Data Processing for Industrial Machine Learning

Abstract

The present invention relates to a computer-implemented method (100) for automating the development of industrial machine learning applications, particularly for predictive maintenance, process monitoring, event prediction, or root cause analysis. The method comprises one or more sub-methods that can be executed iteratively depending on the industrial machine learning problem. These sub-methods include at least one of the following: automating data cleaning during machine learning model training (S10) and post-training (S15), labeling time series (particularly signal data) using other timestamp records (S11), feature engineering using process mining (S12), and automated hyperparameter tuning for data segmentation and classification (S14).

Description

[Technical field]

[0001] The present invention relates to a computer-implemented method for data pre-processing for industrial machine learning, which method can be used, for example, for predictive maintenance, process monitoring, event prediction or root cause analysis. The present invention further relates to a data processing system configured to perform the steps of the computer-implemented method, a computer program comprising instructions for causing a data processing system to perform the method, and a computer readable medium having stored thereon such a computer program. [Background technology]

[0002] Machine learning can be used in industry for predictive maintenance, process monitoring, event prediction, or root cause analysis, among others. For example, in the case of predictive maintenance, the state of an industrial asset, such as a motor or a robot, may be predicted to estimate the time when a maintenance action should be performed. Thus, a maintenance action may be scheduled according to a machine learning-based prediction of the state of the industrial asset.

[0003] This offers cost savings over time-based predictive maintenance because maintenance actions are performed only when necessary. Additionally, the probability of unexpected failure of industrial assets is reduced because the condition of the assets is continuously monitored.

[0004] However, applying machine learning techniques for predictive maintenance is not a trivial task. In particular, data from sensors of industrial assets or from control systems of industrial processes or plants typically needs to be pre-processed prior to the application of machine learning models. This pre-processing may include, for example, cleaning of raw sensor data, including, for example, removal of outliers and / or suppression of noise. Furthermore, pre-processing typically involves the derivation of features from time series data. These pre-processing algorithms are critical for the performance that can be achieved by the machine learning models. Another important requirement is to provide a sufficient number of training samples for the training of the machine learning models.

[0005] Therefore, machine learning applications, not only for predictive maintenance but also for other purposes such as process monitoring, event prediction, or root cause analysis, are developed by mixed teams of domain and machine learning experts. Summary of the Invention

[0006] However, machine learning and data science experts are rare and often lack the domain expertise required for industrial machine learning. Furthermore, developing industrial machine learning applications is a time-consuming process, especially the time required for manual data cleaning, feature engineering, data labeling, and hyperparameter tuning. There is no automated method that allows domain experts to develop machine learning applications themselves.

[0007] Existing approaches for supporting domain experts in developing machine learning applications, such as automated machine learning (AutoML), leverage the homogeneous properties of mainstream machine learning applications, such as machine learning on tabular, text, or image data. These approaches rely on the availability of labeled data to establish objective functions for model selection and hyperparameter tuning. However, such labeled data is typically not available in industrial machine learning applications.

[0008] Therefore, it may be desirable to provide improved automation for the development of industrial machine learning applications.

[0009] This is achieved by the subject matter of the independent claims, further embodiments are incorporated in the dependent claims and the following description. It should be noted that any step, feature or aspect of the computer-implemented method described below applies equally to a data processing system, a computer program and a computer-readable medium configured to perform the method steps described below, and vice versa.

[0010] The method for automated development of industrial machine learning applications consists of one or more sub-methods that may be executed iteratively depending on the industrial machine learning problem. The sub-methods may be (a) a method for automating data cleaning in training and subsequent application of machine learning models, (b) a method for labeling time series of data such as sensor signals using other timestamp records, (c) feature engineering using process mining, and (d) automated hyperparameter tuning for data segmentation and classification.

[0011] According to a first aspect of the present disclosure, a computer-implemented method for machine learning is presented. The method includes obtaining first time series data from a sensor of an industrial asset or from a control system for an industrial process or plant. Further, the method includes processing the first time series data to obtain an event log, and applying process mining to the event log to provide relevance analysis and / or bottleneck identification.

[0012] The first time series data may be a discrete time signal from a sensor of an industrial asset, such as a motor or a robot, or from a control system for an industrial process or plant, such as a computerized distributed or centralized control system. Obtaining the first time series data may mean, for example, receiving the first time series data from a sensor or control system, or loading the first time series data from a storage medium. For example, the first time series data may be loaded from a server, such as a remote server. The first time series data may include raw data from a sensor or from a control system, or the first time series data may be processed data, for example, cleaned time series data.

[0013] The steps of obtaining the first time series data, processing the first time series data, and applying process mining may be pre-processing steps that may be performed before training or applying a first machine learning model, which may be utilized for, for example, predictive maintenance or forecasting how a batch process will evolve. In particular, the steps of obtaining the first time series data, processing the first time series data, and applying process mining may be used for feature engineering, i.e., to determine input parameters of the first machine learning model.

[0014] In one example, the computer-implemented method further includes determining a condition indicator of the industrial asset based on the suitability analysis and / or the bottleneck identification.

[0015] The fitness analytics provided by process mining may be quantified into condition indicators for industrial assets. For example, different types of fitness and thresholds can be used and / or optimized. By calculating these condition indicators periodically (e.g., every second, minute, hour, or day), these metrics can be compared to discover anomalous behavior.

[0016] For example, alarm and / or event data from a control system and / or sensor data from a motor may be exploited using process mining, for example, to monitor its condition and predict its behavior. This approach is agnostic to the sensor or control system used, i.e., it may also be applied separately to other industrial assets and control systems, since the normal operation of the asset is inferred as data is collected over time. In other words, no explicit information or working models are required to detect anomalies, such as degradation over time.

[0017] In one example, the computer-implemented method further includes training and / or applying a first machine learning model to determine process deviations, determine potential improvements, perform condition-based monitoring, perform predictive maintenance, and / or predict how the batch process will evolve, where input parameters of the first machine learning model are based on the suitability analysis and / or bottleneck identification.

[0018] When the first machine learning model is trained, the first time series data may be time series data, such as raw or cleaned training time series data. In particular, the training time series data may be historical time series data. In contrast, when the first machine learning model is applied, the first time series data may be a life data stream from an industrial asset or from a control system, such as a computerized distributed or centralized control system.

[0019] The first machine learning model may be trained to determine process deviations, determine potential improvements, perform condition-based monitoring, perform predictive maintenance, and / or predict how a batch process will evolve.

[0020] The input parameters of the first machine learning model may be or may be based on the suitability analysis and / or bottleneck identification. In particular, some or all of the input parameters of the first machine learning model may be or may be based on condition indicators of the industrial asset derived from the suitability analysis and / or bottleneck identification.

[0021] In another example, processing the first time series data to obtain the event log includes encoding the first time series data by applying symbol set approximation or artificial intelligence techniques.

[0022] To perform process mining on time series data, the time series data needs to be transformed into an event log, i.e., a set of discrete events. Such encoding may be done using Symbolic Aggregate Approximation (SAX) or AI techniques.

[0023] In another example, processing the first time series data to obtain the event log further includes performing abstraction on the encoded first time series data.

[0024] Because it may be difficult to perform process mining on raw low-level event logs, these logs may be transformed by performing abstractions. In one example, this may include aggregating the raw low-level events or applying a below-threshold filter. For example, raw low-level events below a threshold may be set to 0 to remove noise. Other abstractions of the raw low-level events are possible.

[0025] In another example, the computer-implemented method further includes obtaining second time series data and cleaning the second time series data to obtain third time series data. Further, the data cleaning machine learning model is trained using a plurality of first training samples, the first training samples including the clean data points from the third time series data and the plurality of raw data points from the second time series data.

[0026] Thus, the computer-implemented method may include training a machine learning model for data cleaning. To train the machine learning model, a first set of training samples may be used, and the first set of training samples may be derived from the second time series data and the third time series data.

[0027] The second time series data may be raw time series data from a sensor of an industrial asset or from a control system for an industrial process or plant.

[0028] The third time series data may be determined manually, for example, by a domain expert or a machine learning expert. Cleaning the second time series data to obtain the third time series data may include handling missing values, removing noise, and / or removing outliers.

[0029] The different first training samples may include different clean data points from the third time series data. Each of the first training samples may further include a plurality of raw data points from the second time series data. Thereby, the raw data points of the second time series data may be included in some of the first training samples. In particular, the first training samples may include the raw data points of the second time series data within a time window that may be centered around the time of the corresponding clean data point. To train the data cleaning machine learning model, the clean data points of the training samples may serve as the desired output of the machine learning model, and the raw data points of the training samples serve as input parameters to the machine learning model.

[0030] After training the machine learning model for data cleaning, the machine learning model may be applied to raw time series data from sensors of an industrial asset or from a control system to provide clean time series data, which may be equal to the first time series data.

[0031] In another example, the computer-implemented method further includes obtaining fourth time series data from a sensor or from a control system and applying a data cleaning machine learning model to the fourth time series data to obtain the first time series data.

[0032] A data cleaning machine learning model may be trained as described above based on the second and third time series data, which may require manual determination of the third time series data, for example by a domain expert.

[0033] The fourth time series data may be different from the second time series data. In other words, the trained data cleaning machine learning model may be applied to new data that is not in the training set of the first training sample. Thus, the data cleaning machine learning model provides a generalized cleaning logic. In particular, the fourth time series data may be a live data stream from a sensor or from a control system. The fourth time series data may include thousands of data points per second that may be cleaned by the data cleaning machine learning model.

[0034] The second and third time series data may also include raw and clean time series data from other applications, i.e., the raw and clean time series data from other applications may be utilized to train the data cleaning machine learning model, which may reduce or avoid the effort of manually determining clean data points for the third time series data.

[0035] Alternatively, a data cleaning machine learning model from another application may be utilized to clean the fourth time series data.

[0036] In another example, a dedicated data cleaning algorithm may be used to clean the fourth time series data. This dedicated data cleaning algorithm may not be based on a machine learning model. This may be required when the data cleaning machine learning model determined above does not provide sufficient data cleaning performance.

[0037] In another example, the computer-implemented method further includes obtaining a first set of labels for training a machine learning model for automatic labeling. Further, one or more data sources are obtained, and a first set of features is extracted from the one or more data sources. The machine learning model for automatic labeling may then be trained using a plurality of second training samples, the second training samples including labels from the first set of labels and one or more features from the first set of features.

[0038] The labels in the first set of labels may have a timestamp. These labels may be used as class labels in the classification process. The labels in the first set of labels may be determined manually.

[0039] The data sources may be unstructured, semi-structured or tabular data sources. Typical examples are alarm and event data, shift book entries, and entries in a computerized maintenance management system (CMMS).

[0040] The features extracted from one or more data sources can include typical natural language processing features (e.g., bag-of-words, recognized named entities), but also sentiment analysis or text classification, statistical values ​​(alarm rates, #operator actions), quality tests from the laboratory, or fault notes for assets in a particular plant area (from the CMMS). Quality tests from the laboratory can be Boolean (e.g., in spec vs. out of spec) or numerical or categorical quality indicators.

[0041] Entries in the data sources may have a timestamp associated with them, or these entries may be equipped with time information (e.g., the time mentioned in a shift book entry). This may be utilized to extract a time range for labeling the process values. One challenge with these data sources is that their timestamps may not exactly match the timestamps of the process values. This problem can be solved by assigning labels with a probability over a time window, where the process values ​​may be data points of the first time series data. However, features of the first machine learning problem, such as condition indicators of an industrial asset, may also be assigned the same labels as the process values ​​from which they are derived.

[0042] The machine learning model for automatic labeling may be a probabilistic network / model, such as a Bayesian network. Thus, the features of the first set of features may be used as input to a probabilistic model that describes the joint probability distribution over the features and the label of interest (e.g., normal vs. abnormal behavior).

[0043] For each probabilistic model, it may be defined which documents or entries from the data source are used to generate the input to the probabilistic model and how the time window (t_start, t_end) is generated for the output label. For example, the probabilistic model may generate a label for a four hour (4h) window from t_start to t_end=t_start+4h. Thereby, for example, alarms and events between t_start and t_end may be used. Additionally or alternatively, for example, shift book entries between t_start and t_start+8h (corresponding approximately to one shift) may be used, or shift book entries from t_start to the end of the shift may be used. Additionally or alternatively, for example, CMMS data between t_start-12h and t_start+12h may be used.

[0044] The notion of a label generated by a machine learning model for automatic labeling may not be that the label is likely to exist during the entire time window between t_start and t_end, but rather that the label is likely to exist during at least some time between t_start and t_end.

[0045] After training the machine learning model for automatic labeling, the model may be used to label previously unlabeled time windows based on corresponding data in a shift book, an alarm list, an event list, and / or a CMMS.

[0046] In another example, the computer-implemented method further includes extracting a second set of features from the one or more data sources and determining a second set of labels by applying a machine learning model for automatic labeling to features from the second set of features.

[0047] The second set of features may be extracted from later entries of the data source compared to the first set of features. It is possible that there is overlap, and thus some entries of the data source may be used to extract features of both the first and second sets of features.

[0048] Given features from the second set of features, the probability of a label value may be inferred by a machine learning model for automatic labeling. Thus, a time-stamped label in the second set of labels may be determined by selecting the label value with the maximum probability. This may be utilized to label historical processes with a label from the second set of labels.

[0049] In another example, instead of a single label, multiple labels can be assigned to the process values. In addition, multiple machine learning models, such as multiple probabilistic models, can be used. For example, one probabilistic model can be used per data source. Furthermore, algorithms for the implementation of real industrial monitoring and control tasks can be used, which can be configured to handle inconsistent class labels.

[0050] In another example, the first machine learning model is trained using a plurality of third training samples, the third training samples including labels from the first or second set of labels and / or condition indicators of the industrial asset.

[0051] More specifically, for training the first machine learning model, the labels of the first and / or second set of labels may be utilized as desired output values ​​of the first machine learning model. Further, condition indicators of the industrial asset may be utilized as input values ​​of the first machine learning model.

[0052] According to the present disclosure, a data processing system is also provided, the data processing system being configured to perform any of the steps of the method according to the present invention.

[0053] The data processing system may comprise a storage medium for storing, among others, the first, second, third, and / or fourth time series data. The data processing system may further comprise a processor, such as a microprocessor having one or more processor cores. In addition, the data processing system may comprise a graphics processing unit, which may be used to efficiently train the first machine learning model, the machine learning model for data cleaning, and / or the machine learning model for automatic labeling. The data processing system may also comprise a communication means, such as a LAN, a WLAN, or a cellular communication modem. The data processing system may be connected to sensors of an industrial asset or to a control system of an industrial process or plant via the communication means. The data processing system may further be connected to one or more servers, which may store training samples or perform one or more steps of a computer-implemented method, such as training the first machine learning model, the machine learning model for data cleaning, and / or the machine learning model for automatic labeling. Furthermore, the data processing system may comprise peripheral devices, such as a screen.

[0054] According to the present disclosure, a computer program is also presented, which comprises instructions for causing a data processing system to perform any one of the methods according to the invention as claimed in the independent claims, when the computer program is executed on a data processing system.

[0055] According to the present disclosure a computer readable medium is also presented, which stores a computer program as defined in the independent claims.

[0056] It is to be understood that the computer-implemented method for machine learning, the data processing system configured to perform the steps of the method, the computer program for causing a data processing system to perform the method, and the computer readable medium having stored such a computer program have similar and / or identical preferred embodiments, in particular as defined in the dependent claims. It is further to be understood that a preferred embodiment of the invention may also be any combination of the dependent claims with the respective independent claims.

[0057] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter. [Brief description of the drawings]

[0058] Exemplary embodiments of the invention are described below with reference to the accompanying drawings. [Figure 1] FIG. 1 illustrates a method for automating the development of industrial machine learning applications. [Diagram 2] FIG. 2 illustrates a method for training and applying a data cleaning model to achieve automated data cleaning on raw data received online from an industrial asset. [Diagram 3] FIG. 3 shows a method for automatically determining labels by applying a machine learning model for automatic labeling. [Figure 4] FIG. 4 illustrates a method for training a machine learning model for automatic labeling. [Diagram 5] Figure 5 illustrates how process mining can be performed on time series data. [Figure 6] Figure 6 illustrates the workflow from scenario selection to model export. [Figure 7] FIG. 7 illustrates the process of generating unsupervised models for anomaly and process phase detection. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0059] FIG. 1 illustrates a method 100 for automating the development of industrial machine learning applications, particularly for predictive maintenance, process monitoring, event prediction, or root cause analysis.

[0060] In step S10, an automatic data cleaning algorithm is applied to the historical data. A machine learning model for data cleaning may be applied thereto. In step S11, labels are determined, which may be performed by a machine learning model for automatic labeling. In the final pre-processing step, step S12, feature engineering is performed by process mining. In step S13, conventional training of the machine learning model is performed. This machine learning model may be configured for applications such as predictive maintenance, process monitoring, event prediction, or root cause analysis. The training data may include or be based on the labels determined in step S11 and the features determined in step S12.

[0061] In step S14, automated machine learning orchestration is performed for steps S10 to S12. This process is iterative, and one or more of steps S10 to S12 may be revisited depending on the measured performance of the machine learning model obtained from step S13. In some embodiments, one or more of steps S10 to S12 may be performed at least in part manually, e.g., initial data cleaning. The machine learning orchestration may also be performed manually. It is also possible that one or more of steps S10 to S12 and S14, e.g., automated data labeling or feature engineering steps, are skipped.

[0062] When an iteration of the machine learning orchestration algorithm is completed, a final data cleaning algorithm in step S10, a final feature pre-processing algorithm in step S12, and a final machine learning model in step S13 may be provided for application to new data, as illustrated by steps S15 to S17.

[0063] In step S15, a final data cleaning algorithm is applied to the live data stream from the industrial facility. In step S16, a final feature determination algorithm is applied to the cleaned data obtained from step S15. In step S17, a trained machine learning model is applied to the features determined in step S16.

[0064] The order of the data cleaning, labelling and feature engineering steps S10, S11 and S12, respectively, may be changed in different embodiments.

[0065] FIG. 2 illustrates a method 200 for training and applying a data cleaning model to achieve automated data cleaning on raw data received online from an industrial asset.

[0066] In step S20, raw data from the industrial asset is received and cleaned. Thereby, raw data points in the received raw time series data may be mapped onto clean data points in the clean time series data. The mapping from raw data points to clean data points may be performed, at least in part, manually, for example, by a machine learning expert. The cleaning of the received raw data may include handling missing values. For example, a missing value may be set to an average of a preceding data point and a following data point. Furthermore, the cleaning of the received raw data may include removing noise. For example, removing noise may be achieved by setting data points below a threshold to 0. Furthermore, the cleaning of the received raw data may include removing outliers.

[0067] In step S21, the cleaned data points may be used as labels to train a machine learning model for data cleaning. The complete set of raw data is available as regressors. It is also possible to select a subset of the complete set of raw data as regressors for the cleaned data points using metadata such as topological connections between measurements or other types of measurements (temperature, level, pressure). Thus, a training sample for training a machine learning model for data cleaning may include the cleaned data points and a subset of data points of the raw dataset. The machine learning model for data cleaning may be trained to predict values ​​of the cleaned data points from a subset of raw data points in the corresponding training sample. Training of this model may be done in a traditional way using manual tuning or may be automated using concepts such as hyperparameter tuning. The output may be a machine learning model or several machine learning models capable of generating cleaned data points based on multiple raw data points.

[0068] In step S22, the machine learning model for data cleaning obtained from step S21 is applied to the data stream from the industrial process, i.e., to the time series data, to clean the raw online data and make it suitable as input for a subsequent monitoring and / or control model. The output of the monitoring and / or control model may be displayed on a human machine interface (HMI). Additionally or alternatively, the output of the monitoring and / or control model may trigger some action on the technical system, for example when used as a model in a model predictive controller.

[0069] When a sufficient number of training samples for data cleaning are already available from other applications, step S20 may be skipped. Then, the training samples from these other applications may be utilized to train a machine learning model for data cleaning. In this case, human effort to determine training data is no longer required.

[0070] Alternatively, machine learning models for data cleaning may be obtained from other applications.

[0071] In one embodiment, training of an improved machine learning model for data cleaning may be performed even if a sufficient number of training samples for data cleaning or a machine learning model for data cleaning may be available from other applications. This may involve labeling additional raw data points (designating clean data points) in an active learning process. The active learning process may selectively request labels from a machine learning developer or domain expert to provide further information for the training process.

[0072] In another embodiment, hyperparameter optimization and other AutoML techniques are used in the training process to find the best possible hyperparameter settings and machine learning model architectures for learning the data cleaning logic.

[0073] 3 shows a method 300 for automatically determining labels using unstructured, semi-structured, or tabular data sources with time stamps. Exemplary data sources are alarm and / or event lists, shift books, or CMMS.

[0074] In step S30, features are extracted from the data entries of the different data sources. For example, in step S30a, features may be extracted from data entries in a shift book. In step S30b, features may be extracted from data entries in an alarm and / or event list. In step S30c, features may be extracted from data entries in a CMMS. The features extracted may be typical natural language processing features (e.g., bag of words, recognized named entities), but also sentiment analysis or text classification, statistical values ​​(alarm rate, #operator actions), quality tests from a lab, or fault notes for assets in a particular plant area (from the CMMS).

[0075] Entries in a data source may have a timestamp associated with them or may contain time information. From the timestamp associated with an entry in a data source or from the time information in the entry itself (e.g., the time listed in a shift book), a time range can be extracted for labeling the process value. One challenge with data sources such as shift books, alarm and / or event lists, and CMMS is that their timestamps cannot be precisely mapped to the timestamps of process values. This problem can be addressed, for example, by assigning labels with probabilities over a time window.

[0076] In step S31, the extracted features are used as input to a probabilistic model, e.g., a Bayesian network, which can describe the joint probability distribution over the features and the target label. For example, the target label can indicate anomaly or normal behavior. Given the features, the probability of the label value can be inferred, and a time-stamped label can be created by selecting the label with the maximum probability.

[0077] In step S32, the labels determined in step S31 are assigned to process values, e.g., to data points of time series data or to quantities derived from one or more process values, such as condition indicators of an industrial asset. The determined labels together with the features determined in step S12 of Figure 1 may form training samples for training a machine learning model in step S14 of Figure 1.

[0078] For each probabilistic model, it is defined which documents or entries from the data source are used to generate the input to the probabilistic model and how a time window (t_start, t_end) is generated for the output label.

[0079] In one exemplary embodiment, the probabilistic model can generate labels for a 4 hour window (t_start=t, t_end=t_start+4 hours) using alarms and events between t_start and t_end, shift book entries from t_start to t_start+8 hours (approximately corresponding to one shift) or from t_start to the end of the shift, and CMMS entries between t_start-12 hours and t_end+12 hours.

[0080] The notion of a generated label may not be that the label is likely to exist during the entire time window between t_start and t_end, but rather that the generated label is likely to exist for at least some time between t_start and t_end.

[0081] 4 shows a method 400 for training a machine learning model for automatic labeling. In step S40, features are extracted from data entries of different data sources. For example, in step S40a, features may be extracted from data entries of a shift book. In step S40b, features may be extracted from data entries of an alarm and / or event list. In step S40c, features may be extracted from data entries in a CMMS. The processing of the shift book, the alarm / event list, and the data entries in the CMMS to extract features may be similar or identical to the processing of steps S30a to S30c.

[0082] In step S41, a machine learning model for automatic labeling is trained. The machine learning model for automatic labeling may be a probabilistic model such as a Bayesian network. To train the machine learning model for automatic labeling, the time-stamped labels are used as class labels in the classification process.

[0083] The trained probabilistic model may be used in steps S11 and S31 to determine labels for previously unlabeled time windows based on data entries in the shift book, alarm / event lists, and / or CMMS.

[0084] In one embodiment, instead of a single label, multiple labels may be determined for each time window and / or process value. In addition, several probability models may be used, possibly even one probability model per data source or multiple machine learning models. In this case, algorithms for the implementation of real industrial monitoring and / or control tasks that can handle inconsistent class labels may be used.

[0085] FIG. 5 illustrates a method 500 for performing process mining on time series data, which may be utilized for feature engineering, particularly machine learning models for condition-based monitoring or predictive maintenance for industrial assets.

[0086] Process mining provides the ability to perform conformance analysis. Such conformance reports may be quantified into condition indicators for industrial assets. For example, different types of conformance and thresholds may be used and / or optimized. By calculating these condition indicators periodically (e.g., every second, minute, hour, or day), these metrics can be compared to discover anomalous behavior.

[0087] For example, alarm and / or event data from a control system and / or sensor data of an industrial asset, such as a motor, may be exploited using process mining to monitor its condition and predict its behavior. This approach is agnostic to the sensor or control system used, i.e., it may be applied separately to other industrial assets and control systems (e.g., to robot data), since the normal operation of the asset is inferred as data is collected over time. In other words, no explicit information or working models are required to detect anomalies, such as degradation over time.

[0088] When reporting anomalies to a domain expert, explanations for detecting new data as anomalous can be easily provided as status indicators, while the actual historical event logs can all be easily retrieved.

[0089] Indeed, such methods need not be limited to state-based monitoring. As more data is collected and used for process mining, this collection of historical data can be continuously used to train machine learning models to make predictions of future state indicators and other statistics (e.g., frequency of occurrence of different events). For example, in the case of a batch process, by taking real-time batch data as input, it is possible to predict how the process will continue to evolve.

[0090] In step S50 of Figure 5, time series data is obtained. The time series may be a raw time series from a sensor of an industrial asset, such as a motor or robot, or from a control system, such as a distributed or centralized control system for an industrial process or plant. Alternatively, the time series may be a processed time series from a sensor or from a control system. For example, a cleaned time series may be obtained from a sensor or from a control system.

[0091] In step S51, the acquired time series data is encoded, for example using symbolic set approximation (SAX) or artificial intelligence techniques, thereby converting the time series data into a raw low-level event log, i.e., a set of discrete raw low-level events.

[0092] In optional step S52, relevant events may be extracted from the raw low-level event log. Additionally or alternatively, abstraction may be performed on the raw low-level event log. This may include performing aggregation or filtering on the raw low-level event log. For example, filtering of the raw low-level event log may be performed to remove noise. This may be achieved by setting values ​​below a threshold to 0. Step S52 provides the low-level event log.

[0093] In step S53, process mining is applied to the low-level event logs to provide conformance analysis and / or bottleneck identification. In particular, bottlenecks in batch processes and / or deviations from standard operating procedures may be discovered.

[0094] Process mining in step S53 allows to focus on cases of interest for investigation. For these cases of interest, further data analysis may be performed in step S54. This allows to take into account contextual information such as the operator's workload at the time by looking in more detail at processes that deviate from the normal workflow. As a result, different actions can be taken to improve process efficiency and safety, for example by providing training to operators, adapting standard operating procedures, etc.

[0095] One simple example of how process mining can be applied is the reaction to an alarm. There may be alarms of different priorities. After the activation of an alarm, an operator's positive response may be expected. Furthermore, depending on the alarm priority, an operator's action may be expected within a time limit, which may depend on the priority of the alarm. If a large deviation is detected, for example when a reaction to a priority 1 alarm occurs later than 5 minutes after the alarm, this can be used to reprioritize the alarm or retrain the operator to act faster. These action sequences with a quick return to normal should become the standard response to an alarm. In other words, the action sequences may be optimized for the shortest time to return to normal.

[0096] FIG. 6 shows a workflow 600 from scenario selection to model export.

[0097] In step S60, a scenario is selected.

[0098] In step S61, data is provided.

[0099] In step S62, a machine learning model is determined using AutoML, which may include determining an unsupervised machine learning model by AutoML (step S62a), determining a supervised machine learning model by AutoML (step S62b), and automated machine learning orchestration by a model manager (step S62c).

[0100] Starting from raw process / time series data, the method targets two problem classes: anomaly detection and segmentation of time series data into phases. For both problems, an ensemble of unsupervised machine learning models is run to find the best unsupervised machine learning model for both tasks. On top of these results, sequential pattern mining can be applied to derive association rules that can, for example, assist with root cause analysis. Association rules can help, for example, to identify situations in which certain anomalies tend to occur or in which process productivity is compromised (e.g., "90% of the time when phase A was less than 15 minutes, an anomaly occurred in a subsequent phase").

[0101] In step S63, a report is generated. A number of results can be presented to the user, such as segmentation of the time series into phases, anomalies in the time series data, and a list of mined rules / patterns. A confidence threshold for all results may be selected by the user so that only those results for which the machine learning model has a high degree of confidence are displayed.

[0102] The user can then export the machine learning model for productive use, e.g., for monitoring or troubleshooting (step S64), or provide feedback on the results, i.e., true / false (or more detailed labels) for detected anomalies, higher / lower granularity for detected phases (and optionally labels) (step S65). Based on the feedback, the unsupervised machine learning model is improved or a supervised machine learning model is created using AutoML (step S62b), and the results of the unsupervised machine learning model and user feedback are used to generate labels. The process may be repeated until the user accepts the machine learning model for export. This can be either a supervised or unsupervised machine learning model.

[0103] 7 illustrates a process 700 for generating an unsupervised machine learning model for anomaly and process phase detection. Thus, the process of FIG. 7 may be used for time series segmentation and / or anomaly detection. Furthermore, association rules for segments or association rules for anomalies can be derived.

[0104] In step S70, data (pre-processing) processing is performed, for example using symbol set approximation or dynamic time warping.

[0105] In step S71, cluster mining is performed, optionally via ensemble learning.

[0106] In step S72, a model and data stability check is performed.

[0107] It should be noted that the embodiments of the present invention are described with reference to different subject matters. However, a person skilled in the art will infer from the above and following description that, unless otherwise indicated, any combination of features belonging to one type of subject matter, as well as any combination between features relating to different subjects, are also considered to be disclosed in the present application. However, all features can be combined to provide synergistic effects that are greater than the simple sum of the features.

[0108] While the invention has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description is to be considered exemplary or illustrative and not restrictive. The invention is not limited to the disclosed embodiments. Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.

[0109] In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage. Any reference signs in the claims should not be interpreted as limiting the scope. The invention as described in the claims of the original application is set forth below. [C1] A computer-implemented method (100) for machine learning, comprising: Obtaining first time series data from a sensor of an industrial asset or from a control system for an industrial process or plant (S12, S16, S50); Processing the first time series data to obtain an event log (S12, S16, S51, S52); and applying process mining to the event log (S12, S16, S53) to provide relevance analysis and / or bottleneck identification. [C2] The computer-implemented method (100) of C1 further comprising determining a condition indicator of the industrial asset based on the suitability analysis and / or bottleneck identification. [C3] The computer-implemented method (100) of C1 or C2, further comprising training (S13) and / or applying (S17) a first machine learning model to determine process deviations, determine potential improvements, perform condition-based monitoring, perform predictive maintenance, and / or predict how a batch process will evolve, wherein input parameters to the first machine learning model are based on the suitability analysis and / or bottleneck identification. [C4] The computer-implemented method (100) of any one of C1 to C3, wherein processing the first time series data to obtain the event log includes encoding the first time series data by applying symbol set approximation or artificial intelligence techniques (S12, S16, S51). [C5] The computer-implemented method (100) of C4, wherein processing the first time series data to obtain the event log further includes performing abstraction on the encoded first time series data (S12, S16, S52). [C6] The computer-implemented method (100) of C5, wherein the abstraction performed on the encoded first time series data includes data aggregation and / or a noise suppression filter. [C7] Obtaining second time series data; Cleaning the second time series data (S10, S20) to obtain third time series data; and training a data cleaning machine learning model using the plurality of first training samples (S10, S21). 7. The computer-implemented method (100) of any one of C1 to C6, wherein a first training sample includes clean data points from the third time series data and a plurality of raw data points from the second time series data. [C8] The computer-implemented method (100) of C7, wherein cleaning the second time series data includes handling missing values, removing noise, and / or removing outliers. [C9] acquiring fourth time series data from the sensor or from the control system; The computer-implemented method (100) of any one of C1 to C8, further comprising applying a data cleaning machine learning model to the fourth time series data (S10, S15, S22) to obtain the first time series data. [C10] Obtaining a first set of labels for training a machine learning model for automatic labeling; Obtaining one or more data sources; Extracting a first set of features from the one or more data sources (S11, S40); and training a machine learning model for automatic labeling using the second training samples (S11, S41). The computer-implemented method (100) of any one of C1 to C9, wherein a second training sample includes a label from the first set of labels and one or more features from the first set of features. [C11] The one or more data sources include at least one of a shift book, an alarm list, an event list, and / or a data source from a computerized maintenance management system; and / or The computer-implemented method (100) of C10, wherein the machine learning model for the automatic labeling is a probabilistic model. [C12] Extracting a second set of features from the one or more data sources (S11, S30); and The computer-implemented method (100) of C10 or 11, further comprising applying (S11, S31) a machine learning model for automatic labeling to features from the second set of features to obtain a second set of labels. [C13] the first machine learning model is trained using a third plurality of training samples; The computer-implemented method (100) of C2, 3 and 12, wherein a third training sample includes labels from the first set or the second set of labels and / or the condition indicators of the industrial asset. [C14] A data processing system comprising means for performing the steps of the method according to any one of claims 1 to 13. [C15] A computer program comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method according to any one of claims 1 to 13. [C16] A computer readable medium containing instructions which, when executed by a computer, cause the computer to perform the steps of the method of any one of C1 to C13.

Claims

1. A computer-implemented method (100) for machine learning, comprising: Obtaining first time series data from an industrial process (S12, S16, S50); processing the first time series data (S12, S16, S51, S52) to obtain an event log, wherein the event log comprises a set of discrete events; and applying process mining to the event log to provide a fitness analysis that determines the fitness of the industrial process and a bottleneck identification that identifies bottlenecks in the industrial process (S12, S16, S53); training (S13) and / or applying (S17) a first machine learning model for processing input parameters obtained based on the suitability analysis and the bottleneck identification to at least one of: determine process deviations of the industrial process, determine potential improvements to the industrial process, perform condition-based process monitoring, perform predictive maintenance to predict maintenance actions, and, if the industrial process comprises a batch process, predict how the batch process will evolve by taking real-time batch data as input; A computer-implemented method (100), comprising:

2. The industrial process includes a sensor of an industrial asset, the first time series data is obtained from the sensor of the industrial asset; The computer-implemented method (100) of claim 1 , further comprising determining a condition indicator of the industrial asset based on the suitability analysis and / or bottleneck identification.

3. The industrial process includes a control system, The computer-implemented method of claim 1 , wherein the first time series data is obtained from the control system.

4. 4. The computer-implemented method (100) of claim 1, wherein processing the first time series data to obtain the event log comprises encoding the first time series data by applying symbol set approximation or artificial intelligence techniques (S12, S16, S51).

5. 5. The computer-implemented method (100) of claim 4, wherein processing the first time series data to obtain the event log further comprises performing abstraction on the encoded first time series data (S12, S16, S52).

6. The computer-implemented method (100) of claim 5, wherein the abstraction performed on the encoded first time series data comprises a data aggregation and / or a noise suppression filter.

7. Obtaining second time series data; Cleaning the second time series data to obtain third time series data (S10, S20); and training a data cleaning machine learning model using the plurality of first training samples (S10, S21). The computer-implemented method (100) of any one of claims 1 to 6, wherein a first training sample comprises clean data points from the third time series data and a plurality of raw data points from the second time series data.

8. The computer-implemented method (100) of claim 7, wherein cleaning the second time series data comprises handling missing values, removing noise, and / or removing outliers.

9. acquiring fourth time series data from the industrial process; The computer-implemented method (100) of claim 1 , further comprising: applying a data cleaning machine learning model to the fourth time series data (S10, S15, S22) to obtain the first time series data.

10. Obtaining a first set of labels for training a machine learning model for automatic labeling; Obtaining one or more data sources; Extracting a first set of features from the one or more data sources (S11, S40); and training a machine learning model for automatic labeling using the second training samples (S11, S41). The computer-implemented method (100) of any one of claims 1 to 9, wherein a second training sample includes a label from the first set of labels and one or more features from the first set of features.

11. The one or more data sources include at least one of a shift book, an alarm list, an event list, and / or a data source from a computerized maintenance management system; and / or The computer-implemented method (100) of claim 10, wherein the machine learning model for the automatic labeling is a probabilistic model.

12. Extracting a second set of features from the one or more data sources (S11, S30); and The computer-implemented method (100) of claim 10 or 11, further comprising applying (S11, S31) a machine learning model for automatic labeling to features from the second set of features to obtain a second set of labels.

13. the first machine learning model is trained using a third plurality of training samples; The computer-implemented method (100) of claim 12 when dependent on claim 2, wherein a third training sample comprises labels from the first set or the second set of labels and / or the condition indicators of the industrial assets.

14. A data processing system comprising means for carrying out the steps of the method according to any one of claims 1 to 13.

15. A computer program comprising instructions which, when said program is executed by a computer, cause said computer to carry out the steps of the method according to any one of claims 1 to 13.

16. A computer readable storage medium containing instructions which, when executed by a computer, cause the computer to perform the steps of the method of any one of claims 1 to 13.

Images