Method and system for causal inference and root cause identification in industrial processes
Patent Information
- Application Number
- IN202221041919
- Authority / Receiving Office
- IN · IN
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-07-21
- Publication Date
- 2026-08-07
- Estimated Expiration
- 2042-07-21
AI Technical Summary
Current methods for real-time fault detection and diagnosis in industrial processes are impractical for domain experts due to the complexity of identifying causal networks and root causes, with knowledge-based approaches requiring significant initial effort and data-driven methods struggling with nonlinear and non-stationary processes.
A processor-implemented method using deep neural networks with stochastic proximal gradient descent and adaptive learning rate for causal inference, which processes multivariate time-series data to identify root cause variables and fault propagation paths by preprocessing data, detecting faults, and generating causal structures.
Enables real-time causal network learning for nonlinear and non-stationary industrial processes without prior knowledge, improving fault detection and diagnosis accuracy and efficiency.
Abstract
Description
TECHNICAL FIELD
[001] The disclosure herein generally relates to the field of industrial dataanalytics and specifically, to a method and system for a causal network learningand identifying one or more root cause variables and one or more fault propagationpaths for faults in industrial processes and equipment.BACKGROUND
[002] Effective fault detection and diagnosis is a key step towardspredictive and prescriptive maintenance of industrial processes and equipment.Here, fault diagnosis is an umbrella term that typically involves identification ofkey variables / sensors bearing the fault signature (fault localization / isolation),classification of the detected fault into one or more known fault classes (faultclassification) and detecting the root cause / source of the fault (causality or RCA).Of these, real-time root cause identification of faults or abnormal events is a keyask from industries as it gives plant operators and managers an opportunity toaddress the problem in real-time before the fault progresses and leads to a failure.For example, in coke oven batteries, whenever there is a change in coke quality(process fault), operators would want to know if it is due to changes in cokingparameters, health of the coke ovens or the chemical composition of coal blend sothat appropriate corrective action can be taken.
[003] Considering the growing need for real-time process monitoring andreliable fault detection and diagnosis in industries and given the large number ofprocess variables, it is not practical for domain experts or plant operators to localizethe fault or identify the causal network of variables responsible for the fault in real25time. Most of the time, causal networks are generated using prior processknowledge or manually by subject matter experts (SMEs). There are two types ofapproaches for learning causal networks and root cause identification (RCI) offaults, namely knowledge-based and data-driven approaches. Knowledge-basedmethods (such as FMEA) require apriori knowledge of faults / failures and therelationship between faults and observations (symptoms). While such knowledgecan be derived from fundamental understanding of the process, sources of domainknowledge and experience with the process, the initial effort required for thisapproach is significant and the gathered knowledge may not be exhaustive leadingto missed or incorrect identifications in some cases. On the other hand, data-drivenmethods rely entirely on historical and current operating data and can be appliedwith minimum initial effort. However, existing data-driven techniques cannoteffectively identify nonlinear relationships among the variables and cannot dealwith non-stationarity in the data. Therefore, these techniques are not effective forlearning causal networks of nonlinear and nonstationary industrial processes.SUMMARY
[004] Embodiments of the disclosure present technological improvementsas solutions to one or more of the above-mentioned technical problems recognizedby the inventors in conventional systems. For example, in one embodiment, asystem and method for causal network learning and identifying one or more rootcause variables and one or more fault propagation paths for faults in industrialprocesses is provided.
[005] In one aspect, a processor-implemented method for causal networklearning and identifying one or more root cause variables and one or more faultpropagation paths for faults in industrial processes is provided. The processor20implemented method includes one or more steps such as receiving a multivariatetime-series data of one or more sensors from a plurality of sources, pre-processingthe received multivariate time-series data using a plurality of pre-processingtechniques to remove noise and spurious values, imputing / interpolating missingvalues of the plurality of received data and resampling the received multivariatetime series data to a uniform frequency. Further, the processor-implementedmethod comprises detecting at least one fault in the pre-processed multivariate timeseries data and a plurality of fault variables responsible for the at least one faultusing a fault detection and diagnosis technique, dividing the preprocessedmultivariate time-series data of the plurality of fault variables along the time axisinto a plurality of windows of a predefined dimension, training a deep neuralnetwork for causal inference using each of the plurality of windows and a stochasticproximal gradient descent (SPGD) technique with an adaptive learning rate,extracting one or more encoding parameters from one or more encoding layers ofthe deep neural network and one or more lag decomposition parameters from oneor more lag decomposition layers of the deep neural network, computing causalityscores for each of the one or more sources corresponding to the target variable usingFrobenius norm of one or more encoding parameters to obtain causality score forone or more source-target variable pairs, applying a causality significance test onthe computed one or more causality scores to identify a threshold value and selectone or more source-target variable pairs having a causality score above theidentified threshold in the multivariate time-series data, determining a temporal lagbased on a contribution to the lag decomposition parameters corresponding to eachof the one or more selected source-target variable pairs having a causality scoreabove the threshold, generating a causal structure of fault propagation from theselected one or more source-target variable pairs and the determined temporal lagsand identifying one or more root cause variables and one or more fault propagationpaths based on traversing the generated causal structure.
[006] In another aspect, a system for causal network learning andidentifying one or more root cause variables and one or more fault propagationpaths for faults in industrial processes is provided. The system includes aninput / output interface configured to receive a multivariate time-series data of oneor more sensors from a plurality of sources and preprocessing the receivedmultivariate time-series data using a plurality of pre-processing techniques toremove noise and spurious values, imputing / interpolating missing values of theplurality of received data and resampling the received multivariate time series datato a uniform frequency. Further, the system is configured to detect at least one faultin the pre-processed multivariate time series data and a plurality of fault variablesresponsible for the at least one fault using a fault detection and diagnosis technique,divide the preprocessed multivariate time-series data of the plurality of faultvariables along the time axis into a plurality of windows of a predefined dimension,train a deep neural network for causal inference using each of the plurality ofwindows and a stochastic proximal gradient descent (SPGD) technique with anadaptive learning rate, extract one or more encoding parameters from one or moreencoding layers of the deep neural network and one or more lag decompositionparameters from one or more lag decomposition layers of the deep neural network,compute causality scores for each of the one or more sources corresponding to thetarget variable using Frobenius norm of one or more encoding parameters to obtaincausality score for one or more source-target variable pairs, apply a causalitysignificance test on the computed one or more causality scores to identify athreshold value and select one or more source-target variable pairs having acausality score above the identified threshold in the multivariate time-series data,determine a temporal lag based on a contribution to the lag decompositionparameters corresponding to each of the one or more selected source-target variablepairs having a causality score above the threshold, generate a causal structure offault propagation from the selected one or more source-target variable pairs and thedetermined temporal lags and identify one or more root cause variables and one ormore fault propagation paths based on traversing the generated causal structure.
[007] In yet another aspect, one or more non-transitory machine-readableinformation storage mediums are provided comprising one or more instructions,which when executed by one or more hardware processors causes a method ofcausal network learning and identifying one or more root cause variables and oneor more fault propagation paths for faults in industrial processes is provided. Theprocessor-implemented method includes one or more steps such as receiving amultivariate time-series data of one or more sensors from a plurality of sources, andpreprocessing the received multivariate time-series data using a plurality of preprocessingtechniques to remove noise and spurious values, imputing / interpolatingmissing values of the plurality of received data and resampling the receivedmultivariate time series data to a uniform frequency. Further, the processorimplementedmethod detecting at least one fault in the pre-processed multivariatetime series data and a plurality of fault variables responsible for the at least onefault using a fault detection and diagnosis technique, dividing the preprocessedmultivariate time-series data of the plurality of fault variables along the time axisinto a plurality of windows of a predefined dimension, training a deep neuralnetwork for causal inference using each of the plurality of windows and a stochasticproximal gradient descent (SPGD) technique with an adaptive learning rate,extracting one or more encoding parameters from one or more encoding layers ofthe deep neural network and one or more lag decomposition parameters from oneor more lag decomposition layers of the deep neural network, computing causalityscores for each of the one or more sources corresponding to the target variable usingFrobenius norm of one or more encoding parameters to obtain causality score forone or more source-target variable pairs, applying a causality significance test onthe computed one or more causality scores to identify a threshold value and selectone or more source-target variable pairs having a causality score above theidentified threshold in the multivariate time-series data, determining a temporal lagbased on a contribution to the lag decomposition parameters corresponding to eachof the one or more selected source-target variable pairs having a causality scoreabove the threshold, generating a causal structure of fault propagation from theselected one or more source-target variable pairs and the determined temporal lagsand identifying one or more root cause variables and one or more fault propagationpaths based on traversing the generated causal structure.
[008] It is to be understood that both the foregoing general description andthe following detailed description are exemplary and explanatory only and are notrestrictive of the invention, as claimed.BRIEF DESCRIPTION OF THE DRAWINGS
[009] The accompanying drawings, which are incorporated in andconstitute a part of this disclosure, illustrate exemplary embodiments and, togetherwith the description, serve to explain the disclosed principles:
[010] FIG. 1 illustrates an exemplary system for causal network learningand identifying one or more root cause variables and one or more fault propagationpaths for faults in industrial processes and equipment, in accordance with someembodiments of the present disclosure.
[011] FIG. 2 illustrates an schematic diagram illustrating components ofthe system for a causal network learning (CNL) for causal inference in industrialmanufacturing processes and equipment, in accordance with some embodiments ofthe present disclosure.
[012] FIG. 3 is a schematic diagram to illustrate the architecture of a DeepNeural Model (DNM) for causal inference, in accordance with some embodimentsof the present disclosure.
[013] FIG. 4 is a flow chart to illustrate a method for causal networklearning, in accordance with some embodiments of the present disclosure.
[014] FIG. 5 illustrates a processor-implemented method for causalnetwork learning and identifying one or more root cause variables and one or morefault propagation paths for faults in industrial processes, in accordance with someembodiments of the present disclosure.
[015] FIG. 6 shows the results obtained from CNL module for the fault inthe Tennessee Eastman Process. FIG. 6(a) represents the weight matrix where eachvalue in the matrix represents the magnitude of significant causal contributionbetween source and target variables. FIG. 6(b) illustrates the obtained causalnetwork, in accordance with some embodiments of the present disclosure.
[016] It should be appreciated by those skilled in the art that any blockdiagrams herein represent conceptual views of illustrative systems and devicesembodying the principles of the present subject matter. Similarly, it will beappreciated that any flow charts, flow diagrams, and the like represent variousprocesses, which may be substantially represented in computer readable mediumand so executed by a computer or processor, whether or not such computer orprocessor is explicitly shown.DETAILED DESCRIPTION OF EMBODIMENTS
[017] Exemplary embodiments are described with reference to theaccompanying drawings. In the figures, the left-most digit(s) of a reference numberidentifies the figure in which the reference number first appears. Whereverconvenient, the same reference numbers are used throughout the drawings to referto the same or like parts. While examples and features of disclosed principles aredescribed herein, modifications, adaptations, and other implementations arepossible without departing from the spirit and scope of the disclosed embodiments.It is intended that the following detailed description be considered as exemplaryonly, with the true scope and spirit being indicated by the following claims.
[018] There is a growing need for real-time monitoring and reliable faultdetection and diagnosis in industries. In case of a fault / failure, given the largenumber of process variables, it is not practical for domain experts or plant operatorsto identify causal chain that led to the fault and pin-point the root causes. There arebroadly two approaches to root cause identification, namely knowledge-based anddata-driven approaches. Knowledge-based methods (such as FMEA) require apriori knowledge of faults / failures and the relationship between faults andobservations (symptoms). While such knowledge can be derived from fundamentalunderstanding of the process, sources of domain knowledge and experience withthe process, the initial effort required for this approach is significant and thegathered knowledge may not be exhaustive leading to missed identifications insome cases. On the other hand, data-driven methods rely entirely on historical andcurrent operating data and can be applied with minimum initial effort.
[019] The embodiments herein provide a method and system foridentifying one or more root cause variables and one or more fault propagationpaths for faults in industrial processes. The present invention is a method andsystem for real-time causal network learning for manufacturing processes andequipment faults. One of the aspects of disclosure is to learn the causal network(graph structure) for a dynamic non-stationary and nonlinear complex process orsystem fault using observed data without any prior process knowledge. The causalnetworks of faults are identified in real-time using a deep learning-based causalnetwork learning technique.
[020] Referring now to the drawings, and more particularly to FIG. 1through 5, where similar reference characters denote corresponding featuresconsistently throughout the figures, there are shown preferred embodiments andthese embodiments are described in the context of the following exemplary systemand / or method.
[021] FIG. 1 illustrates an exemplary system (100) for causal networklearning and identifying one or more root cause variables and one or more faultpropagation paths for faults in industrial processes and equipment, in accordancewith an example embodiment. Although the present disclosure is explainedconsidering that the system (100) is implemented on a server, it may be understoodthat the system (100) may comprises one or more computing devices (102), such asa laptop computer, a desktop computer, a notebook, a workstation, a cloud-basedcomputing environment and the like. It will be understood that the system (100)may be accessed through one or more input / output interfaces 104-1, 104-2... 104-N, collectively referred to as I / O interface (104). Examples of the I / O interface(104) may include, but are not limited to, a user interface, a portable computer, apersonal digital assistant, a handheld device, a smartphone, a tablet computer, aworkstation, and the like. The I / O interface (104) are communicatively coupled tothe system (100) through a network (106).
[022] In an embodiment, the network (106) may be a wireless or a wirednetwork, or a combination thereof. In an example, the network (106) can beimplemented as a computer network, as one of the different types of networks, suchas virtual private network (VPN), intranet, local area network (LAN), wide areanetwork (WAN), the internet, and such. The network (106) may either be adedicated network or a shared network, which represents an association of thedifferent types of networks that use a variety of protocols, for example, HypertextTransfer Protocol (HTTP), Transmission Control Protocol / Internet Protocol(TCP / IP), and Wireless Application Protocol (WAP), to communicate with eachother. Further, the network (106) may include a variety of network devices,including routers, bridges, servers, computing devices, storage devices. Thenetwork devices within the network (106) may interact with the system (100)through communication links. Further, the system (100) comprising one or morehardware processors (108), a memory (110) in communication with the one or morehardware processors (108) and a database (112).
[023] The system (100) supports various connectivity options such asBLUETOOTH, USB, ZigBee and other cellular services. The networkenvironment enables connection of various components of the system (100) usingany communication link including Internet, WAN, MAN, and so on. In anexemplary embodiment, the system (100) is implemented to operate as a standalonedevice. In another embodiment, the system (100) may be implemented towork as a loosely coupled device to a smart computing environment. Thecomponents and functionalities of the system (100) are described further in detail.
[024] FIG. 2 is a schematic diagram (200) illustrating the components ofthe system (100) for a causal network learning (CNL) for causal inference and rootcause identification (RCI) in industrial manufacturing processes and equipment, inaccordance with some embodiments of the present disclosure. The system (100)comprises a plurality of modules for performing various functions. It comprises aplurality of industrial plant data sources, a plant automation system or distributedcontrol system (DCS), a server, real-time CNL and RCI modules, a modelrepository, a knowledge database, and a plurality of databases. It should beappreciated that the knowledge database and the plurality of databases could be thepart of the data repository. It would be appreciated that the multivariate time seriesdata from all the databases from industrial plant is brought to the system (100) viathe server.
[025] Herein, the input / output interface (104) is configured to receive aplurality of data. The plurality of data includes a real-time and a non-real-time datafrom various databases at a pre-determined frequency (e.g., 1 / second, 1 / minute,1 / hour, etc.), where the frequency is configurable by the user. The real-time dataincludes operations data and environment data. The operations data is recorded bysensors mounted on the industrial assets and includes temperatures, pressures, flowrates, liquid levels, vibrations, etc. from processes and equipment in the units of theindustrial plant. The operations data is obtained from a distributed control system(DCS), OPC server, etc. and could be stored in an operations database or historian.The environment data such as ambient temperature, atmospheric pressure, ambienthumidity, rainfall, etc. is also recorded by sensors and could be stored in anenvironment database.
[026] The non-real-time data includes data from the laboratories andmaintenance activities. The laboratory data comprises characteristics (e.g.,chemical composition, size distribution, concentration, density, viscosity, calorificvalue, microstructural composition, etc.) of raw materials, products, byproducts,solid and liquid waste, and emissions that are tested at the laboratory. Thelaboratory data is typically stored and retrieved from a laboratory informationmanagement system (LIMS), relational database (RDB) or SQL database. Theinformation related to the condition of the process and equipment, plant runningstatus, maintenance activities performed on the industrial units, etc. is stored andretrieved from a maintenance database.
[027] In another embodiment, the system (100) is configured to pre10process the plurality of data, received from multiple databases of the industrialplant. The pre-processing involves removal of redundant data, unification ofsampling frequency, filtering of data, outlier identification & removal, imputationof missing data, synchronization of data by incorporating appropriate lags, andintegration of variables from various data sources.
[028] In yet another embodiment, the system (100) is configured tocompute a plurality of soft-sensor data using the preprocessed plurality of data ofthe plurality of industrial assets. It is to be noted that sometimes the direct use ofphysical sensors may not be possible in some applications or environments due toconstraints such as access requirements, extreme operating conditions, disturbancesin environment, limitations of the sensors, etc. Further, the information collectedfrom physical sensors may not be sufficient to capture the behavior of the assets. Insuch cases, soft sensors (either physics-based soft sensors or data-driven softsensors) are used to generate additional information about the plant. Examples ofsoft sensors include temperature in the firing zone of a furnace, concentration ofproduct or byproducts inside a reactor, etc. Soft sensors (physics-based and datadriven)are stored in soft sensor database.
[029] In another embodiment, the system (100) is configured to run one ormore fault detection models for identifying faulty data instances. The faulty datainstances are identified using data-driven models developed using statistical,machine learning and deep learning techniques such as Mahalanobis Distance,principal component analysis (PCA), Local Outlier Factor, Isolation Forest, EllipticEnvelope, One-class Support Vector Machine, K-Nearest Neighbors, artificialneural networks (ANNs) and their variants such as Auto-Encoders (AEs) includingLSTM-AE, Dense-AE and CNN-AE.
[030] Using real-time data and fault detection models, a fault score ispredicted for the current time as well as forecasted for future time instances (e.g.,next 1 hour). If a fault is detected using the forecasted fault score, then faultlocalization, generation of causal structure(s) and identification of root causevariables can be done in advance and appropriate recommendations related to shutdown, maintenance or optimization can be provided to the operators. The detectedfaulty data instances are used by the system (100) to identify variables responsiblefor the detected fault. Further, the fault localization is carried out for the faultyinstances and candidate variables carrying the signature of abnormal event / faultsare identified. Complete decomposition contribution (CDC), leave-one-variableout(LOVO) and reconstruction error-based fault localization techniques are usedfor identifying the candidate faulty variables along with their contribution scores tothe fault.
[031] The adaptive learning module of the system (100) is configured toretrain the fault detection and localization models based on newer process andequipment data in case of performance degradation of the models due to changesin raw materials, operating strategy of the plant or wear and tear of the equipmentwith time.
[032] The identified list of faulty variables and the data corresponding tothese variables are sent to the causal network learning module of the system (100)to learn a causal network of fault propagation. The root cause identification moduleof the system (100) is configured to identify fault propagation paths and root causevariables for the fault by visual inspection or by traversing the causal network usinggraph traversal techniques. The recommendation module of the system (100) isconfigured to suggest one or more corrective actions to the plant personnel basedon the identified root cause variables and fault propagation paths.
[033] Referring FIG. 3, a schematic diagram (300) to illustrate thearchitecture of a Deep Neural Model (DNM) for causal inference, in accordancewith some embodiments of the present disclosure. It would be appreciated that theDNM takes windowed / segmented data as an input for training in order tooptimize / minimize a predetermined cost function.. The DNM is generated based onthe window length, number of variables and hyperparameters describing thenumber of layers. Herein, the DNM is designed to forecast the output windowyt1 to ytp comprising the values of all P variables for the next time instance, t given theinput window comprising the previous values (t-1 to t-m) of all variables. The inputwindow is passed through a series of network layers, namely the encoding layer,the time lag decomposition layer, and the forecasting layer to generate the outputwindow. The encoding layer comprises a series of RNNs which project each inputvariable to a different dimension as an encoded input. The encoded input is thenfed to the time lag decomposition layer which applies 1-D convolution across thelength of the window. Convolutions are performed to generate a single vectorrepresentation of the input variable which is named as the 'context vector'. Contextvectors from all input layers are concatenated and passed through a forecastinglayer which utilizes multilayer perceptron (MLP) to predict the output window. TheDNM is capable of learning the non-stationarity in the input multivariate timeseriesdata. The cost function for the DNM can be one of mean squared error(MSE), root mean squared error (RMSE) or mean absolute error (MAE) along withan additional causal penalty term. The weights and biases in the DNM are updatedduring model training such that the value of the cost function is as low as possible.
[034] Referring FIG. 4, a flow diagram (400) illustrating the working ofthe causal network learning module of the system (100), in accordance with someembodiments of the present disclosure. Herein, the causal network learning moduleof the system (100) comprises data segmentation, model training and causalnetwork generation steps.
[035] The input multivariate time-series data is obtained from the faultysensors comprising P number of variables recorded for time instances T. Thesystem (100) divides input data into windows of fixed length m along the time-axis(temporal dimension). Further, each window, is split into source X and target ywherein the source consists of values of the variables at previous time steps, i.e.,[Xt-1 i , Xt-2 i , … , Xt-m i ] for each variable Xi and the target y comprises values of thevariables at the present timestep, i.e., yti for each variable yi.
[036] The DNM is trained to minimize an error metric for forecast withsparsity inducing penalty in the causal parameters. The error metric can be one ofroot mean squared error (RMSE), mean squared error (MSE), and mean absoluteerror (MAE). The causal parameters include the input encoding layer (Wh) and lagdecomposition layer (Wτ) parameters. This allows relative difference in magnitudeof parameters among causally significant and insignificant variables and theirrespective temporal lags. The penalty is tuned by hyperparameters λ1 and λ2. Theresulting loss function is defined with mean squared error (MSE) as error metric asfollows:Formula(1)
[037] The loss function in equation (1) is nonconvex in nature and difficultto converge. The gradients for the loss function are computed via backpropagationthrough time (BPTT) and network parameters are updated using stochasticproximal gradient descent (SPGD) with adaptive learning rate. Proximal gradientdescent ensures sparsity in causal parameters which is important for causalinterpretation and adaptive learning rate allows better convergence. It is to be notedthat after network training, the system (100) is configured to extract causalinformation from the network parameters.
[038] The DNM input encoding layer and time lag decomposition layerparameters, Wh and Wτ respectively, are extracted from the trained network. Wh isused to interpret significant causal connections among variables while Wτ is toidentify most significant temporal lag for a causal relationship. Hyperparametercausal network density ((element of)) is tuned to get the threshold and significant causalvariables K'. The causality significance test to identify K'is defined as:Formula(2)wherein Q(element of)(x) represents percentile range (PR) from (element of) to 1.
[039] The most significant temporal lag τ' for a causal variable K' isidentified as the lag with maximum contribution in the time lag decompositionlayer.Formula(3)wherein, ||. ||F is frobenius norm of a matrix.
[040] The set of causal connections and temporal lags obtained fromequations (2) & (3) are used to generate the causal network of the detected processfault. The faulty variables are represented as nodes while the set of causalconnections generated from Eq. (2) serve as directed edges between the nodes,representing causality and direction of fault propagation. The temporal lags areobtained for each edge from Eq. (3) represent time delays in fault propagation. Aweight matrix is representation of causal connections generated from Eq (2), whereeach value represents the magnitude of causal contribution between two variablesis also generated as an output.
[041] Referring FIG. 5, to illustrate a processor-implemented method(500) for causal network learning and identifying one or more root cause variablesand one or more fault propagation paths for faults in industrial processes isprovided. .
[042] Initially, at the step (502), receiving, via an input / output interface, amultivariate time-series data for one or more sensors from a plurality of sources,wherein the multivariate time-series data comprising a plurality of variablesrecorded for a plurality of time instances. The plurality of sources is associated withthe manufacturing and process industries that include iron and steel, power, cement,semiconductor, oil and gas refineries, fine chemicals, etc. comprising a sensordatabase, a laboratory database, an environment database, and a maintenancedatabase.
[043] At the next step (504), preprocessing the received multivariate timeseriesdata using a plurality of pre-processing techniques to remove noise andspurious values, imputing / interpolating missing values of the plurality of receiveddata and resampling the received multivariate time series data to a uniformfrequency.
[044] At the next step (506), detecting at least one fault in the preprocessedmultivariate time series data and a plurality of fault variables responsiblefor the at least one fault using a fault detection and localization technique. The faultdetection technique comprises one of the statistical, machine learning and deeplearning techniques including Principal component analysis (PCA), Mahalanobisdistance, isolation forest, elliptical envelope, k-nearest neighbors, multilayerperceptron, long short term memory autoencoder (LSTM-AE), and convolutionnetwork autoencoder (CNN-AE) while the fault localization technique comprisesone of complete decomposition contribution (CDC), leave-one-variable-out(LOVO) and reconstruction error based technique.
[045] At the next step (508), dividing the preprocessed multivariate timeseriesdata of the plurality of fault variables along the time axis into a plurality ofwindows of a predefined dimension. Each of the plurality of windows is furthersplit into one or more sources and a target variable. The source comprising themultivariate time series data corresponding to one or more previous time instances.The target comprising the multivariate time series data corresponding to a currenttime instance.
[046] At the next step (510), training a deep neural network for causalinference using each of the plurality of windows and a stochastic proximal gradientdescent (SPGD) technique with an adaptive learning rate. The deep neural model(DNM) comprising one or more encoding layers and one or more time lagdecomposition layers with one or more penalized causal parameters and one ormore forecasting layers.
[047] At the next step (512), extracting one or more encoding parametersfrom one or more encoding layers of the deep neural model and one or more timelag decomposition parameters from one or more time lag decomposition layers ofthe deep neural model.
[048] At the next step (514), computing causality scores for each of theone or more sources corresponding to the target variable using Frobenius norm ofone or more encoding parameters to obtain causality scores for one or more sourcetargetvariable pairs.
[049] At the next step (516), applying a causality significance test on thecomputed one or more causality scores for one or more source-target variable pairsto identify a threshold value to select one or more source-target variable pairshaving a causality score above the identified threshold in the multivariate timeseriesdata.
[050] At the next step (518), determining a temporal lag based on acontribution to the lag decomposition parameters corresponding to each of the oneor more selected source-target variable pairs having a causality score above thethreshold.
[051] At the next step (520), generating a causal structure of faultpropagation from the selected one or more source-target variable pairs and thedetermined temporal lags.
[052] At the last step (522), identifying one or more root cause variablesand one or more fault propagation paths based on traversing the generated causalstructure.Experiment:
[053] The industrial Tennessee Eastman process (TEP) is considered forthe case study. Operating data from the process is sent to FDD module whichdetected a nonstationary fault and 9 candidate faulty variables viz. X4, X7, X8,X13, X18, X19, X20, X45 and X50 are localized using the LSTM-AE technique.The identified candidate variables are passed as input to the CNL module. In theCNL module, the input multivariate data is segmented into overlapping timewindowsof length 10. The windowed data is further split into source and targetinstances and fed to the DNM for modelling the data. The network is trained usingMSE as the error metric and LSTMs for the encoding layer because of their abilityto model complex temporal interactions. Weights for the encoding and timedecompositionlayers are penalized and the network is trained with SPGD algorithmwith an adaptive learning rate in the range 10-6 to 10-2 . The causal network for thefault is generated by extracting parameters from the trained DNM and performingsteps as shown in Fig. 5. Causality significance test from Eq. (2) is applied on theencoding layer parameters to identify causal relationships between variables. Theresult is represented as a weight matrix shown in FIG. 6 where each value in thematrix represents the magnitude of significant causal contribution between sourceand target variables.
[054] Significant lags for the causal connections are identified as per Eq.(3). The inferred set of causal connections and temporal lags are graphicallyrepresented as a directed causal network graph where each edge signifies thedirection of causal connection and the edge weight represents the time delay in foreach interaction. Fig. 6 shows the causal network obtained from CNL module forthe fault in the TEP. Causal connections between various candidate variables areidentified along with time lagged information. The generated causal network is sentto the RCI module for detection of fault propagation paths and root cause variables.
[055] The written description describes the subject matter herein to enableany person skilled in the art to make and use the embodiments. The scope of thesubject matter embodiments is defined by the claims and may include othermodifications that occur to those skilled in the art. Such other modifications areintended to be within the scope of the claims if they have similar elements that donot differ from the literal language of the claims or if they include equivalentelements with insubstantial differences from the literal language of the claims.
[056] The embodiments of present disclosure herein address unresolvedproblem of learning causal networks for faults in nonlinear and nonstationaryindustrial processes and equipment. A method and system for causal networklearning and identifying one or more root cause variables and one or more faultpropagation paths for faults in industrial processes and equipment is provided.
[057] It is to be understood that the scope of the protection is extended tosuch a program and in addition to a computer-readable means having a messagetherein; such computer-readable storage means contain program-code means forimplementation of one or more steps of the method, when the program runs on aserver or mobile device or any suitable programmable device. The hardware devicecan be any kind of device which can be programmed including e.g., any kind ofcomputer like a server or a personal computer, or the like, or any combinationthereof. The device may also include means which could be e.g., hardware meanslike e.g., an application-specific integrated circuit (ASIC), a field-programmablegate array (FPGA), or a combination of hardware and software means, e.g., anASIC and an FPGA, or at least one microprocessor and at least one memory withsoftware modules located therein. Thus, the means can include both hardwaremeans, and software means. The method embodiments described herein could beimplemented in hardware and software. The device may also include softwaremeans. Alternatively, the embodiments may be implemented on different hardwaredevices, e.g., using a plurality of CPUs.
[058] The embodiments herein can comprise hardware and softwareelements. The embodiments that are implemented in software include but are notlimited to, firmware, resident software, microcode, etc. The functions performedby various modules described herein may be implemented in other modules orcombinations of other modules. For the purposes of this description, a computerusableor computer readable medium can be any apparatus that can comprise, store,communicate, propagate, or transport the program for use by or in connection withthe instruction execution system, apparatus, or device.
[059] The illustrated steps are set out to explain the exemplaryembodiments shown, and it should be anticipated that ongoing technologicaldevelopment will change the manner in which particular functions are performed.These examples are presented herein for purposes of illustration, and not limitation.Further, the boundaries of the functional building blocks have been arbitrarilydefined herein for the convenience of the description. Alternative boundaries canbe defined so long as the specified functions and relationships thereof areappropriately performed. Alternatives (including equivalents, extensions,variations, deviations, etc., of those described herein) will be apparent to personsskilled in the relevant art(s) based on the teachings contained herein. Suchalternatives fall within the scope and spirit of the disclosed embodiments. Also, thewords "comprising," "having," "containing," and "including," and other similarforms are intended to be equivalent in meaning and be open ended in that an itemor items following any one of these words is not meant to be an exhaustive listingof such item or items or meant to be limited to only the listed item or items. It mustalso be noted that as used herein and in the appended claims, the singular forms"a," "an," and "the" include plural references unless the context clearly dictatesotherwise.
[060] Furthermore, one or more computer-readable storage media may beutilized in implementing embodiments consistent with the present disclosure. Acomputer-readable storage medium refers to any type of physical memory on whichinformation or data readable by a processor may be stored. Thus, a computerreadablestorage medium may store instructions for execution by one or moreprocessors, including instructions for causing the processor(s) to perform steps orstages consistent with the embodiments described herein. The term "computer15readable medium" should be understood to include tangible items and excludecarrier waves and transient signals, i.e., be non-transitory. Examples includerandom access memory (RAM), read-only memory (ROM), volatile memory,nonvolatile memory, hard drives, CD ROMs, DVDs, flash drives, disks, and anyother known physical storage media.
[061] It is intended that the disclosure and examples be considered asexemplary only, with a true scope of disclosed embodiments being indicated by thefollowing claims.
Claims
1. A processor-implemented method (500) for causal network learning and identifying one or more root cause variables and corresponding one or more fault propagation paths for faults in industrial processes comprising steps of: receiving (502), via an input / output interface, a multivariate timeseries data of one or more sensors from a plurality of sources, wherein the multivariate time-series data comprising a plurality of variables recorded for a plurality of time instances; pre-processing (504), via one or more hardware processors, the received multivariate time-series data using a plurality of preprocessing techniques to remove noise and spurious values, imputing / interpolating missing values, and resampling the received multivariate time series data to a uniform frequency; detecting (506), via the one or more hardware processors, at least one fault in the pre-processed multivariate time series data and a plurality of fault variables responsible for the at least one fault using a fault detection and diagnosis technique; dividing (508), via one or more hardware processors, the preprocessed multivariate time-series data of the plurality of fault variables along the time axis into a plurality of windows of a predefined dimension, wherein each of the plurality of windows is further split into one or more sources and a target variable; training (510), via the one or more hardware processors, a deep neural network for causal inference using each of the plurality of windows and a stochastic proximal gradient descent (SPGD) technique with an adaptive learning rate; extracting (512), via the one or more hardware processors, one or more encoding parameters from one or more encoding layers of the deep neural network and one or more time lag decomposition parameters from one or more time lag decomposition layers of the deep neural network; computing (514), via the one or more hardware processors, causality scores for each of the one or more sources corresponding to the target variable using Frobenius norm of one or more encoding parameters to obtain causality score for one or more source-target variable pairs; applying (516), via the one or more hardware processors, a causality significance test on the computed one or more causality scores to identify a threshold value to select one or more source-target variable pairs having a causality score above the identified threshold in the multivariate time-series data; determining (518), via the one or more hardware processors, a temporal lag based on a contribution to the lag decomposition parameters corresponding to each of the one or more selected source-target variable pairs having a causality score above the threshold; generating (520), via the one or more hardware processors, a causal structure of fault propagation from the selected one or more sourcetarget variable pairs and the determined temporal lags; and identifying (522), via the one or more hardware processors, one or more root cause variables and one or more fault propagation paths for faults in industrial processes based on traversing the generated causal structure.
2. The processor-implemented method (500) of claim 1, wherein the deep neural network comprising one or more encoding layers and one or more time lag decomposition layers with one or more penalized causal parameters and one or more forecasting layers.
3. The processor-implemented method (500) of claim 1, wherein the fault detection and diagnosis technique comprise one of a statistical, machine learning and deep learning techniques comprising a principal component analysis (PCA), a Mahalanobis distance, an isolation forest, an elliptical envelope, a K-nearest neighbors, a multilayer perceptron, a long-short term memory autoencoder (LSTM-AE), a convolution network autoencoder (CNN-AE).
4. The processor-implemented method (500) of claim 1, wherein the one or more source variables comprising the multivariate time series data corresponding to one or more previous time instances.
5. The processor-implemented method (500) of claim 1, wherein the target variable comprising the multivariate time series data corresponding to a current time instance.
6. The processor implemented method (500) of claim 1, wherein one or more causal parameters are penalized while training the deep neural network to isolate one or more source-target variable pairs having a causality score above the identified threshold in the multivariate time-series data.
7. The processor implemented method (500) of claim 1, wherein the nonconvex optimization function is trained using SPGD algorithm with adaptive learning.
8. A system (100) for causal network learning and identifying one or more root cause variables and corresponding one or more fault propagation paths for faults in industrial processes comprising: an input / output interface (104) to receive a multivariate time-series data of one or more sensors from a plurality of sources, wherein the multivariate time-series data comprising a plurality of variables recorded for a plurality of time instances; a memory (110) in communication with the one or more hardware processors (108), wherein the one or more hardware processors are configured to execute programmed instructions stored in the memory to: preprocess the received multivariate time-series data using a plurality of pre-processing techniques to remove noise and spurious values, imputing / interpolating missing values and resampling the received multivariate time series data to a uniform frequency; detect at least one fault in the pre-processed multivariate time series data and a plurality of fault variables responsible for the at least one fault using a fault detection and diagnosis technique; divide the preprocessed multivariate time-series data of the plurality of fault variables along the time axis into a plurality of windows of a predefined dimension, wherein each of the plurality of windows is further split into one or more sources and a target variable; train a deep neural network for causal inference using each of the plurality of windows and a stochastic proximal gradient descent (SPGD) technique with an adaptive learning rate; extract one or more encoding parameters from one or more encoding layers of the deep neural network and one or more time lag decomposition parameters from one or more time lag decomposition layers of the deep neural network; compute causality scores for each of the one or more sources corresponding to the target variable using Frobenius norm of one or more encoding parameters to obtain one or more source-target variable pairs; apply a causality significance test on the computed one or more causality scores to identify a threshold value to select one or more source-target variable pairs having a causality score above the identified threshold in the multivariate timeseries data; determine a temporal lag based on a contribution to the time lag decomposition parameters corresponding to each of the one or more selected source-target variable pairs having a causality score above the threshold; and generate a causal structure of fault propagation from the selected one or more source-target variable pairs and the determined temporal lags; and identify one or more root cause variables and one or more fault propagation paths for faults in industrial processes based on traversing the generated causal structure.
9. The system (100) of claim 8, wherein one or more causal parameters are penalized while training the deep neural network to isolate one or more source-target variable pairs having a causality score above the identified threshold in the multivariate time-series data.
10. A non-transitory computer readable medium storing one or more instructions which when executed by one or more processors on a system, cause the one or more processors to perform method comprising: receiving, via an input / output interface, a multivariate time-series data of one or more sensors from a plurality of sources, wherein the multivariate time-series data comprising a plurality of variables recorded for a plurality of time instances; preprocessing, via the one or more hardware processors, the received multivariate time-series data using a plurality of preprocessing techniques to remove noise and spurious values, imputing / interpolating missing values and resampling the received multivariate time series data to a uniform frequency; detecting, via the one or more hardware processors, at least one fault in the pre-processed multivariate time series data and a plurality of fault variables responsible for the at least one fault using a fault detection and diagnosis technique; dividing, via one or more hardware processors, the preprocessed multivariate time-series data of the plurality of fault variables along the time axis into a plurality of windows of a predefined dimension, wherein each of the plurality of windows is further split into one or more sources and a target variable; training, via the one or more hardware processors, a deep neural network for causal inference using each of the plurality of windows and a stochastic proximal gradient descent (SPGD) technique with an adaptive learning rate; extracting, via the one or more hardware processors, one or more encoding parameters from one or more encoding layers of the deep neural network and one or more time lag decomposition parameters from one or more time lag decomposition layers of the deep neural network; computing, via the one or more hardware processors, causality scores for each of the one or more sources corresponding to the target variable using Frobenius norm of one or more encoding parameters to obtain one or more source-target variable pairs; applying, via the one or more hardware processors, a causality significance test on the computed one or more causality scores to identify a threshold value to select one or more source-target variable pairs having a causality score above the identified threshold in the multivariate time-series data; determining, via the one or more hardware processors, a temporal lag based on a contribution to the time lag decomposition parameters corresponding to each of the one or more selected source-target variable pairs having a causality score above the threshold; and generating, via the one or more hardware processors, a causal structure of fault propagation from the selected one or more sourcetarget variable pairs and the determined temporal lags; and identifying, via the one or more hardware processors, one or more root cause variables and one or more fault propagation paths for faults in industrial processes based on traversing the generated causal structure.
