Industrial Process Model Generation System

Abstract

The present invention relates to an industrial process model generation system, comprising: an input unit; and a processing unit. The input unit is configured to receive multiple input value trajectories, the input value trajectories including operational input value trajectories and simulation input value trajectories related to the industrial process. The processing unit is configured to implement a simulator for the industrial process. The processing unit is configured to generate multiple industrial process behavior data, wherein the industrial process behavior data is generated for at least some of the multiple input value trajectories, and wherein the generation of the industrial process behavior data for at least some of the multiple input value trajectories includes utilizing a simulator. The processing unit is configured to implement a machine learning algorithm for modeling the industrial process. The processing unit is configured to train the machine learning algorithm. The processing unit is configured to process first behavior data.

Description

Technical Field

[0001] The present invention relates to an industrial process model generation system, an industrial process model selection and generation system, an industrial process model generation method, and an industrial process model selection and generation method. Background Art

[0002] Industrial applications of machine learning often suffer from the availability of sufficient (labeled) training data for training machine learning models. High-fidelity simulators can be used to generate additional data to improve the quality of such machine learning models. However, running high-fidelity simulations is time-consuming and computationally expensive. Blind generation of data is very expensive and has no guaranteed impact on the performance of machine learning algorithms.

[0003] Improved model development for industrial processes is needed Summary of the Invention

[0004] Therefore, it would be advantageous to have an improved technique for modeling industrial processes.

[0005] The objects of the invention are solved by the subject-matter of the independent claims, wherein further embodiments are incorporated in the dependent claims.

[0006] In a first aspect, an industrial process model generation system is provided, comprising:

[0007] input unit; and

[0008] Processing unit.

[0009] The input unit is configured to receive a plurality of input value trajectories, the plurality of input value trajectories including operational input value trajectories and simulated input value trajectories related to an industrial process. The processing unit is configured to implement a simulator for the industrial process. The processing unit is configured to generate a plurality of industrial process behavior data. Generating the industrial process behavior data for at least some of the plurality of input value trajectories, and generating the industrial process behavior data for the at least some of the plurality of input value trajectories includes utilizing a simulator. The processing unit is configured to implement a machine learning algorithm for modeling the industrial process. The processing unit is configured to train the machine learning algorithm. The processing unit is configured to process first behavior data from the plurality of behavior data using the machine learning algorithm to determine a first modeling result. The processing unit is configured to determine whether to train the machine learning algorithm using the first behavior data or not, the determination including comparing the first modeling result with a performance condition. The processing unit is configured to process second behavior data from the plurality of behavior data using the machine learning algorithm to determine a second modeling result. The processing unit is configured to determine whether to use the second behavioral data to train or not to train the machine learning algorithm, or to further train or not to further train the machine learning algorithm using the second behavioral data, the determination comprising comparing the second modeling result with the performance condition.

[0010] In one example, the input value trace is real input data or simulated input data.

[0011] Thus, a set of industrial process behavior data is generated by a simulator based on input value trajectories that can be real values ​​or simulated, and a machine learning algorithm is trained on a subset of this data, wherein the system decides which subsets of the data are actually used in the training.

[0012] In addition to being trained on the industrial process behavior data generated by the simulator, the machine learning algorithm can be trained on "real" industrial process behavior data, as the plurality of industrial process behavior data can then include both the industrial process behavior data generated by the simulator and the "real" data. However, all of the plurality of industrial process behavior data can be the industrial process behavior data generated by the simulator.

[0013] In this way, as new input traces are provided, i.e., real input data or simulated input data, the associated behavioral data generated from the simulator can be used to train or untrain the machine learning model of the industrial process. This continues step by step as the behavioral data is generated, with the machine learning algorithm continuously trained with appropriate data but not trained with inappropriate data.

[0014] The operational input value track is any input to the production process by the facility operator. These include setpoints (target values ​​for automation / control loop(s), parameters for actuators (such as the percentage of open or closed of a valve), and digital inputs (pump on / off).

[0015] In addition to the inputs that operators will make (operational input value traces), there are also inputs used to configure and control the simulation. Examples include the initial facility state at the start of the simulation, raw material composition / quality, and the simulation of certain types of failures, such as valve failures (leaking, sticking, etc.) and rotating equipment failures (pumps, compressors). These are called simulation input value traces.

[0016] Thus, the input value trajectory is the input to the control (feedforward) simulation process, and the simulator can generate industrial process behavior data, which includes, for example, one or more of the following: process data (such as simulated temperature, pressure, level, flow value), actuator data (simulated valve position, heat exchanger flow, motor current, etc.), set point (such as the target value of a PID (proportional integral derivative) controller).

[0017] It should be noted that "processor" and "the processor" do not mean that the system must use only one processor. For example, a processor can implement a simulator to generate data, while a second processor implements a machine learning algorithm.

[0018] In one example, the plurality of behavioral data includes one or more of: process data; temperature data; pressure data; flow data; level data; voltage data; current flow data; power data; actuator data; valve data; sensor data; controller data.

[0019] Thus, the input value trajectory is the input to the control (feedforward) simulation process, and the simulator can generate industrial process behavior data, which includes, for example, one or more of the following: initial simulation state (e.g., simulated temperature, pressure, level, flow value), actuator data (simulated valve position, heat exchanger flow, motor current, etc.), set point (e.g., target value of a PID (proportional integral derivative) controller).

[0020] In one example, the modeling results include control or monitoring outputs of an industrial process.

[0021] Thus, a machine learning algorithm is trained to produce a machine learning model that, for example, implements task control or monitors the actual industrial process being simulated.

[0022] In one example, the processing unit is configured to select at least some of the plurality of input value trajectories.

[0023] Therefore, the system can select or decide which input value trajectories to provide to the simulator to generate new behavior data that can be used for training, which optimizes computational efficiency because the simulator may not need to be called.

[0024] In one example, the selection of at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0025] In one example, determining whether to train or not train the machine learning algorithm using the first behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0026] In one example, determining whether to train or not train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0027] In one example, determining whether to further train or not further train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0028] In one example, determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0029] In one example, the processing unit is configured to determine to stop training of the machine learning algorithm, the determination comprising determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0030] Therefore, the machine learning algorithm (model) is analyzed to identify which operational input traces and simulated input value traces should be used to generate new behavioral data for training the machine learning algorithm, so that when the machine learning algorithm is (additionally) trained on the new data, the newly generated simulated data will significantly change the machine learning model "a lot." And, when the machine learning algorithm does not change significantly, the training can be stopped, and the final trained machine learning algorithm can be provided to model the process.

[0031] In a second aspect, an industrial process model selection and generation system is provided, comprising:

[0032] input unit; and

[0033] Processing unit.

[0034] The input unit is configured to receive a plurality of input value trajectories including operational input value trajectories and simulated input value trajectories. The processing unit is configured to implement a simulator for an industrial process. The processing unit is configured to generate a plurality of industrial process behavior data. Generating the industrial process behavior data for at least some of the plurality of input value trajectories, and generating the industrial process behavior data for at least some of the plurality of input value trajectories includes utilizing a simulator. The processing unit is configured to implement a plurality of machine learning algorithms for modeling the industrial process. The processing unit is configured to process a first behavior data from the plurality of behavior data using a first machine learning algorithm from the plurality of machine learning algorithms to determine a first modeling result of the first machine learning algorithm. The processing unit is configured to determine whether to train the first machine learning algorithm using the first behavior data or to implement a second machine learning algorithm from the plurality of machine learning algorithms, the determination including comparing the first modeling result of the first machine learning algorithm with a performance condition.

[0035] In one example, a processing unit is configured to process second behavioral data from a plurality of behavioral data using a first machine learning algorithm to determine a second modeling result of the first machine learning algorithm. The processing unit is configured to determine whether to use the second behavioral data to train the first machine learning algorithm or implement the second machine learning algorithm, and the determination includes comparing the second modeling result of the first machine learning algorithm to a performance condition.

[0036] In one example, a processing unit is configured to process the first behavioral data using a second machine learning algorithm to determine a first modeling result of the second machine learning algorithm. The processing unit is configured to determine whether to train the second machine learning algorithm using the first behavioral data or to implement a third machine learning algorithm from a plurality of machine learning algorithms, the determination comprising comparing the first modeling result of the second machine learning algorithm to a performance condition.

[0037] In this way, as new input traces are provided, either real input data or simulated input data, the associated behavioral data generated from the simulator can be used to train a machine learning model or to select and train a different machine learning model for the industrial process. This continues step by step as the behavioral data is generated, with the machine learning algorithm continuously trained with the appropriate data, or a different machine learning algorithm is selected and trained with the appropriate data.

[0038] In one example, the processing unit is configured to select at least some of the plurality of input value trajectories.

[0039] Therefore, the system can select or decide which input value trajectories to provide to the simulator to generate new behavior data that can be used for training, which optimizes computational efficiency because the simulator may not need to be called.

[0040] In one example, the selection of at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0041] In one example, determining to use the behavioral data to train an existing machine learning algorithm or implement a new machine learning algorithm includes determining a sensitivity of the existing trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0042] In one example, the processing unit is configured to determine to stop training of an existing machine learning algorithm, the determination comprising determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0043] In one example, determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0044] Therefore, the machine learning algorithm (model) is analyzed to identify which operational input traces and simulated input value traces should be used to generate new behavioral data for training the machine learning algorithm, so that when the machine learning algorithm is (additionally) trained on the new data, the newly generated simulated data will significantly change the machine learning model "a lot." Furthermore, when the machine learning algorithm does not perform as desired, a more complex machine learning algorithm can be implemented, and when the machine learning algorithm does not change significantly, training can be stopped, and a final trained machine learning algorithm can be provided to model the process.

[0045] In one example, the first machine learning algorithm is the simplest machine learning algorithm among the plurality of machine learning algorithms.

[0046] In one example, the second machine learning algorithm is the second simplest machine learning algorithm among the plurality of machine learning algorithms.

[0047] In one example, the third machine learning algorithm is the third simplest machine learning algorithm among the plurality of machine learning algorithms.

[0048] In a third aspect, a method for generating an industrial process model is provided, comprising:

[0049] receiving a plurality of input value trajectories including an operational input value trajectory and a simulated input value trajectory associated with an industrial process;

[0050] A simulator for implementing the industrial process by a processing unit;

[0051] generating, by a processing unit, a plurality of industrial process behavior data, wherein the industrial process behavior data is generated for at least some of the plurality of input value trajectories, and wherein generating the industrial process behavior data for at least some of the plurality of input value trajectories includes utilizing a simulator;

[0052] implementing, by the processing unit, a machine learning algorithm that models the industrial process, wherein the processing unit is configured to train the machine learning algorithm;

[0053] Processing, by a processing unit, first behavioral data from the plurality of behavioral data using a machine learning algorithm to determine a first modeling result;

[0054] determining, by the processing unit, whether to train or not train the machine learning algorithm using the first behavioral data, the determination comprising comparing the first modeling result with a performance condition;

[0055] Processing, by the processing unit, second behavioral data from the plurality of behavioral data using a machine learning algorithm to determine a second modeling result;

[0056] The processing unit determines whether to train or not train the machine learning algorithm using the second behavioral data, or to further train or not further train the machine learning algorithm using the second behavioral data, the determination including comparing the second modeling result with the performance condition.

[0057] In one example, the method includes selecting, by the processing unit, at least some of the plurality of input value trajectories.

[0058] In one example, selecting at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0059] In one example, determining whether to train or not train the machine learning algorithm using the first behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0060] In one example, determining whether to train or not train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0061] In one example, determining whether to further train or not further train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0062] In one example, determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0063] In one example, a method includes determining, by a processing unit, to stop training a machine learning algorithm, the determination including determining a sensitivity of the trained machine learning algorithm to at least a portion of a plurality of behavioral data.

[0064] In a fourth aspect, a method for selecting and generating an industrial process model is provided, comprising:

[0065] receiving a plurality of input value trajectories including an operational input value trajectory and a simulated input value trajectory associated with an industrial process;

[0066] A simulator for implementing the industrial process by a processing unit;

[0067] generating, by a processing unit, a plurality of industrial process behavior data, wherein the industrial process behavior data is generated for at least some of the plurality of input value trajectories, and wherein generating the industrial process behavior data for at least some of the plurality of input value trajectories includes utilizing a simulator;

[0068] implementing, by the processing unit, a first machine learning algorithm of a plurality of machine learning algorithms for modeling the industrial process;

[0069] Processing, by a processing unit, first behavioral data among the plurality of behavioral data using a first machine learning algorithm among the plurality of machine learning algorithms to determine a first modeling result of the first machine learning algorithm; and

[0070] The processing unit determines whether to use the first behavior data to train a first machine learning algorithm or to implement a second machine learning algorithm among multiple machine learning algorithms, and the determination includes comparing a first modeling result of the first machine learning algorithm with a performance condition.

[0071] In one example, the method includes: a processing unit using a first machine learning algorithm to process second behavioral data among multiple behavioral data to determine a second modeling result of the first machine learning algorithm; and the processing unit determining to use the second behavioral data to train the first machine learning algorithm or implement the second machine learning algorithm, the determination including comparing the second modeling result of the first machine learning algorithm with a performance condition.

[0072] In one example, the method includes: utilizing a processing unit to process first behavioral data using a second machine learning algorithm to determine a first modeling result of the second machine learning algorithm; and determining by the processing unit whether to use the first behavioral data to train the second machine learning algorithm or to implement a third machine learning algorithm among multiple machine learning algorithms, the determination including comparing the first modeling result of the second machine learning algorithm with a performance condition.

[0073] In one example, the method includes selecting, by the processing unit, at least some of a plurality of input value trajectories.

[0074] In one example, selecting at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0075] In an example, determining to use the behavioral data to train an existing machine learning algorithm or implement a new machine learning algorithm includes determining a sensitivity of the existing trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0076] In one example, the processing unit is configured to determine to stop training of an existing machine learning algorithm, the determination comprising determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0077] In one example, determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0078] Therefore, the machine learning algorithm (model) is analyzed to identify which operational input traces and simulated input value traces should be used to generate new behavioral data for training the machine learning algorithm, so that when the machine learning algorithm is (additionally) trained on the new data, the newly generated simulated data will significantly change the machine learning model "a lot." Furthermore, when the machine learning algorithm does not perform as desired, a more complex machine learning algorithm can be implemented, and when the machine learning algorithm does not change significantly, training can be stopped, and a final trained machine learning algorithm can be provided to model the process.

[0079] The above aspects and examples will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Exemplary embodiments will be described below with reference to the following drawings:

[0081] Figure 1 The overall workflow of the new process of model training and training data generation is shown;

[0082] Figure 2 The overall workflow of the new process of searching for a suitable model architecture (i.e., searching for a suitable machine learning algorithm) is shown;

[0083] Figure 3 Shows the detailed workflow of the standard process of model training and training data generation;

[0084] Figure 4 A detailed workflow showing the new process of model training and training data generation; and

[0085] Figure 5 The detailed workflow of the new process for searching for a suitable model architecture is shown. DETAILED DESCRIPTION

[0086] Figure 1-Figure 5 The invention relates to an industrial process model generation system, an industrial process model selection and generation system, an industrial process model generation method, and an industrial process model selection and generation method.

[0087] Industrial Process Model Generation System

[0088] An example of an industrial process model generation system includes an input unit and a processing unit. The input unit is configured to receive multiple input value trajectories, including multiple operational input value trajectories and multiple simulated input value trajectories associated with an industrial process. The processing unit is configured to implement a simulator for the industrial process. The processing unit is configured to generate multiple industrial process behavior data. Generating the industrial process behavior data for the multiple input value trajectories includes utilizing the simulator. The processing unit is configured to implement a machine learning algorithm for modeling the industrial process. The processing unit is configured to train the machine learning algorithm. The processing unit is configured to process first behavior data from the multiple behavior data using the machine learning algorithm to determine a first modeling result. The processing unit is configured to determine whether to use the first behavior data to train the machine learning algorithm or not, the determination including comparing the first modeling result with a performance condition. The processing unit is configured to process second behavior data from the multiple behavior data using the machine learning algorithm to determine a second modeling result. The processing unit is configured to determine whether to train or not train the machine learning algorithm using the second behavioral data, or to further train or not further train the machine learning algorithm using the second behavioral data, the determination comprising comparing the second modeling result with the performance condition.

[0089] In other words, it can be determined that the machine learning algorithm can be improved by further training it with the generated data, or it can be determined that no generated data is needed to improve the machine learning model.

[0090] In one example, the performance conditions include one or more of: a target accuracy, a target false positive rate, a target false negative rate.

[0091] According to an example, the plurality of input value traces include one or more of: process data; temperature data; pressure data; flow data; level data; voltage data; current data; power data; actuator data; valve data; sensor data; controller data.

[0092] According to one example, the modeling results include control or monitoring outputs of the industrial process.

[0093] According to an example, the processing unit is configured to select at least some of the plurality of input value trajectories.

[0094] In one example, the selection of at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0095] According to one example, determining whether to train or not train the machine learning algorithm using the first behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0096] According to one example, determining whether to train or not train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0097] According to one example, determining whether to further train or not further train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0098] According to one example, determining the sensitivity of the training machine learning algorithm to the plurality of behavioral data includes analyzing a loss function of the training machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0099] According to one example, the processing unit is configured to determine to stop training the machine learning algorithm, the determination comprising determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0100] Industrial process model selection and generation system

[0101] An example of an industrial process model selection and generation system includes an input unit and a processing unit. The input unit is configured to receive multiple input value trajectories including operational input value trajectories and simulated input value trajectories. The processing unit is configured to implement a simulator for an industrial process. The processing unit is configured to generate multiple industrial process behavior data. Generating the industrial process behavior data for the multiple input value trajectories, and generating the industrial process behavior data for the multiple input value trajectories includes utilizing the simulator. The processing unit is configured to implement multiple machine learning algorithms for modeling the industrial process. The processing unit is configured to utilize a first machine learning algorithm from among the multiple machine learning algorithms to process first behavior data from among the multiple behavior data to determine a first modeling result of the first machine learning algorithm. The processing unit is configured to determine whether to train the first machine learning algorithm using the first behavior data or to implement a second machine learning algorithm from among the multiple machine learning algorithms, the determination including comparing the first modeling result of the first machine learning algorithm with a performance condition.

[0102] In this way, the best machine learning algorithm for process monitoring can be selected and trained.

[0103] In one example, the performance conditions include one or more of: a target accuracy, a target false positive rate, a target false negative rate.

[0104] According to one example, a processing unit is configured to process second behavioral data from a plurality of behavioral data using a first machine learning algorithm to determine a second modeling result of the first machine learning algorithm. The processing unit is configured to determine whether to use the second behavioral data to train the first machine learning algorithm or to implement the second machine learning algorithm, and the determination includes comparing the second modeling result of the first machine learning algorithm to a performance condition.

[0105] According to one example, a processing unit is configured to process first behavioral data using a second machine learning algorithm to determine a first modeling result of the second machine learning algorithm. The processing unit is configured to determine whether to train the second machine learning algorithm using the first behavioral data or to implement a third machine learning algorithm from a plurality of machine learning algorithms, the determination comprising comparing the first modeling result of the second machine learning algorithm to a performance condition.

[0106] According to an example, the processing unit is configured to select at least some of the plurality of input value trajectories.

[0107] In one example, the selection of at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0108] According to an example, determining to use the behavioral data to train an existing machine learning algorithm or implement a new machine learning algorithm includes determining a sensitivity of an existing trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0109] According to one example, the processing unit is configured to determine to stop training an existing machine learning algorithm, the determination comprising determining a sensitivity of the trained machine learning algorithm to at least a portion of a plurality of behavioral data.

[0110] According to one example, determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0111] According to one example, the first machine learning algorithm is the simplest machine learning algorithm among the plurality of machine learning algorithms.

[0112] According to an example, the second machine learning algorithm is the second simplest machine learning algorithm among the plurality of machine learning algorithms.

[0113] According to an example, the third machine learning algorithm is the third simplest machine learning algorithm among the plurality of machine learning algorithms.

[0114] Industrial process model generation method

[0115] Examples of industrial process model generation methods include:

[0116] receiving a plurality of input value trajectories including an operational input value trajectory and a simulated input value trajectory associated with an industrial process;

[0117] A simulator for implementing the industrial process by a processing unit;

[0118] generating, by a processing unit, a plurality of industrial process behavior data, wherein the industrial process behavior data is generated for a plurality of input value trajectories, and wherein generating the industrial process behavior data for the plurality of input value trajectories includes utilizing a simulator;

[0119] implementing, by the processing unit, a machine learning algorithm that models the industrial process, wherein the processing unit is configured to train the machine learning algorithm;

[0120] Processing, by a processing unit, first behavioral data from the plurality of behavioral data using a machine learning algorithm to determine a first modeling result;

[0121] determining, by the processing unit, whether to train or not train the machine learning algorithm using the first behavioral data, the determination comprising comparing the first modeling result to the performance condition;

[0122] Processing, by the processing unit, second behavioral data from the plurality of behavioral data using a machine learning algorithm to determine a second modeling result; and

[0123] The processing unit determines whether to train or not train the machine learning algorithm using the second behavioral data, or to further train or not further train the machine learning algorithm using the second behavioral data, the determination including comparing the second modeling result with the performance condition.

[0124] In one example, determining whether to train or not train the machine learning algorithm using the first behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0125] In one example, determining whether to train or not train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0126] In one example, determining whether to further train or not further train the machine learning algorithm using the second behavioral data includes determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0127] In one example, determining a sensitivity of a trained machine learning algorithm to the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0128] In an example, a method includes determining, by a processing unit, to stop training a machine learning algorithm, the determining including determining a sensitivity of the trained machine learning algorithm to at least a portion of a plurality of behavioral data.

[0129] In one example, the method includes selecting, by the processing unit, at least some of the plurality of input value trajectories.

[0130] In one example, selecting at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0131] Industrial process model selection and generation methods

[0132] Examples of industrial process model selection and generation methods include:

[0133] receiving a plurality of input value trajectories including an operational input value trajectory and a simulated input value trajectory associated with an industrial process;

[0134] A simulator for implementing the industrial process by a processing unit;

[0135] generating, by a processing unit, a plurality of industrial process behavior data, wherein the industrial process behavior data is generated for a plurality of input value trajectories, and wherein generating the industrial process behavior data for the plurality of input value trajectories includes utilizing a simulator;

[0136] implementing, by the processing unit, a first machine learning algorithm from a plurality of machine learning algorithms that models the industrial process;

[0137] Processing, by a processing unit, first behavioral data among the plurality of behavioral data using a first machine learning algorithm among the plurality of machine learning algorithms to determine a first modeling result of the first machine learning algorithm; and

[0138] A processing unit determines whether to use the first behavioral data to train a first machine learning algorithm or to implement a second machine learning algorithm among a plurality of machine learning algorithms, the determination comprising comparing a first modeling result of the first machine learning algorithm with a performance condition.

[0139] In one example, the method includes: a processing unit using a first machine learning algorithm to process second behavioral data among multiple behavioral data to determine a second modeling result of the first machine learning algorithm; and the processing unit determining whether to use the second behavioral data to train the first machine learning algorithm or implement the second machine learning algorithm, the determination including comparing the second modeling result of the first machine learning algorithm with a performance condition.

[0140] In one example, the method includes: utilizing a processing unit to process first behavioral data using a second machine learning algorithm to determine a first modeling result of the second machine learning algorithm; and determining by the processing unit whether to use the first behavioral data to train the second machine learning algorithm or to implement a third machine learning algorithm among multiple machine learning algorithms, the determination including comparing the first modeling result of the second machine learning algorithm with a performance condition.

[0141] In one example, the method includes selecting, by the processing unit, at least some of the plurality of input values ​​in the plurality of input value trajectories.

[0142] In one example, selecting at least some of the plurality of input value trajectories includes utilizing a determined sensitivity of a trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0143] In one example, determining to use the behavioral data to train an existing machine learning algorithm or implement a new machine learning algorithm includes determining a sensitivity of an existing trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0144] In one example, the processing unit is configured to determine to stop training of an existing machine learning algorithm, the determination comprising determining a sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data.

[0145] In one example, determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of behavioral data includes analyzing a loss function of the trained machine learning algorithm with respect to at least a portion of the plurality of behavioral data.

[0146] The industrial process training data generating system, the industrial process monitoring system, the method for generating industrial process training data, and the industrial process monitoring method are now described in further detail with reference to specific detailed embodiments, wherein reference is again made to Figure 1-Figure 2 .

[0147] Figure 1 An overview of the integrated workflow of model training and training data generation using a high-fidelity simulator is shown.

[0148] First, there are a number of predefined operator input value trajectories and disturbance trajectories, for example, a set of step changes on the relevant set points of the process or the activation or deactivation of an actuator fault (such as a valve fault). For simplicity, the operator input value trajectories and the disturbance trajectories can be collectively referred to as input value trajectories. Such step change experiments are well known from system identification. Using the input value trajectories and the disturbance trajectories, labels can be created together with the simulation data. The data in which the disturbance is introduced in the simulation is marked as an anomaly or a fault. In the case of a system identification task, the data after the set point change (response) is used as labeling information for the machine learning process.

[0149] These input value trajectories are used to control the process over multiple simulation runs. A simulation run captures the system behavior and response to a trajectory of input values. At the end of these simulation runs, the initial training dataset is created.

[0150] In the next step, a machine learning algorithm (here, algorithm can mean more than one algorithm running together) is trained by presenting training data that changes the parameters of the underlying model based on the loss function (it has been established that machine learning algorithms can be based on deep learning networks, linear regression, random forests, support vector machines, etc.). At the end of this step, the first model is trained. During this step, the training data set can be divided into a training data set and a test (hold-out) data set. In addition, the training data set can be repeatedly divided into a training data set and a validation data set (cross-validation).

[0151] In the next step, the system extracts the sensitivity of the machine learning algorithm(s) to the entire space of possible inputs to the algorithm(s). One way is to analyze the loss function (a measure of the model's prediction error) of the model on the training set or validation set samples. However, it is known how to perform this in other ways.

[0152] Generate the most informative new inputs based on additional constraints (e.g., minimum and maximum setpoint values, or thresholds for system tripping). When using existing samples in a training or validation dataset, new inputs can be generated to create a trajectory of input values ​​in the neighborhood of the original data points.

[0153] The new set point and disturbance trajectory (input value trajectory) are used as input for a new simulation run. These runs generate new data points. First, the existing model is tested on the new data points. If the performance is good enough (e.g., based on the target accuracy or target false positive / false negative rate), the process ends and the created machine learning model can be used online.

[0154] If the model is not good enough, new data points are added to the training dataset and the machine learning model is retrained to improve it.

[0155] In a variation of the above process, more than one machine learning model can be trained, but a set of machine learning algorithms or models can be trained. The most informative points are then a mixture of the most informative points from all models. Models that produce significantly worse results (e.g., based on statistical tests) are retained and other models can then be eliminated from the process.

[0156] In another variation of the drafting process described above, the training data set also includes historical data from actual facilities used for initial training.

[0157] Figure 2 Shown Figure 1 The process is external to the one shown. Here, the search process starts with a very simple machine learning model (e.g., a logistic regression or an artificial neural network with a few layers). When the model performance saturates at a poor accuracy (or some other performance measure) or even begins to degrade when presented with more informative data, the search process changes to a more complex model—either from a predefined list of machine learning models or by adding additional degrees of freedom to the model architecture (e.g., by adding layers). A combination of these two approaches is also possible—first increasing the complexity of a given class of, for example, deep learning networks, and then changing to a more complex network architecture.

[0158] As described above, a workflow has been developed in which a machine learning model, or its performance on individual samples in a training or validation dataset, is used to determine new simulation inputs, which results in highly informative simulation runs against the machine learning model. New data is generated using this simulation system. The trained algorithm is tested on the new, unseen data. If the algorithm's performance is adequate, the process stops. If not, new data is added to the training data, the machine learning model is retrained, and the process continues.

[0159] To further explain the industrial process model generation system, industrial process model selection and generation system and related methods, refer to Figure 3-Figure 5 .

[0160] Figure 3 A detailed example of an existing simulation workflow is shown.

[0161] Simulations of industrial processes use two types of inputs: [1] inputs made by the operator of the process and [2] inputs that are possible only in the simulation and are outside the operator's control in reality, such as equipment failure, raw material quality, external temperature, etc.

[0162] To generate data for training the machine learning model, we need to define trajectories of these inputs to control the simulation. Trajectories could define, for example, “At minute 5, the operator opens the valve” and “At minute 20, the valve starts leaking 50% of its flow rate.”

[0163] The outputs of a process simulator are all the values ​​typically available in a process control system (set points, sensor readings, actuator values, such as valve positions).

[0164] The method is used to train machine learning algorithms, for example, for (1) detecting process anomalies, (2) detecting equipment failures, (3) predicting the future behavior of the process, and (4) selecting the best possible next controller output.

[0165] The output of training is a model that can be fed new and previously unseen data and can perform, for example, one of tasks (1)-(4).

[0166] The model is connected to an actual industrial process (which produces the same type of data points as the simulator version) and will perform the task and produce the corresponding output.

[0167] In this process, a human will define the input trajectory, examine the results of the ML algorithm (the performance of the model), and decide what input value trajectory to use in the next iteration.

[0168] Figure 4 A detailed example of the development performed by the inventors is shown.

[0169] Essentially, the new development is to analyze the models generated by machine learning algorithms in order to find which new data will produce the best improvement of the generated model.

[0170] If the model is good enough for use in the actual process, no new data will be generated and the model will be used.

[0171] In this process, a system or method tests the model, analyzes it, and decides what new input value trajectory to use in the next iteration.

[0172] Figure 5 A detailed example of further development performed by the present inventors is shown.

[0173] This extension involves introducing another loop in addition to the repeated data generation and training.

[0174] After the model acceptance check, the improvement checker tests whether the model is still improving by adding new data to the dataset. This can be done by directly analyzing the sensitivity of the model on the input (if the sensitivity is low on all inputs, the model cannot be improved further) or by simply tracking whether the model has seen improvements exceeding a threshold in the last n versions. Performance is measured by some performance measure such as F1 score (for classification) or RMSE (root mean square error) (for regression).

[0175] If the model stops improving (and will not improve based on the output of the sensitivity layer), the model manager maintains a list of algorithms starting with the simplest ML algorithms (e.g., linear regression (regression) or logistic regression (classification)) and moving towards more complex algorithms (e.g., support vector machines and deep learning artificial neural networks). Each time an ML model stops improving, the next more complex model is selected and training begins again—always starting with all the training data available at that time.

[0176] The complexity of an artificial neural network is determined by the complexity of the architecture, for example as measured by the number of hidden layers and the number of nodes in a hidden layer.

[0177] During this process, the system or method still tests the model, analyzes the model, and decides what new input value trajectory to use in the next iteration. In addition, the system or method determines whether the currently used machine learning algorithm can still be improved and can achieve the desired performance, or whether a more complex algorithm should be used.

[0178] Note: Do not use the most complex algorithm in the beginning because it may "overfit" the training data (essentially creating a lookup table for the training data) and will not generalize to data outside of the training dataset.

[0179] exist Figure 3-Figure 5 In the embodiment shown, different computers or processing units of such computers can perform different functions. This may be the case, but if required, a single processing unit can also perform all the different functions.

[0180] Sensitivity analysis

[0181] Here are three examples of analyzing the sensitivity of a machine learning algorithm or model:

[0182] One example of analyzing the sensitivity of a machine learning model is to select n pieces of industrial process behavior data for which the machine learning model has the largest difference from actual values ​​(this difference is also called the prediction error). To generate input value trajectories that contribute to model improvement, the n input value trajectories used to generate the n industrial process behaviors with the largest prediction errors are modified, for example, by randomly changing the initial state of the simulation or randomly changing some input values ​​within the input value trajectory.

[0183] In another example, another machine learning model algorithm is trained to predict the prediction error of the first machine learning algorithm using the initial state and other characteristics of the input value trajectory as predictor variables. This machine learning model can be, for example, a decision tree that will return a decision sequence for which the input value trajectory value will produce a poor prediction on the current machine learning model. This information can be used to generate a new input value trajectory using exactly these conditions, such as the specific initial state for the simulation, the specific type of operator input (e.g., set point change, manual actuator change) and the parameters of these inputs (e.g., initial set point and new set point), or the type of fault.

[0184] In another example specifically for artificial neural networks, new high-information data points are generated as follows: an optimization algorithm (e.g., a genetic algorithm, a reinforcement learning algorithm, or a Bayesian optimization algorithm) generates a trajectory of input values, which is used to generate behavioral data using a process simulation system. For each behavioral data point, the derivative (i.e., sensitivity) of the neural network prediction (model response) with respect to the model parameters can be evaluated without going through the entire training and validation cycle. This means that the gradient information can be evaluated for a large number of candidate points in a short period of time. The resulting points are "high-information" points in the following sense:

[0185] During subsequent training, a gradient descent-type algorithm will take large steps towards a better solution if the prediction error is high (i.e., if the model can be improved).

[0186] During subsequent training, if the prediction error is low, that is, if the model generalizes well to newly found input trajectories, the gradient descent-type algorithm will stay near its current solution.

[0187] Therefore, the input trajectory that maximizes the sensitivity of the model response will allow the model to improve or validate the current model. In special cases, when the input to the artificial neural network (predictor variables) only includes information that is part of the input value trajectory and the output of the process simulation (process behavior data) is part of the output (predicted value), the process simulation step can be avoided and a large number of candidate input value trajectories can be directly evaluated using the derivatives of the neural network predictions (model responses) with respect to the model parameters. In this case, the time to evaluate candidate input trajectories is even shorter.

[0188] Although the present invention has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description should be considered illustrative or exemplary rather than restrictive. The present invention is not limited to the disclosed embodiments. Other variations of 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.

Claims

1. An industrial process model generation system, comprising: Input unit; as well as processing unit; wherein the input unit is configured to receive a plurality of input value trajectories, the input value trajectories comprising real input data and / or simulated input data related to the industrial process, and wherein the input value trajectories are used to control the industrial process in a plurality of simulation runs to generate an initial training data set; wherein the processing unit is configured to implement a simulator of the industrial process; wherein the processing unit is configured to implement a machine learning algorithm for modeling the industrial process; wherein the processing unit is configured to train a machine learning algorithm, the training comprising utilizing the initial training data set; wherein the processing unit is configured to generate a plurality of industrial process behavior data using the simulator, wherein the industrial process behavior data is generated for at least some of the plurality of input value trajectories, wherein the processing unit is configured to select the at least some of the plurality of input value trajectories, wherein the selection of the at least some of the plurality of input value trajectories comprises utilizing a determined sensitivity of the machine learning algorithm to at least a portion of the plurality of industrial process behavior data, and wherein the industrial process behavior data comprises one or more of: process data, actuator data, and set points; wherein the processing unit is configured to train the machine learning algorithm in a step-by-step process comprising: As part of the stepwise process, the processing unit is configured to process first behavioral data from the plurality of industrial process behavioral data using the machine learning algorithm to determine a first modeling result; As part of the stepwise process, the processing unit is configured to determine whether to train or not train the machine learning algorithm using the first behavioral data, the determination comprising a comparison of the first modeling result to a performance condition; As part of the stepwise process, the processing unit is configured to process second behavioral data from the plurality of industrial process behavioral data using the machine learning algorithm to determine a second modeling result; As part of the stepwise process, after determining whether to train or not train the machine learning algorithm using the first behavioral data, the processing unit is configured to determine whether to train or not train the machine learning algorithm using the second behavioral data, the determination comprising a comparison of the second modeling result with the performance condition; and As part of the stepwise process, the processing unit is configured to continue training the machine learning algorithm using behavioral data and associated modeling results until the machine learning algorithm has not changed significantly in a final step of training, and wherein when the training has stopped, a final trained machine learning algorithm is provided; and The final trained machine learning algorithm is configured to be fed new data to perform one or more of the following: (1) detect process anomalies, (2) detect equipment failures, (3) predict future behavior of the process, and (4) select an optimal next controller output.

2. The system of claim 1 , wherein the plurality of input value traces comprises one or more of: process data; temperature data; pressure data; flow data; level data; voltage data; current data; power data; actuator data; valve data; sensor data; controller data.

3. The system of any one of claims 1 to 2, wherein the determination of whether to train or not train the machine learning algorithm using the first behavioral data comprises: A sensitivity of the trained machine learning algorithm to at least a portion of the plurality of industrial process behavior data is determined.

4. The system of any one of claims 1 to 2, wherein the determination of whether to use the second behavioral data to train or not to train the machine learning algorithm comprises: A sensitivity of the trained machine learning algorithm to at least a portion of the plurality of industrial process behavior data is determined.

5. The system of any one of claims 1 to 2, wherein the determination of whether to further train or not further train the machine learning algorithm using the second behavioral data comprises: A sensitivity of the trained machine learning algorithm to at least a portion of the plurality of industrial process behavior data is determined.

6. The system of claim 5, wherein determining the sensitivity of the trained machine learning algorithm to at least a portion of the plurality of industrial process behavior data comprises: A loss function of the trained machine learning algorithm with respect to at least the portion of the plurality of industrial process behavior data is analyzed.

7. The system according to any one of claims 1 to 2, wherein the processing unit is configured to determine to stop training the machine learning algorithm, the determination comprising: A sensitivity of the trained machine learning algorithm to at least a portion of the plurality of industrial process behavior data is determined.

8. A method for generating an industrial process model, comprising: receiving a plurality of input value trajectories, the input value trajectories comprising real input data and / or simulated input data associated with an industrial process, and wherein the input value trajectories are used to control the industrial process in a plurality of simulation runs to generate an initial training data set; A simulator of the industrial process is implemented by a processing unit; implementing, by the processing unit, a machine learning algorithm for modeling the industrial process; training the machine learning algorithm by the processing unit, wherein the training includes utilizing the initial training data set; generating, by the processing unit, a plurality of industrial process behavior data, wherein the industrial process behavior data is generated for at least some of the plurality of input value trajectories, wherein the processing unit selects the at least some of the plurality of input value trajectories, wherein the selecting of the at least some of the plurality of input value trajectories comprises utilizing a determined sensitivity of the machine learning algorithm to at least a portion of the plurality of industrial process behavior data, and wherein the industrial process behavior data comprises one or more of: process data, actuator data, and a set point; Training the machine learning algorithm by the processing unit in a step-by-step process includes: As part of the stepwise process, processing, by the processing unit, first behavioral data from the plurality of industrial process behavioral data using the machine learning algorithm to determine a first modeling result; determining, by the processing unit as part of the stepwise process, whether to train or not train the machine learning algorithm using the first behavioral data, the determination comprising comparing the first modeling results to performance conditions; As part of the stepwise process, processing, by the processing unit, second behavior data from the plurality of industrial process behavior data using the machine learning algorithm to determine a second modeling result; as part of the stepwise process, after determining whether to train or not train the machine learning algorithm using the first behavioral data, determining, by the processing unit, whether to train or not train the machine learning algorithm using the second behavioral data, the determining comprising comparing the second modeling result to the performance condition; and continuing to train the machine learning algorithm using the behavioral data and associated modeling results by the processing unit as part of the stepwise process until the machine learning algorithm has not changed significantly in a final step of training, and wherein a final trained machine learning algorithm is provided when the training has ceased; and The final trained machine learning algorithm is configured to be fed new data to perform one or more of the following: (1) detect process anomalies, (2) detect equipment failures, (3) predict future behavior of the process, and (4) select an optimal next controller output.

Images