Learning system and learning method for prediction model

Abstract

The present disclosure relates to a learning system that successively learns predictive models for manufacturing devices having different product requirements. A learning system is provided with: one or more processors; and a storage device that holds, as learning data, a parameter representing a learning value calculated on the basis of data acquired at the time of manufacturing by means of the probability distribution, and a timing at which the parameter is last updated. The one or more processors use the prediction model corrected on the basis of the learning data to calculate a setting value for controlling the manufacturing apparatus, acquire actual performance values for manufacturing corresponding to predicted values of the prediction model, calculate a parameter representing a probability distribution of the actual performance values on the basis of the acquired one or more actual performance values, and control the manufacturing apparatus based on the calculated parameter. The learning data is updated and stored in the storage device.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to a learning system and a learning method for successively learning a prediction model. BACKGROUND

[0002] In a manufacturing process of a product having different product requirements such as a rolling process, in order to achieve various product requirements (steel grade, size, etc.) accepted from a higher-order computer, a setting device calculates various setting values (amount of injection of cooling water, roll gap of a rolling mill, etc.) of each device in the manufacturing process.

[0003] In a case where the manufacturing process is a rolling process, the setting device calculates appropriate setting values capable of predicting appropriate rolling results using a rolling model that predicts a rolling result based on a rolling condition. Therefore, maintaining and improving the prediction accuracy of the rolling model is indispensable to achieve stable operation and to achieve a high level of quality requirements. However, there are a large number of combinations of various product requirements manufactured in the steel field, and a variety of equipment and operations, and therefore it is not realistic to make a rolling model that encompasses all situations and manually manage the accuracy thereof. Therefore, a method is generally adopted in which past performance data of rolling is compared with an operation result using a rolling model, and a learning value for maintaining the accuracy of the rolling model is automatically calculated based on the comparison result. For example, after rolling, an index value (for example, ratio or difference) indicating an error of a performance value from an operation value using a rolling model is calculated, and the index value is stored as a learning value. Then, when an operation value of a setting value before rolling is calculated, the setting value after the influence of the error is corrected is calculated by reflecting the learning value in the operation value using the rolling model. The learning value is stored in a hierarchical classification table that is distinguished according to various product requirements (for example, steel grade, sheet thickness, sheet width, etc.), and is updated at each time of rolling. As an updating method of the learning value used at this time, a method using an exponential smoothing method is typical.

[0004] Patent Literature 1 proposes a method of adjusting a learning gain based on the number of times of learning and history information of a learning value, or a change in an operation mode, and the like, and dividing an error into a component dependent on a product requirement and a component dependent on a change in time. On the other hand, Patent Literature 2 proposes a method of evaluating not only a corresponding cell but also recency, saturation, and stability of a plurality of learning values based on history information of a learning value, and correcting a learning value used in rolling by polynomial interpolation from the plurality of learning values that are well evaluated.

[0005] PRIOR ART DOCUMENTS PATENT LITERATURE Patent Literature 1: Japanese Patent No. 4543684 Patent Literature 2: Japanese Patent No. 6233423 SUMMARY

[0006] Problem to be solved by the Invention In the case where the learning value is managed by the layer classification table divided into a plurality of cells, in order to learn different errors with high accuracy according to various conditions, the division of the layer classification table is preferably fine. This is because, since the learning value is a discrete table value, the more the division is refined, the smaller the deviation of the learning value between cells caused by discretization. However, since a different rolling condition is assigned to each cell, the more the division is fine, the less the learning opportunity of each cell, and the slower the proficiency of learning. One method to accelerate the proficiency of learning is to increase the smoothing coefficient for learning, that is, the learning gain. However, in the case where a large amount of noise is contained in the observed performance value, if the learning gain is simply increased, the accuracy of the learning value can be reduced due to the confusion of the error (the error between the performance value and the calculated value using the rolling model) and the noise. In addition, there is an error in the error that does not depend on the product requirement but depends on the time variation. For example, the change in the operation mode of the upstream process, the cooling water temperature, and the like correspond to this. In the case where learning is performed without distinguishing the error depending on the product requirement from the error depending on the time variation, a learning value with high accuracy cannot be obtained.

[0007] In the method described in the above-described Patent Document 1, the learning gain is calculated based on the number of times of learning and the history information of the learning value. However, in the case where a large amount of noise is contained in the observed performance value, it is not possible to separate the noise from the change in the learning value depending on the product requirement in the history information. Therefore, confusion of the component depending on the time variation and the component depending on the product requirement can occur. In addition, since the convergence condition of the learning value varies depending on the level of the noise, it is difficult to say that it is appropriate to judge the accuracy of the learning value according to the number of times of learning.

[0008] In the method described in the above-described Patent Document 2, the three evaluation indexes of recency, saturation, and stability are calculated and stored with respect to each cell based on the history information of the learning value. Therefore, the required calculation amount is large, and the storage capacity required to store the history information and the three evaluation indexes with respect to each cell is also large. In addition, the correction of the learning value used in the rolling is performed by polynomial interpolation, and therefore the variable of each cell divided in the layer classification table is basically a quantitative variable, and at least a qualitative variable must be sequential. However, a typical layer classification table is divided according to the product category (for example, steel type information such as low carbon steel, silicon steel, high manganese steel, and the like). Therefore, it is difficult to say that the polynomial interpolation performed according to the division of the qualitative variable without sequentiality is appropriate.

[0009] The present application has been achieved in view of the above problems, and aims to provide a learning system of a prediction model and a learning method, which can achieve stable operation and meet a high quality requirement even in a case where a noise of an actual performance value is large, in a case where a component depending on a time change exists, or in a case where maturation of learning is slow.

[0010] Means for solving the problems A first aspect of the present disclosure relates to a learning system that successively learns a prediction model for manufacturing apparatuses having different product requirements. The learning system includes one or more processors, and a storage device that holds, as learning data, parameters that represent learning values calculated based on data acquired at the time of manufacturing by a probability distribution, and a timing at which the parameters are last updated. The one or more processors calculate a set value for controlling the manufacturing apparatuses using a prediction model corrected based on the learning data, acquire actual performance values of manufacturing corresponding to predicted values of the prediction model, calculate parameters that represent a probability distribution of the actual performance values based on the acquired one or more actual performance values, and update and store the learning data in the storage device.

[0011] A second aspect of the present disclosure relates to a learning method that successively learns a prediction model for manufacturing apparatuses having different product requirements. The learning method includes a step of acquiring learning data related to parameters that represent learning values calculated based on data acquired at the time of manufacturing by a probability distribution, and a timing at which the parameters are last updated, a step of calculating a set value for controlling the manufacturing apparatuses using a prediction model corrected based on the learning data, a step of acquiring actual performance values of manufacturing corresponding to predicted values of the prediction model, and a step of calculating parameters that represent a probability distribution of the actual performance values based on the acquired one or more actual performance values, and updating the learning data.

[0012] Effects of the Invention According to the present application, in advance setting of manufacturing apparatuses having different product requirements, by recapturing learning values calculated based on data acquired at the time of manufacturing as probability variables, even in a case where a noise of an actual performance value is large, in a case where a component depending on a time change exists, or in a case where maturation of learning is slow, it is possible to correct errors of prediction models of objects with high precision, achieve stable operation, and meet a high quality requirement. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a block diagram showing a configuration example of a learning system of an embodiment of the present disclosure.

[0014] Figure 2 is a block diagram showing an example of a functional configuration of a learning system of a first embodiment of the present disclosure.

[0015] Figure 3It is a graph that shows the operation of updating the probability distribution of the learning values ​​using the probability distribution of the actual performance values.

[0016] Figure 4 It is a graph that shows the operation of integrating the probability distributions of multiple learning values ​​into a pre-defined probability distribution of the learning values ​​to be used.

[0017] Figure 5 It is a graph that shows the operation of integrating the probability distributions of multiple learning values ​​into a pre-defined probability distribution of the learning values ​​to be used.

[0018] Figure 6 This is a block diagram illustrating examples of the functional configuration of the learning system according to the third and fifth embodiments of this disclosure.

[0019] Figure 7 It is a graph that shows the operation of updating the probability distribution of multiple learning values ​​using the probability distribution of the actual performance values.

[0020] Figure 8 It is a graph that shows the operation of updating the probability distribution of multiple learning values ​​using the probability distribution of the actual performance values. Detailed Implementation

[0021] The embodiments of this disclosure will be described with reference to the accompanying drawings.

[0022] 1. Summary This disclosure relates to a learning system and a learning method for the learning system, which, in a pre-set manufacturing apparatus (e.g., a rolling process) with different product requirements, corrects errors in a predictive model of an object based on the results of batch manufacturing (e.g., in the order of 5 products in product group A, 4 products in product group B, 11 products in product group C, and 3 products in product group B).

[0023] 2. First Implementation Method Hereinafter, the learning system of the first embodiment of the present disclosure and the learning method executed by the learning system will be described with reference to the accompanying drawings.

[0024] 3. The Composition of a Learning System Figure 1 This is a block diagram illustrating an example of the hardware configuration of the learning system 100 according to the first embodiment of this disclosure. The learning system 100 is, for example, a system that, in rolling processes where product requirements differ on a per-plate basis, successively obtains batch manufacturing results and learns from them, and based on the learning results, corrects errors in the prediction model for the next manufactured object. Hereinafter, in this embodiment, the application of the learning system 100 to a rolling process will be described as an example. However, the manufacturing process for which the learning system 100 is applied is not limited to a rolling process; it can be any manufacturing process for products with different batch requirements.

[0025] The learning system 100 controls one or more mechanical elements 200 included in the manufacturing apparatus. The learning system 100 is connected to the one or more mechanical elements 200 via a control network 20.

[0026] The learning system 100 includes a processor (processing circuit) 101, a program memory 102, a data storage memory 103, a communication module 104, and a user interface 105. The processor 101 can be a CPU, RISC, DSP, FPGA, ASIC, PLD, or other processing unit, or a combination of two or more of these, or a dedicated processor for the learning system 100. In this embodiment, there is one processor 101, but the learning system 100 may also have multiple processors 101.

[0027] Program memory 102 is communicatively coupled to processor 101. Program memory 102 stores a program consisting of multiple INST instructions executable by processor 101. The program consisting of INST instructions can be retrieved using computer-readable, non-transitory storage media or via a network. Alternatively, program memory 102 may be integrated into processor 101.

[0028] The data storage memory 103 is communicatively coupled to the processor 101. Learning data DATA is registered in the data storage memory 103. The learning data DATA is data used to correct the settings of the manufacturing apparatus. The learning data DATA is calculated based on data obtained during past manufacturing processes and is updated each time manufacturing is repeated. Alternatively, the data storage memory 103 may also be integrated into the processor 101.

[0029] The communication module 104 is communicatively integrated with the processor 101. The communication module 104 is configured to communicate with external devices, including a host computer 1 that determines product requirements.

[0030] 4. Learning the system's actions Figure 2 An example of the functional configuration of the learning system 100 is shown. The learning system 100 includes a setting unit 2, a control unit 3, and a learning unit 10. These functional units are implemented by reading the instruction INST from the program memory 102 and executing it by the processor 101. These functional units can be constructed by independent hardware, or a single hardware unit can have multiple functions.

[0031] The processing of each functional unit of the learning system 100 is explained according to a series of actions performed by the steel plate unit. This series of actions consists of a setting step, a control step, and a learning step. The setting step is performed at a timed interval before the front end of the steel plate reaches the rolling process. The control step is performed at a timed interval from the start of rolling at the front end of the steel plate to the end of rolling at the tail end. The learning step is performed at a timed interval after the required measurement values ​​for learning are obtained. Alternatively, a portion of the timing may be repeated between the learning step and the control step.

[0032] In the setting step, the setting unit 2 calculates various setting values ​​for the rolling process equipment based on product requirements (e.g., steel grade, plate thickness, etc.) provided by the host computer 1. A prediction model is used in the calculation of these setting values. At this time, the setting unit 2 uses past similar product requirements and calculated values ​​obtained during the prediction model-based calculation process (e.g., setting speed, setting temperature, etc.) as search keywords to obtain learning values ​​from the learning unit 10 for correcting the calculation results of the prediction model. These learning values ​​are numerical values ​​used to correct errors contained in the prediction model, i.e., errors arising between predicted and actual values, and are used by the setting unit 2 in the calculation of the various setting values. Then, the setting unit 2 outputs the calculated setting values ​​to the control device of the corresponding manufacturing apparatus.

[0033] The prediction model is represented by a function as shown in Equation (1). The function can be a multi-input function or contain nonlinear elements. In addition, the function can also include computations such as neural networks. As an example of correction based on learning values, additive correction as shown in Equation (2) can be cited. Furthermore, the prediction model is a collection of multiple prediction models, and the prediction value of one prediction model is used as the input of other prediction models in turn. In the example shown in Equation (3), the prediction value of prediction model _1 is used as the input of prediction model _2.

[0034] Here, X1 and X2 are the input vectors for prediction model 1 and prediction model 2, respectively. The input vectors are a portion of the product requirements provided from the host computer 1, necessary for calculating the functions of the prediction models, and may also include table values ​​indexed based on a portion of the product requirements. Additionally, Y1... prd and Y2 prd These are the predicted values ​​from prediction model 1 and prediction model 2, respectively. Additionally, Y1... set These are the settings for prediction model_1. f1 and f2 are functions of prediction model_1 and prediction model_2, respectively. z1 is the learning value for correcting prediction model_1.

[0035] In the learning step, if the control unit 3 records the actual performance value corresponding to the input of the prediction model, the setting unit 2 performs prediction calculation again based on the most recent actual performance value, as shown in equation (4). Hereinafter, the value calculated by re-predicting based on the most recent measured value will be called the performance recalculation value. Furthermore, the performance recalculation value is not limited to the performance recalculation value calculated directly using the actual performance value. For example, the performance value indirectly calculated based on the performance recalculation value obtained from prediction model 1 through re-predicting based on prediction model 2 is also included in the performance recalculation value.

[0036] In equation (4), X2 act Y1 is the input vector of prediction model 2 based on actual performance values. act Y2 is the actual performance value of prediction model 1. rprd It is the recalculated value of the actual performance of prediction model 2.

[0037] In the setting step, the control unit 3 obtains the setting value from the setting unit 2. Then, in the control step, the control unit 3 continuously controls the manufacturing device based on various measured values ​​and instruction values ​​from the operator, and continuously records the various measured values ​​as actual values.

[0038] The learning unit 10 stores the learning distribution, which is the probability distribution of the learning values, and the learning timing of the learning distribution. In the setting step, the setting unit 2 provides the learning unit 10 with product requirements and calculated values ​​obtained during the calculation process based on the prediction model. Based on the stored learning distribution, the learning unit 10 predicts the current learning distribution using this information and outputs the predicted learning distribution to the setting unit 2. Then, in the learning step, the learning unit 10 obtains the performance value from the control unit 3 and updates the stored learning distribution and its learning timing based on the obtained performance value, the learning distribution predicted in the setting step, the stored learning distribution, and its learning timing. The learning unit 10 repeats the above process each time a steel sheet is manufactured, updating the learning distribution based on the obtained performance value and successively correcting the prediction model.

[0039] 5. Learning Methods 5-1. Functions of the Study Department The learning unit 10 includes a learning distribution prediction unit 11, a performance distribution calculation unit 12, a learning distribution calculation unit 13, a learning distribution storage unit 14, and a learning timing storage unit 15. The functions constituting the learning unit 10 will be explained.

[0040] 5-2. Storage of Learning Information The learning distributed storage unit 14 and the learning timed storage unit 15 manage information by classifying calculated values ​​from the product requirements and prediction model calculation process based on specified conditions. Here, we assume there are three conditions. Let's define the three conditions as condition A, condition B, and condition C. Let the distinction value for condition A be NA, the distinction value for condition B be NB, and the distinction value for condition C be NC. Then, management is performed in a hierarchical classification table with cells of NA × NB × NC (hereinafter, the cells are referred to as learning batches). A specific learning batch is a group of values ​​(which may not be numerical) corresponding to conditions A, B, and C. Furthermore, the number of conditions can be arbitrary, and the hierarchical classification table becomes a tensor of order corresponding to the number of conditions.

[0041] The learning distribution storage unit 14 stores the probability density function Pr(Z) that represents the probability distribution of the learned values. old The parameters of the probability mass function are stored in the learning distribution storage unit 14. Furthermore, the learning values ​​are not limited to continuous probability variables; they can also be discrete probability variables. When the learning values ​​are discrete probability variables, the learning distribution storage unit 14 stores the parameters of the probability mass function.

[0042] The following explanation addresses the case where a normal distribution is assumed in the probability distribution. However, the probability distribution is not limited to a normal distribution and can be extended to a general probability distribution. In the case of a general probability distribution, computation can be performed, for example, using the Markov chain Monte Carlo method.

[0043] The learning distributed storage unit 14 has the average value μ of the learning value. old The relevant hierarchical classification table and the variance σ with respect to the learning values old 2 The relevant stratified classification table, the probability density function is expressed by equation (5).

[0044] The learning timing storage unit 15 has a hierarchical classification table that stores the previous learning timing. The learning timing is a variable (hereinafter referred to as the learning timing value) belonging to a space where the distance between the current value and the previous value can be defined by a function (distance function). The learning timing value can be, for example, an integer incremented each time manufacturing, the manufacturing time, etc.

[0045] 5-3. Calculation of learning values ​​for correction The calculation of the learning value is performed in the setting step during each rolling. Based on the product requirements provided by the setting unit 2 and the calculated values ​​from the prediction model's calculation process, the learning distribution prediction unit 11 indexes the corresponding learning batch from the hierarchical classification table of the learning distribution storage unit 14, and calculates the probability density function Pr(Z) of the indexed parameters. old The average or most frequent value is calculated and used as the predicted value z.set Output to setting unit 2. Furthermore, in the case of a normal distribution, since the mean or most frequent value is consistent with the mean used to represent the probability density function, only the mean μ of the learned value is indexed. old A relevant hierarchical classification table is sufficient. The output predicted value z set In setting section 2, the learning value z1 of equation (2) is used for calculation.

[0046] 5-4. Calculation of the learning value in this exercise The calculation of the current learning value is performed at each learning step. The performance distribution calculation unit 12 obtains the performance value from the control unit 3 and the performance recalculation value from the setting unit 2, and calculates the current learning value based on them. Regarding the calculation, for example, if it is an additive correction learning value as shown in equation (2), it can be calculated as shown in equation (6). The calculated current learning value is stored until the learning distribution is updated.

[0047] In equation (6), Z cur This is the learning value for this session.

[0048] 5-5. Calculation of Performance Distribution The performance distribution is calculated by performing a new set of learning values ​​in the learning steps when rolling of products classified as the same learning batch has been performed a predetermined number of times, or when the rolled products have been switched to products from different learning batches.

[0049] The performance distribution calculation unit 12 calculates the probability density function Pr(Z) based on the newly obtained set of learning values. cur Probability density function Pr(Z) cur The parameters of ) can be calculated based on equations (7) to (9).

[0050] Here, μ cur It is the average of the performance distribution, σ cur 2 It is the variance of the performance distribution, σ 2 It is the pre-defined variance, Z. cur [i] represents the learning values ​​contained in the current set, and N represents the number of learning values ​​contained in the current set. The performance distribution calculation unit 12 outputs the calculated parameters as the performance distribution to the learning distribution calculation unit 13.

[0051] 5-6. Learning Distribution Update The learning distribution processing unit 13 indexes the corresponding learning batch from the hierarchical classification table of the learning distribution storage unit 14 and the learning timing storage unit 15, based on the probability density function Pr(Z) of the indexed parameters. old ) and learning timing, and the probability density function Pr(Z) obtained from the performance distribution calculation unit. cur ), calculate the probability density function Pr(Z) associated with the updated learned value. new Probability density function Pr(Z) new The parameters of ) can be calculated using the following equations (10) to (13).

[0052] Here, K is the learning gain, μ new It is the updated probability density function Pr(Z) new The average value of σ new 2 It is the updated probability density function Pr(Z) new The variance of ). Additionally, the variance inflation coefficient Q. time It is an adjustment factor used to represent the increase in uncertainty caused by changes over time. Variance inflation coefficient Q time For example, based on equation (13), the current time t and the previous learning time t are compared. old The function f obtained by difference operation time Alternatively, you can refer to the table values.

[0053] Figure 3 This is a graph representing the images based on the operations of equations (10) to (13). The graph above shows the probability distribution Pr(Z) of the learning values ​​before the update obtained from the learning distribution storage unit 14. old The probability distribution Pr(Z) after time variance inflation is shown in the chart below. old The probability distribution Pr(Z) of the actual performance value cur ), and the updated probability density function Pr(Z) calculated based on them. new ).

[0054] Finally, the learning distribution operation unit 13 indexes the corresponding learning batch from the hierarchical classification table of the learning distribution storage unit 14 to update the probability density function Pr(Z) associated with the calculated updated learning value. new The parameters are used to index the corresponding learning batch from the hierarchical classification table of the learning timing storage unit 15 to update the learning timing value.

[0055] 6. Second Implementation Method (Hierarchical Bayesian Prediction) The first embodiment has been described above. Next, the learning system 100 and learning method of the second embodiment will be described. Furthermore, in the second embodiment or the fifth embodiment described below, the hardware configuration and operation timing of the learning system 100 are the same as in the first embodiment. Additionally, the functional configuration of the learning system 100 in the second embodiment is also the same as in the first embodiment. Descriptions of these common parts are omitted.

[0056] In the second embodiment, the learning distribution prediction unit 11, in addition to indexing the corresponding learning batch from the hierarchical classification table of the learning distribution storage unit 14, also obtains parameters of M learning batches located near the corresponding learning batch. Furthermore, the learning distribution prediction unit 11 obtains learning timing values ​​from the learning timing storage unit 15 corresponding to the M learning batches indexed in the learning distribution storage unit 14. Moreover, the obtained learning batches are not limited to those located near the corresponding learning batch in the hierarchical classification table; they can also be learning batches with learning timings close to the current time. Furthermore, the learning distribution prediction unit 11 can also calculate the integration weight w (described later). j Afterwards, repeated choices were made to collect w j Large batches of learning.

[0057] The learning distribution prediction unit 11 will be based on M+1 probability density functions Pr(Z) representing the parameters of the multiple learning batches obtained. 0 old ), Pr (Z) 1 old ), ...Pr (Z) M old (Z) 0 old The probability density function Pr(Z) obtained by integrating the corresponding learning batches set The average value or Pr(Z) can be calculated using numerical integration or Markov chain Monte Carlo methods. set (become the largest Z) set , as the predicted value z set Output to setting unit 2.

[0058] Figure 4 This shows the relationship between the corresponding learning batch and nearby learning batches. Figure 5 The graph of the integration operation of the probability density function is shown. The integration of the probability density function can be performed by using the operations of equations (14) to (16).

[0059] Here, σ 2 old j It is the probability density function Pr(Z) oldj The variance of w j It is Z old j The integrated weight, w' j It is a normalized integrated weight. In addition, the variance inflation coefficient Q... lot j It is a coefficient used to increase the variance as the relationship between the corresponding learning batch and the j-th learning distribution decreases, and to decrease the ensemble weight w as the relationship increases. j Q lot j For example, according to the corresponding batch l in the stratified classification table shown in equation (17) set With the j-th learning batch l j The distance operation function value f lot Alternatively, you can refer to the table values.

[0060] The method for calculating the predicted values ​​in the second embodiment can also be called hierarchical Bayes prediction.

[0061] 7. Third Implementation (Hierarchical Bayesian Prediction: Nearest Neighbor Selection Based on Graph) Figure 6 This is a diagram illustrating the learning system 100 according to the third embodiment. In the third embodiment, hierarchical Bayes prediction is performed in the same way as in the second embodiment. Furthermore, in the third embodiment, in addition to the configuration of the second embodiment, the learning unit 10 also includes a learning batch graph storage unit 16.

[0062] The learning batch graph storage unit 16 stores, for learning batches with one or all conditions, similarity between learning batches based on data analysis results and the manager's domain knowledge. For example, the similarity can be obtained using data analysis results based on variance analysis, cluster analysis, etc., regarding the differences between each learning batch. The learning distribution prediction unit 11 indexes the corresponding learning batch from the learning batch graph storage unit 16 and obtains similarity information between it and other learning batches. Then, the learning distribution prediction unit 11 obtains the parameters of M learning batches in descending order of similarity, replacing the learning batches located near the corresponding learning batches in the hierarchical classification table of the learning distribution storage unit 14 and the learning timing storage unit 15, based on the similarity information. Furthermore, when performing the integrated calculation of the probability density function, the learning distribution prediction unit 11 uses the similarity to replace the learning batch Z represented by formula (16). old j Integrated weight w j The operation.

[0063] 8. Fourth Implementation Method (Hierarchical Bayesian Learning) In the fourth embodiment, the learning distribution calculation unit 13 indexes the corresponding learning data from the hierarchical classification table of the learning distribution storage unit 14, which is the same as in the first to third embodiments. In addition, in the fourth embodiment, the learning distribution calculation unit 13 also obtains parameters for M learning batches located near the corresponding learning batch. Then, the learning distribution calculation unit 13 obtains the learning timing values ​​corresponding to the M learning batches indexed in the learning distribution storage unit 14 from the learning timing storage unit 15.

[0064] The learning distribution operation unit 13 is based on M+1 probability density functions Pr(Z) representing the parameters of the acquired multiple learning batches. old 0 ), Pr (Z) old 1 ), ..., Pr (Z) old M (Z) old 0 (the corresponding learning batch) and the probability density function Pr(Z) obtained from the performance distribution calculation unit 12. cur The operation is performed by M+1 probability density functions Pr(Z) representing the parameters of multiple updated learning batches. new 0 ), Pr (Z) new 1 ), ..., Pr (Z) new M ).

[0065] Figure 7 This shows the relationship between the corresponding learning batch and nearby learning batches. Figure 8 The graph of the probability density function operation is shown. If the probability distribution is set as a normal distribution, the parameters of the probability density distribution can be calculated by equations (18) to (21).

[0066] Here, K j It is Z old j The learning gain, μ new j It is the updated probability density function Pr(Z) new j The average value of σ 2 new j It is the updated probability density function Pr(Z) new j The variance of ). Furthermore, since this is not the actual performance distribution for nearby learning batches, a spatial variance inflation coefficient Q is used as a simulated performance distribution. j lotThe calculation is performed by performing corrections.

[0067] Finally, the learning distribution computation unit 13 indexes the corresponding learning batch from the hierarchical classification table of the learning distribution storage unit 14 to update the probability density function Pr(Z) for the calculated updated multiple learning batches. new 0 ), Pr (Z) new 1 ), ..., Pr (Z) new M The parameters of the learning timing unit 15 are used to index the corresponding learning batch from the hierarchical classification table to update the learning timing value. This learning method in the fourth embodiment can also be called hierarchical Bayesian learning.

[0068] 9. Fifth Implementation Method (Hierarchical Bayesian Learning: Graph-Based Nearest Neighbor Selection) The fifth embodiment is also an embodiment for performing hierarchical Bayesian learning. In addition to the configuration of the learning system 100 in the fourth embodiment, the learning system 100 in the fifth embodiment also includes a learning batch graph storage unit 16. The learning distribution calculation unit 13 indexes the corresponding learning batch from the learning batch graph storage unit 16 and obtains similarity information between it and other learning batches. The learning distribution calculation unit 13, based on the similarity information, obtains the parameters of M learning batches in descending order of similarity, replacing the learning batches located near the corresponding learning batches in the hierarchical classification table of the learning distribution storage unit 14 and the learning timing storage unit 15. The calculation of the learning gain can be performed using equation (18) in the same way as in the fourth embodiment. However, the reciprocal of the similarity is used instead of Z in equation (18). old j Spatial variance expansion coefficient Q j lot .

[0069] Explanation of reference numerals in the attached figures 1...Host computer; 2...Settings Department; 3...Control Department; 10...Study Department; 11... Learning Distribution Prediction Department; 12...Performance Distribution Calculation Department; 13...Study the distributed computing department; 14...Study distributed storage; 15...Study the timed storage section; 16...Study batch chart storage department; 20...Control network; 100... learning system; 101... processor; 102...Program memory; 103...Data storage memory; 104... Communication module; 105... User Interface; 200...Mechanical elements.

Claims

1. A learning system that successively learns a predictive model for a manufacturing apparatus with different product requirements, characterized in that, have: One or more processors; and The storage device holds parameters representing the learned values ​​calculated from data acquired during manufacturing using probability distributions, as well as the timing of the last parameter update, as learning data. The one or more processors are configured as follows: Using the prediction model corrected based on the learned data, the setpoints for controlling the manufacturing apparatus are calculated. Obtain the actual manufacturing performance value corresponding to the predicted value of the prediction model. Based on one or more performance values ​​obtained, parameters representing the probability distribution of the performance values ​​are calculated, the learning data is updated, and stored in a storage device.

2. The learning system according to claim 1, characterized in that, The storage device maintains multiple sets of learning data in a hierarchical classification table, categorized according to product requirements or calculated values ​​obtained during the prediction process. The one or more processors are configured as follows: When calculating the setpoints for controlling the manufacturing apparatus using the prediction model corrected based on the learned data. Based on product requirements or calculated values ​​obtained during the prediction process, the learning data is indexed and retrieved from the hierarchical classification table. Using the prediction model corrected based on the acquired learning data, setpoints for controlling the manufacturing apparatus are calculated.

3. The learning system according to claim 1, characterized in that, The storage device maintains multiple sets of learning data in a hierarchical classification table, categorized according to product requirements or calculated values ​​obtained during the prediction process. The one or more processors are configured as follows: When calculating the setpoints for controlling the manufacturing apparatus using the prediction model corrected based on the learned data. Based on product requirements or calculated values ​​obtained during the prediction process, multiple sets of learning data are indexed and retrieved from the hierarchical classification table. Based on the acquired learning data, a probability distribution for generating new learning values ​​is used, and a prediction model corrected based on the generated probability distribution is used to calculate the setpoint for controlling the manufacturing apparatus.

4. The learning system according to claim 3, characterized in that, The storage device maintains similarity information between the elements of the hierarchical classification table. The one or more processors are configured as follows: When generating a probability distribution for new learning values ​​based on the acquired learning data, Based on the similarity information, a probability distribution for generating new learning values ​​is used, and a prediction model corrected based on the generated probability distribution is used to calculate the setpoints for controlling the manufacturing apparatus.

5. The learning system according to any one of claims 2 to 4, characterized in that, The one or more processors are configured as follows: When updating the learning data, parameters representing the probability distribution of the performance values ​​are calculated based on one or more performance values ​​obtained. Index and retrieve multiple sets of learning data from the storage device, and update the retrieved multiple sets of learning data.

6. The learning system according to claim 5, characterized in that, The storage device maintains similarity information between the elements of the hierarchical classification table. The one or more processors are configured as follows: When updating the acquired learning data, the acquired learning data is indexed and retrieved from the storage device based on the similarity information, and then updated.

7. A learning method for successively learning a predictive model for a manufacturing apparatus with different product requirements, characterized in that, include: The steps of obtaining learning data related to the parameters that represent the learning values ​​calculated from data obtained during manufacturing through probability distributions, and the timing of the last parameter update; The step of calculating the setpoints for controlling the manufacturing apparatus using a prediction model corrected based on the learning data; The steps to obtain the actual manufacturing performance value corresponding to the predicted value of the prediction model; and The step of updating the learning data is to calculate parameters representing the probability distribution of the obtained performance values ​​based on one or more performance values.

8. The learning method according to claim 7, characterized in that, The learning data is managed in a hierarchical classification table, categorized according to product requirements or calculated values ​​obtained during the prediction process. When calculating the setpoints for controlling the manufacturing apparatus using the prediction model corrected based on the learned data, the following steps are included: The steps of indexing and retrieving the learning data from the hierarchical classification table based on product requirements or calculated values ​​obtained during the prediction process; and The step of calculating the setpoints for controlling the manufacturing apparatus using a prediction model corrected based on the acquired learning data.

9. The learning method according to claim 7, characterized in that, The learning data is managed in a hierarchical classification table, categorized according to product requirements or calculated values ​​obtained during the prediction process. When calculating the setpoints for controlling the manufacturing apparatus using the prediction model corrected based on the learned data, the following steps are included: The steps of indexing and retrieving multiple learning data from the hierarchical classification table based on product requirements or calculated values ​​obtained during the prediction process; and The steps include generating a probability distribution of new learning values ​​based on the acquired learning data, and using a prediction model corrected based on the generated probability distribution to calculate the setpoints for controlling the manufacturing apparatus.

10. The learning method according to claim 9, characterized in that, include: The step of obtaining similarity information between elements in the hierarchical classification table; as well as The steps include generating a probability distribution of new learning values ​​based on the acquired learning data, generating a probability distribution of new learning values ​​based on the similarity information, and using a prediction model corrected based on the generated probability distribution to calculate the setpoint for controlling the manufacturing apparatus.

11. The learning method according to any one of claims 8 to 10, characterized in that, include: The steps include indexing and retrieving multiple sets of learning data when updating the learning data, and updating the retrieved sets of learning data.

12. The learning method according to claim 11, characterized in that, include: The step of obtaining similarity information between elements in the hierarchical classification table; as well as The steps for updating the learning data include indexing based on the similarity information and obtaining multiple sets of learning data, and then updating the multiple sets of learning data obtained.

Images