Performance prediction method, storage medium and device

Abstract

The present invention addresses the problem of providing a performance prediction method, a storage medium, and a device that improve the accuracy of performance prediction. [Solution] A performance prediction method according to one embodiment of the present disclosure is provided with a conversion step, a training step, and a prediction step. The conversion step converts one-dimensional arrangement data including a plurality of parameter values arranged in a one-dimensional manner into two-dimensional arrangement data including a plurality of parameter values arranged in a two-dimensional manner by copying and disposing the plurality of parameter values. The training step trains the machine learning model on the basis of training data including two-dimensional arrangement data and performance data corresponding to the one-dimensional arrangement data, and generates a learned model in which the two-dimensional arrangement data is input and predicted performance data is output. The prediction step applies the two-dimensional arrangement data of the object to be processed to the learned model, and generates predicted performance data of the object to be processed.

Description

[0001] Cross-references to related applications

[0002] This application is based on and claims priority to Japanese Patent Application No. 2024-151334, filed on September 3, 2024, the entire contents of which are incorporated herein by reference. Technical Field

[0003] Embodiments of the present invention relate to performance prediction methods, storage media, and apparatus. Background Technology

[0004] There are neural network models that predict performance parameters such as airflow, static pressure, and efficiency based on blade shape parameters related to the blower blades, such as blade size. Since the blade shape parameters are input into the neural network as a one-dimensional arrangement of data, the network sometimes fails to learn the local correlations between these parameters. Therefore, there is room for improvement in the accuracy of performance prediction.

[0005] Existing technical documents:

[0006] Patent Document 1: Japanese Patent Application Publication No. 2023-028246

[0007] Patent Document 2: Japanese Patent Application Publication No. 2020-126601

[0008] Patent Document 3: Japanese Patent Publication No. 2022-500745

[0009] Non-patent document 1: "Surface Air Blower Utilization Performance Prediction Technology", Toshiba Review Volume 78 No. 2 (March 2023), page 24 Summary of the Invention

[0010] The present invention was made to solve the above-mentioned problems, and its purpose is to provide a performance prediction method, storage medium and device that improve the accuracy of performance prediction.

[0011] The performance prediction method of this implementation includes a conversion step, a training step, and a prediction step. In the conversion step, one-dimensional data comprising a one-dimensional arrangement of multiple parameter values ​​is converted into two-dimensional data comprising a two-dimensional arrangement of the multiple parameter values ​​by copying and configuring all or part of the parameter values. The training step trains a machine learning model based on training data including the two-dimensional data and performance data corresponding to the one-dimensional data, generating a fully trained model that takes the two-dimensional data as input and outputs predicted performance data. The prediction step applies the two-dimensional data of the object to be processed to the fully trained model to generate predicted performance data for the object.

[0012] The performance prediction method described above can improve the accuracy of performance prediction. Attached Figure Description

[0013] Figure 1 This is a diagram showing the hardware configuration of a performance prediction device.

[0014] Figure 2 This is a diagram showing the processing order of performance prediction.

[0015] Figure 3 It is a diagram that schematically represents the transformation from one-dimensional permutation parameter values ​​to two-dimensional permutation data.

[0016] Figure 4 This is a diagram representing the network structure of a CNN model.

[0017] Figure 5 It is a diagram that shows the processing order of transforming data from one-dimensional arrangement to two-dimensional arrangement.

[0018] Figure 6 It is a schematic representation Figure 5 A diagram illustrating a specific example of the processing sequence.

[0019] Figure 7 This is a diagram illustrating the process of comparing the prediction accuracy of CNN and DNN models.

[0020] Figure 8 This is a diagram illustrating the training of the DNN model in Example 1.

[0021] Figure 9 This is a diagram illustrating the training of the CNN model in Example 1.

[0022] Figure 10 This is a graph representing the prediction accuracy of the DNN model in Example 1.

[0023] Figure 11 This is a graph representing the prediction accuracy of the CNN model in Example 1.

[0024] Figure 12 It is a graph showing the relationship between the predicted and true values ​​of the performance data of the CNN-based model in Example 2, arranged by each pattern and by performance parameter values ​​(pressure, efficiency, and noise).

[0025] Figure 13 This is a diagram illustrating the training of the one-dimensional CNN model in Example 3.

[0026] Figure 14 It is a graph showing the relationship between the predicted and true values ​​of the performance data of Example 3, for each model and according to the performance parameter values ​​(pressure, efficiency, and noise).

[0027] Figure 15This is a diagram illustrating the process of comparing the prediction accuracy of the CNN model and the DNN model in Example 4.

[0028] Figure 16 This is a graph showing the prediction accuracy of the DNN model and CNN model in Example 4.

[0029] Figure 17 This is a diagram showing the processing order of the model training process for the variant example. Detailed Implementation

[0030] The performance prediction method of this implementation includes a conversion step, a training step, and a prediction step. In the conversion step, one-dimensional data comprising a one-dimensional arrangement of multiple parameter values ​​is converted into two-dimensional data comprising a two-dimensional arrangement of the multiple parameter values ​​by copying and configuring all or part of the parameter values. The training step trains a machine learning model based on training data including the two-dimensional data and performance data corresponding to the one-dimensional data, generating a fully trained model that takes the two-dimensional data as input and outputs predicted performance data. The prediction step applies the two-dimensional data of the object to be processed to the fully trained model to generate predicted performance data for the object.

[0031] Hereinafter, the performance prediction method, storage medium, and apparatus of this embodiment will be described with reference to the accompanying drawings.

[0032] Figure 1 This is a diagram illustrating the hardware configuration of the performance prediction device 1 in this embodiment. (As shown...) Figure 1 As shown, the performance prediction device 1 is a computer having a processor 11, ROM (Read Only Memory) 12, RAM (Random Access Memory) 13, auxiliary storage device 14, input device 15, display device 16, and communication device 17. Data and various signal transmission and reception between the processor 11, ROM 12, RAM 13, auxiliary storage device 14, input device 15, display device 16, and communication device 17 are performed via a bus.

[0033] Processor 11 is an integrated circuit that controls the overall operation of performance prediction device 1. For example, processor 11 may have a CPU (Central Processing Unit), GPU (Graphics Processing Unit), DSP (Digital Signal Processor), and / or FPU (Floating-Point Unit). Processor 11 may also have internal memory and I / O interfaces. Processor 11 executes various processes by interpreting and processing pre-stored programs such as ROM 12 or auxiliary storage device 14. Furthermore, part or all of processor 11 may be implemented using hardware such as ASIC (Application Specific Integrated Circuit) or FPGA (Field Programmable Gate Array).

[0034] ROM 12 is a non-volatile memory that stores various types of data. For example, ROM 12 stores data, settings, etc., used by processor 11 when performing various processes. ROM 12 may also be a non-transitory computer-readable storage medium that stores programs executed by processor 11.

[0035] RAM13 is a volatile memory used for reading and writing data. RAM13 temporarily stores data used by the processor 11 while it performs various processes. RAM13 provides the working area for the processor 11.

[0036] Auxiliary storage device 14 is a non-volatile memory that stores various types of data. For example, auxiliary storage device 14 stores data used by processor 11 during various processes, set values, and data generated by various processes within processor 11. Auxiliary storage device 14 may be composed of HDD (Hard Disk Drive), SSD (Solid State Drive), integrated circuit storage devices, etc. Furthermore, auxiliary storage device 14 may also have a non-transitory computer-readable storage medium that stores programs executed by processor 11.

[0037] Input device 15 accepts various inputs from the operator. Input device 15 can be a keyboard, mouse, various switches, touchpad, touch panel display, etc. An electrical signal corresponding to the accepted input is supplied to processor 11.

[0038] Display device 16 displays various data under the control of processor 11. Display device 16 can be a CRT (Cathode-Ray Tube) display, liquid crystal display, organic EL (Electroluminescence) display, LED (Light-Emitting Diode) display, plasma display, or any other type of display. Display device 16 can also be a projector.

[0039] The communication device 17 includes a communication interface such as a network interface card (NIC) for data communication with various devices connected to the performance prediction device 1 via a network. Furthermore, electrical signals can be supplied from a computer connected to the communication device 17 or an input device of that computer, and various data can be displayed on a display device or the like of a computer connected to the communication device 17. However, for the sake of simplicity in the following description, unless otherwise specified, the source of the electrical signals is assumed to be the input device 15, and the destination for displaying various data is assumed to be the display device 16. The input device 15 can be replaced by a computer connected to the communication device 17 or an input device of that computer, and the display device 16 can be replaced by a display device or the like of a computer connected to the communication device 17.

[0040] The performance prediction device 1 does not need to include all of the processor 11, ROM 12, RAM 13, auxiliary storage device 14, input device 15, display device 16, and communication device 17. Alternatively, it may omit some of the ROM 12, RAM 13, auxiliary storage device 14, input device 15, display device 16, and communication device 17 as needed. Any additional hardware devices useful for executing the processing of this embodiment may also be included in the performance prediction device 1. The performance prediction device 1 does not need to be physically composed of a single computer; it may be composed of a computer system including multiple computers that are communicatively connected via wired or network lines. The allocation of the series of processes involved in this embodiment to the multiple processors 11 installed on the multiple computers can be arbitrarily set. All processors 11 can execute all processes in parallel, or specific processes can be assigned to one or a portion of the processors 11, and the series of processes involved in this embodiment can be executed as a whole by the computer system.

[0041] Processor 11 executes at least a conversion process, a training process, and a prediction process. In the conversion process, one-dimensional data comprising multiple parameter values ​​is converted into two-dimensional data comprising multiple parameter values ​​by copying and configuring multiple parameter values. The one-dimensional data is input data comprising multiple parameter values ​​arranged in a one-dimensional order. Multiple parameter values ​​mean multiple values ​​corresponding to process parameters for the target product. Performance data includes the values ​​of one or more performance parameters that evaluate the performance of the product.

[0042] In the training process, a machine learning model is trained based on training data including two-dimensional permutation data and performance data corresponding to the one-dimensional permutation data. This generates a fully trained model that takes the two-dimensional permutation data as input and outputs predicted performance data. Predicted performance data refers to the performance data output from the fully trained model. Performance data and / or predicted performance data include parameter values ​​for one or more performance parameters used to evaluate the product's performance.

[0043] Machine learning models are typically envisioned as CNN (Convolutional Neural Network) models. In this embodiment, a CNN model means a neural network model with at least convolutional layers. A CNN model can also have network layers other than those mentioned above, such as pooling layers and fully connected layers. Furthermore, the training process can also train a DNN (Deep Neural Network) model as a machine learning model. In this embodiment, a DNN model means a neural network model with fully connected layers but no convolutional layers.

[0044] In the prediction process, the two-dimensional permutation data of the processed object is applied to the learned model to generate predictive performance data for the processed object. The two-dimensional permutation data of the processed object only needs to be converted from the one-dimensional permutation data of the processed object through the transformation process.

[0045] The following describes the processing sequence of the performance prediction process executed by the processor 11 of the performance prediction device 1.

[0046] Figure 2 This is a diagram showing the processing order of performance prediction. Figure 2 The performance prediction method shown begins by the processor 11 reading a performance prediction program from an auxiliary storage device 14, etc., and interpreting and executing the read performance prediction program.

[0047] First, the processor 11 obtains one-dimensional permutation data and the corresponding performance data (step SA1). The combination of the one-dimensional permutation data and the performance data constitutes training samples for training the machine learning model. In step SA1, multiple training samples are obtained.

[0048] When performing step SA1, the processor 11 converts the one-dimensional permutation data into two-dimensional permutation data (step SA2). Step SA2 is performed on multiple training samples separately. In step SA2, the processor 11 copies and configures all of the multiple process parameter values, thereby converting them into two-dimensional permutation data in which the multiple process parameter values ​​are arranged in two dimensions.

[0049] Figure 3 This is a schematic diagram illustrating the transformation from one-dimensional arrangement parameters to two-dimensional arrangement data. Here, process parameter values ​​are labeled "X". n “n” is the subscript for identifying the type of process parameter, and is 1≤n≤N. “N” is the total number of types of process parameters, and is a natural number greater than 2. Figure 3 This example illustrates the case where the total number of process parameter types is 4 (N=4).

[0050] like Figure 3 As shown, a one-dimensional arrangement parameter is N-dimensional data comprising N process parameter values ​​Xn arranged in a one-dimensional arrangement. In this embodiment, the arrangement direction of the process parameter values ​​Xn in the one-dimensional arrangement parameter is set to the horizontal column (row) direction. In this case, "N" means the number of elements in the horizontal direction of the one-dimensional arrangement parameter, and the one-dimensional arrangement data has N process parameter values ​​Xn arranged along the horizontal column direction without repetition of parameter types.

[0051] Two-dimensional arrangement data consists of N × M parameter values ​​Xn arranged in a two-dimensional space defined by the horizontal direction and the vertical (column) direction orthogonal to that horizontal direction. "M" signifies the replication number of the N process parameter values ​​Xn; in other words, it represents the number of elements in the vertical direction of the two-dimensional arrangement data. "M" is a natural number greater than 2. Figure 3 As a column, let 4 be the same as "N".

[0052] In step SA2, the processor 11 can copy multiple process parameter values ​​Xn in the column direction and rearrange them in the horizontal direction. At this time, the process parameter values ​​Xn can be configured in a way that prevents the parameter types from repeating in both the column and horizontal directions. For example, Figure 3The process parameter values ​​Xn in the first row shown are arranged from left to right as (X2, X3, X1, X4), but this arrangement does not exist in any of the second, third, or fourth rows. In other examples, the process parameter values ​​Xn in the second column are arranged from top to bottom as (X3, X1, X2, X4), but this arrangement does not exist in any of the first, third, or fourth columns. By avoiding repetition of parameter types in both rows and columns, the variation in the arrangement pattern of parameter types within the applicable range (receptive field) of the convolutional layers of the CNN model described later can be increased, thereby improving the prediction accuracy of the CNN model. Furthermore, the details of the transformation process from one-dimensional arranged data to two-dimensional arranged data will be described later.

[0053] The permutation patterns of the parameter types associated with the two-dimensional permutation data are common to all training samples. The permutation patterns of the parameter types are stored in auxiliary storage device 14.

[0054] When performing step SA2, processor 11 trains a CNN model based on two-dimensional permutation data and performance data (step SA3). First, the network structure of the CNN model is explained.

[0055] Figure 4 This is a diagram representing the network structure of a CNN model. For example... Figure 4 As shown, a CNN model takes two-dimensionally arranged data as input and outputs prediction performance data. A CNN model consists of an input layer, convolutional layers, pooling layers, a flattening layer, and an output layer.

[0056] The input layer takes in two-dimensional arranged data. The convolutional layer uses multiple filters to convolve the input data, outputting a two-dimensional distribution of feature values, i.e., multiple feature maps. The number of output feature maps is the same as the number of filters. The receptive field 41 of each filter is also called the receptive field. For each of the multiple filters, the convolutional layer applies the filter to the input data while varying the receptive field 41 by a specified stride to perform the convolution operation. The convolutional layer extracts, emphasizes, or suppresses the local two-dimensional arrangement patterns specific to each filter from the two-dimensional arrangement of parameter values ​​included in the receptive field 41. The matrix size of the receptive field 41 is not limited to 2×2 and can be set to any size. The weights at multiple locations within the same filter are shared.

[0057] The pooling layer, while modifying the pooling region 42 of the feature map output from the convolutional layer, compresses the feature values ​​within that region using a prescribed compression method, generating a two-dimensional distribution of compressed values, i.e., a compressed map. As an example of a compression method, average pooling, which calculates the average of multiple feature values ​​within the pooling region 42, can be used. Furthermore, the compression method is not limited to average pooling; other arbitrary compression methods, such as max pooling (calculating the maximum value of multiple feature values ​​within the pooling region 42) or min pooling (calculating the minimum value), can also be used.

[0058] The flattening layer converts the two-dimensional compressed map into a one-dimensional flattening process, generating one-dimensional compressed value data. The output layer performs performance calculations based on the one-dimensional compressed value data, generating predicted performance data Y. The predicted performance data Y is output as a vector of parameter values ​​with one or more performance parameters.

[0059] The CNN model described above is just one example; it can be any network configuration capable of performing convolutional processing on two-dimensional data and outputting prediction performance data. For instance, a CNN model can also have a configuration that includes concatenated blocks of convolutional and pooling layers. Additionally, to expand the matrix size of the feature map or compressed map, zero-padding network layers can be used.

[0060] In step SA3, the processor 11 performs supervised learning on the CNN model based on multiple training samples to generate a fully learned CNN model. Specifically, the two-dimensional arrangement data of the multiple training samples are input into the CNN model before it has finished learning, and convolution processing based on convolutional layers, compression processing based on pooling layers, flattening processing based on flattening layers, and performance calculation processing based on the output layer are performed in sequence to calculate the prediction performance data.

[0061] Next, processor 11 calculates the error between the predicted performance data and the correct prediction data included in the training samples, and updates the network parameters of the CNN model based on the calculated error. Specifically, the network parameters include weights or biases assigned to network layers such as convolutional layers. Processor 11 repeatedly performs the calculation of predicted performance data and the updating of network parameters while changing the training samples until a specified termination condition is met. The termination condition can be set as the number of iterations reaching a specified value, the accuracy of the predicted performance data reaching a specified value, or any other arbitrary condition. The optimization method can use stochastic gradient descent, Adam, or any other arbitrary method. The learning parameters such as learning rate, batch size, and activation function can also be arbitrarily set.

[0062] If the termination condition is met, the network parameters from that number of iterations are saved as fully learned network parameters in auxiliary storage device 104. The CNN model with the fully learned network parameters is used as a fully learned CNN model. The fully learned CNN model can also be saved in auxiliary storage device 104.

[0063] When step SA3 is performed, processor 11 obtains one-dimensional permutation data of the objects to be processed (step SA4). The one-dimensional permutation data of the objects to be processed means the one-dimensional permutation data of the objects to be calculated for predicting performance data.

[0064] During step SA4, the processor 11 converts the one-dimensional permutation data of the object to be processed into two-dimensional permutation data (step SA5). In step SA5, the processor 11 reads the permutation pattern of parameter types from the auxiliary storage device 14, and converts the one-dimensional permutation data of the object to be processed into two-dimensional permutation data according to the read permutation pattern. This allows the use of a common permutation pattern during training and inference, improving the computational accuracy of prediction performance data. The converted two-dimensional permutation data is referred to as the two-dimensional permutation data of the object to be processed.

[0065] When performing step SA5, the processor 11 applies the two-dimensional arrangement data of the object to the learned CNN model to generate the prediction performance data of the object (step SA6). In step SA6, the processor 11 reads the learned CNN model from the auxiliary storage device 14, inputs the two-dimensional arrangement data of the object to the read learned CNN model, and sequentially performs convolution processing based on convolutional layers, compression processing based on pooling layers, flattening processing based on flattening layers, and performance calculation processing based on the output layer to calculate the prediction performance data of the object.

[0066] When performing step SA6, processor 11 displays the predicted performance data of the processed object on display device 16 (step SA7). The method of displaying the predicted performance data of the processed object is not particularly limited. Processor 11 may also display one-dimensional and / or two-dimensional arrangement data of the processed object together with the predicted performance data. Alternatively, processor 11 may append labels indicating that the predicted performance data of the processed object is predicted by a CNN model to the predicted performance data of the processed object.

[0067] The performance prediction process for processor 11 is now complete.

[0068] Next, the details of the conversion process from one-dimensional permutation data to two-dimensional permutation data performed by the processor 11 will be explained. In the conversion process, M permutations are generated sequentially by arranging N parameter values ​​in M ​​columns of a two-dimensional space defined by the first and second directions according to a permutation order based on a prescribed algorithm. Here, the parameter values ​​of the first or second permutation are rearranged in such a way that the parameter types do not overlap in a prescribed direction other than the first direction (hereinafter, the checking direction) between adjacent first and second permutations, thereby generating two-dimensional permutation data.

[0069] Figure 5 It is a diagram illustrating the process of converting one-dimensional data into two-dimensional data. Figure 6 It is a schematic representation Figure 5 A diagram illustrating a specific example of the processing procedure is shown. Figure 5 and Figure 6 In the middle, the first direction is horizontal, the second direction is vertical, and the inspection direction is vertical. Figure 6 In this configuration, the total number of parameter types N is 4, and the total number of columns M is 4. That is, the two-dimensional space defined by the first and second directions has 4×4 pixels.

[0070] like Figure 5 As shown, processor 11 will arrange the N process parameter values ​​X of the one-dimensional data. n Arranged in the first row (step SB1). The process parameter values ​​X arranged in the first row. n The order can be any order, including the original order. For example, such as... Figure 6 As shown in Figure 1, from column 1 to column 4, N process parameter values ​​X are arranged in the first horizontal column in the order of (X2, X3, X1, X4). n .

[0071] When performing step SB1, processor 11 uses a random number generator to generate N process parameter values ​​X n Copy and configure it in the m-th row (step SB2). m is set sequentially from 2 to M. The initial setting is m = 2. When m = 2, the processor 11 generates a process parameter value X of any one of the four parameter types using a random number generator. n The generated process parameter value X n The blank pixels are arranged in the first column of the second row. Next, the processor 11 generates a process parameter value X for any of the remaining three parameter types using a random number generator. n The generated process parameter value X n The blank pixels are configured in the second column of the second row. Similarly, the processor 11 configures the process parameter value X in the blank pixels of the third column and the fourth column of the second row.n Therefore, the parameter value X is configured in the second row according to the arrangement order of the random number generator. n The result is as follows: Figure 6 As shown in Figure 2, N parameter values ​​X are arranged in the second row in the order of (X1, X3, X4, X2). n .

[0072] When performing step SB2, processor 11 determines whether there are duplicate parameter types in the (m-1)th and mth columns in the vertical direction (checking direction) (step SB3). Duplicates are checked in all columns. For example, such as... Figure 6 As shown in Figure 2, regarding the second column, the parameter type of both the first and second rows is X3, which is the same, therefore it is determined that there is a duplicate. On the other hand, regarding the other first, third, and fourth columns, the parameter types of the first and second rows are different, therefore it is determined that there is no duplicate.

[0073] If it is determined that there is a repetition in at least one column (step SB4: yes), the processor 11 rearranges the process parameter value X of the m-th column in a manner to avoid repetition in the vertical direction. n (Step SB5). There are no particular limitations on the method of rearranging. As an example, processor 11 can rearrange the N process parameter values ​​X in the second row. n The process parameter value X can be rearranged again according to the order of the random number generator, or it can be replaced between pixels identified as duplicates and pixels that are horizontally adjacent to those pixels. n For example, such as Figure 6 As shown in Figure 3, the process parameter value X is replaced in the pixels of the first column and the second column of the second horizontal column. n To avoid duplication with the types of parameters in the vertical direction.

[0074] If step SB5 has been performed or if it is determined that there are no duplicates (step SB4: No), processor 11 determines whether the m-th column of the processed object is the last M-th column, i.e., whether m = M (step SB6). If it is determined that m = M (step SB6: No), processor 11 shifts the m-th column of the processed object downwards by 1 column, i.e., executes m = m + 1 (step SB7). Processor 11 repeats steps SB2 to SB6 similarly for the m-th column of the processed object until it is determined in step SB6 that the m-th column of the processed object is the last M-th column. Therefore, two-dimensional permutation data is output.

[0075] For example, such as Figure 6 As shown in Figure 4, the processor 11 replicates N process parameter values ​​X using a random number generator. nAnd it is configured in the third row, and in step SB3, it is determined whether there is a repetition of parameter type in the vertical direction (step SB2). Figure 6 In case 4, it is determined that there is a duplicate of parameter types in the second and fourth columns (step SB4: yes), therefore, as Figure 6 As shown in Figure 5, processor 11 pairs the N process parameter values ​​X in the third row. n Rearrange (step SB5). Then, as... Figure 6 As shown in Figure 6, the processor 11 replicates N process parameter values ​​X using a random number generator. n And configure it in the last fourth row (step SB2). In Figure 6 In case 6, it is determined that there is no vertical repetition in all columns (step SB4: No).

[0076] Then, if it is determined that the m-th column of the object to be processed is the last M-th column (step SB6: yes), the processor 11 outputs the two-dimensional arrangement data (step SB8).

[0077] Therefore, the conversion process from one-dimensional permutation data to two-dimensional permutation data is complete.

[0078] By configuring N parameter values ​​X in two-dimensional space according to the above algorithm. n This method can generate two-dimensional data with unique parameter types in both the horizontal and vertical directions. By avoiding repetition of parameter types in both directions, the applicability of convolution processing can be improved. (41) n The diversity of the arrangement can improve the prediction accuracy of performance data.

[0079] Furthermore, the processing order of the above conversion is just one example; it can perform deletion, addition, and / or modification of various procedures. As an example, with N parameter values ​​X... n The first column is not limited to the first column; it can be the Mth column or any other column. As another example, N parameters X... n It can be configured along the horizontal column, but it can also be configured along the vertical column.

[0080] Regarding other viewpoints, in the above conversion process, the direction for checking for duplicate parameter types is vertical, but it can also be a diagonal direction. Furthermore, the checking direction is not limited to one direction; it can be two directions, such as vertical and diagonal. The range for checking for duplicate parameter types (hereinafter, the duplication detection range) is not limited to two adjacent pixels in the checking direction as described above; it can also be set to any rectangular area consisting of a number of pixels less than or equal to the number of parameter types. In this case, the processor 11 arranges multiple parameter values ​​in the two-dimensional arrangement data in a way that the arrangement pattern of parameter types included in a rectangular area of ​​a specified size does not repeat at any position. For example, when N=4, a square area of ​​2 pixels vertically × 2 pixels horizontally can be set as the duplication detection range. In this case, duplicate parameter types can be avoided within the duplication detection range, thus, for example, avoiding duplicate parameter types only in the vertical direction. Figure 5 and Figure 6 The algorithm allows for repetition of tilt directions.

[0081] Next, an embodiment of this implementation will be described.

[0082] <Example 1>

[0083] Example 1 compares the prediction accuracy of performance data between CNN models and DNN models.

[0084] Figure 7 This diagram schematically illustrates the processing steps for comparing the prediction accuracy of CNN and DNN models. The product in Example 1 is a centrifugal blower. For example... Figure 7 As shown, the process parameters are five: air volume Q, outlet angle θ, number of blades N, blade length L, and blade height h. The performance parameters are three: pressure, efficiency, and performance. Based on the five process parameter values, the three performance parameter values ​​are derived respectively.

[0085] A total of 659 samples were prepared, consisting of combinations of process and performance parameter values. 454 of these were used as training samples, and the remaining 165 were used as test samples. DNN and CNN models were trained based on the training samples, and the prediction accuracy of the DNN and CNN models was verified based on the test samples.

[0086] Figure 8 This is a diagram illustrating the training process of the DNN model in Example 1. (See diagram for details.) Figure 8As shown, the DNN model has two intermediate layers. The intermediate layers are fully connected layers with 7 nodes. Weights, serving as network parameters, are assigned to the connections between nodes. The DNN model is trained using supervised learning based on 454 training samples, taking five process parameter values ​​X (airflow Q, outlet angle θ, number of blades N, blade length L, blade height h) as input and outputting three performance parameter values ​​Y (pressure, efficiency, noise). The learning rate is 0.05, the batch size is 26, and the activation function for the intermediate layers is ReLU.

[0087] Figure 9 This is a diagram illustrating the training process of the CNN model in Example 1. (See diagram for details.) Figure 9 As shown, the CNN model has convolutional layers and pooling layers. Two blocks are connected as a combination of convolutional and pooling layers. The first block has 16 filters in the convolutional layers, and the second block has 32 filters. The receptive field matrix size of the convolutional layers is 3×3, and the pooling region matrix size of the pooling layers is 2×2. The CNN model is trained using supervised learning based on 454 training samples, taking five process parameter values ​​X (airflow Q, outlet angle θ, number of blades N, blade length L, blade height h) as input and outputting three performance parameter values ​​Y (pressure, efficiency, noise). The input two-dimensional permutation data has a pattern of permutations (distributed permutation) with no repetition of parameter types in the vertical and horizontal directions. Repetition is allowed in a portion of the diagonal direction. The learning rate is 0.05, the batch size is 26, and the ReLU activation function is used.

[0088] Figure 10 This is a graph representing the prediction accuracy of the DNN model. Figure 11 This is a graph representing the prediction accuracy of the CNN model. Figure 10 and Figure 11 The 1 represents the coefficient of determination R for each training iteration using training samples, categorized by performance parameter type. 2 The curve graph, Figure 10 and Figure 11 The 2 represents the coefficient of determination R for each training iteration using the test samples, categorized by performance parameter type. 2 The curve graph. Figure 10 and Figure 11 Figure 3 is a graph showing the relationship between the predicted and true values ​​of pressure, which is a performance parameter. Figure 10 and Figure 11 Figure 4 is a graph showing the relationship between the predicted and true values ​​of efficiency, which is a performance parameter. Figure 10 and Figure 11 Figure 5 is a graph showing the relationship between the predicted values ​​and the true values ​​of noise, which are performance parameter values. Furthermore, the true value represents the correct prediction data within the sample.

[0089] WillFigure 10 and Figure 11 A comparison reveals that, between the training and test samples, the coefficient of determination R0 of the CNN model is higher for all performance parameter values. 2 Both are better than DNN models. In particular, the noise determination coefficient R0 is lower. 2 The value is 0.8723 in the DNN model, but 0.9124 in the CNN model, showing a significant improvement.

[0090] <Example 2>

[0091] Example 2 uses a CNN model to verify the dependence of the arrangement pattern of parameter types in two-dimensional permutation data on prediction accuracy. The product, process parameters, and performance parameters in Example 2 are the same as in Example 1.

[0092] Figure 12 It is a graph that shows the relationship between predicted and true values ​​of performance data based on CNN models, arranged by each pattern and by performance parameter type (pressure, efficiency, and noise). Figure 12 The upper part of the arrangement pattern is a dispersed arrangement. A dispersed arrangement is an arrangement pattern in which there is no repetition of parameter types in the vertical and horizontal directions, but the repetition of parameter types is limited to a part of the diagonal direction. Figure 12 The middle section of the pattern is arranged horizontally. A horizontal arrangement is a pattern in which there is no repetition of parameter types in the vertical and diagonal directions, but there is repetition of parameter types in the horizontal direction. Figure 12 The lower section of the pattern is arranged diagonally. A diagonal arrangement is a pattern in which there is no repetition of parameter types in the horizontal and vertical directions, but there is repetition of parameter types in all diagonal directions.

[0093] If with Figure 12 Comparing the performance parameter values ​​of the DNN model with those of the distributed arrangement, the prediction accuracy for noise is improved for any arrangement pattern. For the dispersed arrangement, the prediction accuracy for all performance parameters related to pressure, efficiency, and noise is improved compared to the DNN model, especially for efficiency, where the prediction accuracy is significantly improved. Therefore, it can be concluded that the dispersed arrangement has higher prediction accuracy compared to the horizontal and tilted arrangements.

[0094] <Example 3>

[0095] Example 3 uses a CNN model to verify the difference in prediction accuracy between one-dimensional and two-dimensional permutation data. The product, process parameters, and performance parameters in Example 3 are the same as in Example 1.

[0096] Figure 13 This is a diagram illustrating the training of the one-dimensional CNN model in Example 3. (See diagram for details.) Figure 13 As shown, a one-dimensional CNN model and Figure 9The illustrated 2D CNN model also features an input layer, convolutional layers, pooling layers, a flattening layer, and an output layer. The input layer takes five process parameter values ​​X (airflow Q, outlet angle θ, number of blades N, blade length L, and blade height h) arranged in a one-dimensional array. The convolutional layers perform one-dimensional convolutions. The receptive field matrix size of the convolutional layer is 3. The pooling layers perform one-dimensional pooling. The pooling region matrix size of the pooling layer is 2. The two convolutional and pooling layers are concatenated. The number of filters in both the first and second blocks of the convolutional layers is 64. The 1D CNN model is trained using supervised learning based on 454 training samples, taking the five process parameter values ​​X (airflow Q, outlet angle θ, number of blades N, blade length L, and blade height h) arranged in a one-dimensional array as input and outputting three performance parameter values ​​Y (pressure, efficiency, and noise) as output. The learning rate is 0.05, the batch size is 26, and the activation function used is ReLU.

[0097] Figure 14 It is a graph showing the relationship between predicted and true values ​​of performance data for each model and according to performance parameter values ​​(pressure, efficiency, and noise). Figure 14 The model of the upper section is Figure 8 The DNN model shown is Figure 14 The middle section model is Figure 13 The one-dimensional CNN model shown is Figure 14 The model of the lower segment is Figure 9 The two-dimensional CNN model shown is illustrated. If compared... Figure 14 In the upper and middle sections, the one-dimensional CNN model only improves accuracy in terms of noise compared to the DNN model. If we compare... Figure 14 In the middle and lower sections, the accuracy of pressure, efficiency, and noise is improved in the two-dimensional CNN model compared to the one-dimensional CNN model. One reason for this is that, since the two-dimensional model takes one-dimensional arrangement data as input, while the one-dimensional model takes two-dimensional arrangement data as input, the two-dimensional model can extract various local arrangement patterns of process parameters compared to the one-dimensional model.

[0098] <Example 4>

[0099] Example 4 verifies the prediction accuracy of the DNN and CNN models for products, process parameters, and performance parameters that are different from those in Examples 1, 2, and 3.

[0100] Figure 15 This diagram schematically illustrates the comparison of the prediction accuracy of the CNN model and the DNN model in Example 4. The product of Example 4 is concrete. Figure 15As shown, the process parameters are the mass fractions of component 1 (cement), component 2 (blast furnace slag), component 3 (fly ash), component 4 (water), component 5 (high-performance water-reducing agent), component 6 (coarse aggregate), component 7 (fine aggregate), and the number of years of use. The performance parameter is compressive strength. A single performance parameter value is derived based on these eight process parameter values.

[0101] 1030 samples were prepared as combinations of process parameter values ​​and performance parameter values. 782 of these were used as training samples, and the remaining 248 were used as test samples. DNN and CNN models were trained based on the training samples, and the prediction accuracy of the DNN and CNN models was verified based on the test samples.

[0102] Figure 16 This is a graph showing the prediction accuracy of the DNN model and CNN model in Example 4. Figure 16 The upper section is a graph showing the relationship between the predicted and true values ​​of compression strength, which is a performance parameter of the DNN model. Figure 16 The lower section is a graph showing the relationship between the predicted and true values ​​of compression strength, which is a performance parameter of the CNN model. For example... Figure 16 As shown, the determination coefficient R of the DNN model 2 The determination coefficient R of the CNN model is 0.8802. 2 The value is 0.9006, therefore it can be seen that the CNN model has better prediction accuracy than the DNN model regarding the process parameters and performance parameters of Example 4.

[0103] (Modified Example)

[0104] In the training process of the variant example, a first fully learned model with fully connected layers but no convolutional layers is generated as the fully learned model. If the performance of the first fully learned model does not reach the benchmark, a second fully learned model with convolutional layers is generated as the fully learned model. The training process of the variant example is explained in detail below. The first fully learned model means a DNN model, and the second fully learned model means a CNN model.

[0105] Figure 17 This is a diagram showing the processing order of the model training for the variant examples. For example... Figure 17 As shown, firstly, processor 11 trains a DNN model based on one-dimensional permutation data to generate a fully trained DNN model (step SC1). The training method for the DNN model is the same as described above, which can be carried out through supervised learning based on training samples.

[0106] When performing step SC1, processor 11 evaluates the performance of the learned DNN model (step SC2). Specifically, processor 11 applies samples to the learned DNN model and calculates parameter values ​​for evaluating the performance of the learned DNN model. These parameter values ​​can be expressed using the coefficient of determination R0. 2 Any metric such as accuracy rate.

[0107] When performing step SC2, processor 11 determines whether the performance meets the benchmark (step SC3). Specifically, processor 11 determines whether the parameter values ​​calculated in step SC2 for performance evaluation exceed the benchmark values. If the parameter values ​​exceed the benchmark values, the performance is determined to meet the benchmark; if the parameter values ​​are lower than the benchmark values, the performance is determined to not meet the benchmark.

[0108] If the performance is determined to meet the benchmark in step SC3 (step SC3: Yes), processor 11 outputs the learned DNN model (step SC4). The output learned DNN is used to generate prediction performance data.

[0109] If, in step SC3, the performance is determined not to meet the benchmark (step SC3: No), processor 11 trains a CNN model based on the two-dimensional permutation data, generating a fully trained CNN model (step SC5). The training method for the CNN model is the same as described above, which can be performed through supervised learning based on training samples.

[0110] When step SC5 is performed, processor 11 outputs the learned CNN model (step SC6). The output learned CNN is used to generate prediction performance data. In addition to the learned CNN model, a learned DNN model can also be output.

[0111] When step SC4 or SC6 is performed, the model training process for the variant example ends.

[0112] As mentioned above, the input to a DNN model is a one-dimensional permutation of data, while the input to a CNN model is a two-dimensional permutation of data. Therefore, it is assumed that the training load of a DNN model is lighter than that of a CNN model. Furthermore, the performance of DNN and CNN models can vary depending on the types of process parameters and performance parameters. According to a variation, the DNN model is trained first, and then the CNN model is trained only when the DNN model's performance does not reach the benchmark. Therefore, when the DNN model's performance reaches the benchmark, a machine learning model with good accuracy can be obtained with a lighter workload. Additionally, appropriate machine learning models can be generated based on the types of process parameters and performance parameters.

[0113] Thus, according to the above embodiments, a performance prediction method, storage medium, and apparatus that improve the accuracy of performance prediction can be provided.

[0114] While certain embodiments have been described, these embodiments are presented by way of example only and are not intended to limit the scope of the invention. In fact, the novel embodiments described herein can be embodied in various other ways; furthermore, various omissions, substitutions, and modifications can be made to the embodiments described herein without departing from the spirit of the invention. The appended claims and their equivalents include any modifications or embodiments falling within the scope and spirit of the invention.

Claims

1. A performance prediction method comprising: a conversion step of converting one-dimensional arrangement data including a plurality of parameter values arranged in one dimension into two-dimensional arrangement data in which the plurality of parameter values are arranged in two dimensions, by copying and arranging the plurality of parameter values; a training step of training a machine learning model based on training data including the two-dimensional arrangement data and performance data corresponding to the one-dimensional arrangement data, to generate a learned model that inputs the two-dimensional arrangement data and outputs predicted performance data; and a prediction step of generating predicted performance data of a processing target by applying two-dimensional arrangement data of the processing target to the learned model.

2. The performance prediction method according to claim 1, wherein in the conversion step, the plurality of parameter values are arranged in a manner such that a parameter type is not repeated in each column and each row of the two-dimensional arrangement data.

3. The performance prediction method according to claim 1, wherein in the conversion step, the plurality of parameter values are arranged in the two-dimensional arrangement data in a manner such that an arrangement pattern of parameter types included in a rectangular region of a prescribed size is not repeated at any position.

4. The performance prediction method according to claim 1, wherein the one-dimensional arrangement data has N parameter values arranged in one dimension as the plurality of parameter values, N being a natural number, the two-dimensional arrangement data has N x M parameter values arranged in a two-dimensional space defined by a first direction and a second direction orthogonal to the first direction, M being a natural number, in the conversion step, M arrangements are sequentially generated by arranging the N parameter values in M columns of the two-dimensional space in an arrangement order based on a prescribed algorithm, and in this, the parameter values of a first arrangement or a second arrangement are rearranged in a manner such that a parameter type is not repeated in a prescribed direction other than the first direction between the adjacent first arrangement and the second arrangement, to generate the two-dimensional arrangement data.

5. The performance prediction method according to claim 1, wherein the machine learning model is a convolutional neural network including a convolution layer that performs convolution processing on the two-dimensional arrangement data using a plurality of filters.

6. The performance prediction method according to claim 1, wherein in the training step, a first learned model having a fully connected layer and not having a convolution layer is generated as the learned model, in a case where a performance of the first learned model does not reach a criterion, a second learned model having a convolution layer is generated as the learned model.

7. The performance prediction method according to claim 1, wherein the plurality of parameter values include a plurality of values respectively corresponding to a plurality of process parameters regarding a product, the performance data and / or the predicted performance data include values of one or more performance parameters that evaluate a performance of the product.

8. The performance prediction method according to claim 7, wherein the plurality of process parameters include at least two of a volume of air, an outlet angle, a number of blades, a length of blades, and a height of blades of a centrifugal blower that is the product, the one or more performance parameters include pressure, efficiency, and / or noise. 9.A non-transitory computer-readable storage medium including computer-executable instructions that, when executed by a processor, cause the processor to perform the following procedures: a conversion procedure of converting one-dimensional arrangement data including a plurality of parameter values arranged in one dimension into two-dimensional arrangement data in which the plurality of parameter values are arranged in two dimensions by copying and configuring the plurality of parameter values, a training procedure of training a machine learning model based on training data including the two-dimensional arrangement data and performance data corresponding to the one-dimensional arrangement data, generating a learned model that inputs the two-dimensional arrangement data and outputs predicted performance data, and a prediction procedure of generating predicted performance data of a processing target by applying two-dimensional arrangement data of the processing target to the learned model. 10.A performance prediction device including a processor that performs the following processes: converts one-dimensional arrangement data including a plurality of parameter values arranged in one dimension into two-dimensional arrangement data in which the plurality of parameter values are arranged in two dimensions by copying and configuring the plurality of parameter values, trains a machine learning model based on training data including the two-dimensional arrangement data and performance data corresponding to the one-dimensional arrangement data, generates a learned model that inputs the two-dimensional arrangement data and outputs predicted performance data, generates predicted performance data of a processing target by applying two-dimensional arrangement data of the processing target to the learned model.

Images