Learning model generation system and learning model generation method

Abstract

The present application provides a learning model generation system and a learning model generation method, which can perform control closer to the operation of an operator. A learning model generation system includes a model generation unit that generates a learning model for controlling an operation device having a plurality of operation channels, and a training data generation unit that generates training data for the model generation unit to generate a learning model based on actual data related to the control of the operation device, the training data generation unit having a trimming processing unit that selects an operation channel in which an operation amount indicated by the actual data is less than a set value among the plurality of operation channels as a trimming target channel, and trims the operation amount of the trimming target channel based on data extracted from the actual data.

Description

Technical Field

[0001] This invention relates to a learning model generation system and a learning model generation method. Background Technology

[0002] As a technology related to the control of operating devices such as rolling mills, there is the technology described in Patent Document 1. Patent Document 1 describes the following: "It includes: a new teaching data extraction device that calculates an evaluation value based on actual equipment operation data for the results of operations performed during a predetermined extraction time width from a predetermined start time, uses the evaluation value to determine whether teaching data can be generated, and when it is determined that the teaching data can be generated, extracts teaching data including an input section of teaching data calculated based on the state quantity at the start time and an output section of teaching data calculated based on the operation quantity during the predetermined extraction time width from the start time; and a teaching data database update device that stores the teaching data extracted by the new teaching data extraction device in a database." "Therefore, control rules can be learned offline, and their performance can be verified offline. This improves the reliability of control and enables applications using real-time control with AI."

[0003] Existing technical documents

[0004] Patent Document 1: Japanese Patent Application Publication No. 2019-212146 Summary of the Invention

[0005] However, the actual equipment operation data mentioned above includes the amount of equipment operation data performed by the operator, but the operator can only operate a limited number of operation channels, such as around 10 channels. However, during equipment operation, the operator can only operate a portion of these operation channels at a time.

[0006] Therefore, actual equipment operation data sometimes includes multiple operating modes where only a portion of the operating channels were operated for the same (or similar) evaluation value. In such cases, the teaching data generated using actual equipment operation data also includes operating channels with zero operation volume, as the operation volume is averaged out. Therefore, the control of the device by the learning model generated based on this teaching data becomes a passive control with low operation volume compared to the operator's operation when the actual equipment operation data was obtained.

[0007] Therefore, the purpose of this invention is to provide a learning model generation system and a learning model generation method that can perform control operations that are closer to those of an operator.

[0008] To address the aforementioned issues, for example, a structure described in the scope of the patent claim may be adopted.

[0009] This application includes several means to solve the above-mentioned problems. One example is a learning model generation system comprising: a model generation unit that generates a learning model for controlling an operating device having multiple operating channels; and a training data generation unit that generates training data for the model generation unit to generate the learning model based on actual data related to the control of the operating device. The training data generation unit includes: a trimming processing unit that selects the operating channels among the multiple operating channels whose operating amounts shown by the actual data are less than a set value as trimming target channels, and trims the operating amounts of the trimming target channels based on data extracted from the actual data.

[0010] According to the present invention, a learning model generation system and a learning model generation method are provided that enable control that more closely resembles the operation of an operator. Attached Figure Description

[0011] Figure 1 This is a block diagram of the learning model generation system for the implementation method.

[0012] Figure 2 This is a diagram representing the learning model generated by the learning model generation system in the implementation method.

[0013] Figure 3 This is a diagram (Figure 1) showing the settings screen displayed on the settings terminal.

[0014] Figure 4 This is a diagram (Figure 2) showing the settings screen displayed on the settings terminal.

[0015] Figure 5 This is a diagram (3) showing the settings screen displayed on the settings terminal.

[0016] Figure 6 This is a flowchart illustrating the order in which training data for generating the learning model is produced in accordance with the implementation method.

[0017] Figure 7 This is a flowchart showing the order of the filtering process.

[0018] Figure 8 This is a flowchart showing the sequence of zero-trim processing.

[0019] Figure 9 This is a flowchart showing the order of elimination processes.

[0020] Figure 10 This is a flowchart showing the sequence of trimming processes.

[0021] Figure 11 It is a graph used to illustrate the training data obtained through trimming.

[0022] Figure 12 This is a graph used to illustrate training data without any trimming.

[0023] Figure 13 This is a flowchart showing the sequence of error backpropagation processes implemented during the generation of the learned model. Detailed Implementation

[0024] Hereinafter, embodiments of the learning model generation system and learning model generation method of the present invention will be described with reference to the accompanying drawings. Furthermore, in this embodiment, as an example, the structure of a learning model generation system for controlling a rolling mill will be described. However, the present invention is not limited thereto, and can be applied to learning model generation systems and learning model generation methods for controlling operating devices having multiple control channels.

[0025] Learning Model Generation System

[0026] Figure 1 This is a block diagram of the learning model generation system 1 in the implementation method. As an example, Figure 1 The learning model generation system 1 shown generates a learning model for controlling the rolling mill 10. This learning model generation system 1 includes a control device 20, an operation terminal 30 connected to the control device 20, a model generation device 40, and a setting terminal 50. Next, the structure of each element constituting the rolling mill 10 and the learning model generation system 1 will be described.

[0027] <Rolling Unit 10>

[0028] The rolling mill 10 is an operating device for rolling metal materials into sheet-like steel plates, and has multiple operating channels (operating sections) for controlling the rolling state. For example, the rolling mill 10 has 10 operating channels. Furthermore, the rolling mill 10 is equipped with multiple sensors (not shown) for detecting the shape of the rolled steel plate as it passes through each operating channel. Information from each sensor is sent to the control device 20, which will be described later, for feedback control of each operating channel.

[0029] <Control Device 20>

[0030] The control device 20 is a computer used to control the operation of each operating channel in the rolling mill 10. The computer is hardware used as a so-called computer and has a network interface capable of transmitting and receiving information with the rolling mill 10, the operating terminal 30, and the model generation device 40 via wired or wireless means. This control device 20 includes functional units for a control data management unit 21 and a device control unit 22. These functional units are as follows.

[0031] [Control Data Management Department 21]

[0032] The control data management unit 21 maintains the actual data acquired when the rolling mill 10 is operated. The actual data is a data set consisting of the operation amount of the operation channel at a certain time [t] when the rolling mill 10 is operated, information from each sensor of the rolling mill 10, and the state information of the rolling mill 10 at that time. The information from each sensor is maintained as the shape deviation of each part of the steel plate. The shape deviation is the difference relative to the shape of the target steel plate, for example, the shape deviation at position 32. This shape deviation corresponds to the result data of the operation for multiple operation channels. Furthermore, the state information of the rolling mill 10 includes the conveying speed of the steel plate in the rolling mill 10, the path number, and the position of each roll constituting the rolling mill 10. The control data management unit 21 maintains this data set as a sample, and multiple samples as actual data. Furthermore, while shape deviation is used in this embodiment, it can also be maintained as shape.

[0033] [Device Control Unit 22]

[0034] The device control unit 22 controls the operation amount of each operation channel of the rolling mill 10 based on the learning model obtained from the model generation device 40, the information from each sensor of the rolling mill 10, and the status information of the rolling mill 10.

[0035] <Operating Terminal 30>

[0036] The operation terminal 30 is a terminal device for operator P1 to input settings and instructions related to the forming of steel plates by the rolling device 10.

[0037] <Model Generation Device 40>

[0038] The model generation apparatus 40 is a computer used to generate a learning model for controlling the rolling mill 10. This computer has a network interface capable of transmitting and receiving information with the control device 20 and the setting terminal 50 via wired or wireless means. This model generation apparatus 40 includes functional units such as a training data generation unit 41, a model generation unit 42, a data management unit 43, and a terminal input / output control unit 44. These functional units are as follows.

[0039] [Training Data Generation Section 41]

[0040] The training data generation unit 41 generates training data for generating the learning model in the model generation unit 42, which will be described next. The training data generation unit 41 generates training data based on actual data related to the control of the rolling mill 10 obtained from the control data management unit 21 of the control device 20 and the settings in the setting terminal 50, which will be described later.

[0041] The training data generation unit 41 includes multiple processing units for performing various processes for generating training data based on a program stored in a computer. These processing units are a filtering processing unit 411, a zero-trimming processing unit 412, a discarding processing unit 413, a trimming processing unit 414, and a data synthesis processing unit 415. Details of the processes performed by these processing units for generating training data will be explained later in the section on learning model generation methods.

[0042] [Model Generation Section 42]

[0043] The model generation unit 42 generates a learning model for controlling the operation amount of each operation channel of the rolling mill 10 based on the training data generated by the training data generation unit 41. Figure 2 This is a diagram illustrating the learning model generated by the learning model generation system in this implementation. The learning model generated here is a machine learning model that parses the data input from the input layer 4201 in the intermediate layer 4202 and outputs it from the output layer 4203.

[0044] Input layer 4201 is the layer where actual data is input. This input layer 4201 includes inputs of shape deviations for various parts of the steel plate and the rolling device 10 (see reference). Figure 1 The rolling mill 10 has multiple input channels for status information. These input channels include 32 input channels for inputting shape deviations at 32 locations in the steel plate and 4 input channels for inputting status information of the rolling mill 10. The shape deviations input from the 32 input channels are differences relative to the shape of the target steel plate, calculated based on information from various sensors within the rolling mill 10. The status information of the rolling mill 10 input from the 4 input channels relates to the steel plate's conveying speed, path number, upper intermediate roll position, and lower intermediate roll position.

[0045] The intermediate layer 4202 is a layer with multiple hidden layers; in the attached figure, it has two hidden layers. The number of layers in the intermediate layer 4202 and the number of nodes in each layer are appropriately set considering the computation time and accuracy of the output value. The model generation unit 42 generates a learning model that optimizes the model coefficients, i.e., weights and biases, in the learning model by learning from the training data generated by the training data generation unit 41.

[0046] Output layer 4203 is a layer that outputs the results calculated based on the processing in intermediate layer 4202. This output layer 4203 has multiple output channels that output the operation amounts of each operation channel; in this case, it has 10 output channels corresponding to 10 operation channels. The operation amount is output from each output channel as a difference relative to the current operation amount. Furthermore, the feature unit of the model generation unit 42 in this embodiment that generates the learning model will be described in the following learning model generation method.

[0047] [Data Management Department 43]

[0048] return Figure 1 The data management unit 43 maintains the actual data obtained from the control data management unit 21 of the control device 20 and the data generated by the training data generation unit 41.

[0049] [Terminal Input / Output Control Unit 44]

[0050] The terminal input / output control unit 44 controls the output of the display screen shown on the setting terminal 50 and the input from the setting terminal 50.

[0051] <Terminal 50 Setup>

[0052] The setting terminal 50 has a display unit that displays various setting screens for generating training data by the training data generation unit 41, which are controlled by the terminal input / output control unit 44. Additionally, the setting terminal 50 has an input function for operator P2 to input settings for generating training data according to the setting screens displayed on the display unit.

[0053] Figures 3-5 These are figures (1) to (3) showing the settings screen displayed on the settings terminal 50. Figures 3-5 As shown, the display unit of the setting terminal 50 displays a setting screen for setting each processing unit of the training data generation unit 41 to perform each processing step. Additionally, although the illustration is omitted here, the display unit of the setting terminal 50 also displays a setting screen for the model generation unit 42 to generate the learning model. Furthermore, details of the setting screen displayed on the setting terminal 50 under the control of the terminal input / output control unit 44 will be explained in the following description of the learning model generation method.

[0054] Learning Model Generation Methods

[0055] Next, the learning model generation method of the aforementioned learning model generation system 1 will be explained. Here, firstly, the generation order of the training data will be explained as a learning model generation method, and then the generation of the learning model using the generated training data will be explained.

[0056] -Order of training data generation-

[0057] Figure 6 This is a flowchart illustrating the sequence of generating training data for a learning model according to an embodiment. It describes the sequence implemented using the procedures of each part of the training data generation unit 41 in the learning model generation system 1 described above. Hereinafter, according to... Figure 6 The order shown is based on the previous Figures 1-5And other necessary graphs to illustrate the order in which the training data was generated.

[0058] Step S1: Filtration Process Figure 6 )>

[0059] In step S1, the filtration processing unit 411 (refer to...) Figure 1 The actual data is obtained from the control data management unit 21 of the control device 20, and the actual data is filtered. In addition, the filtering processing unit 411 can also obtain the generated training data from the control data management unit 21 of the control device 20 and use it as actual data to perform the following filtering processing.

[0060] The filtering processing unit 411 performs filtering processing based on the settings in the setting terminal 50 to remove samples from the actual data whose correspondence, represented by the direction of operation direction and shape deviation, is physically inappropriate. Figure 7 This is a flowchart illustrating the order of the filtering process, as implemented below. Additionally, while samples are deleted in this embodiment, it's also possible to reduce the processing amount for inappropriate samples to zero.

[0061] [Step S101( Figure 7 )]

[0062] In step S101, the filtering processing unit 411 selects a sample from the actual data. Here, as described above, the sample is a data set containing the operating amount of each operating channel at a certain moment [t] when the rolling mill 10 is operated, the shape deviation of each part of the steel plate based on information obtained from the sensor, and the state information of the rolling mill 10.

[0063] [Step S102( Figure 7 )]

[0064] In step S102, the filtering processing unit 411 compares the operation amount (including the operation direction) of each operation channel in the selected sample with the magnitude (including the direction) of the shape deviation of each part corresponding to each operation channel, and determines whether the polarity is normal. Here, in this embodiment, normal polarity means that the direction of operation is consistent with the direction of shape change. Regarding normality, other references can also be set according to the rolling apparatus. At this time, the filtering processing unit 411 performs this determination based on the rules preset by the input from the setting terminal 50.

[0065] That is, such as Figure 3 As shown, the setting screen of the setting terminal 50 includes a filter processing setting unit 51. This filter processing setting unit 51 is configured to set the range of acceptable outputs for inputs. Here, input refers to the magnitude of the shape deviation in each channel, which is one of the input channels. Output refers to the operational quantity in each operation channel, which is an output channel. Furthermore, in Figure 3 In this paper, only the setting of the range in the filter processing setting unit 51 is shown, but it is also possible to have a structure that allows setting the range of the operating amount (output) of each operation unit for each shape deviation (input).

[0066] If all shape deviations at point 32 are within the set range, the filter processing unit 411 determines that the polarity is normal (yes) and proceeds to step S103. Otherwise, it determines that the polarity is abnormal (no) and proceeds to step S103a.

[0067] [Step S103a( Figure 7 )]

[0068] In step S103a, the filtering processing unit 411 deletes the sample selected in step S101 from the actual data and proceeds to step S104.

[0069] Step S103 Figure 7 )]

[0070] In step S103, the filtering processing unit 411 saves the sample selected in step S101 as filtered data.

[0071] [Step S104( Figure 7 )]

[0072] In step S104, the filtering processing unit 411 determines whether the processing of all samples in the actual data obtained from the control data management unit 21 of the control device 20 has ended. Then, if it is determined that the processing has ended (yes), the filtering process ends; if it is determined that the processing has not ended (no), it returns to step S101 and repeats the subsequent processing. Furthermore, in the subsequent step S101, samples other than those already selected from the actual data obtained from the control data management unit 21 of the control device 20 are selected.

[0073] Through the above filtering process, training data can be generated based solely on actual data from samples whose operating direction and shape deviation direction are inconsistent. Therefore, the learning model generated using this training data can control the rolling device 10 with high precision.

[0074] Step S2 Figure 6 )>

[0075] In step S2, the zero-trimming processing unit 412 (refer to...) Figure 1 The filtered data saved in step S1 is subjected to zero-trimming processing. At this time, the zero-trimming processing unit 412 performs zero-trimming processing based on the settings in the setting terminal 50.

[0076] Figure 8This is a flowchart showing the sequence of zero-trim processing, as implemented below.

[0077] [Step S201( Figure 8 )]

[0078] In step S201, the zero-trimming processing unit 412 selects a sample from the filtered data.

[0079] Step S202 Figure 8 )]

[0080] In step S202, the zero-trimming processing unit 412 determines whether the shape deviations of the selected sample are within a set range. At this time, the zero-trimming processing unit 412 performs this determination based on the rules preset by the operation of the setting terminal 50.

[0081] That is, such as Figure 3 As shown, the setting screen of the setting terminal 50 includes a zero-adjustment processing setting unit 52. This zero-adjustment processing setting unit 52 is configured to set the maximum value of the shape deviation. The maximum value of the shape deviation set here is the maximum value of the difference relative to the target shape, and is set to a size that is determined to be approximately consistent with the target shape.

[0082] If the size of all shape deviations in the selected sample is within the set range, the zero-adjustment processing unit 412 determines that it is within the set range (Yes) and proceeds to step S203. Otherwise, it determines that it is not within the set range (No) and proceeds to step S205.

[0083] Step S203 Figure 8 )]

[0084] In step S203, the zero trimming processing unit 412 sets the operation quantity of each operation channel, which is the output of the selected sample, to zero.

[0085] Step S204 Figure 8 )]

[0086] In step S204, the zero-trimming processing unit 412 saves the sample after zero-trimming the output in step S203 as zero-trimmed data.

[0087] [Step S205( Figure 8 )]

[0088] In step S205, the zero-trimming processing unit 412 determines whether the processing of all samples in the filtered data has ended. Then, if it determines that the processing has ended (yes), the zero-trimming process is terminated; if it determines that the processing has not ended (no), it returns to step S201 and repeats the subsequent processing. Furthermore, in the subsequent step S201, samples other than those already selected from the filtered data are selected.

[0089] By performing the zero-trimming process described above, the operational quantities of each operation channel in samples with small shape deviations are set to zero. As a result, in the model generation unit, a learning model can be generated by learning that the operational quantities of the operation channels of samples with small shape deviations are zero.

[0090] Step S3 Figure 6 )>

[0091] In step S3, the elimination processing unit 413 (refer to) Figure 1 The filtered data saved in step S1 is discarded. At this time, the discarding processing unit 413 performs the discarding processing according to the settings in the setting terminal 50. Figure 9 This is a flowchart illustrating the steps involved in the elimination process, as follows.

[0092] [Step S301( Figure 9 )]

[0093] In step S301, the elimination processing unit 413 selects a sample from the filtered data.

[0094] Step S302 Figure 9 )]

[0095] In step S302, the rejection processing unit 413 determines whether the shape improvement degree of the selected sample is good. At this time, the rejection processing unit 413 performs this determination according to the rules preset by the input from the setting terminal 50.

[0096] That is, such as Figure 3 As shown, the setting screen of the setting terminal 50 includes a rejection processing setting unit 53. In this rejection processing setting unit 53, a judgment criterion for shape improvement is set. The judgment criterion for shape improvement set here is, for example, the average value of the difference between the shape deviation at the time point when each operation channel in the selected sample is operated and the shape deviation at the time point when the sensor measurement is performed.

[0097] If the difference in all shape deviations among the selected samples exceeds a set minimum value, the elimination processing unit 413 determines that the shape improvement is good (Yes) and proceeds to step S303. Otherwise, it determines that the shape improvement is bad (No) and proceeds to step S303a.

[0098] Step S303 Figure 9 )]

[0099] In step S303, the elimination processing unit 413 determines whether the selected sample has shape diversity. At this time, the elimination processing unit 413 performs this determination according to the rules preset by the input from the setting terminal 50.

[0100] That is, such as Figure 3 As shown, in the elimination processing setting unit 53 of the setting screen of the setting terminal 50, a judgment criterion for shape diversity is set. The judgment criterion for shape diversity set here is, for example, the minimum value of the distance between the shape deviation and other samples, and the sample that is judged to have a large shape deviation distance is set to have a shape diversity value.

[0101] If the shape deviation distance from other samples exceeds a set minimum value, the elimination processing unit 413 determines that shape diversity exists (Yes) and proceeds to step S304. Otherwise, it determines that there is no shape diversity (No) and proceeds to step S303a. Here, "other samples" refers to other samples within the filtered data.

[0102] [Step S303a( Figure 9 )]

[0103] In step S303a, the elimination processing unit 413 deletes the sample selected in step S301 from the filtered data and proceeds to step S305.

[0104] Step S304 Figure 9 )]

[0105] In step S304, the elimination processing unit 413 saves the samples that have good shape improvement in step S303 and are judged to have shape diversity in step S303 as eliminated data.

[0106] [Step S305( Figure 9 )]

[0107] In step S305, the elimination processing unit 413 determines whether the processing of all samples in the filtered data has ended. If it determines that the process has ended (yes), the elimination process ends; if it determines that the process has not ended (no), it returns to step S301 and repeats the subsequent processing. Furthermore, in the subsequent step S301, samples other than those already selected from the filtered data are selected.

[0108] Through the above elimination process, high-value actual data can be retained due to the large deviation and diversity of the operational shape. This enables the generation of efficient learning models that reduce the number of learning samples in the model generation department.

[0109] Step S4 Figure 6 )>

[0110] In step S4, the trimming processing unit 414 (refer to...) Figure 1 The obsolete data saved in step S3 is then modified. At this time, the modification processing unit 414 performs modification processing based on the settings in the setting terminal 50. Figure 10 This is a flowchart showing the sequence of trimming processes, as implemented below.

[0111] [Step S401( Figure 10 )]

[0112] In step S401, the trimming processing unit 414 selects a sample from the obsolete data.

[0113] Step S402 Figure 10 )]

[0114] In step S402, the trimming processing unit 414 determines whether the selected sample is a trimming target sample. At this time, the trimming processing unit 414 performs this determination based on a rule preset by input from the setting terminal 50.

[0115] That is, such as Figure 4 As shown, the setting screen of the setting terminal 50 includes a trimming processing setting unit 54 and an object sample setting unit 541 for determining whether the selected sample is an object sample to be trimmed. This object sample setting unit 541 includes a setting unit 541a for determining whether to set an evaluation threshold [θL] for the object sample, configured to set the maximum value of the shape deviation. This evaluation threshold [θL] is, for example, set to a value that indicates the shape deviation is to some extent acceptable.

[0116] If the shape deviations of all selected samples are less than the maximum value of the set shape deviation, the trimming processing unit 414 determines that it is a target sample (Yes) and proceeds to step S403 to perform subsequent trimming processing. Otherwise, it determines that it is not a target sample (No) and proceeds to step S411.

[0117] Step S403 Figure 10 )]

[0118] In step S403, the trimming processing unit 414 selects an operation channel from the trimming object sample.

[0119] like Figure 4As shown, the trimming setting unit 54 in the setting screen of the setting terminal 50 has an operation channel setting unit 541b. The operation channel setting unit 541b displays 10 operation channels (IMR-Shift1, IMR-Shift2, ...), and by inputting a check mark, the operation channel to be studied as the trimming target can be selected.

[0120] The trimming and processing unit 414 further selects an operation channel as the target from the operation channels selected by the operation channel setting unit 541b.

[0121] Step S404 Figure 10 )]

[0122] In step S404, the trimming processing unit 414 determines whether the operation channel selected in step S403 has been operated. At this time, the trimming processing unit 414 performs this determination based on a rule preset by input from the setting terminal 50.

[0123] like Figure 4 As shown, the trimming setting unit 54 in the setting screen of the setting terminal 50 has an operation range setting unit 541c for the operation channels. This operation range setting unit 541c is configured to set a maximum value for determining whether each operation channel is operated. This maximum value is, for example, set to a value so small that the operation amount can be considered zero.

[0124] If the operation amount of the selected operation channel in step S403 is greater than or equal to the maximum set operation amount, the trimming processing unit 414 determines that there is an operation (Yes) and proceeds to step S409. On the other hand, if the operation amount of the selected operation channel in step S403 is less than the maximum set operation amount, it determines that there is no operation (No) and proceeds to step S405. Here, the operation channel that is determined to have no operation (No) becomes the trimming target channel selected as the trimming target.

[0125] Step S405 Figure 10 )]

[0126] In step S405, the trimming processing unit 414 extracts trimming samples from obsolete data other than the trimming target samples according to the settings.

[0127] like Figure 4As shown, the trimming processing setting unit 54 of the setting screen of the setting terminal 50 includes a trimming sample setting unit 542. This trimming sample setting unit 542 is configured to set search conditions for trimming samples for each operation channel to be trimmed. This trimming sample setting unit 542 includes an operation channel setting unit 542a for setting the operation channel that is the target of the setting. Furthermore, the trimming sample setting unit 542 includes a minimum operation amount setting unit 542b, a shape similarity setting unit 542c, a trimming implementation necessity number setting unit 542d, a maximum number setting unit 542e, and a shape improvement degree setting unit 542f.

[0128] The operation channel setting unit 542a is configured to select an operation channel as the setting target from 10 operation channels.

[0129] The minimum operation amount setting unit 542b sets a minimum operation amount [θH] for the operation channel selected in step S403 so as to select a sample with an operation amount of a certain magnitude as a sample for trimming.

[0130] Furthermore, the shape similarity setting unit 542c sets a shape similarity [θd] for the sample selected in step S401, so as to select a sample with a certain degree of similarity in shape as a sample for trimming. Here, the shape similarity [θd] is set to the maximum value of the difference between the shape deviation and the sample selected in step S401.

[0131] Furthermore, the trimming implementation necessity setting unit 542d sets the necessary number [K] of trimming samples so that trimming is performed only when the number of trimming samples extracted through retrieval reaches a certain level. In addition, the trimming implementation necessity setting unit 542d is configured to perform checks when the trimming implementation necessity is applied during the extraction of trimming samples.

[0132] The maximum number setting unit 542e sets the maximum value [k] of the trimming samples so as to reduce the number of trimming samples when the number of samples retrieved by retrieval exceeds a certain level. When the number of trimming samples retrieved by retrieval exceeds the maximum value [k] set by the maximum number setting unit 542e, the trimming processing unit 414 sequentially excludes samples from the trimming samples starting from those with large shape deviations relative to the trimming target sample.

[0133] The shape improvement setting unit 542f is configured as a judgment criterion for setting the shape improvement degree.

[0134] The criteria for judging shape improvement set here can be the same as those for judging shape improvement during the rejection process. However, the value of the criteria can be different from the value during the rejection process. The shape improvement setting unit 542f is configured to perform the check when the number of necessary adjustments is applied during the extraction of the adjustment sample.

[0135] Furthermore, if the training data generation here is based on processing actual data that does not include the already generated training data, then shape improvement must be applied in step S405. Alternatively, if the training data generation here includes the already generated training data as actual data (where the actual data obtained in step S1 includes the training data), then the application of shape improvement in step S405 is arbitrary.

[0136] In step S405, the trimming processing unit 414 extracts multiple trimming samples from the actual data based on the settings in the trimming processing setting unit 54 as described above.

[0137] Step S406 Figure 10 )]

[0138] In step S406, the trimming processing unit 414 determines whether the number of trimming samples extracted through the retrieval in step S405 is equal to the number set by the trimming implementation necessary number setting unit 542d (see reference). Figure 4 If the required number of repairs to be performed is set to be 4K or higher, the repair processing unit 414 proceeds to step S407 if it determines that the required number of repairs to be performed is 4K or higher (Yes), and proceeds to step S409 if it determines that the required number of repairs to be performed is not 4K or higher (No).

[0139] Step S407 Figure 10 )]

[0140] In step S407, the trimming processing unit 414 calculates the trimming amount for the trimming target channel that was determined to have no operation (no) in step S404. The trimming amount calculated here is the trimming amount of the operation amount. At this time, the trimming processing unit 414 calculates the trimming amount based on the operation amount of the operation channel corresponding to the trimming target channel from all the trimming samples extracted through the retrieval in step S405. Regarding the trimming amount, typically it can be calculated as the average of the operation amounts of the corresponding operation channels in all the extracted trimming samples, but it is not limited to this. For example, it can also be a constant multiple of the average, a percentile value, etc., and several or all of the average, a constant multiple of the average, and a percentile value can be combined to calculate the trimming amount.

[0141] Step S408 Figure 10 )]

[0142] In step S408, the trimming processing unit 414 trims the operation amount of the trimming target channel in the trimming target sample based on the trimming amount calculated in step S407. In this case, the operation amount of the trimming target channel is simply added to the trimming amount calculated in step S407.

[0143] Step S409 Figure 10 )]

[0144] In step S409, for all operation channels in the sample to be repaired, it is determined whether the repair is complete. The repair processing unit 414, for all operation channels in the sample to be repaired, determines whether the repair is complete. Figure 4 If the operation channel setting unit 541b shows that all operation channels with the check mark have finished their adjustments, it determines that the process is complete (Yes) and proceeds to step S410. Otherwise, it returns to step S403 and repeats the subsequent processing. Furthermore, in the subsequent step S403, an operation channel other than the previously selected operation channel is selected.

[0145] Step S410 Figure 10 )]

[0146] In step S410, the trimming processing unit 414 saves the trimmed object sample, which has trimmed the operation amount of all trimmed object channels, as trimmed data.

[0147] [Step S411( Figure 10 )]

[0148] In step S411, the trimming processing unit 414 determines whether the processing of all samples in the eliminated data has ended. Then, if the determination is "complete" (yes), the trimming process ends; if the determination is "not complete" (no), it returns to step S401 and repeats the subsequent processing. Furthermore, in the subsequent step S401, samples from the eliminated data other than those already selected are selected.

[0149] Furthermore, in the repair processes described above, when using... Figure 4 The various settings described for trimming use all the shape deviation data for each of the 32 channels, but it is also possible to select the shape deviation used for trimming settings individually for each operating channel. In this case, such as Figure 5 As shown, the trimming processing setting unit 54 of the setting terminal 50 has a shape deviation setting unit 543 for trimming. In this setting unit 543, the operation channel and the shape deviation channel are selected, and for each selected combination, it is possible to set whether to use trimming. For example, input [1] can be selected when trimming is used, and input [0] can be selected when not using trimming.

[0150] Step S5 Figure 6 )>

[0151] In step S5, the data processing unit 415 (refer to...) Figure 1 ) for step S204 ( Figure 8 The zero-trimmed data saved in step S410 and the zero-trimmed data saved in step S410 Figure 10 The data processing unit 415 integrates the trimmed data stored in the data management unit 43. The integrated data is then stored as training data in the data management unit 43, concluding the training data generation unit 41 (see reference 415). Figure 1 The order in which the training data is generated.

[0152] Figure 11 This is a diagram used to illustrate the training data obtained through the above-described trimming process.

[0153] In the order of training data generation described above, training data (2) is generated that assigns an operation to only the operation channels (channels 3 and 4 in this case) where the operation amount of the actual data (1) is zero.

[0154] Figure 12 This is a graph used to illustrate the training data without the aforementioned trimming processes. For example... Figure 12 As shown, without performing any trimming, operational training data (2) is generated by averaging the operational quantities of actual data A(1) and actual data B(1)' with similar shape deviations. Therefore, in the learning process using the training data generated in this way, a learning model of negative control with fewer operational quantities compared to the operator's operations is generated.

[0155] -Steps for generating a learning model-

[0156] Next, the generation order of the learning models using the training data generated as described above will be explained. The generation order of the learning models described here is achieved through… Figure 1 The learning model generation system 1 shown in the diagram implements the backpropagation of errors in the sequence of procedures of its model generation unit 42 when using training data for machine learning. Figure 13 This is a flowchart illustrating the sequence of error backpropagation processes implemented during the generation of the learned model, as follows.

[0157] Step S21 Figure 13 )]

[0158] In step S21, the model generation unit 42 selects an operation channel.

[0159] Step S22 Figure 13 )]

[0160] In step S22, the model generation unit 42 determines whether the selected operation channel is operational. This determination is performed according to a preset determination criterion. This determination can be performed in the same way as step S404 described in the trimming process, and the determination criterion is set by input from the model setting screen (not shown) displayed on the setting terminal 50.

[0161] If the operation amount of the operation channel selected in step S21 exceeds the maximum set operation amount, the model generation unit 42 determines that there is an operation (yes) and proceeds to step S23.

[0162] On the other hand, if the operation quantity of the operation channel selected in step S21 is below the maximum value of the set operation quantity, it is determined that there is no operation (no) and the process proceeds to step S24.

[0163] [Step S23]

[0164] In step S23, the model generation unit 42 performs backpropagation processing on the operation quantity of the operation channel selected in step S21.

[0165] [Step S24]

[0166] In step S24, the model generation unit 42 determines whether processing has been completed for all operation channels in the sample to be learned. If it determines that processing has been completed for all operation channels (yes), the error backpropagation process for that sample ends. Otherwise, it returns to step S21 and repeats the subsequent processing. Furthermore, in the subsequent step S21, operation channels other than those already selected are selected.

[0167] The model generation unit 42 performs learning, including the above error backpropagation processing, on all samples contained in the training data to generate a learning model that optimizes the model coefficients, i.e., weights and biases, in the learning model.

[0168] The above error backpropagation process can improve the prediction accuracy of the learning model generated by the model generation unit 42.

[0169] Effects of the Implementation Method

[0170] According to the implementation described above, training data can be generated that assigns operational quantities to operation channels that are zero in the samples constituting the actual data, but based on other samples. Therefore, compared to averaging the operational quantities of samples with similar shape deviations, the operational quantities are larger, and training data with operational quantities that correspond to the operator's operations shown in the actual data can be generated. As a result, by learning from this training data, a learning model for control that more closely approximates the operator's actions can be generated.

[0171] Furthermore, the present invention is not limited to the embodiments and modifications described above, but also includes various modifications. For example, the embodiments described above are embodiments that have been explained in detail for the purpose of easily understanding the present invention, and are not limited to having all the structures described. In addition, a part of the structure of a certain embodiment can be replaced with the structure of another embodiment, and it is also possible to add the structure of another embodiment to the structure of a certain embodiment.

[0172] Furthermore, regarding a portion of the structure of each embodiment, other structures can be added, deleted, or replaced.

[0173] Explanation of reference numerals in the attached figures

[0174] 1 Learning Model Generation System

[0175] 10. Rolling Equipment (Operating Device)

[0176] 20 control devices

[0177] 21 Control Data Management Department

[0178] 22 Device Control Section

[0179] 42 Model Generation Department

[0180] 41 Training Data Generation Department

[0181] 411 Filtration Treatment Department

[0182] 412 Repair and Maintenance Department

[0183] 413 Elimination and Disposal Department

[0184] 414 Repair and Treatment Department

[0185] 415 Data Processing Department

[0186] 50. Terminal settings.

Claims

1. A learning model generation system, characterized in that, The learning model generation system has the following features: A model generation unit that generates a learning model for controlling an operating device with multiple operating channels; and The training data generation unit generates training data for the model generation unit to generate a learning model, based on actual data related to the control of the operating device. The training data generation unit includes a trimming processing unit, which selects the operation channel whose operation amount shown by the actual data is less than a set value from the plurality of operation channels as the trimming target channel, and trims the operation amount of the trimming target channel based on the data extracted from the actual data.

2. The learning model generation system according to claim 1, characterized in that, The actual data is composed of multiple samples consisting of data sets, each containing the operational quantities of the multiple operation channels and the result data of the operations performed on the multiple operation channels. The trimming processing unit performs the trimming processing on samples from the plurality of samples where the difference between the result data and the target value for the result data is less than a set value.

3. The learning model generation system according to claim 1, characterized in that, The actual data is composed of multiple samples consisting of data sets, each containing the operational quantities of the multiple operation channels and the result data of the operations performed on the multiple operation channels. The trimming processing unit extracts multiple samples from the multiple samples that have result data whose difference from the result data of the sample containing the trimming target channel is smaller than a set value, and uses these samples as trimming samples. The trimming is performed based on the operation amount of each trimming channel corresponding to the trimming target channel contained in the multiple trimming samples.

4. The learning model generation system according to claim 3, characterized in that, The trimming processing unit performs the trimming based on the operation amount of multiple trimming channels in the trimming channel that is greater than a set value.

5. The learning model generation system according to claim 3, characterized in that, The trimming processing unit performs the trimming based on the operation amount of each trimming channel in multiple trimming samples where the improvement degree of the result data in the trimming sample is greater than a set value.

6. The learning model generation system according to claim 3, characterized in that, The trimming process unit performs the trimming by at least one of the following: the average of the operating amounts of each trimming channel, a constant multiple of the average, and a percentage value.

7. The learning model generation system according to claim 1, characterized in that, The actual data is composed of multiple samples consisting of data sets, each containing the operational quantities of the multiple operation channels and the result data of the operations performed on the multiple operation channels. The training data generation unit includes a filtering processing unit that deletes samples from the plurality of samples whose result data deviates from a set range, or replaces the operation quantity of the operation channel with zero. The trimming processing unit performs the trimming on the actual data that has been deleted or replaced by the filtering processing unit.

8. The learning model generation system according to claim 1, characterized in that, The actual data is composed of multiple samples consisting of data sets, each containing the operational quantities of the multiple operation channels and the result data of the operations performed on the multiple operation channels. The training data generation unit has: The zero-trimming processing unit transforms the operation amount of the multiple operation channels of the multiple samples in which the difference between the result data and the target value for the result data is less than a set value to zero; as well as The data integration processing unit integrates the data transformed by the zero-trimming processing unit and the actual data trimmed by the trimming processing unit.

9. The learning model generation system according to claim 1, characterized in that, The actual data is composed of multiple samples consisting of data sets, each containing the operational quantities of the multiple operation channels and the result data of the operations performed on the multiple operation channels. The training data generation unit includes an elimination processing unit that deletes samples from the actual data whose improvement in the result data is less than a set value. The trimming processing unit performs the trimming on the actual data that has been deleted by the elimination processing unit.

10. The learning model generation system according to claim 1, characterized in that, The training data generation unit selects the trimming object channel from the pre-set operation channels among the plurality of operation channels.

11. The learning model generation system according to claim 1, characterized in that, The training data generation unit includes a setting terminal that sets the setting value for the trimming by the trimming processing unit.

12. The learning model generation system according to claim 1, characterized in that, When the model generation unit generates a learning model that sets the operation quantities of the plurality of operation channels as outputs by learning from the training data generated by the training data generation unit, it performs error backpropagation processing on the operation channels among the plurality of operation channels whose operation quantities are less than a set value.

13. A method for generating a learning model based on a learning model generation system, the learning model generation system comprising: a model generation unit that generates a learning model for controlling an operating device having multiple operating channels; and a training data generation unit that generates training data for the model generation unit to generate the learning model based on actual data related to the control of the operating device, characterized in that... The trimming processing unit of the training data generation unit selects the operation channel whose operation amount shown by the actual data is less than a set value from the plurality of operation channels as the trimming target channel. The operation amount of the target channel is adjusted based on the data extracted from the actual data.

Images