WO2023186317A1 - System, apparatus and method for managing one or more assets - Google Patents

System, apparatus and method for managing one or more assets Download PDF

Info

Publication number
WO2023186317A1
WO2023186317A1 PCT/EP2022/058667 EP2022058667W WO2023186317A1 WO 2023186317 A1 WO2023186317 A1 WO 2023186317A1 EP 2022058667 W EP2022058667 W EP 2022058667W WO 2023186317 A1 WO2023186317 A1 WO 2023186317A1
Authority
WO
WIPO (PCT)
Prior art keywords
assets
knowledge graph
asset
node
digital twin
Prior art date
Application number
PCT/EP2022/058667
Other languages
French (fr)
Inventor
Indrajit Kar
Chethan SEEGEHALLI
Lalit WADHWANI
Original Assignee
Siemens Aktiengesellschaft
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Siemens Aktiengesellschaft filed Critical Siemens Aktiengesellschaft
Priority to PCT/EP2022/058667 priority Critical patent/WO2023186317A1/en
Publication of WO2023186317A1 publication Critical patent/WO2023186317A1/en

Links

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B17/00—Systems involving the use of models or simulators of said systems
    • G05B17/02—Systems involving the use of models or simulators of said systems electric
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/004—Error avoidance
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G06F11/34—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment
    • G06F11/3466—Performance evaluation by tracing or monitoring
    • G06F11/3495—Performance evaluation by tracing or monitoring for systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G06N3/0455—Auto-encoder networks; Encoder-decoder networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/047—Probabilistic or stochastic networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/02—Knowledge representation; Symbolic representation
    • G06N5/022—Knowledge engineering; Knowledge acquisition
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/0703—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation
    • G06F11/079—Root cause analysis, i.e. error or fault diagnosis
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/07—Responding to the occurrence of a fault, e.g. fault tolerance
    • G06F11/0703—Error or fault processing not based on redundancy, i.e. by taking additional measures to deal with the error or fault not making use of redundancy in operation, in hardware, or in data representation
    • G06F11/0793—Remedial or corrective actions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F11/00—Error detection; Error correction; Monitoring
    • G06F11/30—Monitoring
    • G06F11/34—Recording or statistical evaluation of computer activity, e.g. of down time, of input/output operation ; Recording or statistical evaluation of user activity, e.g. usability assessment
    • G06F11/3447—Performance evaluation by modeling

Definitions

  • the present invention relates to a field of data and simulation systems and more particularly relates to management of one or more assets in a technical installation for recommendation and control using a prescriptive digital twin.
  • digital twin of plurality of assets of the technical installation is generated to predict behavior of the assets.
  • the digital twin of the industrial plant may provide insights into several domains such as parameter prediction, remaining useful life prediction, failure or breakdown prediction, anomaly detection and so forth.
  • currently digital twins are based on data from sensors associated with assets in the technical installation.
  • the predictions of the digital twin is based on univariate data only and does univariate tasks.
  • external factors such as requirements of the processes, resource availability, energy availability, manual inputs and so forth are not accounted for in the current digital twin systems. Thereby, the existing digital twins are inaccurate and inefficient.
  • the digital twins are capable of performing tasks such as prediction, anomaly detection based only on univariate data from sensors.
  • tasks based on the data collected from several sources, databases, requirements of users is not possible with current systems.
  • there are many external factors that are not captured by the sensors are disregarded in the current systems which creates data gaps for instance, manual controls and/or unmonitored environmental conditions or load may lead to inherently unpredictable time-series data due to gaps. Therefore, performing prediction, detection and analysis tasks in such scenarios using standard approaches based on mathematical models that rely on stationarity, or prediction models that utilize prediction errors to perform such tasks becomes challenging.
  • the current systems also lack in identifying the root cause of the any failure in the assets.
  • the method comprises receiving multivariate data from one or more sources.
  • the multivariate data comprises sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, and simulation data pertaining to the one or more assets in the technical installation.
  • the term ‘multivariate data’ refers to data as received from multiple sources and comprising more than one variable.
  • the multivariate data ensures that every aspect of sensor data, domain knowledge and business requirements is considered while generating output such as predictions, recommendations, anomaly detection and root cause analysis, prediction, anomaly detection or tasks, self-learning root cause analysis , recommendations for the assets and execute recommended action.
  • the term ‘task’ as used herein refers to as but not limited to parameter prediction, remaining useful life prediction, failure or breakdown prediction, anomaly detection.
  • one or more requirements refer to information related to energy availability, resource availability, asset availability, raw material availability, manual changes to the assets or parameter settings, process demand, business needs, business data and the like.
  • the method comprises extracting plurality of second knowledge graphs derived from the first knowledge graph.
  • each of the plurality of second knowledge graphs comprises information specific to the corresponding one or more assets associated therewith.
  • the second knowledge graphs are distilled knowledge graphs which also includes subgraphs motifs that have been derived from the first knowledge graph for each specific asset.
  • the method of extracting the second knowledge graph from the first knowledge graph comprises associating a primary node pertaining to the second knowledge graph with the asset, wherein the primary node is directly related to the asset.
  • the method comprises associating one or more secondary nodes to the second knowledge graph, wherein the one or more secondary nodes are adjacent to the primary node of the asset.
  • the method comprises determining an embedding for the asset based on the primary node and one or more secondary-nodes.
  • the method comprises generating plurality of second digital twin corresponding to plurality of second graphs.
  • each of the plurality of second digital twins are local digital twin associated with corresponding assets.
  • the second digital twin is aware of neighboring assets and is capable of predicting the parameters accurately.
  • the method comprises providing one or more recommendations for the assets using the plurality of second digital twins based on the prediction of the first digital twin.
  • the method for determining one or more recommendations for the assets comprises generating the one or more corrective actions using the plurality of second digital twins based on the prediction of the first digital twin and the one or more requirements. Further, the method comprises executing the one or more corrective actions on the asset by the plurality of second digital twins deployed on edge.
  • the method further comprises detecting one or more anomalies in the assets using the first digital twin, and wherein detection of one or more anomalies further comprises identification of reconstruction errors in the encoder decoder network.
  • the method of detecting one or more anomalies in the assets comprises classifying the one or more nodes of the first knowledge graph into one or more categories using a first machine learning model.
  • the method comprises determining a status of the one or more nodes based on the determined categories, wherein the status is one of: a failure node, a partial failure node and a functional node.
  • the method comprises selecting at least one node from the one or more nodes based on the status of the nodes, wherein the at least one selected node is either the partial failure node or failure node.
  • the method comprises detecting one or more anomalies in the asset based on the selected at least one node.
  • the method further comprises automatically scanning the first knowledge graph and the second knowledge graph for detection of anomalies in real-time
  • the method comprises determining root cause of the detected anomalies using the plurality of second digital twins.
  • the method comprises rendering the one or more predicted parameters, detected anomalies, and one or more recommendations on a display device.
  • the object of the present invention is also achieved by a system for managing one or more assets in a technical installation.
  • the system comprising one or more sources associated with assets, and an apparatus communicatively coupled to the one or more sources.
  • the apparatus is configured for managing one or more assets of the technical installation according to the abovementioned method.
  • FIG 1 A illustrates a block-diagram of a system for managing one or more of assets in a technical installation, in accordance with an embodiment of the present invention
  • FIG 1 B illustrates a block-diagram of an apparatus for managing one or more of assets in a technical installation, in accordance with an embodiment of the present invention
  • FIG 2 depicts a flowchart of a method for managing one or more of assets in a technical installation, in accordance with an embodiment of the present invention
  • FIG 3 depicts a flowchart of a method for extracting the second knowledge graph from the first knowledge graph, in accordance with an embodiment of the present invention
  • FIG 4 depicts flowchart of a method for detecting one or more anomalies in the assets, in accordance with an embodiment of the present invention
  • FIG 5 is an exemplary block-diagram for managing one or more assets in a technical installation, in accordance with an embodiment of the present invention
  • FIG 6 is a block-diagram 600 representing exchange of data signal between different components of the system deployed in the technical installation, in accordance with an embodiment of the present invention.
  • multivariate data' refers to data from multiple one or more sources having more than one variable.
  • the multivariate data may comprise sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, simulation data, business data stored in databases.
  • the multivariate data may be obtained in different forms, formats, and order and in processed by the processing unit.
  • the one or more sources may be electronic devices configured to obtain and transmit the data to the processing unit. Non-limiting examples of sources include sensing units, controllers, edge devices, databases, user devices, and simulation models.
  • simulation model refers to an analytical model in machine-executable form derived from at least one of a physics-based model, a data-driven model or a hybrid model associated with the asset.
  • the simulation model may be a 1-dimensional(1 D) model, a 2-dimensional(2D) model, a 3-dimen- sional(3D) model or a combination thereof.
  • the simulation model may generate one or more simulation instances based on the set of requirements.
  • the simulation instance may be executed in a simulation environment as one of stochastic simulations, deterministic simulations, dynamic simulations, continuous simulations, discrete simulations, local simulations, distributed simulations, co-simulations or a combination thereof.
  • the system 100 comprises an apparatus 110 communicatively coupled to an edge device 115.
  • the edge device 115 is connected to the apparatus 110 via a network 120, for example, local area network (LAN), wide area network (WAN), WiFi, etc.
  • the edge device 115 is configured to receive multivariate data from one or more sources 125 associated with the one or more assets 105.
  • the multivariate data may be received from one or more sources 125 such as sensing units, controllers, edge devices, databases, user devices, and simulation models. It should be understood that the system 100 also takes into consideration the business requirements of the user and is not only dependent on sensor data information.
  • the dynamic data received from the process and operation of the technical installation is also input to the system 100 and is one or more requirements of the processes ongoing in the technical installation is also input to the system 100 using a user device 130.
  • the one or more requirements may be information related to energy availability, resource availability, asset availability, raw material availability, manual changes to the assets or parameter settings, process demand, business needs and the like.
  • the one or more requirements may be stored in a memory of the edge device 115 or may be input to the edge device 115 by an operator.
  • the edge device 115 may be communicatively coupled to the display device 130.
  • Non-limiting examples of display device 130 include, personal computers, workstations, personal digital assistants, human machine interfaces.
  • the display device 130 may enable the user to input one or more requirements through a web-based interface.
  • the edge device 115 Upon receiving the one or more requirements from the user, the edge device 115 transmits a request for monitoring and managing one or more assets to the apparatus 110.
  • the apparatus 110 can be an edge computing device.
  • edge computing refers to computing environment that is capable of being performed on an edge device (e.g., connected to one or more sources in an industrial setup and one end and to a remote server(s) such as for computing server(s) or cloud computing server(s) on other end), which may be a compact computing device that has a small form factor and resource constraints in terms of computing power.
  • a network of the edge computing devices can also be used to implement the apparatus. Such a network of edge computing devices is referred to as a fog network.
  • the apparatus 110 comprises a processing unit 135, a memory unit 140, a storage unit 145, an input unit 155, an output unit 160 and a standard interface or bus 195, as shown in FIG 1 B.
  • the apparatus 110 can be a computer, a workstation, a virtual machine running on host hardware, a microcontroller, or an integrated circuit.
  • the apparatus 110 can be a real or a virtual group of computers (the technical term for a real group of computers is “cluster”, the technical term for a virtual group of computers is “cloud”).
  • Each of the plurality of second knowledge graph comprises information specific to the corresponding one or more assets 105 associated therewith. Generating plurality of second digital twin corresponding to plurality of second graphs. Each of the plurality of second digital twins are local digital twin associated with corresponding assets 105.
  • the processing unit 135 is configured for providing one or more recommendations for the assets 105 using the plurality of second digital twins based on the prediction of the first digital twin, and the one or more requirements.
  • the memory unit 140 may be volatile memory and non-volatile memory.
  • the memory unit 140 may be coupled for communication with the processing unit 135.
  • the processing unit 135 may execute instructions and/or code stored in the memory unit 140.
  • a variety of computer-readable storage media may be stored in and accessed from the memory unit 140.
  • the memory unit 140 may include any suitable elements for storing data and machine- readable instructions, such as read only memory, random access memory, erasable programmable read only memory, electrically erasable programmable read only memory, a hard drive, a removable media drive for handling compact disks, digital video disks, diskettes, magnetic tape cartridges, memory cards, and the like.
  • the digital twin generation module 180 is configured for generating the first digital twin and the plurality of second digital twins.
  • the digital twin generation module 180 is configured for generating the first digital twin using the first knowledge graph.
  • the first digital twin is a global digital twin for predicting one or more parameters of assets in the technical installation.
  • the digital twin generation module is configured for generating plurality of second digital twins based on second knowledge graphs. Each of the plurality of second digital twins are local digital twins associated with corresponding assets.
  • the root cause analysis module 185 is configured for determining root cause of the detected anomalies using the plurality of second digital twins. Further, the root cause analysis module 185 is configured for tracing the second knowledge graph corresponding to one or more assets 105 associated with detected anomalies.
  • the recommendation module 190 is configured for providing one or more recommendations for the assets 105 using the plurality of second digital twins based on the prediction and graph embeddings of the first digital twin.
  • the action module 192 is configured for executing our one or more recommended actions for the assets 105 using the plurality of second digital twins.
  • the storage unit 145 comprises a non-volatile memory which stores the domain knowledge and multivariate data as receiveied from the one or more sources 125.
  • the storage unit 145 includes the database 150 that comprises the first knowledge and the second knowledge graph, historical data such as previously detected anomalies, previous recommendations and so forth.
  • the bus 195 acts as interconnect between the processing unit 135, the memory unit 140, the storage unit 145, the input unit 155 and the output unit 160.
  • the input unit 155 enables the user to input one or more requirements or raise request for managing the assets 105.
  • the output unit 160 is configured for presenting the status of assets 105, detected anomalies, predicted parameters, generated recommendations, and the like.
  • FIG 1 A and 1 B may vary for different implementations.
  • peripheral devices such as an optical disk drive and the like, Local Area Network (LAN)/ Wide Area Network (WAN)/ Wireless (e.g., Wi-Fi) adapter, graphics adapter, disk controller, input/output (I/O) adapter, network connectivity devices also may be used in addition or in place of the hardware depicted.
  • LAN Local Area Network
  • WAN Wide Area Network
  • Wireless Wireless
  • graphics adapter disk controller
  • I/O input/output
  • network connectivity devices also may be used in addition or in place of the hardware depicted.
  • the depicted example is provided for the purpose of explanation only and is not meant to imply architectural limitations with respect to the present disclosure.
  • the one or more requirements may be information related to energy availability, resource availability, asset availability, raw material availability, manual changes to the assets or parameter settings, process demand, business needs and the like.
  • Such dynamic data pertaining to the process and operation of the technical installation is considered predicting one or more parameters of the assets 105.
  • external factors such as the business requirements of the technical installation are also considered for determining output of the digital twins.
  • the output of the digital twins based on the multivariate data is accurate as it considers all the factors involved in the operations of the assets 105.
  • the multivariate data is received from one or more sources 125.
  • the one or more sources 125 may refer to electronic devices configured to obtain and transmit the data to the processing unit 135.
  • Non-limiting examples of sources 125 include sensors, controllers, edge devices, databases, and simulators.
  • the one or more sources 125 may be sensors such as a temperature sensor, a velocity sensor, an acceleration sensor, a pressure sensor, and a force sensor.
  • the output from the sensors may be in the form of temperature data, velocity data, acceleration data or pressure data.
  • the sensor data are obtained through data acquisition interfaces.
  • the source 125 maybe a simulation model configured to forecast one or more parameters in the assets.
  • the source 125 is a database storing data from sensors, simulation models and manual entries by the user.
  • the source 125 is a display device 130 that enables user to enter requirements and data available through a web-based interface.
  • the first knowledge graph is generated based on the multivariate data.
  • the first knowledge graph comprises a structured arrangement of plurality of nodes.
  • each node pertains to the asset 105 in the technical installation.
  • the first knowledge graph represents a real-world replica of the assets with respect to nodes.
  • a gas turbine or generator has multiple moving components and the data from the components are captured by a plurality of sensors.
  • the knowledge graph aims to capture the influence of multiple moving components by multiple sensors which determines a single output. Therefore, the nodes in the first knowledge graph are not restricted to the asset 105 alone, but also the components of the asset 105.
  • the first knowledge graph captures a connection between different nodes and influence of other neighboring nodes on a particular node. It should be understood that since each asset 105 or a unit in the asset 105 does not operate independently, therefore it is essential to map assets 105 to the nodes and a correlation between the nodes.
  • the first knowledge graph is a directed graph and node features, and vertex features are encapsulated in the first knowledge graph as graph embeddings.
  • the nodes may be assets 105, one or more sources 125 such as sensors, databases, raw data and the edges are relationship between various assets 105, and one or more sources 125.
  • the first knowledge graph may contain one or more linked nodes capturing domain knowledge and indicative of parameters and corresponding responses pertaining to the one or more components of the asset 105, history of failures of the asset 105, historical data pertaining to parameter values of the one or more components of the asset 105, failure predictions and corresponding cause of failures and so forth. It should be noted that each of linked nodes is connected using one or more paths or edges.
  • the first knowledge graph includes asset specific information, data pertaining to conceptualization of the asset 105, data pertaining to designing of the asset 105, data pertaining to manufacturing of the asset 105, data pertaining to assembling of the asset 105, data pertaining to manufacturing of the asset 105, process specific information, general process information, and any other information related to business requirement may be linked to each other in a defined order. Additionally, the first knowledge graph comprises failures and causes of the failures in a linked manner. In an example, a failure in the motor may be linked to reduced lubrication in the rotor, rotor failure may be linked to defects in the bearing and so forth. Additionally, the first knowledge graph also includes standard procedure data that may be specific to a process in the asset 105.
  • the standard procedure data includes a set of pre-defined rules, a standard operating procedure, a pre-defined activity or a process, and the like.
  • the first knowledge graph also comprises physics-based models for specific processes such as, manufacturing, assembling, testing and operation of the asset 105.
  • the first knowledge graph also comprises data pertaining to previously detected anomalies or failures and the actions that were initiated to resolve the previously predicted failures.
  • the first digital twin is generated using the first knowledge graph.
  • the first digital twin is a global digital twin for predicting one or more parameters of assets 105 in the technical installation.
  • the term ‘first digital twin’ as used herein refers to a dynamic virtual replica based on one or more of physics-based models, Computer-Aided Design (CAD) models, Computer-Aided Engineering (CAE) models, one-dimensional (1 D) models, two-dimensional (2D) models, three-dimensional (3D) models, finite-element (FE) models, descriptive models, metamodels, stochastic models, parametric models, reduced-order models, statistical models, heuristic models, prediction models, ageing models, machine learning models, Artificial Intelligence models, deep learning models, system models, knowledge graphs and so on.
  • CAD Computer-Aided Design
  • CAE Computer-Aided Engineering
  • FE finite-element
  • the method comprises deriving motifs which are recurrent and statistically significant subgraphs or patterns of a larger knowledge graph. Each of these sub-graphs, defined by a particular pattern of interactions between vertices, may reflect a framework in which particular functions are achieved efficiently.
  • the term ‘first machine learning model’ as used herein refers to a neural network prediction model given a set of training dataset. In general, each individual sample of the training data is a pair containing a dataset (e.g., operating of the asset 105 and corresponding parameter values of the asset 105) and a desired output value or dataset (e.g., parameter values of the asset 105).
  • the first machine learning model analyzes the training data and produces hierarchical graph embeddings as well as a predictor function.
  • the predictor function once derived through training, is capable of reasonably predicting or estimating the correct output value or dataset.
  • Exemplary machine learning models may include artificial intelligence model such as deep neural network, convolutional neural network and but not limited to graph neural network and the like.
  • the first trained machine learning model is a generative model comprising the encoder decoder network.
  • the encoder decoder network models the actual distribution of a particular asset 105 and learns joint probability distribution p(x,y).
  • the encoder decoder network predicts the conditional probability based on a Bayesian model.
  • Bayesian model allows the formulation of prior (multivariate data), generation of appropriate likelihood functions (prediction of requested parameters) using Bayes theorem.
  • the first trained machine learning model is a self-attention machine learning model.
  • the encoder decoder network in the invention does not require complete sequence information to be captured into a single fixed length vector.
  • the encoder decoder network is a self-attention network.
  • the self-attention network includes both alignment and translation technique. Attention is proposed as a method to both align and translate. Alignment is the process in time series that identifies which parts of the input sequence are relevant to each embedding in the output, whereas translation is the process of using the relevant information to select the appropriate output.
  • the core idea of selfattention technique is to focus on the most relevant parts of the input sequence for each output and eliminate the translation issue.
  • a context vector is developed that is filtered specifically for each output time step.
  • the first digital is configured to detect one or more tasks but not limited to anomalies in the asset 105.
  • the detection of one or more anomalies further comprises identification of reconstruction errors in the encoder decoder network.
  • a time-series as shown in equation (1).
  • the first trained machine learning model is configured to reconstruct the normal time-series. Furthermore, the reconstruction errors are used to obtain the likelihood of a point in a test time-series being anomalous. Notably, for each point x(i), an anomaly score a(i) of the point being anomalous is calculated.
  • the anomaly score comprises of two components a normal component and a residual component. It should be understood that there is minimal or no change in normal component during anomaly. Meanwhile, the residual component changes drastically during anomaly, thereby leading to a higher reconstruction error. Therefore, in case the value of anomaly score has high value of residual component, then an anomaly has been detected.
  • such a technique enables to capture latent information which is not possible using basic statistical models or existing machine learning models.
  • the first digital twin predicts the future parameter values of the asset 105 and consequently detects anomalies in the predicted data, thereby predicting future anomalies in real-time.
  • the method of detecting one or more anomalies further comprises automatically scanning the first knowledge graph and the second knowledge graph for detection of anomalies in real-time.
  • the automatic scanning of the first knowledge for anomaly detection is further explained in FIG 4.
  • the external knowledge is processed by a comprehensive first knowledge graph (sometimes referred to as a “teacher model”) to generate valuable information to teach a simple and efficient second knowledge graph (sometimes referred to as a “student model”).
  • a comprehensive first knowledge graph sometimes referred to as a “teacher model”
  • a simple and efficient second knowledge graph sometimes referred to as a “student model”.
  • the second knowledge graphs are continuously updated with specifics from the corresponding assets 105 at the edge, thereby making the second knowledge graphs aware of the neighboring nodes.
  • the second knowledge graphs are the distilled versions of the decoder in the encoder decoder network which are motifs aware.
  • the present invention employs time delayed Embeddings which are more practical than the adjacency matrix since they pack node properties in a vector with a smaller dimension.
  • the embeddings include special attention on sub graph motifs.
  • the embeddings determine an influence of neighboring assets on the asset.
  • the method comprises extracting the second knowledge graph from the first knowledge graph based on the motifs aware Embeddings distillation.
  • the model predicts the next motif to be attached to first knowledge graph.
  • Entry (/, j) of motif embeddings are the sum of the weights of all motifs containing both nodes / and j.
  • the motif is specified by name and the type of motif instance can be one of but not limited to functional and structural.
  • the weighting scheme can be one of: but not limited to unweighted, mean, and product.
  • the decoder adds one motif at a time, the model needs to predict the attachment configuration also. This way first graph is award of the second graph within and using knowledge distillation we are able to distill the knowledge to a smaller less computation device.
  • the extraction of second knowledge graphs from the first knowledge graph is based on training the student model based on the teacher model.
  • the extraction is also based on motif prediction and attachment.
  • the extraction is based on attention based encoder decoder graph knowledge distillation alpha (AEDGKD- A) technique.
  • AEDGKD- A attention based encoder decoder graph knowledge distillation alpha
  • the AEDGKD-A technique aims at transferring structural knowledge using mutual relations of data examples in the output presentation of the teacher model.
  • the method comprises computing a relational potential i for each n-tuple of data examples and transfers information through the potential from the teacher model to the student model.
  • the plurality of second digital twin corresponding to plurality of second graphs are generated.
  • each of the plurality of second digital twins are local digital twin associated with corresponding assets 105.
  • the term ‘second digital twin’ as used herein refers to a dynamic virtual replica based on one or more of physics-based models, Computer- Aided Design (CAD) models, Computer-Aided Engineering (CAE) models, one-dimensional (1 D) models, two-dimensional (2D) models, three-dimensional (3D) models, finite-element (FE) models, descriptive models, metamodels, stochastic models, parametric models, reduced-order models, statistical models, heuristic models, prediction models, ageing models, machine learning models, Artificial Intelligence models, deep learning models, system models, knowledge graphs and so on.
  • CAD Computer- Aided Design
  • CAE Computer-Aided Engineering
  • FE finite-element
  • the root causes analysis may be based on historical data.
  • the historical data comprises response trends of the component and associated cause for the response trends.
  • the analysis of the parameter values and the historical data is based on the fault tree analysis (FTA) technique.
  • the historic data may be analyzed using one or more of descriptive techniques, exploratory techniques, inferential techniques, predictive techniques, causal techniques, qualitative analysis techniques, quantitative analysis techniques and so on.
  • the method of determining the root cause of the detected anomalies comprises tracing the second knowledge graph corresponding to one or more assets 105 associated with not limited to detected anomalies. It will be appreciated that the second digital twin thus generated is connected to the first digital and is capable of accurately determining cause of predicted anomalies and recommendation based on the determined root cause.
  • one or more recommendations are provided for the assets 105 using the plurality of second digital twins based on the prediction of the first digital twin.
  • the one or more recommendation may also be based on local graph embeddings of the first knowledge graph and the second knowledge graph.
  • the one or more recommendation maybe generated for maintenance of the asset 105, optimizing performance of the asset 105, mitigating anomalies in the asset 105, and the like.
  • the recommendations may be generated based on the second knowledge graph. It will be appreciated that the second digital twin is a local twin specific to a particular asset 105 deployed on the edge and thereby is capable of accurately determining one or more recommendation for resolving the anomalies and/or determining actions based on requirements of the user.
  • determining one or more recommendations for the assets further comprises generating the one or more corrective actions using the plurality of second digital twins based on the prediction of the first digital twin and the one or more requirements.
  • the corrective action may be scheduling a maintenance activity based on a predicted anomaly, optimizing efficiency of the technical installation, activate and deactivate assets 105 based on the requirements, and so forth.
  • recommendation may be associated with scheduling a maintenance of the turbine.
  • the business requirement is to increase production, then the recommendation is to increase running power of the assets 105.
  • the method further comprises executing the one or more corrective actions on the asset by the plurality of second digital twins deployed on the edge. It will be appreciated the corrective measures are executed at the edge seamlessly since the second digital twins are associated with each asset 105.
  • the second digital twin is associated with controller of the assets 105 and is configured to provide instruction to the corresponding controller in order to execute the recommendation.
  • FIG 3 depicts a flowchart of a method 300 for extracting the second knowledge graph from the first knowledge graph, in accordance with an embodiment of the present invention.
  • a primary node pertaining to the second knowledge graph is associated with the asset 105.
  • the primary node is directly related to the asset 105.
  • the primary node may be a motor.
  • associating one or more secondary nodes to the second knowledge graph is adjacent to the primary node of the asset 105.
  • secondary nodes may be one or more components of the motor, sensors associated with motor, power source of the motor, output load of the motor, other assets that may be directly or indirectly connected to the motor.
  • determining embeddings for the asset 105 based on the primary node and one or more secondary nodes.
  • the embeddings determine an influence of neighboring assets on the asset 105.
  • the embeddings define correlation between the primary node and the secondary nodes.
  • the embeddings determine the influence of parameters, operations and condition of the secondary nodes on the primary node associated with asset 105. The Embeddings thereby helps in determining root cause of the detected anomalies and generating one or more recommendations.
  • a loss function is calculated for the asset 105 based on the determined embeddings.
  • the loss function determines how well the student model is being trained by the teacher model. Notably, the lower the loss function, the higher is the accuracy of the student model and the outcome is more reliable.
  • the loss function is calculated by a mathematical formula as represented in equation (2).
  • £ Primary graph training is the total loss as experienced by the teacher model.
  • ⁇ vector angle is the vector angle loss that is determined by matching the distance between the edges of the two selected embeddings.
  • £ JensO nshanon divergence loss is the Jenson Shanon divergence loss as experienced by the teacher model when the student model is learning.
  • ⁇ motiff pr diction is the motiff prediction loss that is determined at the stage of training. The motiff prediction loss is determined by calculating the deviation between the actual outcome from the teacher model and the predicted model.
  • L motif f atatchement prediction is the motiff attachment predcition loss that is determined by calculating the deviation of prediction with respect to the neighboring nodes of the teacher model.
  • £ gra ph. prediction is the graph prediction loss that is determined by calculating the prediction losses of the teacher model.
  • ⁇ •student loss is the student loss that is determined by calculating the prediction losses of the student model.
  • £ seC ondary graph training is the total loss experienced by the student model with respect to the neighboring nodes.
  • £ nei g hbO ur nodeconnection is the neighbor node connection loss that is determined by calculating the loss experienced in neighboring assets or devices.
  • the neighbor node connection loss represents the magnitude of how good a particular node (assets or components) is being explained by the neighboring nodes.
  • ⁇ neighbour node structure is the neighbor node structure loss that is determined by understanding the node structure of the between the primary nodes and secondary nodes and calculating the losses in the node structure, attention is the attention loss that is determined by calculating the deviation of outcome of the neighboring nodes with respect to the primary node.
  • the second knowledge graph is extracted from the first knowledge graph based on the embeddings and the calculated loss function.
  • the second knowledge graph thus extracted has minimum losses and the outcome of the second digital twin based on the second knowledge graph is accurate and reliable.
  • the status is one of: a failure node, a partial failure node and a functional node.
  • the second machine learning model is trained to classify a node as ‘failure node’ in case an anomaly is detected based on the reconstruction errors as disclosed above.
  • the second machine learning model is trained to classify a node as ‘partial failure node in case the node is indirectly affected by neighboring nodes or is recovering from a previously detected anomaly.
  • the second machine learning model is trained to classify a node as ‘function node’ in case the node is unaffected and is fully functional.
  • the status of the scanned nodes may be presented on the user device 130 using color coding pertaining to health to each of the nodes.
  • the failure node maybe color coded ‘red’
  • the partial failure node may be color coded ‘yellow’
  • the functional node may be color coded ‘green’.
  • at least one node is selected from the one or more nodes based on the status of the nodes.
  • the at least one selected node is either the partial failure node or failure node.
  • one or more anomalies are detected in the asset 105 based on the selected at least one node.
  • FIG 5 is a block-diagram of an exemplary system 500 for managing one or more assets 105 in a technical installation, in accordance with an embodiment of the present invention.
  • the system 500 is an exemplary implementation of integrating the proposed invention into an existing physical system 502 in order to generate a multi-purpose prescriptive digital twin.
  • the system comprises a physical system 502, a knowledge database 504, simulation environment 506, a historical database 508, first digital twin 510, second digital twin 512, anomaly detection module 514, root cause analysis module 516, classification module 518, and recommendation and action module 520.
  • the system 500 is only an exemplary implementation and should not construed limiting to the specific components and/or modules as disclosed herein.
  • the first digital twin 510 which is a global digital twin is generated based on the first knowledge graph and the first machine learning model as aforementioned. Further, the first digital twin also comprises historical database 508 comprising previously detected anomalies, corresponding recommendations and so forth. Further, the second digital twin 512 which a local digital twin is generated based on the second knowledge graph that is distilled or extracted from the first knowledge graph.
  • the anomaly detection module 514 determines one or more anomalies in the assts using the reconstruction error analysis based on the first trained machine learning model. Further, the root cause analysis module 514 detects anomalies by tracing the second knowledge graph.
  • the classification module 516 scans the entire first knowledge to classify the nodes based on the status of each of the nodes and detect anomalies in real-time based on the classification.
  • the recommendation and action module 520 recommends one or more corrective measures to be executed by the second digital twin. The corrective actions are executed which can be restarting of a system to changing a database column or raising an alert to an operator for action.
  • the figure represents first or global digital twin as 604 and second or local digital twins as 606, 608 and 610 associated with corresponding assets (not shown here) on the edge.
  • the lines 612 represent input data to the edge, in particular the local digital twins 606, 608, 610.
  • the lines 614 represent input to the database.
  • the lines 616 represent data from digital twins.
  • the lines 618 represent trained parameters shared to global digital twin 604.
  • the lines 620 represent soft prediction from the global digital twin 604.
  • the lines 622 represent prediction from local digital twins 606, 608 and 610.
  • the lines 624 represent self-learning local parameters.
  • the lines 626 represent data for training the digital twin models.
  • the lines 628 represent action that is being recommended to executed by the local digital twins 606, 608, and 610. Subsequently, the prediction, detection, analysis, and recommendations are displayed on the display device 630.
  • the present invention facilitates accurate management of multiple assets 105in any technical installtion using a multi-purpose prescriptive digital twin.
  • the present invention ensures accurate prediction, anomaly detection, root cause analysis, recommendations and correction/control of assets in a timely and reliable manner.
  • the present invention determines reliable outcome using multivariate data from different sources and specifically including business requirements. This technique ensures that the business requirements are integrated in the digital twin.
  • the digital twin is aware of the environment leveraging the knowledge stored in the knowledge graphs.
  • the local digital twin based on distilled knowledge from the global digital twin ensures accurate root cause analysis of detected anomalies.
  • the present invention can be seamlessly integrated into with any asset and is scalable on a large scale.
  • the present invention is not limited to a particular computer system platform, processing unit, operating system, or network.
  • One or more aspects of the present invention may be distributed among one or more computer systems, for example, servers configured to provide one or more services to one or more client computers, or to perform a complete task in a distributed system.
  • one or more aspects of the present invention may be performed on a client-server system that comprises components distributed among one or more server systems that perform multiple functions according to various embodiments. These components comprise, for example, executable, intermediate, or interpreted code, which communicate over a network using a communication protocol.
  • the present invention is not limited to be executable on any particular system or group of system, and is not limited to any particular distributed architecture, network, or communication protocol.

Abstract

Disclosed is a system (100), method (200), and apparatus (110, 602) for managing assets (105) in technical installation. The method comprises receiving multivariate data from one or more sources (125), generating a first knowledge graph based on the multivariate data, generating a first digital twin (510, 604) using the first knowledge graph, extracting plurality of second knowledge graphs derived from the first knowledge graph, generating plurality of second digital twins (512, 606, 608, 610) corresponding to plurality of second graphs, providing one or more recommendations (628) for the assets (105) and controlling or executing actions using the plurality of second digital twins (512, 606, 608, 610) based on the prediction of the first digital twin (510, 604).

Description

Beschreibung / Description
SYSTEM, APPARATUS AND METHOD FOR MANAGING ONE OR MORE ASSETS
The present invention relates to a field of data and simulation systems and more particularly relates to management of one or more assets in a technical installation for recommendation and control using a prescriptive digital twin.
In an industrial environment, digital twin of plurality of assets of the technical installation is generated to predict behavior of the assets. The digital twin of the industrial plant may provide insights into several domains such as parameter prediction, remaining useful life prediction, failure or breakdown prediction, anomaly detection and so forth. However, currently digital twins are based on data from sensors associated with assets in the technical installation. The predictions of the digital twin is based on univariate data only and does univariate tasks. Furthermore, external factors such as requirements of the processes, resource availability, energy availability, manual inputs and so forth are not accounted for in the current digital twin systems. Thereby, the existing digital twins are inaccurate and inefficient.
In the existing approach, the digital twins are capable of performing tasks such as prediction, anomaly detection based only on univariate data from sensors. However, tasks based on the data collected from several sources, databases, requirements of users is not possible with current systems. Furthermore, there are many external factors that are not captured by the sensors are disregarded in the current systems which creates data gaps for instance, manual controls and/or unmonitored environmental conditions or load may lead to inherently unpredictable time-series data due to gaps. Therefore, performing prediction, detection and analysis tasks in such scenarios using standard approaches based on mathematical models that rely on stationarity, or prediction models that utilize prediction errors to perform such tasks becomes challenging. Furthermore, the current systems also lack in identifying the root cause of the any failure in the assets. In addition, another major challenge with existing systems is the lack of understanding the environment and assessing the parameters and requirements of the entire system. Therefore, the current digital twins lack in generating accurate recommendations for mitigating the predicted anomalies and optimizing the parameters to determine a desired outcome as per the requirement. Another major limitation of the existing digital twin systems is the inability to determine the actual point and cause of failure ,self-learn and execute action based on the its own recommendations. In light of the above, there exists a need for managing one or more assets in a technical installation using a digital twin that captures domain knowledge and business requirements in a seamless and accurate manner.
Therefore, it is an object of the present invention to provide a system, apparatus and method for managing one or more assets in a technical installation using a multi-purpose prescriptive digital twin for self-learning tasks, root cause analysis, recommendations for the assets and execution of the recommended actions.
The object of the present invention is achieved by a method managing one or more assets in a technical installation. The term ‘assets' refers to any device, system, instrument or machinery manufactured or used in an industry that may be employed for performing an operation. Example of assets include any machinery in a technical system or technical installa- tion/facility such as motors, gears, bearings, shafts, switchgears, rotors, circuit breakers, protection devices, remote terminal units, transformers, reactors, disconnectors, gear-drive, gradient coils, magnet, radio frequency coils etc. Example technical installation include turbines, large drives, manufacturing units, factory set-ups, etc.
The method comprises receiving multivariate data from one or more sources. The multivariate data comprises sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, and simulation data pertaining to the one or more assets in the technical installation. Throughout the present disclosure, the term ‘multivariate data’ refers to data as received from multiple sources and comprising more than one variable. Advantageously, the multivariate data ensures that every aspect of sensor data, domain knowledge and business requirements is considered while generating output such as predictions, recommendations, anomaly detection and root cause analysis, prediction, anomaly detection or tasks, self-learning root cause analysis , recommendations for the assets and execute recommended action. Throughout the present disclosure, the term ‘task’ as used herein refers to as but not limited to parameter prediction, remaining useful life prediction, failure or breakdown prediction, anomaly detection.
Throughout the present disclosure, the term ‘sources’ refers to may refer to electronic devices configured to obtain and transmit the data to the processing unit. Non-limiting examples of sources include sensors, controllers, edge devices, databases, and simulators.
Throughout the present disclosure, the term ‘one or more requirements’ refer to information related to energy availability, resource availability, asset availability, raw material availability, manual changes to the assets or parameter settings, process demand, business needs, business data and the like.
The method comprises generating a first knowledge graph based on the multivariate data. The first knowledge graph comprises a structured arrangement of plurality of nodes, each node pertaining to an asset in the technical installation. The multivariate data is used to extract node embedding, node embeddings are vectors that reflect properties of nodes in a network Throughout the present disclosure, the term, ‘first knowledge graph’ refers to a directed graph with one or more linked assets as nodes and connection between the assets as edges. Throughout the present disclosure, the term ‘graph embeddings’ refers to as both edge and nodes embeddings. Advantageously, the first knowledge graph represents a real- world replica of the assets with respect to nodes and furthermore to node and edge embeddings
The method comprises generating a first digital twin using the first knowledge graph. Herein, the first digital twin is a global digital twin for predicting one or more parameters of assets in the technical installation. Throughout the present disclosure, the term ‘first digital twin’ refers to refers to a dynamic virtual replica based on the first knowledge graph. According to an embodiment, the method of generating the first digital twin comprises training a machine learning model comprising an encoder decoder network.
The method comprises extracting plurality of second knowledge graphs derived from the first knowledge graph. Herein, each of the plurality of second knowledge graphs comprises information specific to the corresponding one or more assets associated therewith. Advantageously, the second knowledge graphs are distilled knowledge graphs which also includes subgraphs motifs that have been derived from the first knowledge graph for each specific asset.
According to an embodiment, the method of extracting the second knowledge graph from the first knowledge graph comprises associating a primary node pertaining to the second knowledge graph with the asset, wherein the primary node is directly related to the asset. The method comprises associating one or more secondary nodes to the second knowledge graph, wherein the one or more secondary nodes are adjacent to the primary node of the asset. The method comprises determining an embedding for the asset based on the primary node and one or more secondary-nodes.
The method comprises generating plurality of second digital twin corresponding to plurality of second graphs. Herein, each of the plurality of second digital twins are local digital twin associated with corresponding assets. Advantageously, the second digital twin is aware of neighboring assets and is capable of predicting the parameters accurately.
The method comprises providing one or more recommendations for the assets using the plurality of second digital twins based on the prediction of the first digital twin.
According to an embodiment, the method for determining one or more recommendations for the assets comprises generating the one or more corrective actions using the plurality of second digital twins based on the prediction of the first digital twin and the one or more requirements. Further, the method comprises executing the one or more corrective actions on the asset by the plurality of second digital twins deployed on edge.
According to an embodiment, the method further comprises detecting one or more anomalies in the assets using the first digital twin, and wherein detection of one or more anomalies further comprises identification of reconstruction errors in the encoder decoder network.
In an embodiment, the method of detecting one or more anomalies in the assets comprises classifying the one or more nodes of the first knowledge graph into one or more categories using a first machine learning model. The method comprises determining a status of the one or more nodes based on the determined categories, wherein the status is one of: a failure node, a partial failure node and a functional node. The method comprises selecting at least one node from the one or more nodes based on the status of the nodes, wherein the at least one selected node is either the partial failure node or failure node. The method comprises detecting one or more anomalies in the asset based on the selected at least one node.
According to an embodiment, the method further comprises automatically scanning the first knowledge graph and the second knowledge graph for detection of anomalies in real-time
According to an embodiment, the method comprises determining root cause of the detected anomalies using the plurality of second digital twins.
According to an embodiment, the method comprises rendering the one or more predicted parameters, detected anomalies, and one or more recommendations on a display device.
The object of the present invention is also achieved by an apparatus for managing one or more assets in a technical installation. The apparatus comprising one or more processing units and a memory unit communicatively coupled to the one or more processing units. The memory unit comprises an asset management module stored in the form of machine- readable instructions executable by the one or more processing units, wherein the asset management module is configured to perform method steps as mentioned above.
The object of the present invention is also achieved by a system for managing one or more assets in a technical installation. The system comprising one or more sources associated with assets, and an apparatus communicatively coupled to the one or more sources. The apparatus is configured for managing one or more assets of the technical installation according to the abovementioned method.
The object of the present invention is also achieved by a computer-program product having machine-readable instructions stored therein, which when executed by one or more processing units, cause the processing units to perform the method as abovementioned.
The above-mentioned attributes, features, and advantages of this invention and the manner of achieving them, will become more apparent and understandable (clear) with the following description of embodiments of the invention in conjunction with the corresponding drawings. The illustrated embodiments are intended to illustrate, but not limit the invention.
FIG 1 A illustrates a block-diagram of a system for managing one or more of assets in a technical installation, in accordance with an embodiment of the present invention;
FIG 1 B illustrates a block-diagram of an apparatus for managing one or more of assets in a technical installation, in accordance with an embodiment of the present invention;
FIG 2 depicts a flowchart of a method for managing one or more of assets in a technical installation, in accordance with an embodiment of the present invention;
FIG 3 depicts a flowchart of a method for extracting the second knowledge graph from the first knowledge graph, in accordance with an embodiment of the present invention;
FIG 4 depicts flowchart of a method for detecting one or more anomalies in the assets, in accordance with an embodiment of the present invention;
FIG 5 is an exemplary block-diagram for managing one or more assets in a technical installation, in accordance with an embodiment of the present invention; and FIG 6 is a block-diagram 600 representing exchange of data signal between different components of the system deployed in the technical installation, in accordance with an embodiment of the present invention.
Hereinafter, embodiments for carrying out the present invention are described in detail. The various embodiments are described with reference to the drawings, wherein like reference numerals are used to refer to like elements throughout. In the following description, for purpose of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more embodiments. It may be evident that such embodiments may be practiced without these specific details.
FIG 1A illustrates a block-diagram of a system 100 for managing one or more of assets 105 in a technical installation, in accordance with an embodiment of the present invention. Example of the technical installation may be a complex industrial set-up with a plurality of assets 105 such as a power plant, wind farm, power grid, manufacturing facility, process plants and so on. Throughout the present disclosure, the term ‘assets' 105 refer to any device, system, instrument or machinery manufactured or used in an industry that may be employed for performing an operation. Example of assets include any machinery in a technical system or technical installation/facility such as motors, gears, bearings, shafts, switchgears, rotors, circuit breakers, protection devices, remote terminal units, transformers, reactors, disconnectors, gear-drive, gradient coils, magnet, radio frequency coils etc.
Throughout the present disclosure, the term ‘multivariate data' refers to data from multiple one or more sources having more than one variable. The multivariate data may comprise sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, simulation data, business data stored in databases. The multivariate data may be obtained in different forms, formats, and order and in processed by the processing unit. The one or more sources may be electronic devices configured to obtain and transmit the data to the processing unit. Non-limiting examples of sources include sensing units, controllers, edge devices, databases, user devices, and simulation models.
Throughout the present disclosure, the term ‘simulation model’ as used herein refers to an analytical model in machine-executable form derived from at least one of a physics-based model, a data-driven model or a hybrid model associated with the asset. The simulation model may be a 1-dimensional(1 D) model, a 2-dimensional(2D) model, a 3-dimen- sional(3D) model or a combination thereof. The simulation model may generate one or more simulation instances based on the set of requirements. The simulation instance may be executed in a simulation environment as one of stochastic simulations, deterministic simulations, dynamic simulations, continuous simulations, discrete simulations, local simulations, distributed simulations, co-simulations or a combination thereof.
The system 100 comprises an apparatus 110 communicatively coupled to an edge device 115. The edge device 115 is connected to the apparatus 110 via a network 120, for example, local area network (LAN), wide area network (WAN), WiFi, etc. The edge device 115 is configured to receive multivariate data from one or more sources 125 associated with the one or more assets 105. The multivariate data may be received from one or more sources 125 such as sensing units, controllers, edge devices, databases, user devices, and simulation models. It should be understood that the system 100 also takes into consideration the business requirements of the user and is not only dependent on sensor data information. The dynamic data received from the process and operation of the technical installation is also input to the system 100 and is one or more requirements of the processes ongoing in the technical installation is also input to the system 100 using a user device 130. The one or more requirements may be information related to energy availability, resource availability, asset availability, raw material availability, manual changes to the assets or parameter settings, process demand, business needs and the like.
The one or more requirements may be stored in a memory of the edge device 115 or may be input to the edge device 115 by an operator. For example, the edge device 115 may be communicatively coupled to the display device 130. Non-limiting examples of display device 130 include, personal computers, workstations, personal digital assistants, human machine interfaces. The display device 130 may enable the user to input one or more requirements through a web-based interface. Upon receiving the one or more requirements from the user, the edge device 115 transmits a request for monitoring and managing one or more assets to the apparatus 110.
In the present embodiment, the apparatus 110 is deployed in a cloud computing environment. As used herein, “cloud computing environment” refers to a processing environment comprising configurable computing physical and logical resources, for example, networks, servers, storage, applications, services, etc., and data distributed over the network 120, for example, the internet. The cloud computing environment provides on-demand network access to a shared pool of the configurable computing physical and logical resources. The apparatus 110 may include an asset management module (shown in FIG 1 B) configured for managing, monitoring, detecting anomalies .recommending actions based on predictions and executing actions based on recommendation. Additionally, the apparatus 110 may include a network interface for communicating with the one or more edge devices 115 via the network 120.
In another embodiment, the apparatus 110 can be an edge computing device. As used herein “edge computing” refers to computing environment that is capable of being performed on an edge device (e.g., connected to one or more sources in an industrial setup and one end and to a remote server(s) such as for computing server(s) or cloud computing server(s) on other end), which may be a compact computing device that has a small form factor and resource constraints in terms of computing power. A network of the edge computing devices can also be used to implement the apparatus. Such a network of edge computing devices is referred to as a fog network.
The apparatus 110 comprises a processing unit 135, a memory unit 140, a storage unit 145, an input unit 155, an output unit 160 and a standard interface or bus 195, as shown in FIG 1 B. The apparatus 110 can be a computer, a workstation, a virtual machine running on host hardware, a microcontroller, or an integrated circuit. As an alternative, the apparatus 110 can be a real or a virtual group of computers (the technical term for a real group of computers is “cluster”, the technical term for a virtual group of computers is “cloud”).
The term ‘processing unit’ 135, as used herein, means any type of computational circuit, such as, but not limited to, a microprocessor, microcontroller, complex instruction set computing microprocessor, reduced instruction set computing microprocessor, very long instruction word microprocessor, explicitly parallel instruction computing microprocessor, graphics processor, digital signal processor, or any other type of processing circuit. The processing unit 135 may also include embedded controllers, such as generic or programmable logic devices or arrays, application specific integrated circuits, single-chip computers, and the like. In general, a processing unit 135 may comprise hardware elements and software elements. The processing unit 135 can be configured for multi-threading, i.e. the processing unit 135 may host different calculation processes at the same time, executing the either in parallel or switching between active and passive calculation processes.
The processing unit 135 is configured for receiving multivariate data from one or more sources 125. The multivariate data comprises sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, and simulation data pertaining to the one or more assets 105 in the technical installation. The processing unit 135 is configured for generating a first knowledge graph based on the multivariate data. The first graph comprises a structured arrangement of plurality of nodes, each node pertaining to an asset 105 in the technical installation. The processing unit 135 is configured for generating a first digital twin using the first knowledge graph. The first digital twin is a global digital twin for predicting one or more parameters of assets 105 in the technical installation. The processing unit 135 is configured for extracting plurality of second knowledge graphs derived from the first knowledge graph. Each of the plurality of second knowledge graph comprises information specific to the corresponding one or more assets 105 associated therewith. Generating plurality of second digital twin corresponding to plurality of second graphs. Each of the plurality of second digital twins are local digital twin associated with corresponding assets 105. The processing unit 135 is configured for providing one or more recommendations for the assets 105 using the plurality of second digital twins based on the prediction of the first digital twin, and the one or more requirements.
The memory unit 140 may be volatile memory and non-volatile memory. The memory unit 140 may be coupled for communication with the processing unit 135. The processing unit 135 may execute instructions and/or code stored in the memory unit 140. A variety of computer-readable storage media may be stored in and accessed from the memory unit 140. The memory unit 140 may include any suitable elements for storing data and machine- readable instructions, such as read only memory, random access memory, erasable programmable read only memory, electrically erasable programmable read only memory, a hard drive, a removable media drive for handling compact disks, digital video disks, diskettes, magnetic tape cartridges, memory cards, and the like.
The memory unit 140 comprises the asset management module 165 in the form of machine- readable instructions on any of the above-mentioned storage media and may be in communication to and executed by the processing unit 135. The asset management module 165 comprises a data acquisition module 170, a knowledge graph generation module 175, a digital twin generation module 180, an anomaly detection module 182, a root cause analysis module 185, a recommendation module 190, and action module 192.
The data acquisition module 170 is configured for obtaining multivariate data from one or more sources. Herein, the multivariate data comprises sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, business data and simulation data pertaining to the one or more assets in the technical installation. Further, the data acquisition module 170 is configured for pre-processing the multivariate data in a format that can be processed by the further modules. The pre-processing may include data normalization, data translation, data transformation, missing values check, data cleaning, noise filtering, and so forth. The knowledge graph generation module 175 is configured for a generating the first knowledge graph and the plurality of second knowledge graph. The knowledge graph generation module 175 is configured for generating the first knowledge graph based on the multivariate data. Further, the knowledge graph generation module 175 is configured for extracting the plurality of second knowledge graphs from the first knowledge graph. The knowledge graph generation module 175 is configured for storing the first knowledge graph and the second knowledge graph in a database 150.
The digital twin generation module 180 is configured for generating the first digital twin and the plurality of second digital twins. The digital twin generation module 180 is configured for generating the first digital twin using the first knowledge graph. The first digital twin is a global digital twin for predicting one or more parameters of assets in the technical installation. Further, the digital twin generation module is configured for generating plurality of second digital twins based on second knowledge graphs. Each of the plurality of second digital twins are local digital twins associated with corresponding assets.
The anomaly detection module 182 is configured for detecting one or more anomalies in the assets 105 using the first digital twin. The anomaly detection module 182 is further configured for identification of reconstruction errors in the encoder decoder network and identification of anomalies using the first machine learning model. The anomaly detection module 182 comprises continuous screening of the first knowledge graph to detect anomalies in assets 105 in real-time. Further, the anomaly detection module comprises classifying the nodes of the first knowledge graph into different categories for identifying a health status of the nodes and assets in the first knowledge graph using the second machine learning model.
The root cause analysis module 185 is configured for determining root cause of the detected anomalies using the plurality of second digital twins. Further, the root cause analysis module 185 is configured for tracing the second knowledge graph corresponding to one or more assets 105 associated with detected anomalies.
The recommendation module 190 is configured for providing one or more recommendations for the assets 105 using the plurality of second digital twins based on the prediction and graph embeddings of the first digital twin.
The action module 192 is configured for executing our one or more recommended actions for the assets 105 using the plurality of second digital twins. The storage unit 145 comprises a non-volatile memory which stores the domain knowledge and multivariate data as recevied from the one or more sources 125. The storage unit 145 includes the database 150 that comprises the first knowledge and the second knowledge graph, historical data such as previously detected anomalies, previous recommendations and so forth. The bus 195 acts as interconnect between the processing unit 135, the memory unit 140, the storage unit 145, the input unit 155 and the output unit 160. The input unit 155 enables the user to input one or more requirements or raise request for managing the assets 105. The output unit 160 is configured for presenting the status of assets 105, detected anomalies, predicted parameters, generated recommendations, and the like.
Those of ordinary skilled in the art will appreciate that the hardware depicted in FIG 1 A and 1 B may vary for different implementations. For example, other peripheral devices such as an optical disk drive and the like, Local Area Network (LAN)/ Wide Area Network (WAN)/ Wireless (e.g., Wi-Fi) adapter, graphics adapter, disk controller, input/output (I/O) adapter, network connectivity devices also may be used in addition or in place of the hardware depicted. The depicted example is provided for the purpose of explanation only and is not meant to imply architectural limitations with respect to the present disclosure.
FIG 2 depicts a flowchart of a method 200 for managing one or more of assets 105 in a technical installation, in accordance with an embodiment of the present invention. At step 202, multivariate data is recevied from one or more sources. As aforementioned, the multivariate data comprises sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, business data pertaining to the one or more assets 105 in the technical installation and simulation data pertaining to the one or more assets 105 in the technical installation. The multivariate data may be received from one or more sources 125 such as sensing units, controllers, edge devices, databases, user devices, and simulation models. The one or more requirements may be information related to energy availability, resource availability, asset availability, raw material availability, manual changes to the assets or parameter settings, process demand, business needs and the like. Such dynamic data pertaining to the process and operation of the technical installation is considered predicting one or more parameters of the assets 105. It will be appreciated that external factors such as the business requirements of the technical installation are also considered for determining output of the digital twins. Advantageously, the output of the digital twins based on the multivariate data is accurate as it considers all the factors involved in the operations of the assets 105. The multivariate data is received from one or more sources 125. The one or more sources 125 may refer to electronic devices configured to obtain and transmit the data to the processing unit 135. Non-limiting examples of sources 125 include sensors, controllers, edge devices, databases, and simulators. In an example, the one or more sources 125 may be sensors such as a temperature sensor, a velocity sensor, an acceleration sensor, a pressure sensor, and a force sensor. The output from the sensors may be in the form of temperature data, velocity data, acceleration data or pressure data. In an embodiment, the sensor data are obtained through data acquisition interfaces. In another example, the source 125 maybe a simulation model configured to forecast one or more parameters in the assets. In another example, the source 125 is a database storing data from sensors, simulation models and manual entries by the user. In yet another example, the source 125 is a display device 130 that enables user to enter requirements and data available through a web-based interface.
At step 204, the first knowledge graph is generated based on the multivariate data. The first knowledge graph comprises a structured arrangement of plurality of nodes. Herein each node pertains to the asset 105 in the technical installation. It should be understood that invention does not focus on one-on-one mapping of assets to the nodes alone. Advantageously, the first knowledge graph represents a real-world replica of the assets with respect to nodes. In an example, a gas turbine or generator has multiple moving components and the data from the components are captured by a plurality of sensors. The knowledge graph aims to capture the influence of multiple moving components by multiple sensors which determines a single output. Therefore, the nodes in the first knowledge graph are not restricted to the asset 105 alone, but also the components of the asset 105. The first knowledge graph captures a connection between different nodes and influence of other neighboring nodes on a particular node. It should be understood that since each asset 105 or a unit in the asset 105 does not operate independently, therefore it is essential to map assets 105 to the nodes and a correlation between the nodes. Notably, the first knowledge graph is a directed graph and node features, and vertex features are encapsulated in the first knowledge graph as graph embeddings. Herein, the nodes may be assets 105, one or more sources 125 such as sensors, databases, raw data and the edges are relationship between various assets 105, and one or more sources 125.
The first knowledge graph may contain one or more linked nodes capturing domain knowledge and indicative of parameters and corresponding responses pertaining to the one or more components of the asset 105, history of failures of the asset 105, historical data pertaining to parameter values of the one or more components of the asset 105, failure predictions and corresponding cause of failures and so forth. It should be noted that each of linked nodes is connected using one or more paths or edges. The first knowledge graph includes asset specific information, data pertaining to conceptualization of the asset 105, data pertaining to designing of the asset 105, data pertaining to manufacturing of the asset 105, data pertaining to assembling of the asset 105, data pertaining to manufacturing of the asset 105, process specific information, general process information, and any other information related to business requirement may be linked to each other in a defined order. Additionally, the first knowledge graph comprises failures and causes of the failures in a linked manner. In an example, a failure in the motor may be linked to reduced lubrication in the rotor, rotor failure may be linked to defects in the bearing and so forth. Additionally, the first knowledge graph also includes standard procedure data that may be specific to a process in the asset 105. The standard procedure data includes a set of pre-defined rules, a standard operating procedure, a pre-defined activity or a process, and the like. Furthermore, the first knowledge graph also comprises physics-based models for specific processes such as, manufacturing, assembling, testing and operation of the asset 105. Furthermore, the first knowledge graph also comprises data pertaining to previously detected anomalies or failures and the actions that were initiated to resolve the previously predicted failures.
At step 206, the first digital twin is generated using the first knowledge graph. Herein, the first digital twin is a global digital twin for predicting one or more parameters of assets 105 in the technical installation. Throughout the present disclosure, the term ‘first digital twin’ as used herein refers to a dynamic virtual replica based on one or more of physics-based models, Computer-Aided Design (CAD) models, Computer-Aided Engineering (CAE) models, one-dimensional (1 D) models, two-dimensional (2D) models, three-dimensional (3D) models, finite-element (FE) models, descriptive models, metamodels, stochastic models, parametric models, reduced-order models, statistical models, heuristic models, prediction models, ageing models, machine learning models, Artificial Intelligence models, deep learning models, system models, knowledge graphs and so on. The digital twin thus generated based on the first knowledge graph maintains high accuracy. The first digital twin is generated using the information as created in the first knowledge graph. Notably, the first digital twin is global digital twin for accurately predicting one or more parameters of the assets 105. It will be appreciated that the first digital twin thus generated accurately predicts one or more parameters based on multivariate data as designed in the first knowledge graph in an orderly fashion. Herein, the ‘first digital twin' is a ‘global digital twin' as it is capable of prediction parameters for each of the assets 105 in the technical installation as requested by the user. According to an embodiment, generating the first digital twin comprises using the first trained machine learning model comprising an encoder decoder network. Further, the method comprises deriving motifs which are recurrent and statistically significant subgraphs or patterns of a larger knowledge graph. Each of these sub-graphs, defined by a particular pattern of interactions between vertices, may reflect a framework in which particular functions are achieved efficiently. The term ‘first machine learning model’ as used herein refers to a neural network prediction model given a set of training dataset. In general, each individual sample of the training data is a pair containing a dataset (e.g., operating of the asset 105 and corresponding parameter values of the asset 105) and a desired output value or dataset (e.g., parameter values of the asset 105). The first machine learning model analyzes the training data and produces hierarchical graph embeddings as well as a predictor function. The predictor function, once derived through training, is capable of reasonably predicting or estimating the correct output value or dataset. Exemplary machine learning models may include artificial intelligence model such as deep neural network, convolutional neural network and but not limited to graph neural network and the like.
In particular, the first trained machine learning model is a generative model comprising the encoder decoder network. The encoder decoder network models the actual distribution of a particular asset 105 and learns joint probability distribution p(x,y). The encoder decoder network predicts the conditional probability based on a Bayesian model. As known in the art, Bayesian model allows the formulation of prior (multivariate data), generation of appropriate likelihood functions (prediction of requested parameters) using Bayes theorem. The first trained machine learning model is a self-attention machine learning model. Advantageously, the encoder decoder network in the invention does not require complete sequence information to be captured into a single fixed length vector. This technique overcomes the issues in encoding long range dependencies in the encoder decoder network and enables the first machine learning model to work efficiently with very long time series data. Herein, the encoder decoder network is a self-attention network. The self-attention network includes both alignment and translation technique. Attention is proposed as a method to both align and translate. Alignment is the process in time series that identifies which parts of the input sequence are relevant to each embedding in the output, whereas translation is the process of using the relevant information to select the appropriate output. The core idea of selfattention technique is to focus on the most relevant parts of the input sequence for each output and eliminate the translation issue. Using the first trained machine learning model, a context vector is developed that is filtered specifically for each output time step. This method ensures eliminates the requirement of encoding the input sequence into a single fixed context vector. According to an embodiment, the first digital is configured to detect one or more tasks but not limited to anomalies in the asset 105. The detection of one or more anomalies further comprises identification of reconstruction errors in the encoder decoder network. In an example, consider a time-series as shown in equation (1).
X = {x(1), x(2) x(n)} (1)
The first trained machine learning model is configured to reconstruct the normal time-series. Furthermore, the reconstruction errors are used to obtain the likelihood of a point in a test time-series being anomalous. Notably, for each point x(i), an anomaly score a(i) of the point being anomalous is calculated. The anomaly score comprises of two components a normal component and a residual component. It should be understood that there is minimal or no change in normal component during anomaly. Meanwhile, the residual component changes drastically during anomaly, thereby leading to a higher reconstruction error. Therefore, in case the value of anomaly score has high value of residual component, then an anomaly has been detected. Advantageously, such a technique enables to capture latent information which is not possible using basic statistical models or existing machine learning models. In an exemplary scenario, the first digital twin predicts the future parameter values of the asset 105 and consequently detects anomalies in the predicted data, thereby predicting future anomalies in real-time.
According to an embodiment, the method of detecting one or more anomalies further comprises automatically scanning the first knowledge graph and the second knowledge graph for detection of anomalies in real-time. The automatic scanning of the first knowledge for anomaly detection is further explained in FIG 4.
At step 208, the plurality of second knowledge graphs are extracted from the first knowledge graph. Herein, each of the plurality of second graphs comprises information specific to the corresponding one or more assets 105 associated therewith. The term ‘second knowledge graphs’ as used herein refers to distilled knowledge graphs that have derived from the first knowledge graph for each specific asset 105 in the technical installation. The method of extracting the second knowledge graph from the first knowledge graph is explained later in FIG 3. In an example, the first knowledge graph is a teacher network, and the second knowledge is a student network. In particular, in the training phase, the external knowledge is processed by a comprehensive first knowledge graph (sometimes referred to as a “teacher model”) to generate valuable information to teach a simple and efficient second knowledge graph (sometimes referred to as a “student model”). It will be appreciated that the second knowledge graphs are continuously updated with specifics from the corresponding assets 105 at the edge, thereby making the second knowledge graphs aware of the neighboring nodes. In particular, the second knowledge graphs are the distilled versions of the decoder in the encoder decoder network which are motifs aware. Furthermore, herein the present invention employs time delayed Embeddings which are more practical than the adjacency matrix since they pack node properties in a vector with a smaller dimension.
In particular, the embeddings include special attention on sub graph motifs. Herein, the embeddings determine an influence of neighboring assets on the asset. The method comprises extracting the second knowledge graph from the first knowledge graph based on the motifs aware Embeddings distillation. The model predicts the next motif to be attached to frist knowledge graph. Entry (/, j) of motif embeddings are the sum of the weights of all motifs containing both nodes / and j. The motif is specified by name and the type of motif instance can be one of but not limited to functional and structural. The weighting scheme can be one of: but not limited to unweighted, mean, and product. The decoder adds one motif at a time, the model needs to predict the attachment configuration also. This way first graph is award of the second graph within and using knowledge distillation we are able to distill the knowledge to a smaller less computation device.
In an exemplary embodiment, the extraction of second knowledge graphs from the first knowledge graph is based on training the student model based on the teacher model. The extraction is also based on motif prediction and attachment. For example, the extraction is based on attention based encoder decoder graph knowledge distillation alpha (AEDGKD- A) technique. The AEDGKD-A technique aims at transferring structural knowledge using mutual relations of data examples in the output presentation of the teacher model. The method comprises computing a relational potential i for each n-tuple of data examples and transfers information through the potential from the teacher model to the student model.
At step 210, the plurality of second digital twin corresponding to plurality of second graphs are generated. Herein, each of the plurality of second digital twins are local digital twin associated with corresponding assets 105. The term ‘second digital twin’ as used herein refers to a dynamic virtual replica based on one or more of physics-based models, Computer- Aided Design (CAD) models, Computer-Aided Engineering (CAE) models, one-dimensional (1 D) models, two-dimensional (2D) models, three-dimensional (3D) models, finite-element (FE) models, descriptive models, metamodels, stochastic models, parametric models, reduced-order models, statistical models, heuristic models, prediction models, ageing models, machine learning models, Artificial Intelligence models, deep learning models, system models, knowledge graphs and so on. The second digital twin thus generated based on the second knowledge graph maintains high accuracy as it is being continuously updated with asset specific data and data from other assets related to the particular asset 105. The second digital twin is generated using the information as extracted and updated in the second knowledge graph. Notably, the second digital twin is local digital twin for accurately predicting one or more parameters of the assets 105.
According to an embodiment, further comprising determining root cause of the detected anomalies using the plurality of second digital twins. The root causes analysis may be based on historical data. Herein, the historical data comprises response trends of the component and associated cause for the response trends. In one example, the analysis of the parameter values and the historical data is based on the fault tree analysis (FTA) technique. In another example, the historic data may be analyzed using one or more of descriptive techniques, exploratory techniques, inferential techniques, predictive techniques, causal techniques, qualitative analysis techniques, quantitative analysis techniques and so on. In an embodiment, the method of determining the root cause of the detected anomalies comprises tracing the second knowledge graph corresponding to one or more assets 105 associated with not limited to detected anomalies. It will be appreciated that the second digital twin thus generated is connected to the first digital and is capable of accurately determining cause of predicted anomalies and recommendation based on the determined root cause.
At step 212, one or more recommendations are provided for the assets 105 using the plurality of second digital twins based on the prediction of the first digital twin. The one or more recommendation may also be based on local graph embeddings of the first knowledge graph and the second knowledge graph. The one or more recommendation maybe generated for maintenance of the asset 105, optimizing performance of the asset 105, mitigating anomalies in the asset 105, and the like. In one embodiment, the recommendations may be generated based on the second knowledge graph. It will be appreciated that the second digital twin is a local twin specific to a particular asset 105 deployed on the edge and thereby is capable of accurately determining one or more recommendation for resolving the anomalies and/or determining actions based on requirements of the user. According to an embodiment determining one or more recommendations for the assets further comprises generating the one or more corrective actions using the plurality of second digital twins based on the prediction of the first digital twin and the one or more requirements. The corrective action may be scheduling a maintenance activity based on a predicted anomaly, optimizing efficiency of the technical installation, activate and deactivate assets 105 based on the requirements, and so forth. In an example, if the root cause of the failure of a turbine is worn blades, then recommendation may be associated with scheduling a maintenance of the turbine. In another example, if the business requirement is to increase production, then the recommendation is to increase running power of the assets 105. The method further comprises executing the one or more corrective actions on the asset by the plurality of second digital twins deployed on the edge. It will be appreciated the corrective measures are executed at the edge seamlessly since the second digital twins are associated with each asset 105. In an example, the second digital twin is associated with controller of the assets 105 and is configured to provide instruction to the corresponding controller in order to execute the recommendation.
FIG 3 depicts a flowchart of a method 300 for extracting the second knowledge graph from the first knowledge graph, in accordance with an embodiment of the present invention. At step 302, a primary node pertaining to the second knowledge graph is associated with the asset 105. Herein, the primary node is directly related to the asset 105. In an example, the primary node may be a motor. At step 304, associating one or more secondary nodes to the second knowledge graph. Herein, the one or more secondary nodes are adjacent to the primary node of the asset 105. In an example, if the primary node is a motor, then secondary nodes may be one or more components of the motor, sensors associated with motor, power source of the motor, output load of the motor, other assets that may be directly or indirectly connected to the motor. At step 306, determining embeddings for the asset 105 based on the primary node and one or more secondary nodes. The embeddings determine an influence of neighboring assets on the asset 105. In particular, the embeddings define correlation between the primary node and the secondary nodes. Advantageously, the embeddings determine the influence of parameters, operations and condition of the secondary nodes on the primary node associated with asset 105. The Embeddings thereby helps in determining root cause of the detected anomalies and generating one or more recommendations. At step 308, a loss function is calculated for the asset 105 based on the determined embeddings. The loss function determines how well the student model is being trained by the teacher model. Notably, the lower the loss function, the higher is the accuracy of the student model and the outcome is more reliable. The loss function is calculated by a mathematical formula as represented in equation (2).
Figure imgf000019_0001
Herein £ Primary graph training is the total loss as experienced by the teacher model. ^vector angle is the vector angle loss that is determined by matching the distance between the edges of the two selected embeddings. £JensOnshanon divergence loss is the Jenson Shanon divergence loss as experienced by the teacher model when the student model is learning. ^motiff pr diction is the motiff prediction loss that is determined at the stage of training. The motiff prediction loss is determined by calculating the deviation between the actual outcome from the teacher model and the predicted model. L motif f atatchement prediction is the motiff attachment predcition loss that is determined by calculating the deviation of prediction with respect to the neighboring nodes of the teacher model. £graph. prediction is the graph prediction loss that is determined by calculating the prediction losses of the teacher model. ^•student loss is the student loss that is determined by calculating the prediction losses of the student model.
Furthermore, the loss function experienced by the student model under the influence of the neighboring nodes is calculated by the a mathematical formula as represented in equation (3).
-^secondary graph training- -^neighbour nodeconnection T -^neighbour node structure + -^attention
(3)
Herein, £seCondary graph training is the total loss experienced by the student model with respect to the neighboring nodes. £neighbOur nodeconnection is the neighbor node connection loss that is determined by calculating the loss experienced in neighboring assets or devices. The neighbor node connection loss represents the magnitude of how good a particular node (assets or components) is being explained by the neighboring nodes. ^neighbour node structure is the neighbor node structure loss that is determined by understanding the node structure of the between the primary nodes and secondary nodes and calculating the losses in the node structure, attention is the attention loss that is determined by calculating the deviation of outcome of the neighboring nodes with respect to the primary node.
At step 310, the second knowledge graph is extracted from the first knowledge graph based on the embeddings and the calculated loss function. The second knowledge graph thus extracted has minimum losses and the outcome of the second digital twin based on the second knowledge graph is accurate and reliable.
FIG 4 depicts flowchart of a method 400 for detecting one or more anomalies in the assets 105, in accordance with an embodiment of the present invention. At step 402, the one or more nodes of the first knowledge graph are classified into one or more categories using a second machine learning model. The term ‘second trained machine learning model’ represents, however not limited to a classification model for detecting anomalies in real-time. The term “second machine learning model” as used herein refers to a prediction model given a set of training dataset. In general, each individual sample of the training data is a pair containing a dataset (e.g., parameters pertaining to the node) and a desired output value or dataset (e.g., labelled node determining health of the node). The second machine learning model analyzes the training data and produces, but not limited to a classification function. The classification function, once derived through training, is capable of classifying the nodes into different classes and labelling the node. Exemplary machine learning models may include artificial intelligence model such as deep neural network, convolutional neural network, and the like. The second machine learning model is trained to generate classifiers and label the nodes under one or more categories. At step 404, a status of the one or more nodes is determined based on the determined categories. Notably, the second machine learning model analyzes the status of the nodes based on various factors such as functionality of the asset 105 associated with node, output signatures from the asset 105 associated with the node, influence of other neighboring nodes on the asset 105 associated with the node and
In an instance, the status is one of: a failure node, a partial failure node and a functional node. The second machine learning model is trained to classify a node as ‘failure node’ in case an anomaly is detected based on the reconstruction errors as disclosed above. The second machine learning model is trained to classify a node as ‘partial failure node in case the node is indirectly affected by neighboring nodes or is recovering from a previously detected anomaly. The second machine learning model is trained to classify a node as ‘function node’ in case the node is unaffected and is fully functional. In an embodiment, the status of the scanned nodes may be presented on the user device 130 using color coding pertaining to health to each of the nodes. For example, the failure node maybe color coded ‘red’, the partial failure node may be color coded ‘yellow’, and the functional node may be color coded ‘green’. At step 406, at least one node is selected from the one or more nodes based on the status of the nodes. Herein, the at least one selected node is either the partial failure node or failure node. At step 408, one or more anomalies are detected in the asset 105 based on the selected at least one node.
FIG 5 is a block-diagram of an exemplary system 500 for managing one or more assets 105 in a technical installation, in accordance with an embodiment of the present invention. The system 500 is an exemplary implementation of integrating the proposed invention into an existing physical system 502 in order to generate a multi-purpose prescriptive digital twin. The system comprises a physical system 502, a knowledge database 504, simulation environment 506, a historical database 508, first digital twin 510, second digital twin 512, anomaly detection module 514, root cause analysis module 516, classification module 518, and recommendation and action module 520. It should be understood that the system 500 is only an exemplary implementation and should not construed limiting to the specific components and/or modules as disclosed herein. The existing physical system 502 may be understood as a technical installation with one or more assets 105 and the one or more sources 125. The sensor data, operational data, one or more requirements are acquired from the physical system 502 and is collected to be stored in the knowledge database 504. The knowledge database is also continuously updated with domain knowledge of the physical system. In an example, if the physical system 502 is power plant, then the knowledge database comprises information pertaining multiple aspects in power plants like grids, pres- surizers, turbines, transformers, generators, steam lines, cooling towers and also physics based models associated with the different components. The simulation environment 506 may generate generic simulation data based on the operations of the physical system 502. Furthermore, the first digital twin 510 which is a global digital twin is generated based on the first knowledge graph and the first machine learning model as aforementioned. Further, the first digital twin also comprises historical database 508 comprising previously detected anomalies, corresponding recommendations and so forth. Further, the second digital twin 512 which a local digital twin is generated based on the second knowledge graph that is distilled or extracted from the first knowledge graph. The anomaly detection module 514 determines one or more anomalies in the assts using the reconstruction error analysis based on the first trained machine learning model. Further, the root cause analysis module 514 detects anomalies by tracing the second knowledge graph. The classification module 516 scans the entire first knowledge to classify the nodes based on the status of each of the nodes and detect anomalies in real-time based on the classification. The recommendation and action module 520 recommends one or more corrective measures to be executed by the second digital twin. The corrective actions are executed which can be restarting of a system to changing a database column or raising an alert to an operator for action.
FIG 6 is a block-diagram 600 representing exchange of signals between different components of the system deployed in the technical installation, in accordance with an embodiment of the present invention. The block-diagram 600 is only for the purpose of illustration and should not be construed limiting to the components, elements, and signal flows disclosed herein. As shown herein, an apparatus 602 (such as the apparatus 110) is represented managing the one or assets in the technical installation. The apparatus 602 is configured for generating a multi-purpose prescriptive digital twin and outputting prediction values, detected anomalies, root cause analysis, corresponding recommendations and execute actions. The apparatus 602 is configured to recevied multivariate data from the one or more sources such as storage databases, sensors, user inputs and so forth. The figure represents first or global digital twin as 604 and second or local digital twins as 606, 608 and 610 associated with corresponding assets (not shown here) on the edge. The lines 612 represent input data to the edge, in particular the local digital twins 606, 608, 610. The lines 614 represent input to the database. The lines 616 represent data from digital twins. The lines 618 represent trained parameters shared to global digital twin 604. The lines 620 represent soft prediction from the global digital twin 604. The lines 622 represent prediction from local digital twins 606, 608 and 610. The lines 624 represent self-learning local parameters. The lines 626 represent data for training the digital twin models. The lines 628 represent action that is being recommended to executed by the local digital twins 606, 608, and 610. Subsequently, the prediction, detection, analysis, and recommendations are displayed on the display device 630.
The present invention facilitates accurate management of multiple assets 105in any technical installtion using a multi-purpose prescriptive digital twin. Advantageously, the present invention ensures accurate prediction, anomaly detection, root cause analysis, recommendations and correction/control of assets in a timely and reliable manner. The present invention determines reliable outcome using multivariate data from different sources and specifically including business requirements. This technique ensures that the business requirements are integrated in the digital twin. Advantageously, the digital twin is aware of the environment leveraging the knowledge stored in the knowledge graphs. Furthermore, the local digital twin based on distilled knowledge from the global digital twin ensures accurate root cause analysis of detected anomalies. Furthermore, the present invention can be seamlessly integrated into with any asset and is scalable on a large scale.
The present invention is not limited to a particular computer system platform, processing unit, operating system, or network. One or more aspects of the present invention may be distributed among one or more computer systems, for example, servers configured to provide one or more services to one or more client computers, or to perform a complete task in a distributed system. For example, one or more aspects of the present invention may be performed on a client-server system that comprises components distributed among one or more server systems that perform multiple functions according to various embodiments. These components comprise, for example, executable, intermediate, or interpreted code, which communicate over a network using a communication protocol. The present invention is not limited to be executable on any particular system or group of system, and is not limited to any particular distributed architecture, network, or communication protocol.
While the invention has been illustrated and described in detail with the help of a preferred embodiment, the invention is not limited to the disclosed examples. Other variations can be deducted by those skilled in the art without leaving the scope of protection of the claimed invention.
List of reference numbers:
100 system for managing plurality of assets
105 assets
110 apparatus
120 network
125 one or more sources
130 display device
135 processing unit
140 memory unit
145 storgae unit
150 database
155 input unit
160 output unit
165 asset management module
170 data acquisition module
175 asset mapping module
180 asset mapping module
182 performance indicator determination module
185 workflow definition module
190 machine learning model selection module
192 output module
195 bus
200 method for managing one or more of assets in a technical installation
300 method for extracting the second knowledge graph from the first knowledge graph
400 method for detecting one or more anomalies in the assets

Claims

Patentanspruche / Patent claims
1. A computer-implemented method (200) for managing an asset (105) in a technical installation, the method comprising: receiving, by a processing unit (135), multivariate data from one or more sources (125), wherein the multivariate data comprises sensor data, vibrations, information pertaining to one or more requirements of processes in the technical installation, and simulation data pertaining to the one or more assets (105) in the technical installation; generating, by the processing unit (135), a first knowledge graph based on the multivariate data, wherein the first graph comprises a structured arrangement of plurality of nodes, each node pertaining to a device in the technical installation; generating, by the processing unit (135), a first digital twin using the first knowledge graph, wherein the first digital twin (510, 604) is a global digital twin for predicting one or more parameters of assets (105) in the technical installation; extracting, by the processing unit (135), plurality of second knowledge graphs derived from the first knowledge graph, wherein each of the plurality of second knowledge graph comprises information specific to the corresponding one or more assets (105) associated therewith; generating, by the processing unit (135), plurality of second digital twins (512, 606, 608, 610) corresponding to plurality of second graphs, wherein each of the plurality of second digital twins (512, 606, 608, 610) are local digital twin associated with corresponding assets (105); providing, by the processing unit (135), one or more recommendations (628) for the assets (105) using the plurality of second digital twins (512, 606, 608, 610) based on the prediction of the first digital twin (510, 604).
2. The method (200) according to claim 1 , wherein determining one or more recommendations (628) for the assets (105) further comprises: generating the one or more corrective actions using the plurality of second digital twins (512, 606, 608, 610) based on the prediction of the first digital twin (510, 604) and the one or more requirements; executing the one or more corrective actions on the asset (105) by the plurality of second digital twins (512, 606, 608, 610) deployed on the edge.
3. The method (200) according to claim 1 , wherein generating the first digital twin (510, 604) comprises training a machine learning model comprising an encoder decoder network.
4. The method (200) according to claim 3, further comprises detecting one or more anomalies in the assets (105) using the first digital twin (510, 604), and wherein detection of one or more anomalies further comprises identification of reconstruction errors in the encoder decoder network.
5. The method (200) according to claim 1 , wherein extracting the second knowledge graph from the first knowledge graph comprises: associating a primary node pertaining to the second knowledge graph with the asset (105), wherein the primary node is directly related to the asset (105); associating one or more secondary nodes to the second knowledge graph, wherein the one or more secondary nodes are adjacent to the primary node of the asset (105); determining embeddings for the asset (105) based on the primary node and one or more secondary nodes, wherein the embeddings determine an influence of neighboring assets on the asset (105); calculating a loss function for the asset (105) based on the determined embeddings; and extracting the second knowledge graph from the first knowledge graph based on the determined embeddings and the loss function.
6. The method (200) according to any of the preceding claims, wherein detecting one or more anomalies in the assets (105) comprises: classifying the one or more nodes of the first knowledge graph into one or more categories using a first machine learning model; determining a status of the one or more nodes based on the determined categories, wherein the status is one of: a failure node, a partial failure node and a functional node; selecting at least one node from the one or more nodes based on the status of the nodes, wherein the at least one selected node is either partial failure node or failure node; detecting one or more anomalies in the asset (105) based on the selected at least one node.
7. The method (200) according to claim 6 further comprising automatically scanning the first knowledge graph and the second knowledge graph for detection of anomalies in real-time.
8. The method (200) according to any of the preceding claims, further comprising determining root cause of the detected anomalies using the plurality of second digital twins (512, 606, 608, 610).
9. The method (200) according to claim 8, wherein determining the root cause of the detected anomalies comprises tracing the second knowledge graph corresponding to one or more assets (105) associated with detected anomalies.
10. The method (200) according to any of the preceding claims further comprising rendering the one or more predicted parameters, detected anomalies, and one or more recommendations on a display device (130, 630).
11 . An apparatus (110, 602) for managing an asset in a technical installation, the apparatus comprising: one or more processing units (135); and a memory unit (140) communicatively coupled to the one or more processing units (135), wherein the memory unit (140) comprises a asset management module (165) stored in the form of machine-readable instructions executable by the one or more processing units (135), wherein the asset management module (165) is configured to perform method steps according to any of the claims 1 to 7.
12. A system (100) for managing an asset (105) in a technical installation, the apparatus comprising: one or more sources (125) associated with assets (105) in a technical installation; and an apparatus (110, 602) according to claim 11 , communicatively coupled to the one or more sources (125), wherein the apparatus (110, 602) is configured for managing one or more assets of the technical installments, according to any of the method claims 1 to 11.
13. A computer-program product having machine-readable instructions stored therein, which when executed by one or more processing units (135), cause the processing units (135) to perform a method according to any of the claims 1 to 11.
PCT/EP2022/058667 2022-03-31 2022-03-31 System, apparatus and method for managing one or more assets WO2023186317A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/EP2022/058667 WO2023186317A1 (en) 2022-03-31 2022-03-31 System, apparatus and method for managing one or more assets

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/EP2022/058667 WO2023186317A1 (en) 2022-03-31 2022-03-31 System, apparatus and method for managing one or more assets

Publications (1)

Publication Number Publication Date
WO2023186317A1 true WO2023186317A1 (en) 2023-10-05

Family

ID=81580051

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2022/058667 WO2023186317A1 (en) 2022-03-31 2022-03-31 System, apparatus and method for managing one or more assets

Country Status (1)

Country Link
WO (1) WO2023186317A1 (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190095806A1 (en) * 2017-09-28 2019-03-28 Siemens Aktiengesellschaft SGCNN: Structural Graph Convolutional Neural Network
EP3832413A1 (en) * 2019-12-05 2021-06-09 Siemens Aktiengesellschaft System, device and method for model based analytics
EP3975050A1 (en) * 2020-09-28 2022-03-30 Siemens Aktiengesellschaft Method and system for evaluating consistency of an engineered system

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190095806A1 (en) * 2017-09-28 2019-03-28 Siemens Aktiengesellschaft SGCNN: Structural Graph Convolutional Neural Network
EP3832413A1 (en) * 2019-12-05 2021-06-09 Siemens Aktiengesellschaft System, device and method for model based analytics
EP3975050A1 (en) * 2020-09-28 2022-03-30 Siemens Aktiengesellschaft Method and system for evaluating consistency of an engineered system

Similar Documents

Publication Publication Date Title
US11921566B2 (en) Abnormality detection system, abnormality detection method, abnormality detection program, and method for generating learned model
JP7069269B2 (en) Semi-supervised methods and systems for deep anomaly detection for large industrial surveillance systems based on time series data using digital twin simulation data
AU2019206069B2 (en) Augmented reality (ar) based fault detection and maintenance
Al-Dahidi et al. Remaining useful life estimation in heterogeneous fleets working under variable operating conditions
US20220366244A1 (en) Modeling Human Behavior in Work Environment Using Neural Networks
Molęda et al. From corrective to predictive maintenance—A review of maintenance approaches for the power industry
Guida et al. An Integrated BIM-IoT approach to support energy monitoring
CN115769235A (en) Method and system for providing an alert related to the accuracy of a training function
Wang et al. Linear approximation fuzzy model for fault detection in cyber-physical system for supply chain management
WO2023072724A1 (en) System, apparatus and method for monitoring condition of an asset in technical installation
Li et al. Intelligent reliability and maintainability of energy infrastructure assets
Welte et al. Integration of degradation processes in a strategic offshore wind farm O&M simulation model
Kerboua et al. Fault diagnosis in an asynchronous motor using three-dimensional convolutional neural network
Llopis et al. Manufacturing Maps, a Novel Tool for Smart Factory Management Based on Petri Nets and Big Data Mini-Terms
Liu et al. Residual useful life prognosis of equipment based on modified hidden semi-Markov model with a co-evolutional optimization method
Rodríguez-López et al. Methodology for detecting malfunctions and evaluating the maintenance effectiveness in wind turbine generator bearings using generic versus specific models from SCADA data
Bond et al. A hybrid learning approach to prognostics and health management applied to military ground vehicles using time-series and maintenance event data
WO2023186317A1 (en) System, apparatus and method for managing one or more assets
Gao et al. Semi-supervised small sample fault diagnosis under a wide range of speed variation conditions based on uncertainty analysis
Colace et al. Unsupervised Learning Techniques for Vibration-Based Structural Health Monitoring Systems Driven by Data: A General Overview
Zhao A probabilistic approach for prognostics of complex rotary machinery systems
US11310349B1 (en) Transforming multivariate time series data into image data to generate image-based predictions
CN114450646B (en) System and method for detecting wind turbine operational anomalies using deep learning
Saleem et al. IoT-based Digital Twin Braiding Machinery for Defect Detection
Blanchard et al. Data-driven prognostics and diagnostics for turbofan engine

Legal Events

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

Ref document number: 22720934

Country of ref document: EP

Kind code of ref document: A1