Simulation device, program, simulation method, and information processing device

Abstract

To provide a technique capable of effectively confirming an effect by unsupervised mechanical learning and easily searching for calculation accuracy more suitable for a problem, an assignment, an application or the like to which AI is applied.SOLUTION: In a simulation device (1), a model setting unit (6) performs setting related to a chunk as a mass of data for one time when the chunk is sequentially input to a machine learning model (40) on the basis of read data, and a model computation unit (5) performs unsupervised learning computation and inference computation by sequentially inputting the chunk to the machine learning model. The model setting unit performs setting related to calculation accuracy of the machine learning model.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present disclosure relates to a simulation device, a program, a simulation method, and an information processing device. [Background technology]

[0002] Conventionally, in the field of industrial machinery, the application of AI (artificial intelligence) to condition-based maintenance of machine systems has been progressing with regard to factory equipment maintenance (for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2019 / 035279

[0004] [overview] When introducing AI into the above-mentioned condition management maintenance, it would be useful to be able to confirm the effectiveness of the AI ​​without using actual equipment. A well-known form of machine learning is unsupervised learning, which does not use training data.

[0005] In general, in machine learning, there is a trade-off between improving calculation accuracy (improving the accuracy of the calculation results) and reducing calculation time. In order to achieve a compromise between improving calculation accuracy and reducing calculation time, Bfloat16 has been proposed in recent years as a calculation accuracy (floating-point format) suitable for use in machine learning.

[0006] However, when considering actual applications, Bfloat16 is not necessarily suitable for all machine learning calculations.

[0007] A simulation device according to one aspect of the present disclosure includes a model storage unit storing a machine learning model configured to perform learning and inference, a model calculation unit configured to perform calculation processing using the machine learning model, an operation input unit, a reading unit configured to read data, and a model setting unit configured to configure settings related to the machine learning model based on input from the operation input unit. The model setting unit is configured to configure settings related to chunks, which are blocks of data to be sequentially input to the machine learning model, based on the read data. The model calculation unit is configured to perform unsupervised learning calculations and inference calculations by sequentially inputting the chunks to the machine learning model. The model setting unit is configured to configure settings related to the calculation accuracy of the machine learning model.

[0008] In addition, a program according to one embodiment of the present disclosure is a program for causing a computer to execute the following steps: a first step of reading data; a second step of setting chunks, which are blocks of data to be sequentially input into a machine learning model based on the read data; a third step of setting the calculation accuracy of the machine learning model; and a fourth step of performing unsupervised learning calculations and inference calculations by sequentially inputting the chunks into the machine learning model.

[0009] In addition, a simulation method according to one embodiment of the present disclosure includes a first step of reading data, a second step of configuring chunks as a single set of data to be sequentially input into a machine learning model based on the read data, a third step of configuring settings related to the calculation accuracy of the machine learning model, and a fourth step of performing unsupervised learning calculations and inference calculations by sequentially inputting the chunks into the machine learning model.

[0010] According to another aspect of the present disclosure, there is provided an information processing device including a single processor, wherein the processor is configured to be capable of switching between performing computational processing at a first computational precision using a machine learning model and performing computational processing at a second computational precision different from the first computational precision using the machine learning model. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram illustrating a configuration of a computer according to an exemplary embodiment of the present disclosure. [Figure 2] FIG. 2 is a diagram illustrating a configuration of a simulation device according to an exemplary embodiment of the present disclosure. [Figure 3] FIG. 3 is a diagram illustrating an example of the configuration of a machine learning model. [Figure 4] FIG. 4 is a diagram illustrating a three-layer neural network. [Figure 5] FIG. 5 is a diagram showing the first setting screen. [Figure 6] FIG. 6 is a diagram showing the dialog box. [Figure 7] FIG. 7 is a diagram showing an example of data in a data file (sample file). [Figure 8] FIG. 8 is a diagram showing the first setting screen. [Figure 9] FIG. 9 is a diagram showing an example of preprocessing settings and processing results for the first column on the first setting screen. [Figure 10] FIG. 10 is a diagram showing an example of preprocessing settings and processing results for the second column on the first setting screen. [Figure 11] FIG. 11 is a diagram showing an example of preprocessing settings and processing results for the third column on the first setting screen. [Figure 12] FIG. 12 is a diagram showing an example of preprocessing settings and processing results for the fourth column on the first setting screen. [Figure 13] FIG. 13 is a diagram showing the second setting screen. [Figure 14] FIG. 14 is a diagram showing the second setting screen. [Figure 15] FIG. 15 is a diagram showing the second setting screen. [Figure 16] FIG. 16 is a diagram showing the third setting screen. [Figure 17] FIG. 17 is a diagram showing an example of a preprocessing result on the third setting screen. [Figure 18] FIG. 18 is a diagram showing the learning and inference setting section on the third setting screen. [Figure 19] FIG. 19 is a diagram showing an example of the learning and inference settings. [Figure 20] FIG. 20 is a diagram showing the fourth setting screen. [Figure 21] FIG. 21 is a diagram showing the fourth setting screen. [Figure 22] FIG. 22 is a diagram illustrating an example of repeated learning. [Figure 23] FIG. 23 is a diagram showing the graph screen.

[0012] [Detailed explanation] Hereinafter, exemplary embodiments of the present disclosure will be described with reference to the drawings.

[0013] <Computer configuration> 1 is a diagram showing the configuration of a computer 100 according to an exemplary embodiment of the present disclosure. The computer 100 functions as a simulation device according to the present disclosure, which will be described later. The computer 100 is, for example, a PC (personal computer). When the computer 100 is a PC, it does not matter whether it is a desktop computer or a notebook computer.

[0014] The computer 100 includes a CPU (Central Processing Unit) 100A, a memory 100B, an auxiliary storage device 100C, an operation input unit 100D, and a display unit 100E.

[0015] The CPU 100A has a control device and an arithmetic device (neither of which are shown). The control device interprets program instructions and controls each part of the computer 100. The arithmetic device is a device that performs arithmetic processing.

[0016] The memory 100B is a semiconductor storage device that temporarily stores programs or data. The information stored in the memory 100B is erased when the computer 100 is powered off.

[0017] The auxiliary storage device 100C is configured with a hard disk drive (HDD) or a solid state drive (SSD), and stores programs or data. The programs stored in the auxiliary storage device 100C are loaded into the memory 100B. The CPU 100A executes the programs loaded into the memory 100B.

[0018] The operation input unit 100D is configured with a keyboard, a mouse, or the like, and is a device that provides operation input to the computer 100. Information input from the operation input unit 100D is sent to the memory 100B.

[0019] The display unit 100E is configured by, for example, a liquid crystal display, and converts the information acquired from the memory 100B into an image and outputs it.

[0020] <Configuration of the simulation device> 2 is a diagram illustrating a configuration of a simulation device 1 according to an exemplary embodiment of the present disclosure. The simulation device 1 is a device capable of simulating learning and inference using machine learning (AI).

[0021] The simulation device 1 includes a file storage unit 2, a file reading unit 3, a model storage unit 4, a model calculation unit 5, a model setting unit 6, a display control unit 7, an operation input unit 8, and a display unit 9.

[0022] The file storage unit 2 stores a data file 21 and is configured as an auxiliary storage device 100C of the computer 100. The data file 21 is configured as an Excel (registered trademark) file, for example. The file may be stored in the auxiliary storage device 100C from outside the computer 100, for example, via a USB interface or a network interface in the computer 100 (neither of which is shown in FIG. 1).

[0023] The data file 21 includes data 210. The data 210 is data for performing so-called unsupervised learning and inference by being input into a machine learning model 40 (described later). As will be described later, it is possible to specify which of the data 210 is used for learning and which is used for inference.

[0024] The file reading unit 3 reads the data file 21 from the file storage unit 2 .

[0025] The model storage unit 4 stores a machine learning model 40, and is configured by an auxiliary storage device 100C of the computer 100. The machine learning model 40 is configured as a part of the program P using, for example, MATLAB / Simulink (registered trademark). A specific example of the machine learning model 40 will be described later.

[0026] The functions of the file reading unit 3, the model calculation unit 5, the model setting unit 6, and the display control unit 7 are realized by the CPU 100A executing the program P. The operation input unit 8 and the display unit 9 correspond to the operation input unit 100D and the display unit 100E in the computer 100, respectively.

[0027] The model calculation unit 5 executes a simulation by performing calculation processing on the machine learning model 40 stored in the model storage unit 4. The model setting unit 6 performs settings related to the machine learning model 40 stored in the model storage unit 4 (setting of data to be used for learning and inference, setting of parameters, setting of function types, etc.) in response to input from the operation input unit 8. The simulation by the model calculation unit 5 is performed in accordance with the settings made by the model setting unit 6. The display control unit 7 controls the display of various screens, such as a setting screen (described later), on the display unit 9 in response to input from the operation input unit 8.

[0028] For example, in the industrial machinery field, machine learning is increasingly being applied to the condition management and maintenance of machinery systems in factory equipment maintenance. However, when users try out a specific AI algorithm, they must provide data to the AI ​​developer, such as an AI vendor. In many cases, this data contains confidential information related to the product and manufacturing process, making it difficult to disclose. Therefore, in the past, users had no choice but to refer to various case studies from AI vendors or have the AI ​​solve open problems that generally serve as benchmarks for AI performance (e.g., predicting housing prices in Boston or estimating the physical properties of diesel fuel) to determine whether the algorithm is suitable for solving the user's problem or issue. Naturally, these open problems do not match the user's problem or issue, so when the selected AI algorithm is applied to the user's data, it often turns out to be inappropriate or not very suitable. This can lead to having to reselect the AI ​​algorithm, which is detrimental to both the user and the AI ​​vendor.

[0029] As mentioned above, even if users want to test what AI can do using the data they have acquired, they need to share the data externally (with an AI vendor or an IC manufacturer with AI solutions, etc.). This means there is a risk that confidential data will be released to the outside world. Furthermore, even if the data is provided alone, it is difficult to ensure that it will be properly analyzed externally, and it is difficult to adjust the AI ​​parameters appropriately without advanced knowledge of the domain in which the user acquired the data. It is also extremely time-consuming for users to convey domain knowledge to AI vendors, etc., and communication errors are likely to occur due to differences in background knowledge. Furthermore, the data acquired by users cannot be analyzed on the same day, on the spot.

[0030] In view of the above-described circumstances, the use of the simulation device 1 according to the present disclosure makes it possible to effectively confirm the AI ​​effects of unsupervised learning using data owned by the user. More specifically, the effectiveness of anomaly detection, which is a task of unsupervised learning, can be confirmed. This provides benefits such as eliminating the risk of leaking confidential data to the outside and enabling on-the-spot analysis of acquired data. In particular, as described below, the present disclosure introduces the concept of chunks to the setting of data to be input to a machine learning model, making it easy to input multiple types of input data in multiple pieces. This makes it possible to consider what kind of anomaly detection can be performed when various signals are combined.

[0031] <Machine learning model configuration> Here, we will explain the machine learning model 40. Fig. 3 is a diagram showing an example of the configuration of the machine learning model 40. The machine learning model 40 has a preprocessing unit 401 and a machine learning unit 402.

[0032] The preprocessing unit 401 performs preprocessing before inputting data to the machine learning unit 402. The preprocessing includes normalization processing. The normalization processing is processing that puts data into a range of approximately 0 to 1 (or -1 to +1). Note that, as will be described later, by setting parameters for the normalization processing, it is also possible to set it to perform shift processing or no processing at all.

[0033] Pre-processing also includes envelope processing, which is performed on data that has been normalized. Note that whether or not envelope processing is performed can be set as described below.

[0034] Preprocessing also includes window function processing and FFT (fast Fourier transform) processing. As will be described later, window function processing and FFT processing are performed on data that has been normalized and envelope processed. Furthermore, window function processing and FFT processing are performed for each chunk, as will be described later. The presence or absence of window function processing and FFT processing can be set, as will be described later. As a pattern, it is possible to select to perform only window function processing, only FFT processing, or both window function processing and FFT processing. Note that frequency analysis processing such as wavelet transform may also be used, rather than being limited to FFT processing.

[0035] The machine learning unit 402 performs learning and inference on the input data. As an AI model used in the machine learning unit 402, for example, a three-layer neural network 30 as shown in FIG.

[0036] As shown in Figure 4, the three-layer neural network 10 is an AI model having an input layer 10A, a hidden layer 10B, and an output layer 10C. In general, in the three-layer neural network 10, n-dimensional input data x∈R with a batch size k is input. k×n For n'-dimensional inference result y∈R k×n’ is obtained as y=G(x·α+b)β, where α∈R n×m is the weight connecting the input layer 10A and the hidden layer 10B, and β∈R m×n’ is the weight connecting the hidden layer 10B and the output layer 10C. Also, b∈R m is the bias of hidden layer 10B, and G is the activation function of hidden layer 10B.

[0037] In this embodiment, an algorithm is used that can sequentially train the three-layer neural network 10 with an arbitrary batch size. i The i-th training data {x i ∈R ki×n , t i ∈R ki×n’} is obtained, β that minimizes the error shown in the following equation (1) i It is necessary to seek.

number

[0038] Optimized weight β i is calculated using the following formula (2). P i =P i-1 -P i-1 H i T (I+H i P i-1 H i T ) -1 H i P i-1 β i =β i-1 +P i H i T (t i -H i β i-1 ) (2)

[0039] Here, P0 and β0 are obtained by the following equation (3). P0=(H0 T H0) -1 β0=P0H0 T t0(3)

[0040] The learning algorithm is as follows: (1) The weight α and bias b are initialized with random numbers. (2) Calculate H0 for x0, and calculate P0 and β0. (3) Batch size k i Each time the i-th training data of P is obtained, i and β i It should be noted that instead of using the formula for calculating β0 in equation (3), a value initialized by a random number may be used as β0.

[0041] In this embodiment, learning is performed using an autoencoder. The autoencoder uses input data as training data and learns so that the input data can be reconstructed as an inference result. In other words, in the above example, learning is performed with t = x. Since the autoencoder does not require separate training data to be created, it is a type of unsupervised learning algorithm.

[0042] According to such an AI model in the machine learning unit 402, learning becomes possible using a computing device equivalent to a microcomputer in an edge device. Note that an edge device refers to an information processing device that takes over a trained machine learning model from another device and performs inference calculations using the trained machine learning model. That is, when such on-device learning is introduced into, for example, motor anomaly detection, the effectiveness of the anomaly detection can be confirmed by simulation. Note that the input data x is time-series data when FFT processing is not performed in the pre-processing unit 401, and is frequency-domain data when FFT processing is performed.

[0043] In the machine learning unit 402, the degree of anomaly is calculated using a loss function L(y, t) that represents the error between the inference result y and the training data t. For example, the loss function may be MAE (Mean Absolute Error) or MSE (Mean Squared Error). When the loss function is MAE, the loss function L is expressed as in the following equation (4).

number

number

[0044] Since learning is performed using an autoencoder, the degree of anomaly is calculated as the loss function L(y,t) = L(y,x).

[0045] <gui> Next, a description will be given of a GUI (Graphical User Interface) that enables settings related to simulation in the simulation device 1 according to this embodiment. Examples of various screens described below are displayed on the display unit 9 by the display control unit 7 (FIG. 2). Selection and setting on various setting screens, or screen switching, are performed based on input from the operation input unit 8. The contents set on various setting screens are set by the model setting unit 6.

[0046] <<First setting screen>> When program P is started, the first setting screen shown in Figure 5 is displayed. On the first setting screen, you can load data files and, if necessary, set preprocessing. Tabs are displayed side by side at the top of the first setting screen. You can switch setting screens by pressing a tab. Figure 5 (first setting screen) shows the state when tab TB1 for "1. Data loading" is pressed.

[0047] The first setting screen displays a selection button BT1 for selecting a data file. Pressing the selection button BT1 displays the dialog box shown in FIG. 6. In the dialog box, a list of file names is displayed in a selection section SA1. The file name selected in the selection section SA1 is displayed in the file name display section DA1 below. When the decision button BT2 is pressed in the dialog box, the data file with the file name displayed in the file name display section DA1 is read by the file reading section 3. Here, an Excel file with the extension xlsx can be selected.

[0048] FIG. 7 shows an example of data in a data file (sample file). The data in FIG. 7 is time-series data of various signals in a case where, for example, damage is present in the inner raceway of a bearing supporting a motor output shaft, and the damage gradually worsens over time. Specifically, column A stores time [s], column B stores motor current [A], column C stores x-direction displacement [m], column D stores x-direction acceleration [m / s^2], column E stores y-direction displacement [m], column F stores y-direction acceleration [m / s~2], and column G stores output shaft rotation speed [rpm]. Note that the first row of each column stores a variable name. Also, in FIG. 7, for example, the first 4,000 points of data are normal, and the degree of damage worsens every 4,000 points thereafter. The following description will be given assuming that such a file is read.

[0049] When a data file is loaded, the data contained in the loaded data file is displayed in a table format in the table display section DA2 on the first setting screen, as shown in FIG. 8. The first line of the data file is ignored when the data is loaded. This allows the user to check whether the data has been loaded correctly. The number of lines of the loaded data is displayed in the line number display section DA3, and the number of columns of the loaded data is displayed in the column number display section DA4. All of the loaded data is displayed in the table display section DA2.

[0050] On the other hand, when a data file is loaded, the loaded data is displayed as a graph by column in the graph display section DA5 on the first setting screen, as shown in Figure 8. The graph is displayed with the data number on the horizontal axis and the data value on the vertical axis. In the column selection section SD1, columns 1 to 5 can be selected using the radio buttons. By default, the data in the first column is displayed as a graph, but you can switch the column to be displayed as a graph by pressing the radio button. Note that columns 6 and beyond cannot be displayed as a graph, but this does not mean that the data has not been loaded.

[0051] The pre-processing setting section ST1 is displayed in the center of the first setting screen (Fig. 8) in the horizontal direction. In the pre-processing setting section ST1, settings related to normalization processing and envelope processing can be made. Normalization processing and envelope processing can be set individually for each of the first to fifth columns of the loaded data.

[0052] The normalization process is performed according to the following equation (6).

number

[0053] The parameter setting section PS1 in the preprocessing setting section ST1 can set the parameters (di, si) for each of the first to fifth columns. Depending on the parameter setting, it is also possible to set shift processing or no processing at all.

[0054] Additionally, check boxes BX1 in the preprocessing setting section ST1 are provided for each of the first to fifth columns. Envelope processing is performed for columns with checked check boxes BX1. Envelope processing is not performed for columns with unchecked check boxes BX1.

[0055] When the preprocessing button BT2 on the first setting screen is pressed, preprocessing is performed on the data in columns 1 to 5 according to the settings in the preprocessing setting section ST1. The table display section DA6 on the first setting screen displays the preprocessed data in a tabular format. Here, the data in columns 1 to 5 is displayed. The graph display section DA7 displays the preprocessed data for the column selected with the radio button in the selection section SD2. The display format is the same as that of the graph display section DA5. The row display section DA8 and column display section DA9 display the number of rows and the number of columns of the data displayed in the table display section DA6, respectively. The column display section DA9 displays the number of columns = 5.

[0056] If pre-processing is not required, do not press the pre-processing button BT2, but press the tab TB2 (FIG. 13, described later) to switch the screen.

[0057] FIG. 9 is a diagram showing an example of preprocessing settings and processing results for the first column. As shown in FIG. 9, the parameters d1 and s1 for the first column are set to d1=0 and s1=1, respectively. In this case, no processing is performed. Note that these parameter settings are the default. In FIG. 9, the display of the first column is selected for the graph display sections DA5 and DA7. Comparing the graph display sections DA5 and DA7, it can be seen that there is no change in the data and no processing has been performed.

[0058] Figure 10 is a diagram showing an example of preprocessing settings and processing results for the second column. As shown in Figure 10, the parameters d2 and s2 for the second column are set to d2 = 1.2 and s2 = 1, respectively. In Figure 10, the display of the second column is selected for the graph displays DA5 and DA7. Comparing the graph displays DA5 and DA7, it can be seen that the data has been shifted in the negative direction of the vertical axis after preprocessing.

[0059] FIG. 11 shows an example of the preprocessing settings and processing results for the third column. As shown in FIG. 11, the parameters d3 and s3 for the third column are set to d3=0 and s3=6e-06, respectively. In FIG. 11, the display of the third column is selected for the graph displays DA5 and DA7. Comparing the graph displays DA5 and DA7, it can be seen that after preprocessing, the data has been enlarged in the vertical axis direction and the scale has been changed.

[0060] FIG. 12 shows an example of the preprocessing settings and processing results for the fourth column. As shown in FIG. 12, the parameters d4 and s4 for the fourth column are set to d4=0 and s4=1, respectively. That is, they remain the default settings. However, envelope processing has been selected for the fourth column (checkbox BX1 is checked). In FIG. 12, the display of the fourth column has been selected for the graph display sections DA5 and DA7. Comparing the graph display sections DA5 and DA7, it can be seen that the upper envelope has been extracted after preprocessing.

[0061] <<Second setting screen>> When tab 2 ("2. Input chunk range") is pressed, the second setting screen is displayed as shown in Fig. 13. On the second setting screen, data setting buttons BT31 and BT32 are displayed, and either button is pressed.

[0062] When the data setting button BT31 is pressed, the data (original data) read on the first setting screen is used as is as data to be input to the machine learning model 40. When the data setting button BT32 is pressed, the data obtained by performing the preprocessing set on the first setting screen on the data read on the first setting screen is used as data to be input to the machine learning model 40.

[0063] The data selected for use with either the data setting button BT31 or BT32 is displayed in the table display section DA10 and the graph display section DA13. The data is displayed in a notated format in the table display section DA10. The display in the graph display section DA13 can be switched between columns 1 to 5 by selecting a column with the radio button in the column selection section SD3.

[0064] 14 is a diagram showing the state when the data setting button BT31 is pressed on the second setting screen (i.e., use of original data). On the second setting screen, a chunk setting section ST2 is displayed. A chunk is a block of data for one time input into the machine learning model 40 (machine learning section 402).

[0065] The chunk setting section ST2 includes a column number setting section ST21, a row count setting section ST22, and a column count display section ST23. In the column number setting section ST21, it is possible to input a column number in the use data (data selected by the data setting buttons BT31 and BT32) that indicates the first column of the input data to be input to the machine learning model. In the row count setting section ST22, it is possible to input the number of rows in one chunk in the use data. In the column count setting section ST23, it is possible to input the number of columns in one chunk in the use data.

[0066] In the example of Fig. 14, "2" is input to column number setting unit ST21, "256" to row number setting unit ST22, and "1" to column number setting unit ST23. As a result, the second column of the usage data is set as the first column of the input data, and a block of data with 256 rows and 1 column is set as one chunk. In other words, the data in the second column (motor current in Fig. 7) is selected as the input data.

[0067] Note that by setting the column number setting unit ST23 to a value of 2 or more, it is possible to use multiple columns of data, i.e., multiple types of signal data, as input data. For example, the motor current and x-direction displacement data in FIG. 7 can be used as input data.

[0068] When the check button BT4 is pressed on the second setting screen, the screen shown in FIG. 15 appears, with the first chunk displayed in table format in the table display section DA14. This allows you to check whether the chunks have been set up appropriately. The chunk count display section DA15 also displays the number of chunks calculated from the number of rows of the data used and the number of rows in one chunk (in the example of FIG. 15, the number of chunks = 62). The data count display section DA16 displays the number of data contained in one chunk. This number of data is the same as the product of the number of rows and the number of columns in one chunk (in the example of FIG. 15, the number of data = 256).

[0069] <<Third setting screen>> Clicking tab TB3 ("3. Preprocessing by MCU") displays the third setting screen as shown in Figure 16. On the third setting screen, you can consider applying preprocessing such as FFT that can be performed by the MCU (microcomputer).

[0070] The third setting screen displays a pre-processing setting section ST3. In the pre-processing setting section ST3, whether or not to apply window function processing and FFT processing is set. Note that window function processing can be performed using a Hann window, Hamming window, Gauss window, triangular window, Kaiser window, Chebyshev window, Blackman window, etc. Specifically, the presence or absence of window function processing is set using a check box BX2. The presence or absence of FFT processing is set using a check box BX3. Furthermore, in a display unit selection section SD4, the display unit (Amplitude or dB) of the FFT processing results can be selected using a radio button.

[0071] Note that window function processing and FFT processing are performed for each chunk. FFT processing and other processing are performed using algorithms suitable for calculations on an MCU. Since only the amplitude of the data after FFT processing is used, the number of pieces of data will be (FFT length) / 2+1, where FFT length (the number of pieces of data contained in one chunk) is used. Note that FFT processing may be performed in a manner that allows selection of whether to use only the amplitude, only the phase, or both the amplitude and the phase.

[0072] In the pre-processing setting section ST3, the FFT length is displayed in the FFT length display section DA17.

[0073] When the preprocessing button BT5 is pressed, the preprocessing set in the preprocessing setting unit ST3 is performed, and the processing results are displayed in the result display unit DA18. The processing results are displayed for each chunk. The data number included in the displayed chunk can be set in the data number setting unit ST31 in the preprocessing setting unit ST3. Here, if a chunk has multiple columns, the data number is a consecutive number for all chunks when numbers are assigned sequentially, starting from 1, to data in different columns of the same row, starting from the smallest row. For example, if a chunk has three columns, numbers are assigned in the following order: first, second, and third columns of the first row → first, second, and third columns of the second row →... In the case of Figure 15 mentioned above, the set chunk has 256 rows and one column, so data numbers 1 to 256 represent the same leading chunk.

[0074] In the result display section DA18, a raw data display section DA181, a post-window function processing display section DA182, and a post-FFT processing display section DA183 are displayed from left to right. The raw data display section DA181 displays one chunk of data (data before processing). The post-window function processing display section DA182 displays the data obtained by applying a window function to one chunk of data. The post-FFT processing display section DA183 displays the data obtained by applying an FFT to the data after the window function processing. Note that the raw data display section DA181, the post-window function processing display section DA182, and the post-FFT processing display section DA183 all display graphs with the horizontal axis representing the data number (however, this is a serial number when the first data in the target chunk is numbered 1) and the vertical axis representing the data value.

[0075] However, for example, if window function processing is not applied and only FFT processing is performed, raw data is displayed in the window function processing display unit DA182. If neither window function processing nor FFT processing is applied, raw data is displayed in the raw data display unit DA181, window function processing display unit DA182, and FFT processing display unit DA183. In either case, the data displayed in the FFT processing display unit DA183 is the data input to the machine learning unit 402 (three-layer neural network 10).

[0076] FIG. 17 is an example of the display of the processing results when both window function processing and FFT processing are applied. After FFT processing, the number of data items is reduced. Note that a data number setting section ST32 is displayed in the result display section DA18. In the data number setting section ST32, it is possible to set a data number corresponding to the chunk to be redisplayed in the result display section DA18. When the redisplay button BT6 is pressed, the processing results are redisplayed in the result display section DA18.

[0077] Next, we will describe the learning and inference setting section ST4 (Fig. 16) on the third setting screen. Fig. 18 shows the learning and inference setting section ST4. In the learning and inference setting section ST4, a raw data display section DA19 is displayed. In the raw data display section DA19, a graph is displayed of all chunk data (data before processing) set on the second setting screen. The horizontal axis is the data number and the vertical axis is the data value.

[0078] Below the raw data display section ST4, a first data number display section DA20 and a first chunk number display section DA21 are displayed on the left side. The first data number display section DA20 displays the first number of the data numbers ("1" in FIG. 18), and the first chunk number display section DA21 displays the number of the first chunk ("1" in FIG. 18).

[0079] Below the raw data display section ST4, a final data number display section DA22 and a final chunk number display section DA23 are displayed on the right edge. The final data number is displayed in the final data number display section DA22, and the final chunk number is displayed in the final chunk number display section DA23. In the example of Figure 18, the chunks in the example of Figure 15 are set to 256 rows and 1 column, and the number of data in one column is 16001 as shown in Figure 7, so the number of chunks = 62 (final chunk number), and the final data number is 62 x 256 = 15872.

[0080] Training and prediction in the machine learning unit 402 are performed in chunk units. First, one of the data numbers included in the chunk from which learning is to begin is input and set into the data number setting unit ST41. FIG. 19 is a diagram illustrating the setting of learning and prediction for a raw data sequence similar to that of FIG. 18. In the example of FIG. 7, the motor current, which is the data in the second column, shows an increase in current when the motor starts. If this increase in current is not to be learned, it is effective to start learning from the second chunk CH2 (chunk number = 2). Therefore, if a data number included in the second chunk CH2, for example, 257, is set, learning will begin from the second chunk CH2, and the first chunk CH1 will not be used for learning.

[0081] Next, one of the data numbers included in the chunk from which inference is to be started is input and set in the data number setting unit ST42 (FIG. 18). In the example of FIG. 7, the motor current transitions to an abnormal state from data number 4000. Since the state is normal up to that point, data number 3500 is set here as an example so that only the normal state is used for learning. Note that data number 3500 corresponds to chunk CH14 with chunk number 14 as shown in FIG. 19. Thus, in the example of FIG. 19, chunk CH1 is not used for learning, chunks CH2 to CH13 are used for learning, and chunks CH14 to the last chunk CH62 are used for inference.

[0082] <<Fourth setting screen>> When tab TB4 ("4. AI Settings and Sim") is pressed, the fourth setting screen shown in FIG. 20 is displayed. The fourth setting screen is a screen for setting parameters for the machine learning unit 402, running a simulation, and checking the simulation results.

[0083] The fourth setting screen shown in FIG. 20 displays an AI parameter setting unit ST5. The AI ​​parameter setting unit ST5 includes an input node number setting unit ST51, a hidden layer node number setting unit ST52, and an output node number setting unit ST53. The input node number setting unit ST51 displays the number of nodes in the input layer 10A (i.e., the value of n described above). The number of nodes in the input layer is calculated by multiplying the number of rows and the number of columns in one chunk of input data. That is, one chunk of data corresponds to data with a batch size of 1. Note that in the example of FIG. 20, the number of data in one chunk is reduced from 256 to 129 in order to input data after FFT processing to the machine learning unit 402 as shown in FIG. 17, and this 129 is set as the number of input nodes.

[0084] The number of nodes in the hidden layer 10B (i.e., the value of m) can be input to the hidden layer node number setting unit ST52. The number of nodes in the output layer 10C (i.e., the value of n') is displayed to the output node number setting unit ST53. Because this is an autoencoder, the number of output nodes matches the number of input nodes.

[0085] The AI ​​parameter setting unit ST5 also includes an activation function setting unit ST54, a loss function setting unit ST55, and a forgetting rate setting unit ST56. The activation function setting unit ST54 can select the type of activation function for the hidden layer 10B. The activation function can be set to, for example, Sigmoid or ReLU.

[0086] The loss function setting unit ST55 can select the type of loss function used to calculate the degree of anomaly in the machine learning unit 402. For example, the loss function can be set to MAE or MSE.

[0087] In the forgetting rate setting section ST56, a value of the forgetting rate can be input. The forgetting rate is a parameter that indicates the degree to which learning results are forgotten. Methods for not reflecting learning results include, for example, using past learning results or initializing the learning results. A forgetting rate of 1 indicates that previous learning results are not forgotten at all, and a forgetting rate of 0 indicates that they are completely forgotten.

[0088] The AI ​​parameter setting section ST5 also displays a learning repetition number setting section ST57, which allows the user to set the number of times learning is to be repeated.

[0089] The AI ​​parameter setting unit ST5 also displays a checkbox BX4. When the checkbox BX4 is checked, Bfloat16 is selected as the setting for the calculation precision of the machine learning model 40. Bfloat16 is a floating-point format that allocates 7 bits to the mantissa, which indicates the precision of the floating-point, and 8 bits to the exponent, which indicates the dynamic range of the numeric value.

[0090] On the other hand, when the checkbox BX4 is not checked, double is selected as the setting for the calculation precision of the machine learning model 40. Double is a floating-point format that allocates 52 bits to the mantissa, which indicates the precision of the floating-point, and 11 bits to the exponent, which indicates the dynamic range of the numerical value.

[0091] That is, the model setting unit 6 sets the calculation accuracy of the machine learning model 40. Setting the calculation accuracy of the machine learning model 40 makes it possible to try out multiple calculation accuracies, which makes it easier to find a calculation accuracy that is more suitable for the problem, issue, application, etc. to which AI is applied.

[0092] The settings regarding the calculation accuracy of the machine learning model 40 made by the model setting unit 6 are applied to both the learning calculation and the estimation calculation of the machine learning model 40. Therefore, the CPU 100A, which is a single processor, performs the estimation calculation using the trained machine learning model 40 with the same calculation accuracy as the learning calculation.

[0093] In the present embodiment, the setting for the calculation accuracy of the machine learning model 40 was either Bfloat16 or double, but for example, the model setting unit 6 may select the setting for the calculation accuracy of the machine learning model 40 from a plurality of options including at least one of double, float32, float16, and Bfloat16. For example, the AI ​​parameter setting unit ST5 may be provided with four checkboxes corresponding to double, float32, float16, and Bfloat16, respectively, and the setting for the calculation accuracy of the machine learning model 40 that has the corresponding checkbox checked may be selected from among double, float32, float16, and Bfloat16.

[0094] When all setting items have been set using the above setting screens and the simulation start button BT7 is pressed, the model calculation unit 5 executes a simulation in accordance with the contents set by the model setting unit 6. Here, data is sequentially input to the machine learning unit 402 in chunk units, and β is sequentially updated using the algorithm described above, performing unsupervised learning. When data is input, it is input to each node of the input layer 10A in the order of the data numbers in the chunks. During learning, the machine learning unit 402 also performs inference. After learning is completed, the machine learning unit 402 also performs inference. An inference result is calculated for each chunk. Furthermore, an anomaly degree is calculated for each chunk based on the inference result, input data, and loss function.

[0095] The simulation results are displayed in the result display section RA1 on the fourth setting screen, as shown in FIG. 21, for example. In the result display section RA1, an abnormality degree table display section RA11 and an abnormality degree graph display section RA12 are displayed in the upper section. In the abnormality degree table display section RA11, the abnormality degree of each chunk at the time of learning is displayed in a table format. In the abnormality degree graph display section RA12, the abnormality degree of each chunk at the time of learning is displayed in a graph format. In both the abnormality degree table display section RA11 and the abnormality degree graph display section RA12, chunk numbers and abnormalities are displayed in correspondence with each other.

[0096] The result display section RA1 displays an anomaly degree table display section RA13 and an anomaly degree graph display section RA14 in the lower section. The anomaly degree table display section RA13 displays the anomaly degree of each chunk at the time of inference after learning is complete in table format. The anomaly degree graph display section RA14 displays the anomaly degree of each chunk at the time of inference after learning is complete in graph format. Both the anomaly degree table display section RA31 and the anomaly degree graph display section RA14 display chunk numbers and anomaly degrees in correspondence with each other.

[0097] Here, when the simulation start button BT7 is pressed with the learning repetition count setting unit ST57 being input, learning using all chunks for learning (chunks CH2 to CH13 in the example of FIG. 19) is reused the number of times set in the learning repetition count setting unit ST37. In other words, when the last chunk has been used, learning is resumed by returning to the first chunk.

[0098] In the example of FIG. 21 , the number of times set in the learning repetition number setting unit ST57 is 0, so learning is not repeated. As shown in the abnormality degree graph display unit RA12 of FIG. 21 , the abnormality degree during learning gradually decreases, but learning ends during the decrease. Ideally, a flat state of the abnormality degree, indicating sufficient learning, is desirable. However, it may not be easy to prepare long-term data to achieve such learning. Therefore, by setting the number of times set in the learning repetition number setting unit ST57 to 1 or more, learning can be easily performed until the abnormality degree reaches a flat state. FIG. 22 shows an example of the display in the abnormality degree graph display unit RA12 when the number of times set in the learning repetition number setting unit ST57 is 20. A turn-back line LN for 20 times is displayed, and as learning is repeated in this manner, the abnormality degree eventually approaches a flat state (the number of learning times is 20 + 1 = 21).

[0099] <<Graph screen>> When tab TB5 ("5.Graph") is pressed, a graph screen such as that shown in FIG. 23 is displayed. On the graph screen, graph display section GD1 is displayed in the upper section, and graph display section GD2 is displayed in the lower section. Graph display section GD1 displays the input data values ​​for all data (data consisting of all chunks) input to the machine learning unit 402. The horizontal axis represents the data number. Note that in the example of FIG. 23, the input data is the data obtained by performing window function processing and FFT processing on each chunk on the third setting screen, so one chunk corresponds to the area from a spike-like peak to just before the next peak. Also, due to the FFT processing, the last data number on the third setting screen is 7998, approximately half the number of 15872 (FIG. 18) (last data number display section DA24). Note that the last chunk number is 62, the same as on the third setting screen (last chunk number display section DA25).

[0100] The graph display section GD2 displays the abnormality degree values ​​for all data input to the machine learning section 402. The horizontal axis represents the data number. In the data in FIG. 7, the first 1 / 4 of all data was in a normal state, and the last 3 / 4 was in an abnormal state. The graph display section GD2 displays the abnormality degree for the normal state and three levels of abnormality degree for the abnormal state.

[0101] In the threshold setting section TS1 at the top right of the graph screen, by inputting the threshold value of the abnormality level in the threshold input section TS11 and pressing the setting button BT8, the set threshold value TH is displayed in the graph display section GD2. In the example of Fig. 23, the threshold value is set to 5e-05 so that abnormality level 3 can be identified.

[0102] Furthermore, a graph display section GD3 is displayed at the bottom of the graph screen. In the graph display section GD3, a comparison is displayed for each chunk between the data input to the machine learning section 402 and the inference results by the machine learning section 402. The horizontal axis is displayed as the data number (however, the serial number when the first data in the target chunk is numbered 1). In addition, a chunk number display section DA26 and a data number setting section ST6 are displayed to the left of the graph display section GD3. The chunk number display section DA26 displays the chunk number of the chunk to be displayed in the graph display section GD3. By default, the last chunk number is displayed. In the data number setting section ST6, it is possible to set the data number included in the chunk to be redisplayed in the graph display section GD3. After setting in the data number setting section ST6, pressing the redisplay button BT9 redisplays the graph display section GD3.

[0103] When the graph display section GD3 is displayed for chunks with large anomaly values, the deviation between the input data and the inferred value becomes large.

[0104] <Other> In addition to the above-described embodiments, the various technical features disclosed in this specification can be modified in various ways without departing from the spirit of the technical creation. In other words, the above-described embodiments should be considered to be illustrative and not restrictive in all respects, and the technical scope of the present invention should not be limited to the above-described embodiments, but should be understood to include all modifications that fall within the meaning and scope equivalent to the claims.

[0105] For example, in the above embodiment, chunks are set two-dimensionally by the number of rows and the number of columns, but chunks may be set three or more dimensionally.

[0106] An information processing device (e.g., an edge device equipped with an MCU) that takes over a trained machine learning model from another device and performs inference calculations using the trained machine learning model may include a processor, similar to the above-described simulation device 1, and the single processor may be capable of switching between calculation processing at a first calculation precision using the machine learning model and calculation processing at a second calculation precision different from the first calculation precision using the machine learning model. The above-described information processing device makes it possible to switch between at least the first calculation precision and the second calculation precision and try them out, thereby facilitating the search for a calculation precision more suitable for the problem, issue, application, etc. to which AI is applied.

[0107] In addition, the above-mentioned information processing device may receive setting information for the calculation accuracy in learning calculations using the machine learning model from another device that has handed over the trained machine learning model to the information processing device, and a single processor may perform estimation calculations using the trained machine learning model with the same calculation accuracy as the learning calculations.

[0108] <Additional Notes> As described above, the simulation device (1) according to one aspect of the present disclosure: a model storage unit (4) storing a machine learning model (40) configured to perform learning and inference; a model calculation unit (5) configured to perform calculation processing using the machine learning model; An operation input unit (8); a reading unit (3) configured to read data; a model setting unit (6) configured to set the machine learning model based on an input from the operation input unit; Equipped with the model setting unit is configured to perform settings related to chunks as a set of data for one time input to the machine learning model based on the read data; the model calculation unit is configured to perform unsupervised learning calculations and inference calculations by sequentially inputting the chunks into the machine learning model; The model setting unit is configured to perform settings related to the calculation accuracy of the machine learning model (first configuration).

[0109] In addition, in the simulation device of the first configuration, the model setting unit may be configured to be able to set the chunk to include time-series data of a plurality of types of signals (second configuration).

[0110] In addition, in the simulation device of the first or second configuration, the model setting unit may be configured to be able to set the chunk by setting a column number as the first column, the number of rows, and the number of columns based on the read data (third configuration).

[0111] In addition, in the simulation device having any one of the first to third configurations, the machine learning model has a first pre-processing unit (401) configured to perform a normalization process on the read data using the following formula: The model setting unit may be configured to be able to set parameters of the following formula (fourth configuration).

number

[0112] In addition, in the simulation device having any one of the first to fourth configurations, the machine learning model has a second pre-processing unit (401) configured to perform envelope processing on the read data, The model setting unit may be configured to be able to set whether or not to perform the envelope processing for each column of the read data (fifth configuration).

[0113] In addition, in the simulation device having any one of the first to fifth configurations, the machine learning model has a third pre-processing unit (401) configured to perform window function processing and frequency analysis processing for each chunk, The model setting unit may be configured to be able to set whether or not to perform each of the window function processing and the frequency analysis processing (sixth configuration).

[0114] Furthermore, in a simulation device of any of the above first to sixth configurations, the model setting unit may be configured to be able to set a learning start chunk, which is the chunk that starts learning by the machine learning model, and an inference start chunk, which is the chunk that starts inference by the machine learning model (seventh configuration).

[0115] In the simulation device of the seventh configuration, the learning start chunk and the inference start chunk are set by setting a data number included in the chunk; If the chunk has multiple columns, the data number may be configured as a number assigned in ascending order of column number to data in different columns of the same row, starting from the smallest row (eighth configuration).

[0116] In addition, in the simulation device of any of the first to eighth configurations, the model setting unit may be configured to be able to set items related to a neural network (10) included in the machine learning model (ninth configuration).

[0117] In the simulation device of the ninth configuration, the items may include a forgetting rate, which is a parameter indicating the degree to which a learning result is forgotten (tenth configuration).

[0118] In addition, in the simulation device of any of the first to tenth configurations, the model setting unit may be configured to be able to set the number of times learning using all of the chunks is repeated (eleventh configuration).

[0119] In the simulation device of the first configuration, the model setting unit is configured to be able to set the number of times that learning using all of the chunks is repeated, The configuration may include a first display control unit configured to repeatedly display the values ​​of the loss function for all the chunks, including a turn-around line (LN) from the last chunk to the first chunk, the set number of times (twelfth configuration).

[0120] Furthermore, in the simulation device of any of the first to twelfth configurations above, a second display control unit (7) may be configured to control the display of the loss function values ​​based on the inference results during and after learning by the machine learning model for the serial numbers of the data included in all of the chunks (a thirteenth configuration).

[0121] Furthermore, in the simulation device of the thirteenth configuration, the model setting unit may be configured to be able to set a threshold value for the value of the loss function to be displayed in the display by the third display control unit (fourteenth configuration).

[0122] In addition, in the simulation device of any one of the first to fourteenth configurations, a third display control unit (7) configured to control display of a value of the data and an inference result by the machine learning model for a serial number of the data included in one of the chunks, The model setting unit may be configured to be able to set the chunk to be redisplayed by the third display control unit (15th configuration).

[0123] Furthermore, in the simulation device of any of the first to fifteenth configurations, it may be configured to include a fourth display control unit (7) configured to control the display of the first chunk of the set chunks (sixteenth configuration).

[0124] In addition, in the simulation device of the sixth configuration, a fifth display control unit (7) is provided that is configured to control the display of the processing results of the window function processing and the frequency analysis processing for one of the chunks, The model setting unit may be configured to be able to set the chunk for redisplaying the processing result (seventeenth configuration).

[0125] Furthermore, the simulation device of any of the first to seventeenth configurations may be configured to include a sixth display control unit (7) configured to control the display of the value of a loss function based on the inference results during learning by the machine learning model for each chunk (18th configuration).

[0126] Furthermore, the simulation device of any of the first to eighteenth configurations may be configured to include a seventh display control unit (7) configured to control the display of the value of a loss function based on the inference result after learning using the machine learning model for each chunk (19th configuration).

[0127] Furthermore, in the simulation device of any of the first to nineteenth configurations, the data may be configured to be capable of including time-series data of signals in normal and abnormal states of the motor (twentieth configuration).

[0128] Furthermore, in the simulation device of any of the first to twentieth configurations above, the model setting unit may be configured to select a setting regarding the calculation accuracy of the machine learning model from a plurality of options including at least one of double, float32, float16, and Bfloat16 (21st configuration).

[0129] Furthermore, the program (P) according to one embodiment of the present disclosure includes: The first step is to load the data, A second step of setting chunks as a set of data to be sequentially input into a machine learning model based on the read data; A third step of setting the calculation accuracy of the machine learning model; a fourth step of sequentially inputting the chunks into the machine learning model to perform unsupervised learning and inference; This is a program for causing a computer (100) to execute the above.

[0130] Furthermore, a simulation method according to an aspect of the present disclosure includes: The first step is to load the data, A second step of setting chunks as a set of data to be sequentially input into a machine learning model based on the read data; A third step of setting the calculation accuracy of the machine learning model; and a fourth step of performing unsupervised learning calculations and inference calculations by sequentially inputting the chunks into the machine learning model.

[0131] Furthermore, an information processing device according to an aspect of the present disclosure includes: a processor; The single processor is configured to be able to switch between performing calculations at a first calculation accuracy using a machine learning model and performing calculations at a second calculation accuracy different from the first calculation accuracy using the machine learning model (22nd configuration).

[0132] In addition, in the information processing device of the above-mentioned 22nd configuration, the single processor may be configured to perform estimation calculations with the same calculation accuracy as the learning calculations using the trained machine learning model (23rd configuration). [Industrial Applicability]

[0133] The present disclosure can be used, for example, in unsupervised learning simulations for a variety of applications. [Explanation of symbols]

[0134] 1 Simulation device 2 File storage section 3 File reading section 4 Model storage section 5 Model calculation section 6 Model setting section 7 Display control section 8 Operation input section 9 Display section 10 Three-layer neural network 10A Input Layer 10B Hidden Layer 10C output layer 21 Data Files 40 Machine Learning Models 100 computers 100A CPU 100B memory 100C auxiliary storage 100D Operation input section 100E Display section 210 Data 401 Pretreatment section 402 Machine Learning Department P Program< / gui>

Claims

1. a model storage unit storing a machine learning model configured to perform learning and inference; a model calculation unit configured to perform calculation processing using the machine learning model; an operation input unit; a reading unit configured to read data; a model setting unit configured to perform settings related to the machine learning model based on input from the operation input unit; Equipped with the model setting unit is configured to perform settings related to chunks as a set of data for one time input to the machine learning model based on the read data, the model calculation unit is configured to perform unsupervised learning calculations and inference calculations by sequentially inputting the chunks into the machine learning model; The simulation device, wherein the model setting unit is configured to perform settings related to the calculation accuracy of the machine learning model.

2. 2. The simulation device according to claim 1, wherein the model setting unit is configured to select a setting related to the calculation accuracy of the machine learning model from a plurality of options including at least one of double, float32, float16, and Bfloat16.

3. A first step of reading the data; a second step of setting chunks as a set of data to be sequentially input to a machine learning model based on the read data; a third step of setting the calculation accuracy of the machine learning model; a fourth step of sequentially inputting the chunks into the machine learning model to perform unsupervised learning and inference; A program that causes a computer to execute the following.

4. A first step of reading the data; a second step of setting chunks as a set of data to be sequentially input to a machine learning model based on the read data; a third step of setting the calculation accuracy of the machine learning model; a fourth step of sequentially inputting the chunks into the machine learning model to perform unsupervised learning and inference; A simulation method comprising:

5. a processor; An information processing device, wherein the single processor is configured to be able to switch between performing calculations at a first calculation precision using a machine learning model and performing calculations at a second calculation precision using the machine learning model that is different from the first calculation precision.

6. The information processing device according to claim 5 , wherein the single processor is configured to perform an estimation operation using the trained machine learning model with the same calculation accuracy as that of the learning operation.

Images