigations on cases-of-interest. For these cases-of-interest, further data analytics may be performed in step S54. This allows to take contextual information such as the workload of an operator at the time into account, having a closer look at the processes, which deviated from the normal workflow.
Consequently, different actions could be taken to improve process efficiency and safety, for example, by providing training to operators, adapting standard operating procedures, etc.
One simple example for how process mining may be applied is the reaction to an alarm. There may be alarms of different priorities. After the activation of an alarm, an acknowledge of an operator may be expected. Furthermore, depending on the alarm priority, an action of the operator may be expected within a time limit, wherein the time limit may depend on the priority of the alarm. If large deviations are detected, for example, when the reaction to a priority 1 alarm occurs more than 5 minutes after the alarm, this may be used to either reprioritize the alarm or to retrain the operators to act faster. Those action sequences with a fast return to normal should become standard responses for the alarm. In other words, the action sequence may be optimized for shortest time to return to normal.
Figure 6 shows a workflow 600 from scenario selection to model export.
In step S60, the scenario is selected.
In step S61, data is provisioned.
In step S62, a machine learning model is determined with AutoML. This may include the determination of an unsupervised machine learning model with AutoML (step S62a), the determination of a supervised machine learning model with AutoML (step S62b), and the automated machine learning orchestration by a model manager (step S62c).
Starting with raw process / time series data, the method targets two problem classes: Anomaly detection and the segmentation of the time series of data into phases. For both problems, ensembles of unsupervised machine learning models are run to find the best unsupervised machine learning models for both tasks. On top of these results, sequential pattern mining may be applied to derive association rules that may assist with, e.g., root cause analysis.
Association rules may help to identify situations, in which, e.g., specific anomalies tend to occur, or in which productivity of the process suffers (e.g., "in 90% of the cases when phase A
was shorter than 15 minutes, an anomaly occurred in the subsequent phase").
In step S63, a report is generated. A number of results may be presented to the user: a segmentation of the time series into phases, anomalies within the time series of data, and a list of mined rules/patterns. Confidence thresholds for all results may be selected by the user so that only those results are displayed where the machine learning models are highly confident.
The user can then either export (step S64) the machine learning models for productive use, e.g., for monitoring or troubleshooting, or provide feedback (step S65) to the results: true/false (or more detailed labels) for the detected anomalies, higher/lower granularity (and optionally a label) for the detected phases. Based on the feedback, either the unsupervised machine learning model is improved, or a supervised machine learning model is created with AutoML
(step S62b), where the results of the unsupervised machine learning model and the user feedback are used to generate the labels. The process may be repeated until the user accepts a machine learning model for export. This can be either a supervised or unsupervised machine learning model.
Figure 7 illustrates a process 700 to generate unsupervised machine learning models for anomaly and process phase detection. Thus, the process of Fig. 7 may be used for time series segmentation and/or for anomaly detection. In addition, association rules on segments or association rules for anomalies may be derived.
In step S70, a data (pre)processing is performed using for example symbolic aggregate approximation or dynamic time warping.
In step S71, a cluster mining is performed, optionally via ensemble learning.
In step S72, a model and data stability check is performed.
It has to be noted that embodiments of the invention are described with reference to different subject matters. However, a person skilled in the art will gather from the above and the following description that, unless otherwise notified, in addition to any combination of features belonging to one type of subject matter also any combination between features relating to different subject matters is considered to be disclosed with this application.
However, all features can be combined providing synergetic effects that are more than the simple summation of the features.
While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary arid not restrictive. The invention is not limited to the disclosed embodiments.
Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing a claimed invention, from a study of the drawings, the disclosure, and the dependent claims.
In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or other unit may fulfil the functions of several items re-cited in the claims. The mere fact that certain measures are re-cited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be construed as limiting the scope.
Claims (16)
1. A computer-implemented rnethod (100) for machine learning, the method comprising:
acquiring (S12, S16, S50) a first time series of data from a sensor of an industrial asset or from a control system for an industrial process or plant;
processing (S12, S16, S51, S52) the first time series of data to obtain an event log; and applying (S12, S16, S53) process mining to the event log to provide a conformity analysis and/or bottleneck identification.
acquiring (S12, S16, S50) a first time series of data from a sensor of an industrial asset or from a control system for an industrial process or plant;
processing (S12, S16, S51, S52) the first time series of data to obtain an event log; and applying (S12, S16, S53) process mining to the event log to provide a conformity analysis and/or bottleneck identification.
2. The computer-irnplemented method (100) of clairn 1, further comprising determining a condition indicator of the industrial asset based on the conformity analysis and/or bottleneck identification.
3. The computer-irnplemented method (100) of any of claims 1 or 2, further comprising, training (S13) and/or applying (S17) a first machine learning model to determine process deviations, to determine potential improvements, to perform condition-based monitoring, to perform predictive maintenance, and/or to predict how a batch process will evolve, wherein input parameters to the first machine learning model are based on the conformity analysis and/or bottleneck identification.
4. The computer-implemented method (100) of any of the preceding claims, wherein the processing of the first time series of data to obtain the event log comprises encoding (S12, S16, S51) the first time series of data by applying the symbolic aggregate approximation or artificial intelligence techniques.
5. The computer-implemented method (100) of claim 4, wherein the processing of the first time series of data to obtain the event log further comprises performing abstractions (S12, S16, S52) on the encoded first time series of data.
6. The computer-implemented method (100) of claim 5, wherein the abstractions performed on the encoded first time series of data comprise data aggregations and/or noise suppression filters.
7. The computer-implemented method (100) of any of the preceding claims, further comprising acquiring a second time series of data;
cleaning (S10, S20) the second time series of data to obtain a third time series of data;
training (S10, S21) a data cleaning machine learning model using a plurality of first training samples;
wherein a first training sample comprises a clean data point from the third time series of data and a plurality of raw data points from the second time series of data.
cleaning (S10, S20) the second time series of data to obtain a third time series of data;
training (S10, S21) a data cleaning machine learning model using a plurality of first training samples;
wherein a first training sample comprises a clean data point from the third time series of data and a plurality of raw data points from the second time series of data.
8. The computer-implemented method (100) of claim 7, wherein the cleaning of the second time series of data comprises handling missing values, removing noise, and/or removing outliers.
9. The computer-implemented method (100) of any of the preceding claims, further comprising acquiring a fourth time series of data from the sensor or from the control system; and applying (S10, S15, S22) a data cleaning machine learning model to the fourth time series of data to obtain the first time series of data.
10. The computer-implemented method (100) of any of the preceding claims, further comprising acquiring a first set of labels for training a machine learning model for automatic labelling;
acquiring one or more data sources;
extracting (S11, 840) a first set of features from the one or more data sources;
training (S11, S41) the machine learning model for automatic labelling using a plurality of second training samples;
wherein a second training sample comprises a label from the first set of labels and one or more features from the first set of features.
acquiring one or more data sources;
extracting (S11, 840) a first set of features from the one or more data sources;
training (S11, S41) the machine learning model for automatic labelling using a plurality of second training samples;
wherein a second training sample comprises a label from the first set of labels and one or more features from the first set of features.
11. The computer-implemented method (100) of claim 10, wherein the one or more data sources comprise at least one of a shift book, an alarm list, an events list, and/or a data source from a computerized maintenance management system; and/or wherein the machine learning model for automatic labelling is a probabilistic model.
12. The computer-implemented method (100) of any of claims 10 or 11, further comprising extracting (S11, S30) a second set of features from the one or more data sources;
applying (S11, S31) the machine learning model for automatic labelling to features from the second set of features to obtain a second set of labels.
applying (S11, S31) the machine learning model for automatic labelling to features from the second set of features to obtain a second set of labels.
13. The computer-implemented method (100) of claims 2, 3, and 12, wherein the first machine learning model is trained using a plurality of third training samples; and wherein a third training sample comprises a label from the first or second sets of labels and/or the condition indicator of the industrial asset.
14. A data processing system comprising means for carrying out the steps of a method according to any of claims 1 to 13.
15. A computer program comprising instructions, which, when the program is executed by a computer, cause the computer to carry out the steps of a method according to any of claims 1 to 13.
16. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of a method according to any of claims 1 to 13.
Applications Claiming Priority (3)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EPPCT/EP2020/059135 | 2020-03-31 | ||
| EPPCT/EP2020/059135 | 2020-03-31 | ||
| PCT/EP2021/056093 WO2021197782A1 (en) | 2020-03-31 | 2021-03-10 | Data processing for industrial machine learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA3173398A1 true CA3173398A1 (en) | 2021-10-07 |
Family
ID=74858462
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA3173398A Pending CA3173398A1 (en) | 2020-03-31 | 2021-03-10 | Data processing for industrial machine learning |
Country Status (6)
| Country | Link |
|---|---|
| US (1) | US20230019404A1 (en) |
| EP (1) | EP4128087A1 (en) |
| JP (1) | JP2023520066A (en) |
| CN (1) | CN115380294A (en) |
| CA (1) | CA3173398A1 (en) |
| WO (1) | WO2021197782A1 (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114776304B (en) * | 2022-05-30 | 2022-11-04 | 广州海洋地质调查局 | Method, device, equipment and medium for identifying abnormities of deep sea mineral products |
| CN117009751A (en) * | 2023-10-07 | 2023-11-07 | 南方电网数字电网研究院有限公司 | Time sequence data cleaning method and device based on self-adaption |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH08202444A (en) * | 1995-01-25 | 1996-08-09 | Hitachi Ltd | Method and device for diagnosing abnormality of machine facility |
| JP2011145846A (en) * | 2010-01-14 | 2011-07-28 | Hitachi Ltd | Anomaly detection method, anomaly detection system and anomaly detection program |
| JP5808605B2 (en) * | 2011-08-17 | 2015-11-10 | 株式会社日立製作所 | Abnormality detection / diagnosis method and abnormality detection / diagnosis system |
| JP5996384B2 (en) * | 2012-11-09 | 2016-09-21 | 株式会社東芝 | Process monitoring diagnostic device, process monitoring diagnostic program |
| US10936564B2 (en) * | 2017-04-19 | 2021-03-02 | Xerox Corporation | Diagnostic method and system utilizing historical event logging data |
| JP2020027424A (en) * | 2018-08-10 | 2020-02-20 | 東京エレクトロンデバイス株式会社 | Learning data generating device, discrimination model generating device, and program |
| JP7167992B2 (en) * | 2018-09-20 | 2022-11-09 | 日本電気株式会社 | label correction device |
-
2021
- 2021-03-10 CN CN202180026316.6A patent/CN115380294A/en active Pending
- 2021-03-10 JP JP2022560024A patent/JP2023520066A/en active Pending
- 2021-03-10 CA CA3173398A patent/CA3173398A1/en active Pending
- 2021-03-10 WO PCT/EP2021/056093 patent/WO2021197782A1/en unknown
- 2021-03-10 EP EP21710006.4A patent/EP4128087A1/en active Pending
-
2022
- 2022-09-29 US US17/956,117 patent/US20230019404A1/en active Pending
Also Published As
| Publication number | Publication date |
|---|---|
| EP4128087A1 (en) | 2023-02-08 |
| WO2021197782A1 (en) | 2021-10-07 |
| US20230019404A1 (en) | 2023-01-19 |
| CN115380294A (en) | 2022-11-22 |
| JP2023520066A (en) | 2023-05-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108073497B (en) | Multi-index transaction analysis method based on data center data acquisition platform | |
| US20230019404A1 (en) | Data Processing for Industrial Machine Learning | |
| US10410135B2 (en) | Systems and/or methods for dynamic anomaly detection in machine sensor data | |
| Manco et al. | Fault detection and explanation through big data analysis on sensor streams | |
| US11080127B1 (en) | Methods and apparatus for detection of process parameter anomalies | |
| US9208209B1 (en) | Techniques for monitoring transformation techniques using control charts | |
| US11409962B2 (en) | System and method for automated insight curation and alerting | |
| KR20180108446A (en) | System and method for management of ict infra | |
| CN111177485B (en) | Parameter rule matching based equipment fault prediction method, equipment and medium | |
| EP3315465A1 (en) | Method and system for dynamically managing waste water treatment process for optimizing power consumption | |
| Liulys | Machine learning application in predictive maintenance | |
| JP2022546771A (en) | Using Irrelevant Filters to Facilitate Efficient RUL Analysis of Utility System Assets | |
| Becherer et al. | Intelligent choice of machine learning methods for predictive maintenance of intelligent machines | |
| Züfle et al. | A predictive maintenance methodology: predicting the time-to-failure of machines in industry 4.0 | |
| US11775714B2 (en) | Rational decision-making tool for semiconductor processes | |
| US20230376026A1 (en) | Automated real-time detection, prediction and prevention of rare failures in industrial system with unlabeled sensor data | |
| CN116468423A (en) | Operation and maintenance emergency coordination method, system and terminal equipment | |
| Torikka et al. | Predictive Maintenance Service Powered by Machine Learning and Big Data | |
| EP3706048A1 (en) | Anomaly prediction in an industrial system | |
| Bellini et al. | A Deep Learning Approach for Short Term Prediction of Industrial Plant Working Status | |
| EP3864506A1 (en) | Systems and methods for data-driven process improvement | |
| Husain et al. | Predictive maintenance of single phase ac motor using iot sensor data and machine learning (simple linear regression and multiple linear regression algorithms) | |
| Lee et al. | Enhanced Anomaly Detection in Manufacturing Processes through Hybrid Deep Learning Techniques | |
| CN117216701B (en) | Intelligent bridge monitoring and early warning method and system | |
| US11243776B2 (en) | Systems and/or methods for generating complex event processing (CEP) events and query definitions for real-time decomposition of resource usage data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| EEER | Examination request |
Effective date: 20220926 |
|
| EEER | Examination request |
Effective date: 20220926 |
|
| EEER | Examination request |
Effective date: 20220926 |
|
| EEER | Examination request |
Effective date: 20220926 |
|
| EEER | Examination request |
Effective date: 20220926 |
|
| EEER | Examination request |
Effective date: 20220926 |