Prediction method, prediction device, and program

Abstract

This provides a method for predicting the output of a controlled object. [Solution] A state value acquisition unit 141 acquires one or more state values ​​in the target sample of the controlled object, and a prediction processing unit 143 inputs the one or more state values ​​into a trained model and obtains a predicted value of the output value in the target sample as a prediction result from the trained model. The trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data expressed in difference format between different training samples in a training interval including multiple training samples. The control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object.

Description

[Technical Field]

[0001] This disclosure relates to a prediction method, a prediction device, and a program for predicting the output of a controlled object. [Background technology]

[0002] In the manufacturing sector, machine learning is used to predict the quality of products and to implement control based on those predictions. To predict quality with greater accuracy, it is expected that a large amount of historically acquired process data will be used. However, such a large amount of process data is affected by factors such as changes in material lots, aging of manufacturing equipment, and seasonal temperature changes. If machine learning is applied to this raw data without any processing, the accuracy may decrease.

[0003] For example, Patent Document 1 discloses the application of a method called the JIT (Just-In-Time) model to causal analysis. In the analysis method to which the JIT model is applied, only samples that are close to the current state are extracted from historical data, and by continuously learning from the extracted samples, it is possible to perform analysis that can handle distribution changes and complex nonlinearities. As a result, for example, according to the analysis method described in Patent Document 1, it is possible to analyze local causal relationships even for time series data where the causal relationships are non-stationary and nonlinear. [Prior art documents] [Patent Documents]

[0004] [Patent Document 1] Japanese Patent Publication No. 2023-173391 [Overview of the project] [Problems that the invention aims to solve]

[0005] Unfortunately, the methods described in Patent Document 1 and others have the problem that, if data similar to the input data that can be predicted has not been accumulated in the past, it is difficult to make predictions that take into account the influence of fluctuating components of the controlled object, such as changes in material lot numbers, deterioration of manufacturing equipment over time, and seasonal temperature changes.

[0006] This disclosure has been made in view of the above, and aims to provide a prediction method, prediction device, and program that can accurately predict the output of a controlled object. [Means for solving the problem]

[0007] A prediction method according to one aspect of the present disclosure involves a computer performing the following actions: acquiring one or more state values ​​in a target sample of a controlled object; inputting the one or more state values ​​into a trained model to obtain a predicted value of the output value in the target sample as a prediction result from the trained model; the trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval including multiple training samples, where the training data is expressed in difference format between training samples where one or more state values ​​of the controlled object and the output value of the controlled object are different; and the control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object.

[0008] A prediction device according to one aspect of the present disclosure is a prediction device that predicts the output of a controlled object, comprising a memory and a processor, wherein the processor executes instructions stored in the memory to acquire one or more state values ​​at a sample to be predicted in the controlled object, inputs the one or more state values ​​to a trained model, and obtains a predicted value of the output value at the sample to be predicted as a prediction result from the trained model, wherein the trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a learning interval including multiple training samples, where the training data is expressed in difference format between training samples where one or more state values ​​of the controlled object and the output value of the controlled object are different, wherein the control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object.

[0009] A program according to one aspect of this disclosure is a program for causing a computer to perform the following actions: acquire one or more state values ​​in a target sample of a controlled object; input the one or more state values ​​into a trained model and acquire a predicted value of the output value in the target sample as a prediction result from the trained model, wherein the trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval including multiple training samples, where the training data is expressed in difference format between training samples where one or more state values ​​of the controlled object and the output value of the controlled object are different, wherein the control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object. [Effects of the Invention]

[0010] According to this disclosure, the output of the controlled object can be predicted with high accuracy. Brief Description of the Drawings

[0011] [Figure 1] Figure 1 is an explanatory diagram showing an example of a schematic configuration of a system 1 according to an embodiment of the present disclosure. [Figure 2] Figure 2 is a block diagram showing an example of a schematic configuration of a prediction device 100 according to the first embodiment. [Figure 3] Figure 3 is a diagram showing a specific example of nonlinear mapping by a neural network used in the nonlinear control model according to the first embodiment. [Figure 4] Figure 4 is a diagram showing a specific example of nonlinear mapping by a neural network used in a difference learning method applied to the nonlinear control model according to the first embodiment. [Figure 5] Figure 5 is a diagram schematically showing a specific example of the relevance of various factors used in prediction processing. [Figure 6] Figure 6 is a flowchart showing the flow of prediction processing in the prediction device 100. [Figure 7] Figure 7 is a diagram for explaining the difference in variation of bias b(n) between the operation example and the comparative example. [Figure 8] Figure 8 is a block diagram showing an example of a schematic configuration of a prediction device 1000 according to the second embodiment. Modes for Carrying Out the Invention

[0012] Hereinafter, an embodiment according to the present invention will be described with reference to the accompanying drawings. The embodiment described below is an example, and is not to be restrictively interpreted by this description. Each aspect disclosed in the present specification can be combined with any other feature disclosed in the present specification. Note that in the present specification and the drawings, elements that can be described in the same manner are denoted by the same reference numerals, whereby redundant descriptions may be omitted.

[0013] The description will be given in the following order. 1. Outline of Embodiments of the Invention 2. System Configuration 3. First Embodiment 3.1. Configuration of the prediction device 3.2. Example of Operation 4. Second Embodiment 4.1. Configuration of the prediction device 4.2. Example of Operation 5. Other Embodiments

[0014] <<1. Overview of Embodiments of the Invention>> First, an overview of the embodiments of the present invention will be described.

[0015] (1)Technical issues In the manufacturing sector, machine learning is used to predict the quality of products and to implement control based on those predictions. To predict quality with greater accuracy, it is expected that a large amount of historically acquired process data will be used. However, such a large amount of process data is affected by factors such as changes in material lots, aging of manufacturing equipment, and seasonal temperature changes. If machine learning is applied to this raw data without any processing, the accuracy may decrease.

[0016] Specifically, time series data from plants and other sources include seasonal shifts in sensor values ​​due to temperature changes, as well as trend components due to the aging of the plant. These trend components can reduce the accuracy of predictive models.

[0017] To address such problems, for example, Patent Document 1 discloses the application of a method called the JIT (Just-In-Time) model to causal analysis. In the analysis method to which the JIT model is applied, only samples that are close to the current state are extracted from historical data, and by continuously learning from the extracted samples, it is possible to perform analysis that can handle distribution changes and complex nonlinearities.

[0018] Specifically, the analysis method described in Patent Document 1 uses the JIT method described above to extract neighboring data of the input data to be analyzed from previously accumulated data, and then performs causal search using LiNGAM, assuming that the causal relationships between the features of that neighboring data are linear. As a result, according to the analysis method described in Patent Document 1, it is possible to analyze local causal relationships even for time series data where the causal relationships are non-stationary and non-linear.

[0019] Unfortunately, the methods described in Patent Document 1 and others have the problem that, if data similar to the input data that can be predicted has not been accumulated in the past, it is difficult to make predictions that take into account the influence of fluctuating components of the controlled object, such as changes in material lot numbers, deterioration of manufacturing equipment over time, and seasonal temperature changes.

[0020] Furthermore, the method described in Patent Document 1 only derives causal relationships between features of the extracted neighboring data. For example, assuming that four features—temperature, pressure, CO2 concentration, and set value—are given as the data to be analyzed, causality can only be explained between these features. Such derived causal relationships do not take into account seasonal shifts in sensor values ​​due to temperature changes, trend components due to aging of the plant, and the effects of product lot changes, as described above, and therefore have the following problems.

[0021] First, the controlled object can be modeled based on the causal relationship between explanatory variables, which represent features detectable by sensors, and the target variable, which represents the output of the controlled object. However, such control models do not adequately consider bias elements, which are model elements independent of the explanatory variables. Here, bias elements are factors that may have relatively gradual variations between samples but can affect the accuracy of calculating the target variable. Examples of bias elements include seasonal deviations in sensor values ​​due to temperature changes, trend components due to aging of plants, and lot changes in products.

[0022] The purpose of this embodiment has been taken in view of the above, and aims to accurately predict the output of the controlled object by appropriately considering the bias elements of the controlled object.

[0023] (2) Example of operation In embodiments of this disclosure, for example, a computer performs the following actions: obtain one or more state values ​​in a target sample of the controlled object; input the one or more state values ​​into a trained model and obtain a predicted value of the output value in the target sample as a prediction result from the trained model; the trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval including multiple training samples, where the one or more state values ​​of the controlled object and the output value of the controlled object are represented in difference format between different training samples; and the control model is modeled to derive the output value of the controlled object based on at least a function mapped from the one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object. Specifically, in the control model, the bias element may have slower inter-sample variation compared to, for example, a function mapped from the state values.

[0024] In this way, according to this embodiment, the output of the controlled object can be predicted with high accuracy. More specifically, according to this embodiment, by applying a differential learning method using training data expressed in differential format to a control model that models the controlled object, even if data similar to the input data of the object to be analyzed has not been accumulated in the past, the output of the controlled object can be predicted with high accuracy by taking into account the bias elements described above. According to this embodiment, as specific examples of bias elements, for example, deviations in sensor values ​​with seasonality due to temperature changes, trend components due to aging deterioration of the plant, and the effects of lot changes in products can be taken into account, and the output of the controlled object can be predicted with high accuracy.

[0025] It should be noted that the technical features described above are just one specific example of embodiments of the present invention, and naturally, embodiments of the present invention are not limited to the technical features described above.

[0026] <<2. System Configuration>> Referring to Figure 1, the configuration of System 1 according to an exemplary embodiment of this disclosure will be described. Figure 1 is an explanatory diagram showing an example of a schematic configuration of System 1 according to an embodiment of this disclosure. Referring to Figure 1, System 1 includes a prediction device 100, a plant 200, and a plurality of sensors 301-304 (hereinafter collectively referred to as sensors 300). Note that the number of sensors is not limited to four, and any number of sensors may be included in System 1 depending on the number of objects to be detected.

[0027] According to System 1 configured in this way, Plant 200 is an example of a controlled object according to the embodiment of this disclosure, and is, for example, a process-type manufacturing plant that produces products that meet predetermined quality and conditions by executing one or more manufacturing processes. Each sensor 300 detects a state value that can be detected from Plant 200. The detected state values ​​are provided to the prediction device 100 via a predetermined network line or user input operation. Furthermore, the prediction device 100 uses the state values ​​obtained by each sensor 300 to predict the output value of Plant 200 to be predicted. Such predicted output values ​​are, for example, state values ​​of Plant 200 that are not detected by the sensors 300, or characteristic values ​​for quantitatively evaluating the quality of products manufactured in Plant 200, and can be used as information for the control and maintenance of Plant 200.

[0028] <<3. First Embodiment>> Next, a first embodiment of the present disclosure will be described with reference to Figures 2 to 7.

[0029] <3.1. Configuration of Prediction Device 100> Next, an example of the configuration of the prediction device 100 according to the first embodiment will be described with reference to Figure 2. Figure 2 is a block diagram showing a schematic example of the configuration of the prediction device 100 according to the first embodiment. Referring to Figure 2, the prediction device 100 is a computer that predicts the output value of the plant 200 and comprises a network communication unit 110, a user interface unit 120, a memory 130, and a processor 140.

[0030] (1) Network communication unit 110 The network communication unit 110 communicates with any communication means provided in the plant 200 and the sensor 300 via a communication network, such as a dedicated line or a secure internet line.

[0031] (2) User interface unit 120 The user interface unit 120 includes, for example, a keyboard and a display, and accepts user input and displays the output results from the processor 140 in a way that is visible to the user.

[0032] (3) Memory 130 Memory 130 temporarily or permanently stores programs (instructions), parameters, and various data for the operation of the prediction device 100. The program includes one or more instructions for the operation of the prediction device 100.

[0033] (4) Processor 140 The processor 140 provides various functions of the prediction device 100. The processor 140 includes a state value acquisition unit 141, a prediction processing unit 143, and a learning processing unit 145, which are functions realized by executing instructions stored in the memory 130. The processor 140 may also include other functions; that is, it can perform operations other than those provided by these functions. The specific operations of the state value acquisition unit 141, the prediction processing unit 143, and the learning processing unit 145 will be described later.

[0034] <3.2. Example of Operation> Next, an example of operation of the first embodiment will be described.

[0035] The state value acquisition unit 141 acquires one or more state values ​​for the target sample of the plant 200. The prediction processing unit 143 inputs the above one or more state values ​​into the trained model and obtains a predicted value of the output value for the target sample as a prediction result from the trained model.

[0036] The above-described trained model is trained by applying a difference learning method to a control model that models Plant 200, using the above-described training data, which is expressed in the form of differences between training samples in a training interval containing multiple training samples, where one or more state values ​​of Plant 200 and output values ​​of Plant 200 are different.

[0037] The control model described above is designed to derive the output value of plant 200 based on at least a function mapped from one or more state values ​​of plant 200 and a bias element of plant 200 that is independent of one or more state values ​​of plant 200. Specifically, in the control model described above, the bias element may have less inter-sample variation compared to, for example, the function mapped from the state values.

[0038] More specifically, the trained model described above is modeled to derive a predicted value of the output of the sample to be predicted, based at least on a function mapped from an input representing one or more state values ​​in the difference form between the nth and (n-1)th samples, and the output value of the (n-1)th sample, when the sample to be predicted is the nth sample.

[0039] Furthermore, the control model described above is designed to derive the output value of plant 200 based on at least a function mapped from one or more state values ​​of plant 200, a bias element of plant 200 independent of one or more state values ​​of plant 200, and a noise element of the control model described above. Note that the control model is not limited to the elements described above and may include model elements that have a causal relationship with the output value of plant 200.

[0040] Furthermore, assuming that the above learning interval contains N training samples (where N is an integer greater than 2), the above training data can be specifically represented in the form of the difference between the nth training sample and the (n-1)th training sample by equation (1) below.

number

[0041] Furthermore, one or more state values ​​in the above-mentioned sample for prediction are K state values ​​x k (where k is an integer from 1 to K.) This is expressed as follows: In this case, the above control model is expressed by equation (2) below.

number

number

[0042] This difference format is based on the following relationship: the difference between the output of the nth sample and the (n-1)th sample can be expressed as shown in equation (4) below.

number

number

number

[0043] The trained model, trained using the difference learning method described above, is modeled to derive the predicted value μ(n) of the output value of the sample to be predicted using equation (5) below, when the sample to be predicted is the nth sample.

number

[0044] The trained model represented by equation (5) above is used by the prediction processing unit 143 as follows. First, the prediction processing unit 143 sets the sample to be predicted as the nth sample and obtains the known output value y(n-1) of the (n-1)th sample. For example, the prediction processing unit 143 may obtain the known output value y(n-1) from data within the interval of the training samples. Next, the prediction processing unit 143 calculates the predicted value μ(n) corresponding to the output value y(n) according to equation (5) above.

[0045] Here, the bias b(n) is not explicitly represented in the above calculation formula for the predicted value μ(n). However, the known output value y(n-1) for the (n-1)-th sample implicitly includes the unknown bias b(n-1). Furthermore, under the assumption that the bias changes gradually between samples as described above, [Formula] can be regarded as such. In this manner, the prediction processing unit 143 can perform prediction in consideration of the bias b(n) by using the above calculation formula for the predicted value μ(n), even if the bias b(n) is not explicitly represented.

[0046] Further, the above differential learning method is executed by, for example, the learning processing unit 145. Specifically, the learning processing unit 145 acquires N sets of sample data from past operation history information of the plant 200 stored in the memory 130 such as process data of the plant 200, for example, and converts the acquired N sets of sample data into the above learning data D diff to obtain the above learning data D diff Subsequently, the learning processing unit 145 applies the above differential learning method to the above learning data D diff to train the trained model. Further, the trained model is stored in the memory 130 when the learning processing unit 145 accesses the memory 130.

[0047] Furthermore, the trained model is read out from the memory 130 by, for example, the prediction processing unit 143. When receiving one or more state values of the prediction target sample from the state value acquisition unit 141, the prediction processing unit 143 performs calculation processing using the trained model read out from the memory 130, and acquires a predicted value corresponding to the output value of the prediction target sample as the prediction result.

[0048] Furthermore, the processor 140 may acquire predicted values ​​of the bias element for multiple training samples included in the training interval based on the trained model. In this case, the processor 140 creates image information representing the changes in the acquired predicted values ​​of the bias element between different training samples, and controls the user interface unit 120 to display the image information so that it can be seen by the user. This allows the user to understand the changes in the bias element between samples and to control or maintain the plant 200 more appropriately. The predicted values ​​of the bias element for each training sample may, for example, be calculated by the prediction processing unit 143 based on the trained model, or they may be received from a device other than the prediction device 100.

[0049] (2) Control Model

[0050] The control models described above are, more specifically, models represented in linear form (hereinafter referred to as linear models) or models represented by nonlinear equations (hereinafter referred to as nonlinear models). The following sections will explain the cases of linear and nonlinear models.

[0051] (2-1) Example using a linear model The control model described above is a linear model represented by equation (6) below.

number

[0052] In this case, the above difference format can be given by equation (7) below.

number

number

number

[0053] The above difference learning method follows the least squares method according to equation (8) below, with respect to the squared error L diff By minimizing the above training data D diff From coefficients w1, ..., w k This further includes deriving the following.

number

[0054] The coefficients w1, ..., w were derived in this manner. k These are included in the above-mentioned trained model. Specifically, the training unit 145 processes the coefficients w1,..., w using the least squares method described above. k The coefficients are derived and stored in memory 130 as components of the trained model described above.

[0055] Furthermore, the above difference learning method includes deriving the intercept term w0 according to equation (9) below.

number

[0056] The coefficient w0 derived in this way is included in the trained model described above. Specifically, the training unit 145 derives the intercept term w0 according to equation (9) above for deriving the intercept term w0, and stores the derived intercept term w0 in the memory 130 as a component of the trained model described above.

[0057] Specifically, the above-trained model is modeled to derive the predicted value μ(n) of the output value of the sample to be predicted using equation (10) below.

number

[0058] The trained model represented by equation (10) above is used by the prediction processing unit 143 as follows. First, the prediction processing unit 143 sets the sample to be predicted as the nth sample and obtains the known output value y(n-1) of the (n-1)th sample. For example, the prediction processing unit 143 may obtain the known output value y(n-1) from the data within the interval of the training samples. Next, the prediction processing unit 143 calculates the predicted value μ(n) corresponding to the output value y(n) according to equation (10) above.

[0059] Here, the above formula for calculating the predicted value μ(n) does not explicitly represent the bias b(n). However, the output value y(n-1) for the (n-1)th sample implicitly includes an unknown bias b(n-1). Furthermore, under the assumption that the bias element changes slowly,

number

[0060] Furthermore, the prediction processing unit 143 predicts the bias b(n) of the nth sample in the above learning interval.

number

number

[0061] (2-2) Examples of nonlinear models The control model described above is a nonlinear model in which the function f represents a nonlinear mapping by the neural network shown in equation (12) below, as shown in Figure 3, for example.

number

[0062] In this case, the above difference format can be given to equation (13) below.

number

number

number

[0063] Furthermore, the above difference learning method represents a function that represents a nonlinear mapping by a neural network as shown in equation (14) below, as shown in Figure 4.

number

number

[0064] Furthermore, the above difference learning method is used with the above training data D diff The neural network is trained by applying the difference format according to equation (15) below.

number

[0065] Here, we assume that equation (16) below holds true.

number

[0066] The trained model, which has been trained using the difference learning method described above, is modeled to derive the predicted value μ(n) of the output value of the sample to be predicted using equation (17) below.

number

[0067] The trained model represented by equation (17) above is used by the prediction processing unit 143 as follows: The prediction processing unit 143 sets the sample to be predicted as the nth sample and obtains the known output value y(n-1) of the (n-1)th sample. For example, the prediction processing unit 143 may obtain the known output value y(n-1) from data within the interval of the training samples. Next, the prediction processing unit 143 calculates the predicted value μ(n) corresponding to the output value y(n) according to equation (17) above.

[0068] Here, the above formula for calculating the predicted value μ(n) does not explicitly express the bias b(n). However, the output value y(n-1) for the (n-1)th sample implicitly includes an unknown bias b(n-1). Furthermore, under the assumption that the bias changes gradually between samples,

number

[0069] Furthermore, the prediction processing unit 143 predicts the bias b(n) of the nth sample in the above learning interval.

number

number

[0070] Here,

number

number

number

number

number

number

[0071] (3) Samples to be predicted Specifically, the above-mentioned sample for prediction is the time period at which the output value should be predicted for plant 200.

[0072] In this case, the status value acquisition unit 141 acquires the status value x of the plant at the predicted time. k The prediction processing unit 143 then obtains the above state value x k The data is input into the trained model, and the predicted value of the output at the target time is obtained as the prediction result from the trained model.

[0073] Furthermore, the N training samples mentioned above are samples arranged along the time axis that precede the target time for prediction. The sample intervals for the training samples are, for example, equal. However, the sample intervals for the training samples are not necessarily equal; for example, the sample intervals may change depending on the maintenance period of plant 200.

[0074] Furthermore, the bias b(n) is equal to the state value x k This represents a model element of the temporal changes of plant 200, which is independent of (n). Examples of temporal changes include deviations in sensor values ​​from sensor 300, which exhibit seasonality due to temperature changes, and trend components due to the aging of plant 200.

[0075] The above-mentioned sample to be predicted is not limited to the predicted time period described above. In other words, the sample to be predicted may be determined based on the attributes or type of a predetermined controlled object. In this case, the above-mentioned training sample may be ordered within the training interval based on known attributes according to predetermined rules.

[0076] (4) Specific examples Next, with reference to Figures 5 and 6, a specific example of the prediction process flow in the prediction device 100 will be explained. Figure 5 is a schematic diagram showing a specific example of the relationships between various factors used in the prediction process. Figure 6 is a flowchart showing the prediction process flow in the prediction device 100.

[0077] In this specific example, plant 200 is managed or controlled by various factors in order to manufacture products that meet predetermined quality and conditions according to one or more processes. Specifically, as shown in Figure 5, in this specific example, the outlet temperature 510 of the heating tubes installed inside plant 200 is assumed to be a known explanatory variable.

[0078] Furthermore, as shown in Figure 5, four state values ​​are used: the temperature inside the heating tubes of plant 200 521, the vapor pressure inside the heating tubes of plant 200 523, the charge flow rate 525, and the charge pressure 527. These factors are known explanatory variables, which are detected, for example, by sensors 300 installed in plant 200 and provided to the prediction device 100.

[0079] Furthermore, as shown in Figure 5, the specific heat of the brewing liquid in plant 200 (531), the degree of fouling of the heating pipes in plant 200 (533), and the ambient temperature around plant 200 (535) are used as bias elements. These factors correspond to, for example, seasonal components and trend components, and are treated as unknown variables.

[0080] Furthermore, in the processing flow related to this specific example, the above-mentioned sample to be predicted is assumed to be the time at which the heating tube outlet temperature of 510 should be predicted.

[0081] First, in step S601, the learning processing unit 145 stores the state values ​​x of the N samples of the plant 200 in the memory 130. k Retrieve historical information including (k is an integer from 1 to 4) and the output value y.

[0082] In step S603, the learning processing unit 145 uses the history information acquired in step S601 to create the learning data D. diff Convert to.

[0083] In step S605, the learning processing unit 145 processes the learning data D diff The above-mentioned pre-trained model is trained by applying the above-mentioned differential learning method.

[0084] In step S607, the state value acquisition unit 141 acquires, for example, the state value x at the predicted target time provided from each sensor 300. k (In this specific example, k is an integer from 1 to 4.)

[0085] In step S609, when the prediction processing unit 143 receives the state value for the target time from the state value acquisition unit 141, it performs calculation processing using the trained model read from the memory 130 and obtains the predicted value of the output value for the target sample as the prediction result.

[0086] According to the process shown in Figure 6 above, the heating tube outlet temperature 510 can be accurately predicted by taking into account the influence of bias elements including the specific heat of the plant's brewing liquid 531, the degree of heating tube fouling 533, and the ambient temperature 535 at the target prediction time.

[0087] (5) Time evolution of bias elements As a comparative example of this operation, we will explain the case where a pre-trained model obtained using a learning method different from the differential learning method (hereinafter also referred to as the learning method related to the comparative example) is used (hereinafter referred to as the pre-trained model related to the comparative example).

[0088] First, according to the learning method for the comparative example, the controlled object is modeled linearly by equation (21) below.

number

[0089] Here, if ξ(n) is a Gaussian noise with mean 0 and variance 1, then σξ(n) is a Gaussian noise with mean 0 and variance σ 2 This becomes Gaussian noise.

[0090] Next, in the learning method for the comparative example, the intercept term w0 and coefficients w1, ...w k We will find the least squares value according to equation (22) below.

number

[0091] On the other hand, in the above example of operation, as described above,

number

[0092] Furthermore, in the learning method relating to the comparative example, the predicted value μ(n) corresponding to the output value y(n) is given by equation (23) below.

number

[0093] As is clear from this equation for prediction μ(n), the learning method in the comparative example, unlike the above example which uses a differential learning method, does not take into account the bias b(n) implicitly included in the known output y(n-1).

[0094] Furthermore, in the learning method relating to the comparative example, the bias b(n) in the training sample is calculated using the following equation (24).

number

[0095] Next, with reference to Figure 7, we will explain the difference in the change of bias b(n) between the above example of operation and the comparative example. Figure 7 is a diagram for explaining the difference in the change of bias b(n) between the example of operation and the comparative example. According to Figure 7, graph 710 shows the change in bias b(n) with respect to the time axis (horizontal axis) when the differential learning method related to the above example of operation is applied, graph 720 shows the change in bias b(n) with respect to the time axis (horizontal axis) when the reference learning method related to the comparative example is applied, and graph 730 is a graph that superimposes graphs 710 and 720.

[0096] As is clear from the changes in the vertical axis shown in Graph 730, Graph 710, to which the differential learning method was applied, shows a larger fluctuation component of the bias element than Graph 720, to which the learning method related to the comparative example was applied, and the time variation of the bias element of the controlled plant 200 can be extracted with high accuracy.

[0097] In this way, according to the above example of operation, a control model that more accurately considers bias elements can be constructed. The time changes of such bias elements are displayed to the user in a visible manner, for example, by a display (user interface unit 120). This allows the user to understand in detail, for example, the trend component due to the aging of the plant 200 and the impact of product lot changes, and to perform control or maintenance management of the plant 200 more appropriately.

[0098] In the specific examples shown in Figures 5 and 6, various factors related to the plant 200 were described as being managed or controlled. However, the system to be controlled is not limited to the plant 200 that manufactures products according to one or more manufacturing processes, but may be any control system to which the reference learning method used in the above operation example can be applied.

[0099] <<3. Second Embodiment>> Next, a second embodiment of this disclosure will be described with reference to Figure 8. While the first embodiment described above is a specific embodiment, the second embodiment is a more generalized embodiment.

[0100] <<4.1. Second Embodiment>> Figure 8 is a block diagram showing an example of a schematic configuration of the prediction device 1000 according to the second embodiment. As shown in Figure 8, the prediction device 1000 is a computer that predicts the output value of a control target to be predicted, and comprises a memory 1010 and a processor 1020.

[0101] Memory 1010 temporarily or permanently stores programs (instructions), parameters, and various data for the operation of the prediction device 1000. The program includes one or more instructions for the operation of the prediction device 1000.

[0102] The processor 1020 provides various functions of the prediction device 1000. The processor 1020 includes a state value acquisition unit 1021 and a prediction processing unit 1023, which are functions realized by executing instructions stored in the memory 1010. The processor 1020 may also include other functions; that is, it may perform operations other than those provided by these functions. The specific operations of the state value acquisition unit 1021 and the prediction processing unit 1023 will be described later.

[0103] <<4.2. Example of Operation>> Next, an example of operation of the second embodiment will be described.

[0104] The state value acquisition unit 1021 acquires one or more state values ​​for the control target sample. The prediction processing unit 1023 inputs the one or more state values ​​into the trained model and obtains a predicted value of the output value for the target sample as a prediction result from the trained model.

[0105] The above-mentioned trained model is trained by applying a difference learning method to a control model that models the above-mentioned controlled object, using training data in a training interval containing multiple training samples, where the training data is expressed in the form of differences between training samples, where one or more state values ​​of the controlled object and the output values ​​of the controlled object are different.

[0106] The above control model is designed to derive the output value of plant 200 based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of plant 200 that is independent of the one or more state values ​​of the controlled object.

[0107] - Relationship with the first embodiment For example, the state value acquisition unit 1021 and the prediction processing unit 1023 of the prediction device 1000 according to the second embodiment may perform the operations of the state value acquisition unit 141 and the prediction processing unit 143 of the prediction device 100 according to the first embodiment, respectively. In this case, the description of the first embodiment is also applicable to the second embodiment. However, the second embodiment is not limited to the above example.

[0108] According to the second embodiment described above, the output value of the controlled object can be predicted with high accuracy by considering the bias b(n) representing the model elements of the controlled object.

[0109] <<5. Other Embodiments>> While embodiments of the present disclosure have been described above, the present disclosure is not limited to these embodiments. It will be understood by those skilled in the art that these embodiments are merely illustrative and that various modifications are possible.

[0110] For example, the steps in the process described herein do not necessarily have to be executed chronologically in the order shown in the flowchart. For instance, the steps in the process may be executed in a different order than that shown in the flowchart, or in parallel, as long as the results of the process described above are achieved. Furthermore, some steps in the process may be deleted, or additional steps may be added to the process.

[0111] Furthermore, a prediction method including processing of the components of the prediction device described herein may be provided, and a program for causing a processor to perform the processing of the above components may be provided. A computer-readable non-transitory medium on which such a program is recorded may also be provided. Naturally, such devices, modules, methods, programs, and computer-readable non-transitory mediums are also included in this disclosure.

[0112] The following disclosure is an appendix which is a non-limiting list of exemplary embodiments and combinations of exemplary embodiments of the present disclosure. That is, some or all of the above embodiments may also be described as follows, but are not limited to the following appendix. [Note 1] Obtaining one or more state values ​​from the target sample in the controlled object, The process involves inputting one or more state values ​​into a trained model and obtaining a predicted value of the output value for the target sample as a prediction result from the trained model. The computer executes this, The trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval containing multiple training samples, where each of the training data is expressed in the form of differences between training samples, where one or more state values ​​of the controlled object and the output values ​​of the controlled object are different. A prediction method wherein the control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object. [Note 2] The prediction method described in Appendix 1, wherein the trained model is modeled to derive a predicted value of the output value of the sample to be predicted, based at least on a function mapped from an input representing one or more state values ​​in difference form between the nth and n-1th samples, and the output value of the n-1th sample, when the sample to be predicted is the nth sample. [Note 3] The prediction method according to Appendix 1 or 2, wherein the control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object, a bias element of the controlled object independent of the one or more state values ​​of the controlled object, and a noise element of the control model. [Note 4] If the aforementioned learning interval contains N training samples (where N is an integer greater than 2), then the training data D diff This is expressed in the form of the difference between the nth training sample and the (n-1)th training sample by equation (1) below,

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

[0113] Each configuration and its combination in each embodiment is an example, and additions, omissions, substitutions, and other modifications can be made as appropriate without departing from the spirit of this disclosure. This disclosure is not limited by the embodiments. [Explanation of symbols]

[0114] 1 System 100, 1000 prediction device 110 Network Communications Department 120 User Interface Section 130, 1010 memory 140, 1020 processors 141, 1021 Status value acquisition unit 143, 1023 Prediction Processing Unit 145 Learning Processing Unit

Claims

1. Obtaining one or more state values ​​from the target sample in the controlled object, The process involves inputting one or more of the aforementioned state values ​​into a trained model and obtaining a predicted value of the output value for the target sample as a prediction result from the trained model. The computer executes this, The trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval containing multiple training samples, where each of the training data is expressed in the form of a difference between training samples, where one or more state values ​​of the controlled object and the output values ​​of the controlled object are different. The control model is a method that models the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object.

2. The method according to claim 1, wherein the trained model is modeled to derive a predicted value of the output value of the sample to be predicted, based at least on a function mapped from an input representing one or more state values ​​in difference form between the nth and n-1th samples, and the output value of the n-1th sample, when the sample to be predicted is the nth sample.

3. The method according to claim 1, wherein the control model is modeled to derive an output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object, a bias element of the controlled object independent of the one or more state values ​​of the controlled object, and a noise element of the control model.

4. If the aforementioned learning interval contains N training samples (where N is an integer greater than 2), then the training data D diff This is expressed in the form of the difference between the nth training sample and the (n-1)th training sample by equation (1) below, [Math 1] One or more state values ​​in the aforementioned sample subject to prediction are K state values ​​x ｋ (where k is an integer from 1 to K.) The aforementioned control model is represented by the following equation (2): [Math 2] In equation (2) above, x ｋ (n) is the state value at the nth sample, y(n) is the output value at the nth sample, and f is x ｋ The prediction method according to claim 3, wherein a function represents a mapping with (n) as input, where b(n) represents the bias element at the nth sample and σξ(n) represents the noise element at the nth sample.

5. The difference learning method assumes that b(n) changes gradually between samples in the learning interval, and considers b(n) and b(n-1) to be approximately zero, and uses the difference format shown in equation (3) below for the training data D diff By applying this, including learning, [Math 3] In equation (3) above, σξ(n) has a mean of 0 and a variance of σ ２ This is Gaussian noise, The prediction method according to claim 4.

6. The trained model is designed to derive the predicted value μ(n) of the output value of the sample to be predicted using equation (4) below, when the sample to be predicted is the nth sample. [Math 4] Obtaining the aforementioned predicted value means Setting the aforementioned sample to be predicted as the nth sample, Obtain the known output value y(n-1) of the (n-1)th sample, The prediction method according to claim 5, comprising obtaining a predicted value μ(n) corresponding to the output value y(n) according to equation (4) above.

7. The aforementioned control model is a linear model represented by equation (5) below, [Math 5] The aforementioned differential learning method is, Assuming that b(n) changes slowly with respect to the output value y(n) in the aforementioned learning interval, we consider b(n) and b(n-1) to be approximately zero, and prepare the difference form shown in equation (6) below, [Math 6] according to the least squares method represented by the following formula (7), the aforementioned learning data D diff to derive a coefficient w １ , ..., w ｋ the method further comprising: [Number 7] The trained model is the derived coefficient w １ ,・・・, lol ｋ The prediction method according to claim 4, including the method described in claim 4.

8. The aforementioned difference learning method uses the following equation (8) for the intercept term [Number 8] This includes deriving the following: [Number 9] The trained model is the derived intercept term w ０ The prediction method according to claim 7, further comprising:

9. The aforementioned trained model is modeled to derive the predicted value μ(n) of the output value of the sample to be predicted using equation (9) below, [Number 10] Obtaining the aforementioned predicted value means Setting the aforementioned sample to be predicted as the nth sample, Obtain the known output value y(n-1) of the n-1th sample, The prediction method according to claim 8, comprising obtaining a predicted value μ(n) corresponding to the output value y(n) according to equation (9) above.

10. Predicted value of bias b(n) for the nth sample in the aforementioned learning interval [Math 11] Obtain it according to the following formula (10). [Math 12] The prediction method according to claim 8, further comprising:

11. The control model is a nonlinear model in which the function f represents a nonlinear mapping by a neural network as shown in equation (11) below. [Number 13] The aforementioned differential learning method is, The function representing the nonlinear mapping by the neural network shown in equation (12) below. [Number 14] To prepare, [Number 15] Under the assumption that b(n) changes slowly with respect to the output value y(n) in the aforementioned learning interval, and assuming that b(n) and b(n-1) are approximately zero, the training data D diff By applying the difference format according to the following equation (13), neural network learning is performed, [Number 16] In equation (13), the following equation (14) also holds: [Number 17] The prediction method according to claim 4, including the method described in claim 4.

12. The aforementioned trained model is modeled to derive the predicted value μ(n) of the output value of the sample to be predicted using equation (15) below, [Number 18] Obtaining the aforementioned predicted value means Setting the aforementioned sample to be predicted as the nth sample, Obtain the known output value y(n-1) of the (n-1)th sample, The prediction method according to claim 11, comprising obtaining a predicted value μ(n) corresponding to the output value y(n) according to equation (15) above.

13. Predicted value of bias b(n) for the nth sample in the aforementioned learning interval [Number 19] Output the result according to the following formula (16): [Number 20] In equation (16) above, [Math 21] Assuming that N is a sufficiently large value, y(n) and the mean of y(n) in N sample intervals are calculated according to equation (17) below. [Number 22] Calculate the difference from, [Number 23] Furthermore, according to equation (18) below, [Number 24] of [Number 25] This is derived by substituting the following: [Number 26] The prediction method according to claim 11, further comprising:

14. To acquire the aforementioned training data, The process involves applying the difference learning method to the training data to train the trained model, The prediction method according to any one of claims 1 to 13, further comprising:

15. The prediction method according to any one of claims 1 to 13, wherein the controlled object is a process-type manufacturing plant that performs one or more manufacturing processes.

16. Obtaining one or more state values ​​in the predicted target sample of the controlled object means obtaining one or more state values ​​in the controlled object at the predicted target time, Obtaining a predicted value of the output value at the target sample from the aforementioned trained model as a prediction result means obtaining a predicted value of the output value at the target time for prediction. The aforementioned plurality of training samples are samples arranged in the time axis direction that are prior to the target time for prediction, The prediction method according to any one of claims 1 to 13, wherein the bias element represents a model element of the time-dependent change of the controlled object, independent of one or more state values.

17. Obtaining predicted values ​​of the bias element for multiple training samples included in the training interval based on the trained model, To display to the user, in a way that visually represents the change in the predicted value of the acquired bias element between different training samples, The prediction method according to any one of claims 1 to 13, further comprising the computer performing the following steps.

18. Memory and Processor and A predictive device that predicts the output of a controlled object, The aforementioned processor executes instructions stored in memory, Obtain one or more state values ​​from the target sample in the controlled object, The one or more state values ​​are input to the trained model, and the predicted output value for the sample to be predicted is obtained as a prediction result from the trained model. The trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval containing multiple training samples, where each of the training data is expressed in the form of a difference between training samples, where one or more state values ​​of the controlled object and the output values ​​of the controlled object are different. A prediction device, wherein the control model is modeled to derive the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object.

19. Obtaining one or more state values ​​from the target sample in the controlled object, The process involves inputting one or more of the aforementioned state values ​​into a trained model and obtaining a predicted value of the output value for the target sample as a prediction result from the trained model. It is a program that causes a computer to execute it. The trained model is trained by applying a difference learning method to a control model that models the controlled object, using training data in a training interval containing multiple training samples, where each of the training data is expressed in the form of a difference between training samples, where one or more state values ​​of the controlled object and the output values ​​of the controlled object are different. The control model is a program that models the output value of the controlled object based on at least a function mapped from one or more state values ​​of the controlled object and a bias element of the controlled object that is independent of the one or more state values ​​of the controlled object.

Images