Performance prediction apparatus and program

Abstract

To provide a performance prediction apparatus and program for calculating the accuracy of prediction performance in a short time.SOLUTION: A machine learning model 40 which is trained with multiple pieces of training data 110 representing each of manufacturing conditions of a product and configured to receive, as input, predetermined input data 210, outputs a prediction value 230 indicating property of the product to be manufactured by the manufacturing conditions represented by the input data 210. A dimensional compression unit 50 performs dimensional compression on the multiple training data 110 and the input data 210, respectively. A similarity determination unit 60 determines one piece of the dimensionally compressed training data 110 which is similar to the dimensionally compressed input data 210. An output unit 80 outputs the prediction value 230 and outputs information, as accuracy of the prediction value 230, in accordance with a determination result of the dimensional compression unit 50.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a performance prediction device and a program. [Background technology]

[0002] It is known that predictions are made for given input manufacturing condition data using a machine learning model obtained by machine learning techniques such as neural networks. For example, Patent Document 1 describes a method in which, when an unknown prediction case is input, a set of similar cases, which is a set of cases similar to the prediction case, is extracted from a set of known cases, a confidence level of a certain prediction attribute value is calculated from the set of similar cases, a reliability measure of the confidence level is calculated from the set of similar cases and the confidence level, and the confidence level of a certain prediction attribute value and the reliability measure of the confidence level are output. Furthermore, Patent Document 2 below describes a method in which an input dataset is divided according to specified division conditions, neighborhood data of the input data including feature nodes that represent characteristics of the distribution structure of each divided dataset is generated, and a score representing the relationship between the explanatory variables and the objective variable is calculated based on the explanatory variables of the generated neighborhood data and data on the objective variable obtained by inputting the neighborhood data into a machine learning model. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Laid-Open No. 2003-323601 [Patent Document 2] Japanese Patent Application Publication No. 2019-191895 Summary of the Invention [Problem to be solved by the invention]

[0004] The accuracy of predictions varies from prediction to prediction. If the degree of accuracy could be known for each prediction, it would be useful for the user to decide which prediction to rely on for prototyping. In this regard, if the manufacturing conditions of a product are similar to the learned manufacturing conditions, the accuracy of the prediction performance output by the machine learning model is considered to be high, and conversely, if they are not similar, the accuracy is considered to be low. For this reason, it is desirable to determine the similarity of the manufacturing conditions.

[0005] However, when the manufacturing conditions are multidimensional vector data containing information on many factors, it takes a long time to calculate the similarity.

[0006] The present disclosure has been made in view of the above-mentioned problems, and its purpose is to provide a performance prediction device and a program that can calculate the accuracy of predicted performance in a short time. [Means for solving the problem]

[0007] (1) The performance prediction device according to the present disclosure includes a machine learning model that is trained using a plurality of training manufacturing condition vector data, each of which indicates a manufacturing condition of a product, and that outputs predicted performance data that indicates the properties of a product manufactured under the manufacturing conditions indicated by the input manufacturing condition vector data when given input manufacturing condition vector data is input, a dimensionality reduction unit that reduces the dimensions of each of the plurality of training manufacturing condition vector data and the input manufacturing condition vector data, a similarity determination unit that determines whether the dimensionally reduced input manufacturing condition vector data is similar to any of the plurality of dimensionally reduced training manufacturing condition vector data, and an output unit that outputs the predicted performance data and information corresponding to the determination result of the similarity determination unit as the accuracy of the predicted performance data. This allows the accuracy of the predicted performance to be calculated in a short time.

[0008] (2) In the performance prediction device of (1) above, the dimensionality reduction means may input each of the plurality of training manufacturing condition vector data and the input manufacturing condition vector data to the machine learning model, and obtain an output of an intermediate layer of the machine learning model as the plurality of training manufacturing condition vector data and the input manufacturing condition vector data that have been dimensionally reduced. This makes it possible to obtain the plurality of training manufacturing condition vector data and the input manufacturing condition vector data without performing a separate calculation for dimensionality reduction.

[0009] (3) In the performance prediction device (1) or (2) above, the similarity determination means may perform density estimation of the distribution of the plurality of dimensionally compressed learning manufacturing condition vector data, and determine the similarity based on the result of the density estimation.

[0010] (4) In the performance prediction device according to any one of claims 1 to 3, the product may be a tire. The plurality of learning manufacturing condition vector data and the input manufacturing condition vector data may include one or more elements indicating the material of the tire and one or more elements indicating the structure of the tire. The predicted performance data may indicate the performance of the tire.

[0011] (5) In the performance prediction device according to any one of (1) to (3), the product may be a tire rubber. The plurality of learning manufacturing condition vector data and the input manufacturing condition vector data may include one or more elements indicating the material of the rubber and one or more elements indicating the manufacturing conditions of the rubber. The predicted performance data may indicate physical properties of the rubber.

[0012] (6) A program according to the present disclosure causes a computer to function as a machine learning model that is trained using a plurality of training manufacturing condition vector data, each of which indicates a manufacturing condition of a product, and that outputs predicted performance data that indicates the properties of a product manufactured under the manufacturing conditions indicated by the input manufacturing condition vector data when given input manufacturing condition vector data is input, a dimensionality reduction means that reduces the dimensions of each of the plurality of training manufacturing condition vector data and the input manufacturing condition vector data, a similarity determination means that determines whether the dimensionally reduced input manufacturing condition vector data is similar to any of the plurality of dimensionally reduced training manufacturing condition vector data, and an output means that outputs the predicted performance data and information corresponding to the determination result of the similarity determination means as the accuracy of the predicted performance data. This allows the accuracy of the predicted performance to be calculated using a computer in a short period of time. [Brief explanation of the drawings]

[0013] [Figure 1] 1 is a diagram illustrating a configuration of a performance prediction device according to an embodiment of the present disclosure. [Figure 2] FIG. 2 is a functional block diagram illustrating an example of functions implemented in the performance prediction device. [Figure 3] FIG. 2 is a diagram illustrating an example of learning data and learning-time dimensionality-reduced data. [Figure 4] FIG. 2 is a diagram illustrating an example of input data and input-dimensionally compressed data. [Figure 5] FIG. 10 is a diagram illustrating an example of predicted performance data and accuracy of the predicted performance data. [Figure 6A] FIG. 1 is a diagram illustrating an example of processing performed during learning of a machine learning model. [Figure 6B] FIG. 1 is a diagram illustrating an example of processing performed during learning of a machine learning model. [Figure 7A] FIG. 10 is a diagram illustrating an example of processing performed when making a prediction on input data. [Figure 7B] FIG. 10 is a diagram illustrating an example of processing performed when making a prediction on input data. [Figure 8]FIG. 10 is a functional block diagram illustrating another example of functions implemented in the performance prediction device. DETAILED DESCRIPTION OF THE INVENTION

[0014] An embodiment of the present disclosure will be described below with reference to the drawings. In this embodiment, input data indicating product manufacturing conditions is input to a trained machine learning model, and predicted performance data indicating the properties of a product manufactured under those manufacturing conditions is output, along with information regarding the accuracy of the predicted performance data. In the following, a case will be described in which the product is a tire or tire rubber. However, the product is not limited to a mixture of tires or tire rubber produced using multiple raw materials and a production process, and may be any substance produced based on predetermined conditions.

[0015] [1. Hardware configuration] Fig. 1 is a diagram showing the configuration of a performance prediction device 10 according to an embodiment of the present disclosure. The performance prediction device 10 according to this embodiment is a computer such as a personal computer, a general-purpose computer, or a mobile information terminal, and includes a processor 11, a storage unit 12, a communication unit 13, a display unit 14, and an operation unit 15, as shown in Fig. 1. The performance prediction device 10 may also include an optical disc drive for reading optical discs, a USB (Universal Serial Bus) port, etc.

[0016] The processor 11 is a program-controlled device such as a CPU (Central Processing Unit) that operates according to a program installed in the performance prediction device 10, which is, for example, a computer. The memory unit 12 is a storage element such as a ROM (Read Only Memory) or a RAM (Random Access Memory), or a hard disk drive. The memory unit 12 stores data such as programs executed by the processor 11. The communication unit 13 is a communication interface such as a network board. The display unit 14 is a display device such as a liquid crystal display, and displays various images according to instructions from the processor 11. The operation unit 15 is a user interface such as a keyboard or a mouse, and accepts user operation inputs and outputs signals indicating the contents of the inputs to the processor 11.

[0017] [2. Functional Blocks] The performance prediction device 10 outputs predicted performance data indicating, for example, the predicted performance of a tire or tire rubber (e.g., tire rolling resistance, rubber hardness, etc.), and also outputs information regarding the accuracy of the predicted performance data. The output of the predicted performance data and information regarding the accuracy by the performance prediction device 10 will be described below.

[0018] Fig. 2 is a functional block diagram showing an example of functions implemented in the performance prediction device 10. As shown in Fig. 2, the performance prediction device 10 functionally includes a learning data acquisition unit 20, an input data acquisition unit 30, a machine learning model 40, a dimensionality reduction unit 50, a dimensionality reduced data storage unit 60, a similarity determination unit 70, and an output unit 80. Note that not all of the functions shown in Fig. 2 need to be implemented in the performance prediction device 10, and functions other than the functions shown in Fig. 2 may be implemented.

[0019] [2-1. Learning data acquisition section] The training data acquisition unit 20 acquires training manufacturing condition vector data. The "training manufacturing condition vector data" is data obtained by vectorizing training manufacturing condition data indicating the manufacturing conditions of a product using a given calculation formula, and is training data used for training the machine learning model 40 described below. The training data acquisition unit 20 may acquire data converted from the training manufacturing condition data using a given conversion method other than vectorization as training data, or may acquire the training manufacturing condition data itself. Hereinafter, the data acquired by the training data acquisition unit 20 (in this embodiment, the training manufacturing condition vector data) will also be simply referred to as "training data." The training data acquisition unit 20 may be realized primarily by the processor 11, but may also be realized by a processor of another information processing device connected to the performance prediction device 10 via a network.

[0020] 3 is a diagram showing an example of training data and training-time dimensionality-reduced data. In the example shown in Fig. 3, data identified by an individual ID (1, 2, N, etc.) (one row of data in a table) corresponds to one piece of data.

[0021] 3, the training data 110 includes a plurality of factors. More specifically, factors X1 to X M The training data 110 may include, for example, one or more factors indicating the tire material, or one or more factors indicating the tire structure. In addition, the training data 110 may include one or more factors indicating the tire rubber material, or one or more factors indicating the tire rubber manufacturing conditions.

[0022] The training data 110 also includes predicted performance data indicating the performance (properties) of a product manufactured under the manufacturing conditions indicated therein. The predicted performance data is, for example, data indicating tire performance (such as tire rolling resistance). Alternatively, the predicted performance data may be data indicating the physical properties (such as hardness) of tire rubber. In the example shown in FIG. 3, the predicted performance data is a predicted value, and is indicated by a numerical value. For example, the training data 110 identified by ID "1" indicates a predicted value of "15.0", while the training data 110 identified by ID "2" indicates a predicted value of "14.8".

[0023] Although not shown in the figure, a plurality of factors (factors X1 to X2) included in each of the plurality of learning data 110 are M ) differs between each set of training data. Therefore, the predicted values ​​between each set of training data are also basically different from each other, but there are cases where the predicted values ​​match between multiple sets of training data due to rounding processes such as rounding down the predicted values.

[0024] [2-2. Input data acquisition section] The input data acquisition unit 30 acquires given input manufacturing condition vector data. Similar to the training manufacturing condition vector data, the "input manufacturing condition vector data" refers to data obtained by vectorizing the input manufacturing condition data, which indicates the manufacturing conditions of a product, using a given formula. The input data acquisition unit 30 may acquire data converted from the input manufacturing condition data using a given conversion method other than vectorization, or may acquire the input manufacturing condition data itself. For example, when the training manufacturing condition data itself is used as training data for the machine learning model 40 (described later), the input data acquisition unit 30 may acquire the input manufacturing condition data in an unconverted state. Hereinafter, the data acquired by the input data acquisition unit 30 (input manufacturing condition vector data in this embodiment) will also be simply referred to as "input data." The input data acquisition unit 30 may be primarily implemented by the processor 11, but may also be implemented by a processor of another information processing device connected to the performance prediction device 10 via a network.

[0025] 4 is a diagram showing an example of input data and input time-dimensionally compressed data. In the example shown in FIG. 4, as in FIG. 3, data identified by an individual ID (1, 2, n, etc.) (one row of data in a table) corresponds to one piece of data. Note that, in the example shown in FIG. 4, there are multiple pieces of input data (n pieces), but the number of pieces of input data acquired by the input data acquisition unit 30 and input to a machine learning model 40 (described later) may be one.

[0026] The input data 210 also includes a plurality of factors, similar to the training data 110. More specifically, the factors X1 to X M The input data 210 includes M factors, as in the training data 110. The input data 210 may include one or more factors that indicate the tire material, and may also include one or more factors that indicate the tire structure. In addition, the input data 210 may include one or more factors that indicate the tire rubber material, and may also include one or more factors that indicate the manufacturing conditions of the tire rubber.

[0027] In this embodiment, the number of factors included in each of the plurality of learning data 110 and the input data 210 is M, which is the same number. However, it is sufficient that the plurality of learning data 110 and the input data 210 contain a plurality of common predetermined factors, and the total number of factors included in each does not necessarily have to be the same.

[0028] The input data 210 is intended to predict the properties of a product manufactured under the multiple manufacturing conditions indicated therein. Therefore, unlike the learning data 110, the input data 210 does not include predicted performance data such as predicted values. By creating input data indicating multiple manufacturing conditions and inputting the data to the machine learning model 40 (described later), the user can learn information about the properties predicted by the machine learning model 40 (more specifically, the prediction unit 43).

[0029] [2-3. Machine Learning Model] The machine learning model 40 is a model trained using multiple pieces of training data 110 (training manufacturing condition vector data), each of which indicates a manufacturing condition for a product. When given input data 210 (input manufacturing condition vector data) is input, the model outputs predicted performance data indicating the properties of a product manufactured under the manufacturing conditions indicated by the input manufacturing condition vector data. The performance prediction device 10 may be equipped with multiple machine learning models 40 as functions, each corresponding to a type of predicted performance data to be predicted (e.g., type of property such as tire rolling resistance or rubber hardness). In this case, machine learning may be performed by inputting training data 110 including predicted performance data of a type corresponding to the machine learning model 40 to the machine learning model 40. Furthermore, a user may input input data 210 to one of the multiple machine learning models 40 that corresponds to the type of predicted performance data to be predicted.

[0030] 2, the machine learning model 40 includes a parameter storage unit 41, a learning unit 42, and a prediction unit 43. The machine learning model 40 is a model that is machine-learned using a multi-layer neural network such as a deep neural network (DNN) or a convolutional neural network (CNN). Alternatively, the machine learning model 40 may be a model that is machine-learned using a given statistical method, linear regression, or the like.

[0031] [2-3-1. Parameter storage section] The parameter storage unit 41 stores parameters of the machine learning model 40. More specifically, the parameter storage unit 41 stores parameters of the machine learning model 40 that outputs predicted performance data when input data 210 is input to the machine learning model 40. Note that the parameter storage unit 41 may be realized mainly by the storage unit 12 of the performance prediction device 10, or may be realized by another storage device such as an external storage device or a NAS (Network Attached Storage) connected to the performance prediction device 10 by wire or wirelessly.

[0032] When the machine learning model 40 is realized by a multi-layer neural network, the parameter storage unit 41 may store, as parameters of the machine learning model 40, the multiple nodes constituting the neural network, the weighting of each node, the number of layers, the number of nodes used in each layer, etc. Additionally, for example, the parameter storage unit 41 may store, as parameters of the machine learning model 40, a formula or coefficients of the formula for determining predicted performance data from multiple manufacturing conditions.

[0033] [2-3-2. Learning Department] The learning unit 42 updates the parameters of the machine learning model 40 stored in the parameter storage unit 41 by performing machine learning using the plurality of pieces of learning data 110 acquired by the learning data acquisition unit 20. That is, the parameters of the machine learning model 40 stored in the parameter storage unit 41 are learned (updated) using the plurality of pieces of learning data. Note that the learning unit 42 may be realized mainly by the processor 11 of the performance prediction device 10, or may be realized by a processor of another information processing device connected to the performance prediction device 10 via a network.

[0034] [2-3-3. Prediction Section] When given input data 210 is input to the machine learning model 40, the prediction unit 43 calculates predicted performance data for the input data 210. The prediction unit 43 may be realized mainly by the processor 11 of the performance prediction device 10, or may be realized by a processor of another information processing device connected to the performance prediction device 10 via a network.

[0035] FIG. 5 is a diagram illustrating an example of predicted performance data and the accuracy of the predicted performance data. In the example shown in FIG. 5, the predicted value 230 for the input data 210 identified by the ID "1" is "15.0," and the predicted value 230 for the input data 210 identified by the ID "2" is "13.2." In this manner, the predicted value 230 for each of the multiple input data 210 identified by individual IDs corresponds to an example of predicted performance data. That is, the prediction unit 43 calculates the predicted value 230 for each of the input data 210. Note that the predicted performance data is not limited to a numerical value such as the predicted value 230, and may be anything that can identify the properties of the product. The predicted performance data may be represented, for example, by a number sequence, or by symbols (such as a circle or a cross) or characters (such as hard or soft).

[0036] [2-4. Dimensional Compression] The dimensionality reduction unit 50 reduces the dimensions of each of the plurality of training data 110 (training manufacturing condition vector data) and the input data 210 (input manufacturing condition vector data). "Dimensionality reduction" refers to the process of determining and calculating a new method for expressing certain data from the factors of the data, thereby reducing the number of factors. The dimensionality reduction unit 50 may reduce the dimensions of the plurality of training data 110 and the input data 210 using a given method, such as principal component analysis (PCA), independent component analysis (ICA), or t-Distributed Stochastic Neighbor Embedding (t-SNE). The dimensionality reduction unit 50 may be implemented primarily by the processor 11 of the performance prediction device 10, or may be implemented by a processor of another information processing device connected to the performance prediction device 10 via a network.

[0037] The dimensionality reduction unit 50 reduces the individual factors X1 to X2 contained in each of the plurality of training data 110 as shown in FIG. M New factors A1 to A mBy setting and calculating the above, dimension-reduced training manufacturing condition vector data 120 (hereinafter also referred to as training-time dimension-reduced data 120) is generated. As a result, the number of factors m included in the training-time dimension-reduced data 120 becomes a value smaller than the number of factors M of the training data 110 before the dimension reduction.

[0038] Furthermore, the dimensional compression unit 50 converts the individual factors X1 to X2 contained in the input data 210 into M New factors A1 to A m By setting and calculating the above, dimensionally compressed input manufacturing condition vector data 220 (hereinafter also referred to as input dimensionally compressed data 220) is generated. As a result, the number of factors m included in the input dimensionally compressed data 220 also becomes a value smaller than the number of factors M of the input data 210 before dimensionally compressed.

[0039] In the examples shown in FIGS. 3 and 4, the number of factors included in each of the plurality of training-time dimensional compression data 120 and the input-time dimensional compression data 220 is m, which is the same number, but the total number of factors included in each does not necessarily have to be the same.

[0040] Furthermore, when the machine learning model 40 is realized by a multi-layer neural network such as a DNN (Deep Neural Network), the dimensionality reduction unit 50 may input each of the plurality of training data 110 to the machine learning model 40, thereby acquiring the output of the intermediate layer of the machine learning model 40 as a plurality of training-time dimensionally reduced data 120 (dimensionally reduced training manufacturing condition vector data). Similarly, the dimensionality reduction unit 50 may input the input data 210 to the machine learning model 40, thereby acquiring the output of the intermediate layer of the machine learning model 40 as input-time dimensionally reduced data 220 (dimensionally reduced input manufacturing condition vector data). By using the intermediate layer of the machine learning model 40 in this way, it is possible to acquire a plurality of training-time dimensionally reduced data 120 and input-time dimensionally reduced data 220 without performing a calculation for dimensionality reduction, such as principal component analysis (PCA).

[0041] [2-5. Dimensional compression data storage unit] The dimensionality reduced data storage unit 60 stores multiple pieces of dimensionality reduced data during training 120. The dimensionality reduced data storage unit 60 may store the dimensionality reduced data during training 120 in association with the type of predicted performance data included in the dimensionality reduced data during training 120 (for example, the type of property such as tire rolling resistance or rubber hardness). The dimensionality reduced data storage unit 60 may be realized primarily by the storage unit 12 of the performance prediction device 10, or may be realized by another storage device such as an external storage device or a NAS (Network Attached Storage) connected to the performance prediction device 10 via a wired or wireless connection.

[0042] [2-6. Similarity determination section, output section] The similarity determination unit 70 determines whether the input dimensionality compressed data 220 (dimensionality compressed input manufacturing condition vector data) is similar to any of the multiple training dimensionality compressed data 120 (dimensionality compressed training manufacturing condition vector data). The output unit 80 outputs the predicted performance data calculated by the prediction unit 43 of the machine learning model 40, and also outputs information corresponding to the determination result of the similarity determination unit 70 as the accuracy of the predicted performance data. The similarity determination unit 70 and the output unit 80 may be realized mainly by the processor 11 of the performance prediction device 10, or may be realized by a processor of another information processing device connected to the performance prediction device 10 via a network. The output unit 80 may output the predicted performance data and the accuracy information of the predicted performance data by displaying the predicted performance data and the accuracy information of the predicted performance data on the display unit 14 of the performance prediction device 10 or another display device.

[0043] The similarity determination unit 70, for example, performs density estimation of the distribution of the plurality of training dimension-reduced data 120 and determines the similarity between the plurality of training dimension-reduced data 120 and the input dimension-reduced data 220 based on the results of this density estimation. The similarity determination unit 70 may perform density estimation at a point (position) identified by the input dimension-reduced data 220 within the distribution range of the plurality of training dimension-reduced data 120, and determine that the similarity is high (similar) if the estimated density is high, and low (dissimilar) if the density is low. For example, kernel density estimation (KDE) can be used for such density estimation. Alternatively, the similarity determination unit 70 may determine the similarity between the plurality of training dimension-reduced data 120 and the input dimension-reduced data 220 by utilizing the variance of predicted estimated data based on Gaussian process regression (GP) or the like.

[0044] Alternatively, the similarity determination unit 70 may determine the similarity between the plurality of pieces of dimensionally compressed data at training 120 and the dimensionally compressed data at input 220 by calculating the distance between each of the plurality of pieces of dimensionally compressed data at training 120 stored in the dimensionally compressed data storage unit 60 and the dimensionally compressed data at input 220. For example, the similarity determination unit 70 determines the similarity from the distance between each of the plurality of dimensionally compressed data at training 120 identified by IDs "1," "2," ... "N" in FIG. 3 and the dimensionally compressed data at input 220 identified by ID "1" in FIG. 4. The similarity determination unit 70 may determine that the similarity is high (similar) if the calculated distance is short, and that the similarity is low (dissimilar) if the calculated distance is long.

[0045] In this case, a plurality of factors A1 to A2 included in the training dimension compressed data 120 and the input dimension compressed data 220 in common are mThe similarity determination unit 70 may determine whether the data are similar or not based on the distance between the data 120 and 220. When the distance between the data 120 and 220 for a certain factor (for example, factor A1) is equal to or less than a threshold, the similarity determination unit 70 may determine that the data are similar for that factor, and when the distance exceeds the threshold, the similarity determination unit 70 may determine that the data are not similar for that factor. The similarity determination unit 70 may determine whether the data are similar or not based on three or more criteria such as match, similarity, and mismatch, or may calculate a similarity score according to the distance.

[0046] 5, the output unit 80 outputs a predicted value 230 as predicted performance data, and also outputs an accuracy value 240 as a value indicating the accuracy of the predicted performance data. In the example shown in Fig. 5, it is shown that the accuracy value 240 for the input data 210 identified by the ID "1" is "0.82", and the accuracy value 240 for the input data 210 identified by the ID "2" is "0.77". Note that the information indicating the accuracy of the predicted performance data is not limited to a numerical value such as the accuracy value 240, and may be indicated by symbols (circle, cross, etc.) or letters (high, low), etc.

[0047] In addition, when the similarity determination unit 70 determines the similarity for each of multiple pieces of training-time dimensional compression data 120 by calculating the distance from the input-time dimensional compression data 220, the output unit 80 may output information based on the similarity with the training-time dimensional compression data 120 that is closest to the input-time dimensional compression data 220 (i.e., the training-time dimensional compression data 120 that is determined to be most similar) as the accuracy of the predicted performance data.

[0048] [3. Processing flow] 6A, 6B, 7A, and 7B are diagrams showing an example of the flow of prediction processing performed by the performance prediction device 10. Fig. 6A and Fig. 6B show an example of processing performed when learning the machine learning model 40, and Fig. 7A and Fig. 7B show an example of processing performed when making predictions on the input data 210.

[0049] [3-1. Processing during learning] The processing performed during training of the machine learning model 40 will be described with reference to FIGS. 6A and 6B. First, the training data acquisition unit 20 acquires training data 110, which is manufacturing condition vector data used for training the machine learning model 40 (step S101). Next, the training unit 42 of the machine learning model 40 inputs the training data 110 acquired in step S101 to the machine learning model 40, thereby training the machine learning model 40 (step S102). The training unit 42 updates the parameters of the machine learning model 40 stored in the parameter storage unit 41 in response to the input of the training data 110 so that the prediction unit 43 of the machine learning model 40 outputs predicted performance data included in the training data 110 (e.g., the predicted value 230 shown in FIG. 3 ), thereby training the machine learning model 40. In step S102, the machine learning model 40 is input with the training data 110 including predicted performance data of a corresponding type.

[0050] Next, the dimensionality reduction unit 50 obtains dimensionally reduced training data 120 obtained by reducing the dimensions of the training data 110 obtained in step S101, stores the data in the dimensionally reduced data storage unit 60 (step S103), and ends the process. In step S103, the dimensionality reduction unit 50 may obtain the dimensionally reduced training data 120 by reducing the dimensions of the training data 110 using a given method such as principal component analysis (PCA), or may obtain the data output from the intermediate layer of the machine learning model 40 by the process of step S102 as the dimensionally reduced training data 120.

[0051] Furthermore, in step S103, the dimensionality reduction unit 50 may associate the training-time dimensionally reduced data 120 with the type of predicted performance data included therein (for example, the type of property such as tire rolling resistance or rubber hardness), and store the data in the dimensionality reduced data storage unit 60. By the processes in steps S101 to S103 described above, the dimensionality reduced data storage unit 60 stores a plurality of training-time dimensionally reduced data 120 associated with the type of predicted performance data each time training of the machine learning model 40 is performed.

[0052] [3-2. Processing during prediction] The process of making a prediction for input data 210 using the machine learning model 40 will be described with reference to FIGS. 7A and 7B. First, the input data acquisition unit 30 acquires input data 210, which is manufacturing condition vector data that the user desires to predict (step S201). The input data acquisition unit 30 acquires, for example, input data 210 specified by the user. Note that the number of input data 210 acquired in step S101 may be multiple as shown in FIG. 4, or may be one.

[0053] Next, the prediction unit 43 of the machine learning model 40 calculates, for example, the predicted value 230 shown in FIG. 5 as predicted performance data for the input data 210 acquired in step S201 based on the machine learning model 40 (step S202). The prediction unit 43 inputs the input data 210 to the machine learning model 40, thereby calculating predicted performance data for the input data 210. If the performance prediction device 10 has, as its functions, multiple machine learning models 40 each corresponding to a different type of predicted performance data, the prediction unit 43 may select, for example, a machine learning model 40 in accordance with a user instruction (a machine learning model 40 that outputs predicted performance data of a type that the user wishes to predict), and input the input data 210 to this machine learning model 40.

[0054] Next, the dimensionality reduction unit 50 acquires input-time dimensionally reduced data 220 obtained by dimensionally reducing the input data 210 acquired in step S201 (step S203). In step S203, the dimensionality reduction unit 50 may acquire the input-time dimensionally reduced data 220 by compressing the dimensions of the input data 210 using a given method such as principal component analysis (PCA), or may acquire data output from the intermediate layer of the machine learning model 40 by the processing of step S202 as the input-time dimensionally reduced data 220.

[0055] Next, the similarity determination unit 70 determines the similarity between the input dimensional compressed data 220 acquired in step S203 and the plurality of training dimensional compressed data 120 stored in the dimensional compressed data storage unit 60 in step S103 (step S204). In step S204, the similarity determination unit 70 may determine the similarity between the input dimensional compressed data 220 and the plurality of training dimensional compressed data 120 stored in the dimensional compressed data storage unit 60 and associated with the type of predicted performance data calculated in step S202 (the type corresponding to the machine learning model 40 used in the calculation).

[0056] In step S204, the similarity determination unit 70 may perform density estimation of the distribution of the plurality of training data sets 120 using kernel density estimation (KDE) or the like, and determine the similarity between the plurality of training data sets 120 and the input data set 220 based on the results of this density estimation. Alternatively, the similarity determination unit 70 may determine the similarity between the plurality of training data sets 120 and the input data set 220 by utilizing the variance of predicted estimated data based on Gaussian process regression (GP) or the like. Alternatively, the similarity between the input data set 220 and the plurality of training data sets 120 may be determined by calculating the distance between each of the plurality of training data sets 120 and the input data set 220.

[0057] 5 as the accuracy of the predicted value 230 (predicted performance data) calculated in step S202 and the predicted value 230 according to the similarity determination result determined in step S204, and then ends the process. Note that in step S204, when the similarity between the input dimensional compressed data 220 and the plurality of pieces of training dimensional compressed data 120 is determined by calculating the distance between each of the plurality of pieces of training dimensional compressed data 120 and the input dimensional compressed data 220, the output unit 80 may output, as the accuracy value 240, information based on the similarity with the training dimensional compressed data 120 that is closest to the input dimensional compressed data 220 (i.e., the training dimensional compressed data 120 determined to be most similar).

[0058] [4. Summary] Multiple factors X1 to X included in input data 210 M The likelihood of property prediction based on the above differs for each input data 210 (i.e., for each prediction). In this regard, the similarity determination unit 70 and the output unit 80 output the similarity between the input-dimensionally compressed data 220, which is the dimensionally compressed input data 210, and the plurality of training-dimensionally compressed data 120, which is the plurality of training data 110 that have also been dimensionally compressed, as the likelihood of predicted performance data, so that the user can know the likelihood of the predicted performance data for each prediction.

[0059] Furthermore, in this embodiment, the similarity determination unit 70 determines whether or not the input-time dimension-reduced data 220 obtained by dimensional compression from the input data 210 is similar to the multiple training-time dimension-reduced data 120 obtained by dimensional compression from the multiple training data 110. Therefore, the manufacturing conditions are determined based on a large number of factors (factors X1 to X M ) and the plurality of training data 110. In other words, the accuracy of prediction performance can be calculated in a shorter time than when the input data 210 and the plurality of training data 110 are directly compared.

[0060] [5. Modifications] The present invention is not limited to the above-described embodiments.

[0061] In the embodiment, the dimensionality reduction unit 50 acquires input-time dimensionally reduced data 220 obtained by dimensionally reducing the input data 210 and multiple training-time dimensionally reduced data 120 obtained by dimensionally reducing the multiple training data 110, and the similarity determination unit 70 determines the similarity between the multiple training-time dimensionally reduced data 120 and the input-time dimensionally reduced data 220. However, this is not limiting, and the similarity determination unit 70 may determine the similarity between data in which predetermined factors have been deleted from the multiple training data 110 and data in which predetermined factors have been deleted from the input data 210, and the output unit 80 may output information corresponding to the similarity determined in this manner as the accuracy of the predicted performance data predicted by the machine learning model 40 (more specifically, the prediction unit 43).

[0062] Fig. 8 is a functional block diagram showing another example of functions implemented in the performance prediction device 10. As shown in Fig. 8, the performance prediction device 10 may not include the dimensionality reduction unit 50 and the dimensionality reduced data storage unit 60 as functions, but may instead include a factor deletion unit 90 and a factor deletion data storage unit 100.

[0063] The factor deletion unit 90 deletes predetermined factors from each of the plurality of learning data 110 (learning manufacturing condition vector data) and the input data 210 (input manufacturing condition vector data). The factor deletion unit 90 may be realized mainly by the processor 11 of the performance prediction device 10, or may be realized by a processor of another information processing device connected to the performance prediction device 10 via a network.

[0064] The factor deletion unit 90 deletes a plurality of factors X1 to X2 contained in the learning data 110 acquired by the learning data acquisition unit 20. M (see FIG. 3 ), the factor deletion unit 90 deletes predetermined factors to reduce the number of factors included in the learning data 110. Similarly, the factor deletion unit 90 deletes a plurality of factors X1 to X2 included in the input data 210 acquired by the input data acquisition unit 30. MBy deleting certain factors from among them, the number of factors included in the learning data 110 is reduced. Note that it is desirable for the factor deletion unit 90 to delete factors that have a small influence on the calculation of predicted performance data by the prediction unit 43 of the machine learning model 40. The factor deletion unit 90 may calculate the influence on the calculation of predicted performance data by performing an analysis of variance or the like, and determine factors to be deleted based on this influence.

[0065] The factor-removed data storage unit 100 stores a plurality of training data 110 from which predetermined factors have been removed by the factor removal unit 90. The dimensionally reduced data storage unit 60 may store the training data 110 from which predetermined factors have been removed in association with the type of predicted performance data included in the training data 110 (for example, the type of property such as tire rolling resistance or rubber hardness). The dimensionally reduced data storage unit 60 may be realized primarily by the storage unit 12 of the performance prediction device 10, or may be realized by another storage device such as an external storage device or a NAS (Network Attached Storage) connected to the performance prediction device 10 via a wired or wireless connection.

[0066] The similarity determination unit 70 determines the similarity between a plurality of pieces of learning data 110 stored in the factor-deleted data storage unit 100 from which predetermined factors have been deleted by the factor deletion unit 90, and input data 210 from which predetermined factors have been deleted by the factor deletion unit 90. The similarity determination unit 70 may determine the similarity between the learning data 110 stored in the factor-deleted data storage unit 100 and the input data 210, the learning data 110 being associated with a type corresponding to the machine learning model 40 used to predict (calculate) the predicted performance data.

[0067] The output unit 80 outputs information according to the similarity determined as described above as the accuracy of the predicted performance data. M ) information and the plurality of learning data 110. In other words, the accuracy of prediction performance can be calculated in a short time. [Explanation of symbols]

[0068] 10 Performance prediction device, 11 Processor, 12 Memory unit, 13 Communication unit, 14 Display unit, 15 Operation unit, 20 Learning data acquisition unit, 30 Input data acquisition unit, 40 Machine learning model, 41 Parameter memory unit, 42 Learning unit, 43 Prediction unit, 50 Dimensionality reduction unit, 60 Dimensionality reduced data memory unit, 70 Similarity judgment unit, 80 Output unit, 90 Factor removal unit, 100 Factor-removed data memory unit, 110 Learning data, 120 Dimensionality reduced data during learning, 210 Input data, 220 Dimensionality reduced data during input, 230 Predicted value, 240 Accuracy value.

Claims

1. A machine learning model that is trained using multiple training manufacturing condition vector data representing the manufacturing conditions of a product, and outputs predictive performance data that shows the properties of a product manufactured under the manufacturing conditions indicated by a given input manufacturing condition vector data, A dimensionality reduction means for reducing the dimensionality of each of the aforementioned plurality of learning manufacturing condition vector data and the input manufacturing condition vector data, Similarity determination means that performs density estimation at points identified by the dimensionally compressed input manufacturing condition vector data within the distribution range of the dimensionally compressed plurality of training manufacturing condition vector data, and determines whether the dimensionally compressed input manufacturing condition vector data is similar to the dimensionally compressed plurality of training manufacturing condition vector data based on whether the density estimated by the density estimation is high or low, An output means that outputs the aforementioned prediction performance data and outputs information corresponding to the determination result of the similarity determination means as the accuracy of the prediction performance data, A performance prediction device characterized by including [a certain feature].

2. In the performance prediction device according to Claim 1, The dimensionality reduction means inputs each of the plurality of training manufacturing condition vector data and the input manufacturing condition vector data into the machine learning model, and obtains the output of the intermediate layer of the machine learning model as the dimensionality-reduced plurality of training manufacturing condition vector data and the input manufacturing condition vector data. A performance prediction device characterized by the following features.

3. In the performance prediction device according to claim 1 or 2, The aforementioned product is a tire, The plurality of learning manufacturing condition vector data and the input manufacturing condition vector data each include one or more elements representing the material of the tire and one or more elements representing the structure of the tire. The aforementioned predicted performance data indicates the performance of the tire. A performance prediction device characterized by the following features.

4. In the performance prediction device according to claim 1 or 2, The aforementioned product is tire rubber, The plurality of learning manufacturing condition vector data and the input manufacturing condition vector data each include one or more elements representing the rubber material and one or more elements representing the rubber manufacturing conditions. The aforementioned predictive performance data indicates the physical properties of the rubber. A performance prediction device characterized by the following features.

5. A machine learning model that is trained using a plurality of training manufacturing condition vector data representing the manufacturing conditions of a product, and outputs predictive performance data that represents the properties of a product manufactured under the manufacturing conditions represented by a given input manufacturing condition vector data when given input manufacturing condition vector data. Dimensionality reduction means for reducing the dimensionality of each of the plurality of learning manufacturing condition vector data and the input manufacturing condition vector data, Similarity determination means that performs density estimation at points identified by the dimensionally compressed input manufacturing condition vector data within the distribution range of the dimensionally compressed plurality of training manufacturing condition vector data, and determines whether the dimensionally compressed input manufacturing condition vector data is similar to the dimensionally compressed plurality of training manufacturing condition vector data based on whether the density estimated by the density estimation is high or low, and An output means that outputs the aforementioned prediction performance data and outputs information corresponding to the determination result of the similarity determination means as the accuracy of the prediction performance data, A program that makes a computer function.

Images