Learning system and learning method for prediction model

Abstract

The present disclosure relates to a learning system that sequentially learns a prediction model for a manufacturing device which has different product requirements. This learning system comprises: at least one processor; and a storage device that retains, as learning data, a parameter which uses a probability distribution to express a learning value calculated on the basis of data acquired during manufacturing and the timing at which the parameter was updated last. The at least one processor uses a prediction model corrected on the basis of the learning data to calculate a set value for controlling a manufacturing device, acquires a manufacturing performance value corresponding to a prediction value of the prediction model, computes a parameter expressing a performance value probability distribution on the basis of at least one acquired performance value, updates the learning data, and stores the learning data in a storage device.

Description

Prediction model learning system and learning method

[0001] The present disclosure relates to a learning system and method for incrementally learning a predictive model.

[0002] In manufacturing processes for products with different product requirements, such as rolling processes, a setting device calculates various setting values ​​(amount of cooling water injected, roll gap of the rolling mill, etc.) for each device in the manufacturing process to realize the various product requirements (steel type, dimensions, etc.) received from a host computer.

[0003] When the manufacturing process is a rolling process, the setting device uses a rolling model that predicts rolling results based on rolling conditions and calculates appropriate set values ​​that can predict favorable rolling results. Therefore, maintaining and improving the prediction accuracy of the rolling model is essential for achieving stable operation and high quality requirements. However, in the steel industry, there are a huge number of combinations of product requirements and a wide variety of equipment and operations, making it unrealistic to create a rolling model that covers all situations and manually manage its accuracy. Therefore, a commonly used method is to compare actual data from past rolling operations with calculation results using the rolling model and automatically calculate a learning value to maintain the accuracy of the rolling model based on the comparison results. For example, after rolling, an index value (e.g., ratio or difference) representing the error between the actual value and the value calculated using the rolling model is calculated, and the index value is stored as a learning value. Then, when calculating set values ​​before rolling, the learning value is reflected in the value calculated using the rolling model to calculate set values ​​that correct the influence of the error. The learned values ​​are stored in a stratified table classified by various product requirements (e.g., steel grade, plate thickness, plate width, etc.) and are updated every time rolling is performed. A typical method for updating the learned values ​​is to use the exponential smoothing method.

[0004] Patent Document 1 proposes a method for adjusting a learning gain by separating an error component that depends on product requirements from a component that depends on time changes, based on the number of times of learning and history information of learned values, or changes in operation patterns, etc. On the other hand, Patent Document 2 proposes a method for evaluating the recency, saturation, and stability of not only the cell in question but also multiple learned values ​​based on history information of learned values, and correcting a learned value used in rolling by polynomial interpolation from multiple learned values ​​that have good evaluations.

[0005] Patent No. 4543684 Patent No. 6233423

[0006] When managing learning values ​​in a stratification table divided into many cells, it is desirable to divide the stratification table into smaller cells in order to accurately learn errors that vary depending on various conditions. This is because, since learning values ​​are discrete table values, the finer the divisions, the smaller the discretization reduces the discretization-induced deviations in learning values ​​between cells. However, because different rolling conditions are assigned to each cell, the finer the divisions, the fewer learning opportunities there are per cell, resulting in slower learning. One way to speed up learning is to increase the smoothing coefficient used in learning, i.e., the learning gain. However, if the observed actual values ​​contain a lot of noise, an inordinate increase in the learning gain can result in a decrease in the accuracy of the learned values ​​due to the confusion of errors (errors between actual values ​​and values ​​calculated using a rolling model) with noise. Furthermore, there are errors that do not depend on product requirements but depend on time. Examples include changes in cooling water temperature and the operational patterns of upstream processes. Learning without distinguishing between errors dependent on product requirements and errors dependent on time cannot obtain highly accurate learned values.

[0007] In the method described in Patent Document 1, the learning gain is calculated based on the number of learning times and historical information on the learning values. However, when the observed actual values ​​contain a lot of noise, it is not possible to separate the noise from the changes in the learning values ​​that depend on the product requirements in the historical information. This results in confusion between components that depend on time changes and components that depend on the product requirements. Furthermore, because the convergence status of the learning values ​​changes depending on the noise level, it is difficult to say that it is appropriate to determine the accuracy of the learning values ​​from the number of learning times.

[0008] In the method described in Patent Document 2, three evaluation indices, i.e., recency, saturation, and stability, are calculated and stored for each cell from the history information of the learned values. This requires a large amount of calculation, and the storage capacity required to store the history information and the three evaluation indices for each cell is also large. Furthermore, since the correction of the learned values ​​used in rolling is performed by polynomial interpolation, the variables of each cell classified in the stratification table are basically quantitative variables, and at least the qualitative variables must be ordered. However, a typical stratification table is classified by product type (e.g., steel type information such as low-carbon steel, silicon steel, and high-manganese steel). Therefore, polynomial interpolation performed according to the classification of unordered qualitative variables is hardly appropriate.

[0009] The present invention has been made in consideration of the above-mentioned problems, and aims to provide a predictive model learning system and learning method that can achieve stable operation and meet high quality requirements even when there is a lot of noise in the actual values, when there are components that depend on time changes, or when learning maturity is slow.

[0010] A first aspect of the present disclosure relates to a learning system that sequentially learns a prediction model for manufacturing equipment having different product requirements. The learning system includes one or more processors and a storage device that stores, as learning data, parameters that represent learning values ​​calculated based on data acquired during manufacturing as probability distributions and the timing of the last parameter update. The one or more processors calculate setting values ​​for controlling the manufacturing equipment using the prediction model corrected based on the learning data, acquire actual production values ​​corresponding to the predicted values ​​of the prediction model, calculate parameters that represent the probability distribution of the actual values ​​based on the acquired one or more actual values, update the learning data, and store the updated learning data in the storage device.

[0011] A second aspect of the present disclosure relates to a learning method for sequentially learning a prediction model for manufacturing equipment having different product requirements, the learning method including: acquiring parameters representing learning values ​​calculated based on data acquired during manufacturing as probability distributions and learning data regarding the timing of the last parameter update; calculating setting values ​​for controlling the manufacturing equipment using the prediction model corrected based on the learning data; acquiring actual production values ​​corresponding to the predicted values ​​of the prediction model; and calculating parameters representing the probability distribution of the actual values ​​based on the acquired one or more actual values, thereby updating the learning data.

[0012] According to the present invention, in the pre-setting of manufacturing equipment with different product requirements, the learning values ​​calculated based on data acquired during manufacturing are reinterpreted as random variables. Even when there is a large amount of noise in the actual results, when there are components that depend on time changes, or when learning is slow to mature, the errors in the target prediction model can be accurately corrected, thereby realizing stable operation and meeting high quality requirements.

[0013] FIG. 1 is a block diagram showing an example configuration of a learning system according to an embodiment of the present disclosure. FIG. 2 is a block diagram showing an example configuration of functions possessed by the learning system according to the first embodiment of the present disclosure. FIG. 3 is a diagram showing an operation for updating the probability distribution of learning values ​​with the probability distribution of actual values. FIG. 4 is a diagram showing an operation for ensembling the probability distribution of learning values ​​used in presetting from the probability distributions of multiple learning values. FIG. 5 is a block diagram showing an example configuration of functions possessed by learning systems according to the third and fifth embodiments of the present disclosure. FIG. 6 is a diagram showing an operation for updating the probability distribution of multiple learning values ​​with the probability distribution of actual values. FIG. 7 is a diagram showing an operation for updating the probability distribution of multiple learning values ​​with the probability distribution of actual values.

[0014] Embodiments of the present disclosure will be described with reference to the accompanying drawings.

[0015] 1. Overview The present disclosure relates to a learning system that corrects errors in a target prediction model based on the results of batch production (e.g., in an order such as 5 units of product group A, 4 units of product group B, 11 units of product group C, and 3 units of product group B) in a pre-configured manufacturing device (e.g., a rolling process) with different product requirements, and a learning method using the learning system.

[0016] 2. First Embodiment A learning system and a learning method executed by the learning system according to a first embodiment of the present disclosure will now be described with reference to the accompanying drawings.

[0017] 3. Configuration of the Learning System FIG. 1 is a block diagram showing an example of the hardware configuration of a learning system 100 according to a first embodiment of the present disclosure. The learning system 100 is a system that sequentially acquires and learns from batch-wise production results in a rolling process in which product requirements vary for each steel plate, and corrects errors in a prediction model for the next target to be produced based on the learning results. In the following, this embodiment will be described taking as an example a case in which the learning system 100 is applied to a rolling process. However, the manufacturing process to which the learning system 100 is applied is not limited to a rolling process, and may be any manufacturing process for a product in which product requirements vary for each batch.

[0018] The learning system 100 controls one or more mechanical elements 200 included in the manufacturing equipment. The learning system 100 is connected to the one or more mechanical elements 200 via a control network 20.

[0019] 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 may be a CPU, RISC, DSP, FPGA, ASIC, PLD, or another processing unit, or may be a combination of two or more of these, or may be a dedicated processor for the learning system 100. Although the present embodiment has one processor 101, the learning system 100 may also have multiple processors 101.

[0020] The program memory 102 is communicatively coupled to the processor 101. The program memory 102 stores a program made up of a plurality of instructions INST that can be executed by the processor 101. The program made up of the instructions INST can be acquired using a computer-readable non-transitory storage medium or via a network. The program memory 102 may be built into the processor 101.

[0021] 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 for correcting the setting values ​​of the manufacturing equipment. The learning data DATA is calculated based on data acquired during past manufacturing and is updated each time manufacturing is repeated. Note that the data storage memory 103 may be built into the processor 101.

[0022] A communications module 104 is communicatively coupled to the processor 101. The communications module 104 is provided for communication with external devices, including a host computer 1 that determines product requirements.

[0023] 4. Operation of the Learning System Figure 2 shows an example of the functional configuration of the learning system 100. The learning system 100 includes a setting unit 2, a control unit 3, and a learning unit 10. These functional units are realized by reading instructions INST from a program memory 102 and executing them in a processor 101. These functional units may each be configured by independent hardware, or one piece of hardware may have multiple functions.

[0024] The processing in each functional unit of the learning system 100 will be explained along with a series of operations for each steel plate. The series of operations consists of a setting step, a control step, and a learning step. The setting step is performed before the leading edge of the steel plate reaches the rolling process. The control step is performed from the start of rolling the leading edge of the steel plate to the end of rolling the tail end of the steel plate. The learning step is performed after the measurement values ​​required for learning have been obtained. The timing of the learning step may partially overlap with the control step.

[0025] In the setting step, the setting unit 2 calculates various setting values ​​for the rolling process equipment based on product requirements (e.g., steel type, plate thickness, etc.) provided by the host computer 1. A prediction model is used to calculate the various setting values. In this process, the setting unit 2 acquires learning values ​​for correcting the calculation results of the prediction model from the learning unit 10 using past similar product requirements and calculated values ​​(e.g., set speed, set temperature, etc.) obtained in the calculation process using the prediction model as search keys. The learning values ​​are numerical values ​​for correcting errors contained in the prediction model, i.e., errors occurring between predicted values ​​and actual values, and are used in the calculation of the various setting values ​​by the setting unit 2. The setting unit 2 then outputs the calculated various setting values ​​to the control devices of the corresponding manufacturing devices.

[0026] A prediction model is expressed, for example, by a function such as that shown in equation (1). The function may be a multi-input function or may include a nonlinear element. The function may also include calculations such as a neural network. An example of correction using a learning value is additive correction such as that shown in equation (2). Note that a prediction model is a collection of multiple prediction models, and the predicted value of one prediction model is sequentially used as the input for another prediction model. In the example shown in (3), the predicted value of prediction model_1 is used as the input for prediction model_2.

[0027]

[0028] Here, X １ and X ２ are input vectors for the forecast model_1 and the forecast model_2, respectively. The input vectors are parts of the product requirements provided by the host computer 1 that are necessary for calculating the functions of the forecast models, and may include table values ​​indexed based on parts of the product requirements. １ ｐｒｄ and Y ２ ｐｒｄ are the predicted values ​​of the prediction model_1 and the prediction model_2, respectively. １ ｓｅｔ is the setting value of the prediction model_1. １ and f ２ are functions of the prediction model_1 and prediction model_2, respectively.１ is a learning value that corrects the prediction model_1.

[0029] In the learning step, if an actual value corresponding to the input of the prediction model is recorded in the control unit 3, the setting unit 2 performs a new prediction calculation based on the most recent actual value, as shown in equation (4). Hereinafter, a value calculated by performing a new prediction calculation based on the most recent actual measurement value will be referred to as a recalculated actual value. Note that the recalculated actual value is not limited to an actual recalculated actual value calculated directly using an actual value. For example, an actual value calculated indirectly by performing a new prediction calculation using prediction model_2 based on an actual recalculated actual value obtained by prediction model_1 is also included in the recalculated actual value.

[0030]

[0031] In formula (4), X ２ ａｃｔ is the input vector of the forecast model_2 based on actual values, and Y １ ａｃｔ is the actual value of the prediction model_1, and Y ２ ｒｐｒｄ is the actual recalculated value of forecast model_2.

[0032] In the setting step, the control unit 3 acquires the setting values ​​from the setting unit 2. Then, in the control step, the control unit 3 controls the manufacturing equipment from moment to moment based on various measurement values ​​and command values ​​from the operator, and records the various measurement values ​​from moment to moment as actual values.

[0033] The learning unit 10 stores a learning distribution, which is a probability distribution of 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 in the calculation process using the prediction model. The learning unit 10 predicts the current learning distribution based on the stored learning distribution for the provided information, and outputs the predicted learning distribution to the setting unit 2. Then, in the learning step, the learning unit 10 acquires actual results values ​​from the control unit 3, and updates the stored learning distribution and its learning timing based on the acquired actual results values, the learning distribution predicted in the setting step, and the stored learning distribution and its learning timing. The learning unit 10 repeats the above process each time a steel plate is manufactured, updates the learning distribution based on the acquired actual results values, and sequentially corrects the prediction model.

[0034] 5. Learning Method 5-1. Functions of the Learning Unit 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 of these components that make up the learning unit 10 will be described below.

[0035] 4-2. Storage of Learning Information The learning distribution storage unit 14 and the learning timing storage unit 15 manage information by classifying calculated values ​​in the process of product requirements and prediction models based on specified conditions. Here, the explanation is based on the assumption that there are three conditions. The three conditions are condition A, condition B, and condition C, and the number of categories for condition A is NA, the number of categories for condition B is NB, and the number of categories for condition C is NC. In this case, management is performed using a stratification table having NA x NB x NC cells (hereinafter, the cells are referred to as learning lots). A specific learning lot is a set of values ​​(which do not have to be numerical values) corresponding to condition A, condition B, and condition C. Note that the number of conditions may be arbitrary, and the stratification table becomes a tensor with a rank corresponding to the number of conditions.

[0036] The learning distribution storage unit 14 stores a probability density function Pr(Z ｏｌｄ ) parameters. The learning values ​​are not limited to continuous random variables, but may be discrete random variables. When the learning values ​​are discrete random variables, the learning distribution storage unit 14 stores the parameters of the probability mass function.

[0037] In the following, we will explain the case where the probability distribution is assumed to be a normal distribution. However, the probability distribution is not limited to a normal distribution and can be extended to a general probability distribution. When the probability distribution is a general probability distribution, the calculation can be performed using, for example, the Markov chain Monte Carlo method.

[0038] The learning distribution storage unit 14 stores the average value μ ｏｌｄ Stratification table and variance of learning values ​​σ ｏｌｄ ２ The probability density function is expressed by equation (5).

[0039]

[0040] The learning timing storage unit 15 has a stratified table that stores the previous learning timing. The learning timing is a variable (hereinafter referred to as the learning timing value) that belongs to a space in which the distance between the current value and the previous value can be defined by a function (distance function). The learning timing value may be, for example, an integer that is incremented each time manufacturing is performed or the time of manufacturing.

[0041] 5-3. Calculation of learning values ​​for correction Calculation of learning values ​​is executed in the setting step for each rolling. The learning distribution prediction unit 11 searches for the corresponding learning lot from the stratification table in the learning distribution storage unit 14 based on the product requirements and calculation values ​​in the process of the prediction model given by the setting unit 2, and calculates the probability density function Pr(Z ｏｌｄ ) and calculate the average value or mode value for z ｓｅｔ In the case of a normal distribution, the mean value or the mode value coincides with the mean value for expressing the probability density function, so the mean value μ ｏｌｄ It is sufficient to look up only the stratified table for the output predicted value z ｓｅｔ is the learning value z １ is used in the calculation.

[0042] 5-4. Calculation of Current Learning Value The calculation of the current learning value is performed for each learning step. The actual results distribution calculation unit 12 obtains the actual results from the control unit 3 and the recalculated actual results from the setting unit 2, and calculates the current learning value based on these. For example, in the case of an additive correction learning value shown in equation (2), the calculation may be performed as shown in equation (6). The calculated current learning value is stored until the learning distribution is updated.

[0043]

[0044] In formula (6), Z ｃｕｒ is the learning value this time.

[0045] 5-5. Calculation of performance distribution The calculation of performance distribution is performed in the learning step using a set of newly obtained learning values ​​when rolling of products classified into the same learning lot is performed a predetermined number of times in succession, or when the product to be rolled is switched to a product of a different learning lot.

[0046] The performance distribution calculation unit 12 calculates the probability density function Pr(Z ｃｕｒ ) is calculated. ｃｕｒ ) can be calculated based on equations (7) to (9).

[0047]

[0048] where μ ｃｕｒ is the mean value of the actual distribution, and σ ｃｕｒ ２ is the variance of the actual distribution, and σ ２ is the predefined variance, and Z ｃｕｒ [i] is the learning value included in one set obtained this time, and N is the number of learning values ​​included in one set obtained this time. The performance distribution calculation unit 12 outputs the calculated parameters to the learning distribution calculation unit 13 as a performance distribution.

[0049] 5-6. Updating the Learning Distribution The learning distribution calculation unit 13 searches for the relevant learning lot from the stratification tables in the learning distribution storage unit 14 and the learning timing storage unit 15, and updates the probability density function Pr(Zｏｌｄ ) and the learning timing, and the probability density function Pr(Z ｃｕｒ ) and the probability density function Pr(Z ｎｅｗ ) is calculated. ｎｅｗ ) can be calculated using the following equations (10) to (13).

[0050]

[0051] where K is the learning gain and μ ｎｅｗ is the updated probability density function Pr(Z ｎｅｗ ) and σ ｎｅｗ ２ is the updated probability density function Pr(Z ｎｅｗ ) is the variance of the ｔｉｍｅ is an adjustment factor to express the increase in uncertainty due to changes over time. Variance expansion factor Q ｔｉｍｅ For example, the current time t and the previous learning timing t shown in equation (13) are ｏｌｄ The function f calculated from the difference ｔｉｍｅ Alternatively, it may be a lookup table value.

[0052] 3 is a graph showing an image of the calculations according to the formulas (10) to (13). The upper graph shows the probability distribution Pr(Z ｏｌｄ ) is shown. The bottom graph shows the probability distribution Pr(Z ｏｌｄ ) and the probability distribution of the actual value Pr(Z ｃｕｒ ), and the updated probability density function Pr(Z ｎｅｗ ) is shown.

[0053] Finally, the learning distribution calculation unit 13 calculates the probability density function Pr(Z ｎｅｗ ) is updated by searching for the relevant learning lot in the stratification table of the learning distribution storage unit 14, and the learning timing value is updated by searching for the relevant learning lot in the stratification table of the learning timing storage unit 15.

[0054] 6. Second Embodiment (Hierarchical Bayesian Prediction) The first embodiment has been described above. Next, the learning system 100 and learning method in the second embodiment will be described. Note that in the second embodiment through the fifth embodiment described below, the hardware configuration and operation timing of the learning system 100 are the same as in the first embodiment. Furthermore, in the second embodiment, the functional configuration of the learning system 100 is also the same as in the first embodiment. Description of these common parts will be omitted.

[0055] In the second embodiment, the learning distribution prediction unit 11 not only retrieves the relevant learning lot from the stratification table in the learning distribution storage unit 14, but also acquires parameters of M learning lots located in the vicinity of the relevant learning lot. Furthermore, the learning distribution prediction unit 11 acquires, from the learning timing storage unit 15, learning timing values ​​corresponding to the M learning lots retrieved in the learning distribution storage unit 14. Note that the acquired learning lots are not limited to learning lots located in the vicinity in the stratification table, but may also be learning lots whose learning timing is close to the present. Furthermore, the learning distribution prediction unit 11 uses the ensemble weights w ｊ After calculating w ｊ It is also possible to repeat the selection process to collect training lots with large values.

[0056] The learning distribution prediction unit 11 calculates M+1 probability density functions Pr(Z ０ ｏｌｄ ), Pr(Z １ ｏｌｄ ), ... Pr(Z Ｍ ｏｌｄ ) (Z ０ ｏｌｄ is the probability density function Pr(Z ｓｅｔ ) for the average value, or Pr(Z ｓｅｔ ) is maximized. ｓｅｔ is calculated using numerical integration or the Markov Chain Monte Carlo method, and the predicted value z ｓｅｔ and outputs it to the setting unit 2.

[0057] The relationship between the training lot and neighboring training lots is shown in Figure 4, and an image of the ensemble calculation of probability density functions is shown in Figure 5. The ensemble of probability density functions can be calculated using equations (14) to (16).

[0058]

[0059] where σ ２ ｏｌｄ ｊ is the probability density function Pr(Z ｏｌｄ ｊ ) is the variance of w ｊ is Z ｏｌｄ ｊ is the ensemble weight of ｊ is the normalized ensemble weight. Also, the variance inflation factor Q ｌｏｔ ｊ is a coefficient for expanding the variance as the relationship between the training lot and the jth training distribution becomes smaller. The larger it is, the smaller the ensemble weight w ｊ Make it smaller. Q ｌｏｔ ｊ For example, the lot number in the stratification table shown in equation (17) is ｓｅｔ and the jth training lot l ｊ The function value f calculated from the distance ｌｏｔ , or a lookup table value.

[0060] The method for calculating predicted values ​​in the second embodiment may be called hierarchical Bayesian prediction.

[0061] 7. Third Embodiment (Hierarchical Bayesian Prediction: Graph-Based Neighborhood Selection) Fig. 6 is a diagram showing a learning system 100 according to a third embodiment. In the third embodiment, hierarchical Bayesian prediction is performed, similar to the second embodiment. Furthermore, in the third embodiment, in addition to the configuration of the second embodiment, the learning unit 10 includes a learning lot graph storage unit 16.

[0062] The learning lot graph storage unit 16 stores similarities between learning lots created based on the results of data analysis and the administrator's domain knowledge for learning lots under one or all conditions. For example, the results of data analysis such as variance analysis or cluster analysis between each learning lot can be used as the similarity. The learning distribution prediction unit 11 retrieves the learning lot in question from the learning lot graph storage unit 16 to obtain similarity information with other learning lots. Then, instead of learning lots located near the learning lot in question, the learning distribution prediction unit 11 retrieves parameters of M learning lots in descending order of similarity based on the similarity information from the stratification tables in the learning distribution storage unit 14 and the learning timing storage unit 15. Furthermore, when performing an ensemble calculation of the probability density function, the learning distribution prediction unit 11 calculates the similarity of the learning lot Z expressed by equation (16) ｏｌｄ ｊ The ensemble weights w ｊ Instead of the calculation of (1), similarity is used.

[0063] 8. Fourth Embodiment (Hierarchical Bayesian Learning) The fourth embodiment is similar to the first to third embodiments in that the learning distribution calculation unit 13 searches for the relevant learning from the stratified table in the learning distribution storage unit 14. In the fourth embodiment, in addition to that, the learning distribution calculation unit 13 also acquires parameters of M learning lots located in the vicinity of the relevant learning lot. Then, the learning distribution calculation unit 13 acquires, from the learning timing storage unit 15, learning timing values ​​corresponding to the M learning lots searched in the learning distribution storage unit 14.

[0064] The learning distribution calculation unit 13 calculates M+1 probability density functions Pr(Z ｏｌｄ ０ ), Pr(Z ｏｌｄ １ ),...,Pr(Z ｏｌｄ Ｍ ) (Z ｏｌｄ ０ is the learning lot) and the probability density function Pr(Z ｃｕｒ ) and M+1 probability density functions Pr(Zｎｅｗ ０ ), Pr(Z ｎｅｗ １ ),...,Pr(Z ｎｅｗ Ｍ ) is calculated.

[0065] Figure 7 shows the relationship between the training lot in question and nearby training lots, and Figure 8 shows an image of the calculation of the probability density function. If the probability distribution is a normal distribution, the parameters of the probability density distribution can be calculated using equations (18) to (21).

[0066]

[0067] Here, K ｊ is Z ｏｌｄ ｊ is the learning gain of μ ｎｅｗ ｊ is the updated probability density function Pr(Z ｎｅｗ ｊ ) and σ ２ ｎｅｗ ｊ is the updated probability density function Pr(Z ｎｅｗ ｊ ) is the variance of the actual performance distribution for the nearby learning lot. Therefore, the spatial variance expansion coefficient Q ｊ ｌｏｔ The calculation is carried out after correction.

[0068] Finally, the learning distribution calculation unit 13 calculates the probability density function Pr(Z ｎｅｗ ０ ), Pr(Z ｎｅｗ １ ),...,Pr(Z ｎｅｗ Ｍ ) is updated by looking up the relevant learning lot from the stratification table in the learning distribution storage unit 14, and the learning timing value is updated by looking up the relevant learning lot from the stratification table in the learning timing storage unit 15. Such a learning method in the fourth embodiment may be called hierarchical Bayesian learning.

[0069] 9. Fifth Embodiment (Hierarchical Bayesian Learning: Graph-Based Neighborhood Selection) The fifth embodiment is also an embodiment in which hierarchical Bayesian learning is performed. The learning system 100 in the fifth embodiment has a learning lot graph storage unit 16 in addition to the configuration of the learning system 100 in the fourth embodiment. The learning distribution calculation unit 13 searches the learning lot graph storage unit 16 for the relevant learning lot to obtain similarity information with other learning lots. Instead of the learning lot located in the vicinity of the relevant learning lot, the learning distribution calculation unit 13 obtains parameters of M learning lots in descending order of similarity based on the similarity information from the stratification tables in the learning distribution storage unit 14 and the learning timing storage unit 15. The learning gain can be calculated using equation (18) as in the fourth embodiment. However, Z in equation (18) ｏｌｄ ｊ The spatial dispersion expansion coefficient Q ｊ ｌｏｔ Instead, the inverse of the similarity is used.

[0070] REFERENCE SIGNS LIST 1 Host computer 2 Setting unit 3 Control unit 10 Learning unit 11 Learning distribution prediction unit 12 Performance distribution calculation unit 13 Learning distribution calculation unit 14 Learning distribution storage unit 15 Learning timing storage unit 16 Learning lot graph storage unit 20 Control network 100 Learning system 101 Processor 102 Program memory 103 Data storage memory 104 Communication module 105 User interface 200 Mechanical element

Claims

1. A learning system that sequentially learns a predictive model for manufacturing equipment with different product requirements, comprising one or more processors and a storage device that stores, as learning data, parameters that express learned values ​​calculated based on data acquired during manufacturing as a probability distribution and the timing of the last parameter update, wherein the one or more processors: calculate setting values ​​for controlling the manufacturing equipment using a predictive model corrected based on the learning data; acquire actual manufacturing values ​​that correspond to the predicted values ​​of the predictive model; calculate parameters that express the probability distribution of the actual values ​​based on the acquired one or more actual values; update the learning data and store it in the storage device.

2. A learning system as claimed in claim 1, wherein the storage device holds a plurality of pieces of learning data in a stratified table categorized by calculated values ​​obtained in the process of product requirements or prediction, and the one or more processors, when calculating setting values ​​for controlling manufacturing equipment using a prediction model corrected based on the learning data, retrieve the learning data from the stratified table based on the calculated values ​​obtained in the process of product requirements or prediction, and calculate setting values ​​for controlling the manufacturing equipment using a prediction model corrected based on the retrieved learning data.

3. A learning system as described in claim 1, wherein the storage device holds a plurality of pieces of learning data in a stratified table categorized by calculated values ​​obtained in the process of product requirements or prediction, and the one or more processors, when calculating setting values ​​for controlling manufacturing equipment using a prediction model corrected based on the learning data, retrieve the plurality of pieces of learning data by searching from the stratified table based on the calculated values ​​obtained in the process of product requirements or prediction, generate a probability distribution of new learning values ​​based on the retrieved plurality of pieces of learning data, and calculate setting values ​​for controlling the manufacturing equipment using a prediction model corrected based on the generated probability distribution.

4. A learning system as described in claim 3, wherein the storage device holds similarity information between elements of the stratification table, and the one or more processors, when generating a probability distribution of new learning values ​​based on the acquired multiple pieces of learning data, generate the probability distribution of new learning values ​​based on the similarity information, and calculate setting values ​​for controlling manufacturing equipment using a prediction model corrected based on the generated probability distribution.

5. A learning system according to any one of claims 2 to 4, wherein the one or more processors calculate parameters that represent the probability distribution of actual values ​​based on one or more acquired actual values, and when updating the learning data, index and retrieve multiple pieces of the learning data from the storage device, and update the retrieved multiple pieces of learning data.

6. A learning system as described in claim 5, wherein the storage device holds similarity information between elements of the stratification table, and the one or more processors, when updating the acquired plurality of learning data, index the acquired plurality of learning data from the storage device based on the similarity information, and update the acquired plurality of learning data.

7. A learning method for sequentially learning a prediction model for manufacturing equipment having different product requirements, comprising: acquiring parameters that express learning values ​​calculated based on data acquired during manufacturing as a probability distribution and learning data regarding the timing of the last parameter update; calculating setting values ​​for controlling the manufacturing equipment using a prediction model corrected based on the learning data; acquiring actual manufacturing values ​​corresponding to the predicted values ​​of the prediction model; and calculating parameters that express the probability distribution of actual values ​​based on the acquired one or more actual values, and updating the learning data.

8. A learning method according to claim 7, wherein the learning data is managed in a stratified table classified by calculated values ​​obtained in the process of product requirements or prediction, and when calculating setting values ​​for controlling manufacturing equipment using a prediction model corrected based on the learning data, the learning method comprises: indexing and acquiring the learning data from the stratified table based on the calculated values ​​obtained in the process of product requirements or prediction; and calculating setting values ​​for controlling the manufacturing equipment using the prediction model corrected based on the acquired learning data.

9. A learning method according to claim 7, wherein the learning data is managed in a stratified table classified by calculated values ​​obtained in the process of product requirements or prediction, and when calculating setting values ​​for controlling manufacturing equipment using a prediction model corrected based on the learning data, the learning method includes: indexing and acquiring a plurality of the learning data from the stratified table based on the calculated values ​​obtained in the process of product requirements or prediction; generating a probability distribution of new learning values ​​based on the acquired plurality of the learning data, and calculating setting values ​​for controlling the manufacturing equipment using a prediction model corrected based on the generated probability distribution.

10. A learning method according to claim 9, comprising: acquiring similarity information between elements of the stratification table; and, when generating a probability distribution of new learning values ​​based on the acquired plurality of learning data, generating a probability distribution of new learning values ​​based on the similarity information, and calculating setting values ​​for controlling manufacturing equipment using a prediction model corrected based on the generated probability distribution.

11. A learning method according to any one of claims 8 to 10, characterized in that when updating the learning data, it includes: indexing and acquiring a plurality of said learning data; and updating the acquired plurality of said learning data.

12. A learning method according to claim 11, comprising: acquiring similarity information between elements of the stratification table; and, when updating the learning data, indexing and acquiring a plurality of the learning data based on the similarity information, and updating the acquired plurality of learning data.

Images