Machine learning model construction device, machine learning model construction method and program

Abstract

To provide a machine learning model construction device capable of efficiently constructing a machine learning model having high generalization performance from limited test data, a machine learning model and a program.SOLUTION: A machine learning model construction device has an actual operation database holding actual operation data, and creates a teacher dataset including one or more teacher data on the basis of the actual operation data acquired from the actual operation database. The machine learning model construction device creates a divided teacher dataset including a plurality of divided teacher data by dividing the teacher data such that an error between a characteristic amount of undivided teacher data and the total sum of the plurality of divided teacher data about the teacher data included in the teacher dataset becomes an allowable value or less, and uses the divided teacher dataset to create a machine learning model.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a machine learning model construction device, a machine learning model construction method, and a program. [Background technology]

[0002] In order to minimize the number of reworks that occur during mass production of railway vehicles, it is necessary to predict the damage caused by the load during operation and the ride comfort before mass production of the railway vehicles begins.To do this, it is effective to use actual measurement test data from the operation of an advance prototype vehicle, create a machine learning model through machine learning, and use the machine learning model to predict the load and damage from operation information and route information.

[0003] Patent Document 1 discloses a method for generating training data that can improve the generalization performance of a learning model (hereinafter referred to as "the prior art"). The prior art is a data generation method for generating data for domain generalization in machine learning, which includes a step of a computer augmenting learning data used to train a machine learning model as original data, and a step of a computer extracting a dataset including both the original data and data generated by the data augmentation as a dataset for domain generalization. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Japanese Patent Application Publication No. 2022-124989 Summary of the Invention [Problem to be solved by the invention]

[0005] Since running tests of railway vehicles are conducted on the basis of being carried out on tracks, there is a need to efficiently build machine learning models with high generalization performance from limited test data. Not only for running tests of railway vehicles, but in the construction of machine learning models in general, there is a need to efficiently build machine learning models with high generalization performance from limited test data.

[0006] Increasing the amount of training data by data partitioning is an effective way to build a machine learning model with high generalization performance. When applying a machine learning model to evaluate the characteristics of mechanical systems such as railway vehicles, obtaining physically meaningful results requires not only increasing the amount of data by data partitioning, but also ensuring that the sum of the characteristics obtained from the training data after partitioning remains unchanged from before partitioning. Conventional technologies have mechanisms for increasing the amount of data, but they separate a single source data set into a "target portion that should remain unchanged because it directly affects the task to be learned by machine learning" and a "non-target portion" and augment the non-target portion by adding noise or other factors, which does not guarantee that the sum of the characteristics will be preserved.

[0007] The present invention has been made to solve the above-mentioned problems. That is, one of the objects of the present invention is to provide a machine learning model construction device, a machine learning model, and a program that can efficiently build a machine learning model with high generalization performance from limited test data by "ensuring that the sum of the characteristic quantities obtained from each training data included in the training data set after division does not change from the characteristic quantity that should be retained as one whole data (strictly speaking, within an allowable value range)." [Means for solving the problem]

[0008] In order to solve the above problem, the machine learning model construction device of the present invention is a machine learning model construction device that includes an information processing device that constructs a machine learning model that predicts the characteristics of production operation data of an equipment, wherein the information processing device has a production database that stores the production operation data, acquires the production operation data from the production operation database, creates a teacher dataset including one or more teacher data based on the production operation data, divides the teacher data included in the teacher dataset so that an error between a characteristic amount of the teacher data before division and a sum of characteristic amounts of the multiple teacher data after division is within an allowable value, thereby creating a divided teacher dataset including the multiple teacher data after division, and creates the machine learning model using the divided teacher dataset.

[0009] The present invention provides a machine learning model construction method using an information processing device that constructs a machine learning model that predicts the characteristics of production operation data of an equipment, wherein the information processing device has a production database that stores the production operation data, and the information processing device acquires the production operation data from the production operation database and creates a teacher dataset including one or more teacher data based on the production operation data, divides the teacher data included in the teacher dataset so that the error between the characteristic quantity of the teacher data before division and the sum of the characteristic quantities of the multiple teacher data after division is within an allowable value, thereby creating a divided teacher dataset including the multiple teacher data after division, and creating the machine learning model using the divided teacher dataset.

[0010] The program of the present invention is a program that causes a computer to execute a process of constructing a machine learning model that predicts the characteristics of production operation data of an equipment, and causes the computer to execute the following processes: acquire the production operation data from a production operation database that stores the production operation data; create a teacher dataset including one or more teacher data based on the production operation data; divide the teacher data included in the teacher dataset so that the error between the characteristic amount of the teacher data before division and the sum of the characteristic amounts of the multiple teacher data after division is within an allowable value, thereby creating a divided teacher dataset including the multiple teacher data after division; and create the machine learning model using the divided teacher dataset. [Effects of the Invention]

[0011] According to the present invention, a machine learning model with high generalization performance can be efficiently constructed from limited test data. Note that the effects described herein are not necessarily limited to those described herein, and may be any of the effects described in this disclosure. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram illustrating an example of the configuration and operation of a machine learning model construction device according to a first embodiment of the present invention. [Figure 2] FIG. 2 is a diagram for explaining the operation information. [Figure 3] FIG. 3 is a diagram for explaining the load measurement data. [Figure 4A] FIG. 4A is a diagram for explaining a method for dividing the distance-weight data into data for each predetermined distance window L1. [Figure 4B] FIG. 4B is a diagram for explaining a method for dividing the distance-velocity data into data for each predetermined distance window L1. [Figure 5] FIG. 5 is a diagram for explaining the error function expressed by equation (1). [Figure 6] FIG. 6 is a diagram for explaining the characteristic quantities. [Figure 7]FIG. 7 is a diagram for explaining the similar data reduction unit. [Figure 8] FIG. 8 is a diagram illustrating an example of the hardware configuration of a computer applied to the machine learning model construction device. [Figure 9A] FIG. 9A is a diagram illustrating an example of the configuration and operation of a machine learning model construction device according to the second embodiment. [Figure 9B] FIG. 9B is a flowchart illustrating the operation of constructing a machine learning model according to the second embodiment. [Figure 10] FIG. 10 is a diagram illustrating an example of the configuration and operation of a machine learning model construction device according to the third embodiment of the present invention. [Figure 11] FIG. 11 is a diagram for explaining the environment database unit. [Figure 12] FIG. 12 is a diagram for explaining data division. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, each embodiment of the present invention will be described with reference to the drawings. In all the drawings of the embodiments, the same or corresponding parts may be denoted by the same reference numerals.

[0014] In the following description, various types of information may be described using expressions such as "graphs," but the various types of information may also be expressed using data structures other than these.

[0015] In the following explanation, processing may be described using functional blocks as the subject, but the subject of processing may also be a CPU or device instead of a functional block. The subject of processing performed by executing a program may be a computing unit, and may include a dedicated circuit that performs specific processing. Here, the dedicated circuit may be, for example, an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), or a CPLD (Complex Programmable Logic Device).

[0016] In the following description, a program may be installed on a computer from a program source. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is a program distribution server, the program distribution server may include a processor and storage resources for storing the program to be distributed, and the processor of the program distribution server may distribute the program to be distributed to other computers. In addition, in each embodiment, two or more programs may be realized as one program, or one program may be realized as two or more programs.

[0017] <<First Embodiment>> A machine learning model construction device according to a first embodiment of the present invention will now be described. FIG. 1 is a diagram illustrating an example of the configuration and operation of a machine learning model construction device according to the first embodiment of the present invention. As shown in FIG. 1, the machine learning model construction device according to the first embodiment includes a database unit 1, a training data generation unit 2, a machine learning model construction unit 3, an input unit 4, and an output unit 5.

[0018] The database unit 1 includes an operation data database unit 11. The operation data database unit 11 stores operation information and load measurement data, which are actual operation data of equipment (railroad vehicles in this example).

[0019] FIG. 2 is a diagram for explaining operation information. As shown in FIG. 2, an example of operation information is data DT1 (hereinafter referred to as "distance-speed data DT1") obtained by measuring the speed relative to the distance traveled when a railway vehicle travels in a certain travel section. FIG. 3 is a diagram for explaining load measurement data. As shown in FIG. 3, the load measurement data is data DT2 (hereinafter referred to as "distance-load data DT2") obtained by measuring the load relative to the distance traveled when a railway vehicle travels in a certain travel section.

[0020] A set of distance-speed data DT1 and distance-load data DT2 collected when a railway vehicle travels a certain section is used as training data. One or more sets of training data are referred to as a "training data set." The training data set is used by the machine learning model construction unit 3 to generate a machine learning model.

[0021] The learning data generation unit 2 includes a data division unit 21, an error function calculation unit 22, an allowable value determination unit 23, and a similar data reduction unit 24.

[0022] The data dividing unit 21 acquires the distance-speed data DT1 and distance-load data DT2 for the same travel section from the operation data database unit 11, pairs them together to create one piece of teacher data, and acquires a teacher data set including the created teacher data. Note that the data dividing unit 21 may also acquire multiple pieces of distance-speed data DT1 and distance-load data DT2 for the same travel section from the operation data database unit 11 and pair them together to create multiple pieces of teacher data. The data dividing unit 21 divides each piece of teacher data included in the teacher data set.

[0023] As shown in Fig. 4A, the data dividing unit 21 divides the distance-load data DT2 into data DT2a for each predetermined distance window L1, and as shown in Fig. 4B, the distance-velocity data DT1 also divides the distance-velocity data DT1 into data DT1a for each predetermined distance window L1, thereby dividing the teacher data. Note that Figs. 4A and 4B illustrate an example of dividing one teacher data. When the teacher data set includes one teacher data, the data dividing unit 21 divides the teacher data in this manner for one teacher data piece, and when the teacher data set includes multiple teacher data pieces, the data dividing unit 21 divides the teacher data for each of the multiple teacher data pieces.

[0024] As shown in Fig. 5, the error function calculation unit 22 calculates the value of the error function (E) using the error function expressed by equation (1) in Fig. 5 for evaluating the teacher data set including the teacher data after division. The error function expressed by equation (1) is a function that represents the error in the characteristic quantity between the teacher data before division (teacher data set before division) and the teacher data after division (teacher data set after division). Note that the teacher data used to evaluate the error in the characteristic quantity is the teacher data that is highly relevant to the characteristics of the equipment (in this example, a railway vehicle) for which a machine learning model is to be created (for example, data that is highly relevant to the data output by the machine learning model to be created (in this example, the load frequency distribution) (in this example, the distance-load data DT2)).

[0025] As shown in FIG. 6, an example of the characteristic quantity Dd is expressed by equation (2). An example of the characteristic quantity Da is expressed by equation (3). The characteristic quantity Da expressed by equation (3) is sometimes referred to as the "damage level" indicating the degree of damage to the railway vehicle. Note that in this example, the characteristic quantity Da is the damage level, but is not limited to the damage level and may be another characteristic quantity (a physical quantity that can be derived (evaluated) from data output by the machine learning model to be created).

[0026] The error function calculation unit 22 calculates the error (value of the error function) between the "characteristic quantity Da of the distance-weight data DT2 of the teacher data before division" and the "sum (characteristic quantity Dd) of the characteristic quantities Da of the distance-weight data DT2a of the multiple teacher data after division" by substituting the characteristic quantities Da and Dd into the error function expressed by formula (1). A larger value of the error function indicates a larger error between the characteristic quantity (characteristic quantity Da) of the teacher data before division (teacher data set) and the characteristic quantity (characteristic quantity Dd) of the teacher data after division (teacher data set). Note that if there are multiple teacher data before division, the error expressed by formula (1) is calculated as many times as the number of teacher data.

[0027] In equation (3), the number of occurrences in load range i can be determined by calculating a load frequency distribution (not shown) indicating the frequency (number of occurrences) for each load range i (i=a1, a2, a3, ... an (note that a1 to an each indicate a different range)) from the distance-load data DT2 using a rainflow method or the like, and then determining the calculated load frequency distribution from the calculated load frequency distribution. The life in load range i can be determined in advance by element testing of the components or the like.

[0028] The tolerance determination unit 23 obtains the value of the error function from the error function calculation unit 22 and determines whether the value of the error function is smaller than the error function threshold (tolerance). If there is a plurality of pre-split training data, it determines whether all of the error function values ​​corresponding to each data are smaller than the error function threshold (tolerance). If the value of the error function is equal to or greater than the error function threshold (tolerance), there is a large error in the characteristic quantities of the training data set after splitting compared to the training data set before splitting, which may have a negative impact on the generalization performance of a machine learning model created using the training data set after splitting.

[0029] Therefore, in this case, the tolerance determination unit 23 outputs the teacher data set before division to the data division unit 21. In order to reduce the value of the error function, the data division unit 21 divides the teacher data included in the teacher data set before division again using a distance window Lnew (a distance window Lnew larger or smaller than the previous distance window) of a size different from the previous distance window (L1 initially). In other words, the data division unit 21 divides the teacher data using a number of data divisions different from the number of data divisions used previously. Note that the error in the characteristic quantities between the teacher data set before division and the teacher data set after division differs depending on the size of the distance window.

[0030] The data dividing unit 21 outputs a teacher data set including the plurality of divided teacher data to the error function calculation unit 22. The error function calculation unit 22 calculates the value of the error function for the teacher data set including the divided teacher data divided by the new distance window Lnew. The tolerance determination unit 23 again determines whether the value of the error function of the teacher data set is smaller than the error function threshold (tolerance value).

[0031] If the error function value is smaller than the error function threshold (tolerance value), the error in the characteristic quantities between the teacher dataset before division and the teacher dataset after division is within the tolerance range. Therefore, in this case, the tolerance value determination unit 23 outputs the teacher dataset after division to the similar data reduction unit 24.

[0032] When the similar data reduction unit 24 acquires the divided teacher dataset, it reduces some of the teacher data from the multiple teacher data in the teacher dataset so that the amount of teacher data that is similar to each other in the divided teacher dataset is reduced. In this way, the similar data reduction unit 24 adjusts the number of teacher data included in the teacher dataset so that the teacher dataset does not include only a large amount of specific teacher data and has data diversity, and by using the teacher dataset after similar data reduction to create a machine learning model, it becomes easier to obtain a machine learning model with high generalization performance.

[0033] FIG. 7 is a diagram illustrating the similar data reduction unit 24. The similar data reduction unit 24 includes a similarity calculation unit 24a, a similar data sorting unit 24b, and an excess data reduction unit 24f. When the similar data reduction unit 24 acquires the divided teacher data set, it inputs the divided teacher data set to the similarity calculation unit 24a. The similarity calculation unit 24a calculates the similarity for distance-weight data DT2a among the plurality of teacher data included in the divided teacher data set, and outputs the plurality of teacher data (distance-weight data DT2a and distance-speed data DT1a) with the similarity associated therewith to the similar data sorting unit 24b. Note that the calculation of the similarity may be performed for the distance-speed data DT1a among the plurality of teacher data.

[0034] The similar data sorting unit 24b classifies a large number of teacher data included in the teacher data set into a plurality of similar teacher data groups 24c, 24d, and 24e based on the similarity, and outputs the grouped data to the excess data reducing unit 24f. Each of the similar teacher data groups 24c, 24d, and 24e includes a plurality of teacher data that are similar to each other (for example, a plurality of teacher data whose similarity is within a predetermined threshold).

[0035] The excess data reduction unit 24f reduces the number of teacher data contained in each of the similar teacher data groups 24c, 24d, and 24e from the teacher dataset so that the number of teacher data contained in each of the similar teacher data groups 24c, 24d, and 24e is averaged (for example, so that the number is the same or within a standard number range), and then outputs the teacher dataset after the teacher data reduction to the machine learning model construction unit 3.

[0036] When the machine learning model construction unit 3 acquires the teacher dataset, it uses the teacher dataset to create a machine learning model, for example, by machine learning (for example, a deep learning technique, which is one type of machine learning technique). One example of a machine learning model is a machine learning model using a neural network that receives distance-speed data as input and outputs a load frequency distribution required for assessing the damage level of a railway vehicle.

[0037] The machine learning model construction unit 3 outputs the created machine learning model to the generalization performance evaluation unit 6.

[0038] The generalization performance determination unit 6 evaluates (determines) the generalization performance of the created machine learning model. For example, the generalization performance determination unit 6 calculates the generalization performance (evaluation index) of the machine learning model using, for example, a test dataset prepared in advance, and determines whether the generalization performance (evaluation index) of the machine learning model satisfies a predetermined standard (standard performance) (for example, by comparing the generalization performance with a generalization performance threshold and making a determination based on the comparison result). Examples of evaluation indexes for generalization performance that can be used include MAE (Mean Absolute Error), MAPE (Mean Absolute Percentage Error), WAPE (Weighted Absolute Percentage Error), MSE (Mean Double Error), and RMSE (Root Mean Square Error).

[0039] If it is determined that the generalization performance of the machine learning model satisfies a predetermined standard, the generalization performance determination unit 6 outputs the machine learning model to the output unit 5. If it is determined that the generalization performance of the machine learning model does not satisfy the predetermined standard, the generalization performance determination unit 6 corrects the error function threshold (tolerance) to a different error function threshold (tolerance) from the previously set error function threshold (tolerance) (for example, an error function threshold (tolerance) that makes error judgment stricter) so as to improve the generalization performance, sets the corrected error function threshold (tolerance), obtains the teacher dataset before division, inputs it to the data division unit 21, and redoes the generation of the machine learning model.

[0040] The input unit 4 is an interface that allows a user to input data into the machine learning model construction device.

[0041] The output unit 5 is an interface for presenting data processed by the machine learning model construction device (for example, the created machine learning model) to the user.

[0042] 8 is a diagram showing an example of the hardware configuration of a computer 800 applied to a machine learning model construction device. The computer 800 may be referred to as a "computer" or an "information processing device." The computer 800 includes a CPU 801, a ROM 802, a RAM 803, a non-volatile storage device 804 that can read and write data, a network interface 805, and an input / output interface 806. These are connected to each other via a bus 807 so as to be able to communicate with each other.

[0043] The CPU 801 is a computing device that loads various programs (not shown) stored in the ROM 802 and / or storage device 804 into the RAM 803 and executes the programs loaded into the RAM 803, thereby realizing various functions.

[0044] As described above, the various programs executed by the CPU 801 are loaded into the RAM 803, and data used when the CPU 801 executes the various programs is temporarily stored in the RAM 803. The ROM 802 and / or the storage device 804 are non-volatile storage media, and the ROM 802 and / or the storage device 804 store various programs.

[0045] The network interface 805 is an interface for connecting the computer 800 to a network. The input / output interface 806 is an interface for connecting the computer 800 to an operation device and a display (display device) capable of displaying images.

[0046] 1 corresponds to the database stored in the storage device 804. The learning data generation unit 2 (data division unit 21, error function calculation unit 22, tolerance determination unit 23, and similar data reduction unit 24), the machine learning model construction unit 3, and the generalization performance determination unit 6 are configured by programs stored in the ROM 802 and / or the storage device. The input unit 4 and output unit 5 correspond to input / output interfaces.

[0047] Note that a hardware device configured with a field programmable gate array (FPGA) or the like may be used in place of the computer 800. Such a hardware device may also be referred to as a "computing device."

[0048] The machine learning model construction device may be configured with multiple computers 800, and may be not limited to physical computers 800 but may also be virtual computers 800. The computers 800 may be computational resources and storage resources provided by the cloud, and the functions provided by the machine learning model construction device may be provided by the cloud.

[0049] <Effects> As described above, the machine learning model construction device according to the first embodiment of the present invention can create a machine learning model with high generalization performance even when there is a small amount of training data used to construct the machine learning model.

[0050] <<Second embodiment>> A machine learning model construction device according to a second embodiment of the present invention will be described. FIG. 9A is a diagram illustrating an example of the configuration and operation of the machine learning model construction device according to the second embodiment. As shown in FIG. 9A, the machine learning model construction device includes a generalization performance evaluation unit 91, a learning tendency data storage unit 92, an appropriate learning data storage unit 93, and an allowable calculation count determination unit 94. The rest is the same as the machine learning model construction device shown in FIG. 1. Note that the generalization performance evaluation unit 91, the learning tendency data storage unit 92, and the allowable calculation count determination unit 94 in FIG. 9A are configured by programs stored in the ROM 802 and / or the storage device 804 in FIG. 8. The appropriate learning data storage unit 93 corresponds to the storage device 804 in FIG. 8.

[0051] FIG. 9B is a flowchart illustrating the operation of constructing a machine learning model according to the second embodiment.

[0052] Step 900: The user inputs an error function threshold ef via the input unit 4. th is input to the tolerance determination unit 23, and the generalization performance threshold gp th is input to the generalization performance judgment unit 6, and the allowable number of calculations is input to the allowable number of calculations judgment unit 94.

[0053] Step 901: The tolerance determination unit 23 determines the error function threshold value ef th Based on the error function tolerance ef tol The tolerance determining unit 23 sets the error function tolerance ef tol The error function threshold ef th Set it to half the value.

[0054] Step 902: The data division unit 21 creates a teacher data set based on the actual operation data acquired from the operation data database unit 11, divides each teacher data included in the teacher data set, calculates the value of the error function by the error function calculation unit 22, and determines whether the value of the error function is equal to or less than the error function tolerance value ef by the tolerance determination unit 23. tol The data dividing unit 21 determines whether the value of the error function satisfies the error function tolerance ef tol The tolerance determination unit 23 determines whether the value of the error function satisfies the error function tolerance ef tol If it is determined that the error function satisfies the error function tolerance ef tol The divided teacher data set that satisfies the above is output to the similar data reduction unit 24.

[0055] Step 903: The similar data reduction unit 24 classifies the plurality of teacher data included in the teacher data set into similar teacher data groups using the method described in the first embodiment, and reduces the teacher data from each of the similar teacher data groups.

[0056] Step 904: The machine learning model construction unit 3 creates a machine learning model using the training data set after data reduction.

[0057] Step 905: The generalization performance evaluation unit 91 evaluates the generalization performance gp of the machine learning model.

[0058] Step 906: The generalization performance evaluation unit 91 calculates the error function tolerance ef tol and generalization performance gp, and the learning tendency data (f tol, gp) in the learning tendency data storage unit 92.

[0059] Step 907: The generalization performance determination unit 6 compares the generalization performance gp of the evaluation result with the generalization performance threshold gpth to determine whether the generalization performance gp of the created machine learning model satisfies the standard. If the generalization performance of the machine learning model does not satisfy the standard, the process returns to step 901, and the tolerance determination unit 23 determines the error function tolerance ef toland the corrected error function tolerance ef tol is set (step 901). tol is corrected as follows, for example: The tolerance determination unit 23 uses the learning tendency data (f tol, Based on gp), the error function tolerance ef tol Reduction of error function tolerance ef tol The tolerance determining unit 23 determines which of the two correction methods is more likely to improve the generalization performance. tol For example, the tolerance determination unit 23 modifies the error function tolerance ef tol If a decrease in the error function tolerance ef is determined, tol For example, the tolerance determination unit 23 modifies the error function tolerance ef tol If you decide to increase the error function tolerance ef tol is modified to increase it by a predetermined amount.

[0060] If the generalization performance gp of the machine learning model satisfies the criteria, the generalization performance determination unit 6 proceeds to step 908.

[0061] Step 908: The appropriate learning data storage unit 93 stores the teacher data set (referred to as "appropriate learning data") used to create the machine learning model whose generalization performance gp meets the standard as learning tendency data (f tol, gp) and stores it in the learning tendency data storage unit 92. The allowable calculation count determination unit 94 determines whether the number of determinations made in step 907 as "YES" in the generalization performance determination is equal to or greater than the allowable calculation count. If the number of determinations is less than the allowable calculation count, the allowable calculation count determination unit 94 returns to step 901 and sets the error function allowable value ef tol and the corrected error function tolerance ef tol is set (step 901). After that, steps 902 to 907 are executed again.

[0062] If the number of judgments is equal to or greater than the allowable number of calculations, the allowable number of calculations judgment unit 94 proceeds to step 909, where it determines the appropriate learning data with the best generalization performance gp from among the appropriate learning data stored in the learning tendency data storage unit 92 as the optimal learning data (optimal teacher data set), and outputs the machine learning model created using the determined optimal learning data (optimal teacher data set) to the output unit 5.

[0063] <Effects> As described above, the machine learning model construction device according to the second embodiment of the present invention can create a machine learning model with higher generalization performance even when there is little training data by automatically searching for an appropriate data division.

[0064] <<Third Embodiment>> A machine learning model construction device according to a third embodiment of the present invention will now be described. FIG. 10 is a diagram illustrating an example of the configuration of a machine learning model construction device according to the third embodiment. As shown in FIG. 10, a database unit 1 of the machine learning model construction device includes an environment database unit 12 in which environment data DT3 is stored. The rest is the same as the machine learning model construction device according to the first embodiment of FIG.

[0065] Fig. 11 is a diagram for explaining the environment database unit 12. As shown in Fig. 11, the environment database unit 12 stores branch position information 12a, track curvature information 12b, station position information 12c, track elevation difference information 12d, and bridge position information 12f. Note that the information stored in the environment database is not limited to the above-mentioned 12a to 12f, as long as the environmental information has a correlation with characteristics.

[0066] The branch position information 12a stores data (sometimes referred to as "distance-branch position data") that measures branch positions relative to the travel distance when a railway vehicle travels a certain travel section. Note that the distance-branch position data may represent the presence or absence of a branch position relative to the travel distance using "0" or "1."

[0067] The rail curvature information 12b stores data (sometimes referred to as "distance-rail curvature data") that measures the rail curvature relative to the travel distance when a railway vehicle travels in a certain travel section.

[0068] The station location information 12c stores data (sometimes referred to as "distance-station location data") that measures the location of a station relative to the distance traveled when a railway vehicle travels a certain travel section. Note that the distance-station location data may represent the presence or absence of a station location relative to the travel distance (presence or absence of a station) using "0" and "1."

[0069] The rail elevation difference information 12d stores data (sometimes referred to as "distance-rail elevation difference data") that measures the rail elevation difference relative to the travel distance when a railway vehicle travels a certain travel section. The bridge position information 12f stores data (sometimes referred to as "distance-bridge position data") that measures the bridge position relative to the travel distance when a railway vehicle travels a certain travel section. Note that the distance-bridge position data may represent the presence or absence of a branch position relative to the travel distance using "0" or "1."

[0070] In the machine learning model construction device according to the third embodiment, a set of distance-speed data DT1, distance-load data DT2, and environmental data DT3 (distance-branch position data, distance-rail curvature data, distance-station position data, and distance-rail elevation difference data) collected when a railway vehicle travels a certain travel section is used as training data. The training data set including the plurality of training data is used to generate a machine learning model by the machine learning model construction unit 3. One example of a machine learning model is a machine learning model using a neural network that receives distance-speed data and environmental data as input and outputs a load frequency distribution required for assessing the damage level of a railway vehicle.

[0071] The data dividing unit 21 acquires a teacher data set from the operation data database unit 11 and divides each teacher data set included in the teacher data set. As shown in Fig. 12, the data dividing unit 21 divides the distance-load data DT2 into data DT2a for each predetermined distance window L1, and as shown in Fig. 12, the data dividing unit 21 also divides the environmental data DT3 (in this example, each of the distance-branch position data, distance-rail curvature data, distance-station position data, distance-rail elevation difference data, and bridge position data) into data DT3a for each predetermined distance window L1, and further divides the distance-speed data DT1 into data DT1a for each predetermined distance window L1, thereby dividing the teacher data.

[0072] As in the first embodiment, the error function calculation unit 22 calculates the value of the error function using the error function expressed by equation (1) in FIG. 5 in order to evaluate the error of the teacher data set including the multiple teacher data sets after division.

[0073] The tolerance determination unit 23 obtains the value of the error function from the error function calculation unit 22 and determines whether the value of the error function is smaller than the error function threshold (tolerance value). If the value of the error function is equal to or greater than the error function threshold (tolerance value), the tolerance determination unit 23 outputs the teacher data set before division to the data division unit 21. In order to reduce the value of the error function, the data division unit 21 re-divides the teacher data included in the teacher data set before division using a distance window Lnew (a distance window Lnew larger or smaller than the previous distance window) of a size different from the previous distance window (L1 initially). In other words, the data division unit 21 divides the teacher data using a number of data divisions different from the number of data divisions used previously.

[0074] The data dividing unit 21 outputs a teacher data set including the plurality of teacher data sets after division to the error function calculation unit 22. The error function calculation unit 22 calculates the value of the error function for the teacher data set after division using the new distance window Lnew. The tolerance determination unit 23 again determines whether the value of the error function of the teacher data set is smaller than the error function threshold (tolerance value).

[0075] If the value of the error function is smaller than the error function threshold (tolerance), the error in the characteristic quantities of the divided teacher dataset relative to the pre-division teacher dataset is within the tolerance range. Therefore, in this case, the tolerance determination unit 23 outputs the divided teacher dataset to the similar data reduction unit 24.

[0076] When the similar data reduction unit 24 acquires the divided teacher data set, it reduces some of the teacher data from the divided teacher data set so that there is less similar teacher data.

[0077] After obtaining the training data set after similar data has been reduced, the machine learning model construction unit 3 uses the training data set to create a machine learning model. The machine learning model construction unit 3 outputs the machine learning model to the generalization performance determination unit 6.

[0078] The generalization performance determination unit 6 evaluates (determines) the generalization performance of the machine learning model. For example, the generalization performance determination unit 6 uses a test dataset to determine whether the generalization performance of the machine learning model satisfies a predetermined standard. If it is determined that the generalization performance of the machine learning model satisfies the predetermined standard, the generalization performance determination unit 6 outputs the machine learning model to the output unit 5.

[0079] If it is determined that the generalization performance of the machine learning model does not satisfy a predetermined standard, the generalization performance determination unit 6 modifies the error function threshold (tolerance value) to a different error function threshold (tolerance value) from the error function threshold (tolerance value) set previously (for example, an error function threshold (tolerance value) that makes error judgment stricter) so as to improve the generalization performance, sets the modified error function threshold (tolerance value), obtains the training dataset before division, inputs it to the data division unit 21, and redoes the generation of the machine learning model.

[0080] <Effects> As described above, the machine learning model construction device of the third embodiment can create a machine learning model with higher generalization performance even when there is little training data by predicting loads using environmental information that is correlated with the output (load).

[0081] <<Modifications>> The present invention is not limited to the above-described embodiments, and various modifications can be adopted within the scope of the present invention. Furthermore, the above-described embodiments can be combined with each other without departing from the scope of the present invention.

[0082] In each of the above embodiments, the machine learning model construction device may be configured such that the similar data reduction unit 24 is omitted.

[0083] In each of the above embodiments, the data used as training data may be time-series data, and a time window may be used instead of a distance window. In each of the above embodiments, an example of a machine learning model to be created is a machine learning model using a neural network that inputs distance-speed data and outputs a load frequency distribution of a railway vehicle. However, the machine learning model to be created is not limited to this example, and may be, for example, a machine learning model that predicts the characteristics of actual operation data of equipment other than a railway vehicle. Data appropriate for the machine learning model to be created is used as the training data set (training data). The data used as training data may be data acquired by an acceleration sensor, a sound sensor, or the like, or may be time-series data of load, data obtained by Fourier transforming the time-series data of load, or the like.

[0084] In each of the above embodiments, data that has undergone data processing such as downsampling may be used as training data. Considering actual operation, it is possible that speed data at a high sampling rate, such as in driving tests, cannot always be obtained. For example, it is conceivable that the available operation information will be rough data such as run curves. In this case, when training a machine learning model (neural network) (when creating a machine learning model), it is preferable to use speed time series data (distance-speed data) that has been downsampled to correspond to a run curve, rather than directly using the speed time series data (distance-speed data) of the driving test data. [Explanation of symbols]

[0085] 1...database unit, 2...learning data generation unit, 3...machine learning model construction unit, 4...input unit, 5...output unit, 6...generalization performance judgment unit, 11...operational data database unit, 12...environmental database unit, 21...data division unit, 22...error function calculation unit, 23...tolerance value judgment unit, 24...similar data reduction unit

Claims

1. A machine learning model construction device including an information processing device that constructs a machine learning model that predicts characteristics of actual operation data of equipment, The information processing device includes: a production database for holding the production data; Acquire the production data from the production database, and create a training data set including one or more training data based on the production data; creating a divided teacher data set including the plurality of divided teacher data by dividing the teacher data included in the teacher data set so that an error between a characteristic amount of the teacher data before division and a sum of characteristic amounts of the plurality of teacher data after division is equal to or less than an allowable value; Creating the machine learning model using the divided training dataset. It was configured as follows: Machine learning model building device.

2. The machine learning model construction device according to claim 1, The information processing device includes: Evaluating the generalization performance of the created machine learning model, and if the generalization performance of the machine learning model does not satisfy a predetermined standard performance, modifying the tolerance; The teacher data is divided so that an error between a characteristic amount of the teacher data before division and a sum of characteristic amounts of the plurality of teacher data after division is equal to or less than the corrected allowable value, thereby newly creating a divided teacher data set including the plurality of teacher data after division; Re-creating the machine learning model using the newly created post-division training dataset; It was configured as follows: Machine learning model building device.

3. The machine learning model construction device according to claim 1, The information processing device includes: After creating the divided teacher dataset and before creating the machine learning model, the plurality of teacher data included in the divided teacher dataset are sorted into a plurality of similar teacher data groups each consisting of a plurality of teacher data similar to each other, and a portion of the plurality of teacher data included in the divided teacher dataset is reduced based on the number of the plurality of teacher data included in each of the plurality of similar teacher data groups, thereby improving the diversity of the plurality of teacher data included in the divided teacher dataset. It was configured as follows: Machine learning model building device.

4. The machine learning model construction device according to claim 3, The information processing device includes: reducing a portion of the plurality of teacher data included in the divided teacher data set so as to average the number of the plurality of teacher data included in each of the plurality of similar teacher data groups; It was configured as follows: Machine learning model building device.

5. The machine learning model construction device according to claim 1, The information processing device includes: After dividing the teacher data, a first determination is made to determine whether or not an error between a characteristic amount of the teacher data before division and a sum of the characteristic amounts of the teacher data after division is equal to or less than the allowable value; In the first determination, if the error is greater than the allowable value, changing the number of divisions of the teacher data to a different number of divisions from the previous number of divisions of the teacher data; dividing the teacher data before division using the changed number of divisions of the teacher data; and after dividing the teacher data, performing a second determination to determine whether an error between a characteristic amount of the teacher data before division and a sum of the characteristic amounts of the teacher data after division is equal to or less than the allowable value; and creating the divided teacher data set by repeatedly performing this second determination until the error becomes equal to or less than the allowable value. It was configured as follows: Machine learning model building device.

6. The machine learning model construction device according to claim 1, The information processing device includes: Evaluating the generalization performance of the created machine learning model, and if the generalization performance of the machine learning model does not satisfy a predetermined standard performance, modifying the tolerance; The training data is divided so that an error between a characteristic amount of the training data before division and a sum of characteristic amounts of the plurality of training data after division is equal to or less than the corrected allowable value, thereby creating a new training data set after division including the plurality of training data after division, and recreating the machine learning model using the newly created training data set after division, and this process is repeated until the generalization performance of the machine learning model satisfies a predetermined standard performance. It was configured as follows: Machine learning model building device.

7. 7. The machine learning model construction device according to claim 6, The information processing device includes: Each time the machine learning model is created or recreated, the tolerance and the generalization performance are associated with each other and stored, and the tolerance is corrected based on the stored data trends of the tolerance and the generalization performance. It was configured as follows: Machine learning model building device.

8. The machine learning model construction device according to claim 1, The characteristic quantity is a damage degree. Machine learning model building device.

9. The machine learning model construction device according to claim 1, The information processing device includes: an environment database that stores environment data including information about an environment corresponding to a point where the operational data is acquired; creating the training data set including one or more training data by pairing the production operation data acquired from the production operation database and the environmental data acquired from the environmental database; It was configured as follows: Machine learning model building device.

10. 10. The machine learning model construction device according to claim 9, The environmental data includes information on rail branch positions, information on station positions, information on rail curvature, rail elevation differences, and information on bridge positions. Machine learning model building device.

11. The machine learning model construction device according to claim 1, The information processing device includes: A first evaluation is performed to evaluate the generalization performance of the created machine learning model; If the generalization performance of the machine learning model satisfies a predetermined standard performance in the first evaluation, the created machine learning model and the generalization performance are associated with each other and saved, and then the tolerance is corrected; If the generalization performance of the machine learning model does not satisfy a predetermined standard performance in the first evaluation, modifying the tolerance; After modifying the tolerance, a machine learning model re-creation evaluation process is performed to re-create a post-division teacher dataset by dividing the teacher data before division so that an error between a characteristic amount of the teacher data before division and a sum of characteristic amounts of the plurality of teacher data after division is equal to or less than the corrected allowable value, re-create the machine learning model using the newly created post-division teacher dataset, and perform a second evaluation to evaluate the generalization performance of the re-created machine learning model; If the generalization performance of the machine learning model satisfies a predetermined standard performance in the second evaluation, the recreated machine learning model and the generalization performance are associated and saved, and then the tolerance is corrected, and the machine learning model recreate evaluation process is performed again; If the generalization performance does not satisfy a predetermined standard performance in the second evaluation, the tolerance is corrected and then the machine learning model re-creation evaluation process is performed again; When the number of times that the generalization performance is determined to satisfy a predetermined standard performance in the second evaluation is equal to or greater than a predetermined allowable number of times, the machine learning model is created using the divided teacher dataset with the best generalization performance that is saved each time the machine learning model is created or recreated. It was configured as follows: Machine learning model building device.

12. A machine learning model construction method using an information processing device to construct a machine learning model that predicts characteristics of actual operation data of equipment, the information processing device has a production database for storing the production data; By the information processing device, Acquire the production data from the production database, and create a training data set including one or more training data based on the production data; creating a divided teacher data set including the plurality of divided teacher data by dividing the teacher data included in the teacher data set so that an error between a characteristic amount of the teacher data before division and a sum of characteristic amounts of the plurality of teacher data after division is equal to or less than an allowable value; Creating the machine learning model using the divided training dataset. How to build machine learning models.

13. A program that causes a computer to execute a process of constructing a machine learning model that predicts characteristics of actual operation data of equipment, The computer, Acquire the production data from a production database that stores the production data, and create a teacher data set that includes one or more teacher data based on the production data; creating a divided teacher data set including the plurality of divided teacher data by dividing the teacher data included in the teacher data set so that an error between a characteristic amount of the teacher data before division and a sum of characteristic amounts of the plurality of teacher data after division is equal to or less than an allowable value; A process of creating the machine learning model using the divided training dataset, Let it run, program.

Images