Machine learning device, electronic device, machine learning program, and simulation device

Abstract

The machine learning device has room for further consideration regarding detection of a state abnormality of the mechanical system.SOLUTION: The machine learning device includes a model holding unit that holds a machine learning model, and an arithmetic unit. The processor is configured to calculate a first calculation result by inputting the input data to the machine learning model and performing inference, calculate a second calculation result by inputting output data in the first calculation result to the machine learning model and performing inference, and calculate an intermediate layer error by a loss function based on a first intermediate layer anomaly degree calculated based on data included in the first calculation result and a second intermediate layer anomaly degree calculated based on data included in the second calculation result.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The invention disclosed in this specification relates to a machine learning device, an electronic device, a machine learning program, and a simulation device. [Background technology]

[0002] In the past, in the field of industrial factory equipment maintenance, AI (artificial intelligence) has been increasingly applied to condition-based maintenance of mechanical systems.

[0003] As an example of the related prior art, Patent Document 1 can be mentioned. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] International Publication No. 2019 / 035279

[0005] [overview] The machine learning device disclosed in Patent Document 1 leaves room for further consideration in terms of detecting abnormal conditions in mechanical systems.

[0006] The machine learning device disclosed in this specification includes a model holding unit and a calculation unit. The model holding unit is configured to hold a machine learning model including an input layer, an output layer, and at least one hidden layer disposed between the input layer and the output layer. The calculation unit is configured to input input data to the machine learning model and perform inference to calculate a first calculation result, input output data included in the output layer of the first calculation result to the machine learning model and perform inference to calculate a second calculation result, and calculate a hidden layer error using a loss function based on the first hidden layer anomaly degree calculated based on the data included in the hidden layer of the first calculation result and the second hidden layer anomaly degree calculated based on the data included in the hidden layer of the second calculation result.

[0007] The electronic device described in this specification includes the machine learning device having the above-described configuration.

[0008] The machine learning program disclosed in this specification is intended to function as the machine learning device having the above configuration.

[0009] The simulation device described in this specification calculates the intermediate layer error using the machine learning device configured as described above.

[0010] The anomaly detection method disclosed in this specification is an anomaly detection method using a machine learning device, the anomaly detection method including: a model holding unit configured to hold a machine learning model including an input layer, an output layer, and at least one hidden layer disposed between the input layer and the output layer; and a calculation unit configured to input predetermined input data to the machine learning model, perform inference to calculate a calculation result, and calculate a hidden layer error based on multiple calculation results. This anomaly detection method includes the steps of inputting first input data as input data to the machine learning model, performing inference, and calculating a first calculation result as a calculation result; inputting output data included in the output layer of the first calculation result to the machine learning model, performing inference, and calculating a second calculation result as a calculation result; and calculating a hidden layer error based on data included in the hidden layer of the first calculation result and data included in the hidden layer of the second calculation result. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a diagram showing the configuration of a computer 100. As shown in FIG. [Figure 2] FIG. 2 is a block diagram showing the configuration of the machine learning device 6 according to the first embodiment of the present disclosure. [Figure 3] FIG. 3 is a diagram showing the configuration of the machine learning model 80. [Figure 4] FIG. 4 is a diagram showing a schematic diagram of an aspect in which an input / output error da1 and a second hidden layer error da4 are generated using input data 71 and a machine learning model 80. [Figure 5]FIG. 5 is a graph showing an example of the input data 71. [Figure 6] FIG. 6 is a graph showing an enlarged portion of FIG. 5 from time t1 to time t2. [Figure 7] FIG. 7 is a graph showing an enlarged portion of FIG. 5 from time t4 to time t5. [Figure 8] FIG. 8 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71 is input to the machine learning model 80. [Figure 9] Figure 9 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71 other than the input data 71 illustrated in Figures 5 and 8 is input into the machine learning model 80. [Figure 10] Figure 10 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71 other than the input data 71 illustrated in Figures 5, 8, and 9 is input into the machine learning model 80. [Figure 11] FIG. 11 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71, which is simulation data, is input to the machine learning model 80. [Figure 12] FIG. 12 is a flowchart of an anomaly detection method using the machine learning device 6. [Figure 13] FIG. 13 is a flowchart showing a detailed configuration of the first calculation step. [Figure 14] FIG. 14 is a flowchart showing a detailed configuration of the second calculation step. [Figure 15] FIG. 15 is a flowchart showing a detailed configuration of the third calculation step. [Figure 16] Figure 16 is a diagram schematically showing how the input / output error da1, the first hidden layer anomaly degree da2, the first hidden layer error da3, and the second hidden layer error da4 are calculated in the machine learning device 6 according to the second embodiment.

[0012] [Detailed explanation] <About Computer 100> First, a computer 100 that functions as a machine learning device 6 according to the present disclosure will be described. Next, the machine learning device 6 according to the first embodiment of the present disclosure will be described in detail.

[0013] 1 is a diagram showing the configuration of a computer 100. The computer 100 functions as a machine learning device 6, which will be described later. The computer 100 is, for example, a PC (Personal 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 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] Here, the auxiliary storage device 100C stores a simulation program P. The simulation program P is a program for causing the computer 100 to function as a machine learning device 6, which will be described later. The machine learning device 6 will be described in detail later.

[0019] 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.

[0020] 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.

[0021] <Regarding the machine learning device 6 of the first embodiment> Next, a description will be given of the machine learning device 6 according to the first embodiment of the present disclosure. The machine learning device 6 is configured with an MCU (Micro Controller Unit). The machine learning device 6 is incorporated into a predetermined mechanical system (such as a motor device) and controls this mechanical system. In addition to controlling the mechanical system, the machine learning device 6 can also perform machine learning using various data of this mechanical system as input data.

[0022] 2 is a block diagram showing the configuration of a machine learning device 6 according to the first embodiment of the present disclosure. As shown in FIG. 2, the machine learning device 6 includes a data storage unit 7, a model holding unit 8, a calculation unit 9, and an anomaly detection unit 10.

[0023] The data storage unit 7 stores input data 71 and initial value data 72. The input data 71 is, for example, time-series data output from a mechanical system or the like. This time-series data may be subjected to preprocessing such as normalization or FFT as necessary. The initial value data 72 is set to an initial value determined by the computer 100 as described above.

[0024] The model storage unit 8 stores a machine learning model 80. The machine learning model 80 is a machine learning model that is capable of learning and inference according to input data. Details of the machine learning model 80 will be described later.

[0025] The calculation unit 9 calculates a first calculation result 30 using input data 71 and a machine learning model 80. The first calculation result 30 includes first output data do1, an input / output error da1, a first hidden layer vector ha, and a first hidden layer anomaly degree da2, which will be described later. The calculation unit 9 also calculates a second calculation result 31 using the first output data do1 and the machine learning model 80. The calculation unit 9 will be described in detail as follows.

[0026] The calculation unit 9 includes a learning calculation unit 91, an inference calculation unit 92, and an anomaly degree calculation unit 93. The learning calculation unit 91 performs unsupervised learning using the machine learning model 80, input data 71, and initial value data 72.

[0027] The inference calculation unit 92 performs inference using the machine learning model 80, the input data 71, and the initial value data 72. The inference can be performed during the above-mentioned learning by the learning calculation unit 91 and after the learning is completed.

[0028] The anomaly degree calculation unit 93 calculates the input / output error da1, the first hidden layer anomaly degree da2, the second hidden layer anomaly degree db1, and the second hidden layer error da4 using the input data 71 and the machine learning model 80. The anomaly degree calculation unit 93 transmits the calculated input / output error da1, the first hidden layer anomaly degree da2, the second hidden layer anomaly degree db1, and the second hidden layer error da4 to the anomaly detection unit 10 as a calculation result AS.

[0029] The input / output error da1 is the error between the value contained in the input layer 50A (=input data 71) and the value contained in the output layer 50C (=first output data do1) when input data 71 is input into the machine learning model 80 and inference is performed, and is calculated using the loss function described below.

[0030] The second hidden layer error da4 is the error between the first hidden layer anomaly degree da2 (described later) and the second hidden layer anomaly degree db1 (described later), and is calculated using a loss function (described later).

[0031] The first hidden layer anomaly degree da2 represents the normalized distance between the first hidden layer vector ha and the mean vector of the first hidden layer vector ha. The first hidden layer vector ha is a feature vector that indicates the characteristics of the hidden layer 50B when the input data 71 is input to the machine learning model 80 and inferred.

[0032] The second hidden layer anomaly degree db1 represents the normalized distance between the second hidden layer vector hb and the mean vector of the second hidden layer vector hb. The second hidden layer vector hb is the feature vector of the hidden layer 50B when the first output data do1 is input to the machine learning model 80 and inferred.

[0033] In the case of time-series data in which input data 71 repeats at a predetermined cycle, if the trend of the data changes during repetition, there is a possibility that the input / output error da1, the first hidden layer anomaly degree da2, the second hidden layer anomaly degree db1, and the second hidden layer error da4 will also change. In this case, the input / output error da1, the first hidden layer anomaly degree da2, the second hidden layer anomaly degree db1, and the second hidden layer error da4 may each have a different trend or a similar trend. Detailed methods for calculating the input / output error da1, the first hidden layer anomaly degree da2, the second hidden layer anomaly degree db1, and the second hidden layer error da4 will be described later.

[0034] The anomaly detection unit 10 receives the calculation result AS from the anomaly degree calculation unit 93 and detects whether an anomaly has occurred in the input data 71 from the data included in the calculation result AS. Specifically, the anomaly detection unit 10 refers to the trends of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4, determines whether a change has occurred in the data trend over time, and if a change has occurred, determines that there is an anomaly. The anomaly detection unit 10 outputs the detection result to the display unit 100E. When the computer 100 is made to function as the machine learning device 6 as described above, the anomaly detection unit 10 outputs the detection result to the display unit 100E.

[0035] Next, the machine learning model 80 will be described in detail. FIG. 3 is a diagram showing the configuration of the machine learning model 80. The machine learning model 80 is an inference model that can learn using predetermined training data. As shown in FIG. 3, the machine learning model 80 includes a three-layer neural network 50.

[0036] The three-layer neural network 50 is an AI model having an input layer 50A, a hidden layer 50B, and an output layer 50C. The hidden layer 50B is also called an intermediate layer. Generally, in a three-layer neural network 50, for n-dimensional input data x∈Rk×n with a batch size k, an n'-dimensional inference result y∈Rk×n' is obtained as y=G(x·α+b)β. Here, α∈Rn×m is a weight connecting the input layer 50A and the hidden layer 50B. β∈Rm×n' is a weight connecting the hidden layer 50B and the output layer 50C. Furthermore, b∈Rm is a bias for the hidden layer 50B. G is an activation function for the hidden layer 50B. For example, a Sigmoid or ReLU function can be used as the activation function.

[0037] The three-layer neural network 50 employs an algorithm that allows sequential learning with any batch size. When the i-th training data {xi∈Rki×n, ti∈Rki×n'} of batch size ki is obtained, it is necessary to find βi that minimizes the error shown in the following equation (1).

number

[0038] The optimized weight βi is calculated by the following equation (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] The computational bottleneck in the above equation (2) is (I+H i P i-1 H iT ) -1 However, (I+H i P i-1 H i T ) has a matrix size of k × k, so when k = 1, the inverse matrix operation can be replaced with the reciprocal operation. Therefore, by fixing the batch size to k = 1, the operation becomes easy even for a computing device such as a microcomputer.

[0042] Furthermore, the machine learning model 80 according to this embodiment performs learning using an autoencoder. The autoencoder uses input data as training data and learns to reconstruct the input data as an inference result. In other words, in the above example, it learns with t = x. Because the autoencoder does not require separate training data creation, it is a type of unsupervised learning algorithm. Furthermore, by making the number of nodes in the hidden layer smaller than the number of nodes in the input and output layers, the hidden layer matrix can be considered a dimensionally compressed form of the input data once the error between the input data and the inference result has converged. That is, the encoding result of input data x is H = G(x·γ + b), and the decoding result of H is obtained as y = H·ι.

[0043] FIG. 4 is a diagram showing a schematic diagram of an aspect in which an input / output error da1 and a second hidden layer error da4 are generated using input data 71 and a machine learning model 80.

[0044] As shown in Fig. 4, input data 71 is input to a machine learning model 80 and calculated by a calculation unit 9, thereby obtaining a first calculation result 30. The first calculation result 30 includes first output data do1, an input / output error da1, and a first hidden layer vector ha. Specifically, this is as follows.

[0045] Input data 71 is input to machine learning model 80 and inference is performed by inference calculation unit 92, thereby obtaining first output data do1 as the inference result. At this time, input / output error da1 is calculated by anomaly degree calculation unit 93. At this time, first hidden layer vector ha is obtained as the feature vector of hidden layer 50B.

[0046] Furthermore, the first output data do1 is input to the machine learning model 80 and calculated by the calculation unit 9, thereby obtaining a second calculation result 31. The second calculation result 31 includes the second output data do2 and the second hidden layer vector hb. Specifically, this is as follows.

[0047] By inputting the first output data do1 to the machine learning model 80 and performing inference using the inference calculation unit 92, the second output data do2 is obtained as the inference result. At this time, the second hidden layer vector hb is obtained as the feature vector of the hidden layer 50B.

[0048] The first hidden layer anomaly degree da2 and the second hidden layer anomaly degree db1 are then calculated by the anomaly degree calculation unit 93. The second hidden layer error da4 is then calculated based on the first hidden layer anomaly degree da2 and the second hidden layer anomaly degree db1 by the anomaly degree calculation unit 93. Details of each calculation method will be described later.

[0049] <Example of anomaly detection> Next, a specific example will be used to explain the case where anomaly detection is performed on input data 71. FIG. 5 is a graph showing an example of input data 71. In FIG. 5, the input data 71 is shown as a time series graph with the horizontal axis representing time and the vertical axis representing a predetermined output value. FIG. 6 is a graph showing an enlarged view of the portion from time t1 to time t2 in FIG. 5. FIG. 7 is a graph showing an enlarged view of the portion from time t4 to time t5 in FIG. 5.

[0050] In Fig. 5, the start point of the data is time t0 and the end point of the data is time t6. Also in Fig. 5, the period from time t0 to time t3 is called a normal period T1. Also in Fig. 5, the period from time t3 to time t6 is called an abnormal period T2.

[0051] Normal period T1 is a period during which no abnormality occurs in the input data 71. In other words, normal period T1 is a period during which it is estimated that no particular abnormality occurs in the output value of the mechanical system implementing the machine learning device 6. On the other hand, abnormal period T2 is a period during which an abnormality occurs in the input data 71. In other words, abnormal period T2 is a period during which it is estimated that some kind of abnormality occurs in the output value of the mechanical system implementing the machine learning device 6.

[0052] As shown in Figures 5 and 6, the input data 71 during the normal period T1 has a waveform similar to a sine wave that repeats with a constant cycle. On the other hand, as shown in Figures 5 and 7, the input data 71 during the abnormal period T2 has a waveform that repeats with a constant cycle, but an abnormal portion (hereinafter also referred to as abnormal portion Ap) occurs in the latter half of one cycle of the waveform.

[0053] However, as shown in Fig. 5, it is difficult for a user to determine whether an abnormality has occurred in the input data 71 at a glance. Enlarging the graph of the input data 71 as shown in Fig. 7 may help in making a judgment, but it becomes cumbersome.

[0054] Figure 8 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71 is input to a machine learning model 80. In Figure 8, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 are shown as a time series graph with the horizontal axis representing time and the vertical axis representing a predetermined value.

[0055] The period from time t0 to time ta shown in Figure 8 is referred to as learning period T3. Learning period T3 is a period during which input data 71 is input to machine learning model 80 and learning is performed. The period from time ta to time t6 is referred to as inference period T4. Inference period T4 is a period during which input data 71 is input to machine learning model 80 and inference is performed.

[0056] During the learning period T3, the features of the input data 71 are incorporated. As a result, the data values ​​(=values ​​on the vertical axis) of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 change significantly. During the learning period T3, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 are unstable and sensitive to changes in the input values, making this period unsuitable for detecting abnormal values ​​in the input data 71. For this reason, we will now determine whether an abnormal value has occurred in the input data 71 during the inference period T4.

[0057] As shown in Figure 8, during the inference period T4 and the normal period T1, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 each remain stable and show no significant changes.

[0058] On the other hand, during the abnormal period T2, the input / output error da1 exhibits a significant change in the data trend compared to the normal period T1. Specifically, the data values ​​rise more steeply and more frequently than during the normal period T1.

[0059] Similarly, during abnormal period T2, the first hidden layer anomaly degree da2 drops sharply at time t3, then remains flat at 0. Conversely, during abnormal period T2, the second hidden layer error da4 rises sharply at time t3, then remains flat at approximately 0.1. From these facts, it is possible to infer some kind of change in trend in the input data 71 between normal period T1 and abnormal period T2, and ultimately the occurrence of an abnormality.

[0060] When the anomaly detection unit 10 detects a change in the trend of the input / output error da1, the first hidden layer anomaly degree da2, or the second hidden layer error da4, it outputs the result as a detection result. The output method can be, for example, a method of displaying a warning, alarm, or the like on the display unit 100E.

[0061] By checking and analyzing the detection results output by the anomaly detection unit 10, the user can determine whether the input data 71 contains abnormal values, and ultimately whether an abnormality has occurred in the mechanical system or the like in which the machine learning device 6 is implemented.

[0062] <Another example of anomaly detection> Next, using another data example, we will explain the detection of abnormal values ​​in the input data 71. Here, we will explain an example in which motor current is treated as input data 71, and damage to the inner ring of the motor is detected by detecting an abnormality in the motor current.

[0063] Figure 9 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71 other than the input data 71 illustrated in Figures 5 and 8 is input into the machine learning model 80.

[0064] In FIG. 9, the input data 71, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 are shown in a time series graph with the horizontal axis representing time and the vertical axis representing a predetermined value.

[0065] In FIG. 9, the start point of the data is time t20 and the end point of the data is time t24. The period from time t20 to time t21 is called normal period T1. The period from time t21 to time t22 is called abnormal period T2a. The period from time t22 to time t23 is called abnormal period T2b. The period from time t23 to time t24 is called abnormal period T2c. Abnormal periods T2a to T2c are synonymous with the aforementioned abnormal period T2. However, when attention is paid to the details, there are slight differences in the abnormal conditions occurring in each of abnormal periods T2a to T2c.

[0066] 9, the period from time t20 to time tb is defined as learning period T3. The period from time tb to time t24 is defined as inference period T4. As explained with reference to FIG. 8, it is determined whether an abnormality has occurred in input data 71 during inference period T4.

[0067] As shown in Figure 9, during the inference period T4 and the normal period T1, the first hidden layer anomaly degree da2 and the second hidden layer error da4 remain stable and do not show any significant changes. During the same period, the input / output error da1 fluctuates slightly.

[0068] Next, during the abnormal period T2a, the input / output error da1 fluctuates slightly, to the same extent as during the normal period T1. However, it can also be said that the period of fluctuation is slightly shorter than during the normal period T1. During the abnormal period T2a, the first hidden layer anomaly degree da2 remains stable, just as it did during the normal period T1. During the abnormal period T2a, the second hidden layer error da4 remains stable, just as it did during the normal period T1. However, when looking closely, it can also be said that there is a slight fluctuation in the data value of the second hidden layer error da4.

[0069] Next, during the abnormal period T2b, the input / output error da1 fluctuates slightly, similar to the normal period T1 and the abnormal period T2a. However, the period of the fluctuations can be said to be shorter or more turbulent than during the normal period T1 and the abnormal period T2a. During the abnormal period T2b, the first hidden layer anomaly degree da2 remains stable, just as during the normal period T1 and the abnormal period T2a. During the abnormal period T2b, the second hidden layer error da4 fluctuates in data values ​​compared to the normal period T1. Furthermore, during the abnormal period T2b, the fluctuation of the second hidden layer error da4 increases compared to the abnormal period T2a.

[0070] Next, during the abnormal period T2c, the data value of the input / output error da1 increases compared to the normal period T1, the abnormal period T2a, and the abnormal period T2c. More specifically, the data value rises sharply from time t23 onward. Between time t23 and time t24, the data value repeatedly increases and decreases, with the value of the rising edge at time t23 as the reference.

[0071] In the abnormal period T2c, the first hidden layer anomaly degree da2 remains stable, as in the normal period T1, the abnormal period T2a, and the abnormal period T2b. However, when looking at the details, it can be said that the amount of fluctuation in the data value is larger than in the normal period T1, the abnormal period T2a, and the abnormal period T2b, and that there is a slight disturbance in the fluctuation of the data value.

[0072] In the abnormal period T2c, the second hidden layer error da4 exhibits fluctuations in data values ​​compared to the normal period T1. In addition, in the abnormal period T2b, the fluctuations in the second hidden layer error da4 increase compared to the abnormal periods T2a and T2c.

[0073] From the above, by referring to the input / output error da1, it can be estimated that an abnormality has occurred in the input data 71 (and thus that damage has occurred to the inner ring of the motor) during the period from time t21 to time t24. Looking more closely, it can be estimated that the abnormality that occurred in the input data 71 has shifted from time t21 to time t24, that is, that the damage to the inner ring of the motor is progressing.

[0074] Furthermore, by referring to the first hidden layer anomaly degree da2, it can be estimated that an anomaly has occurred in the input data 71 (and thus that the inner ring of the motor has been damaged) at least during the period from time t23 to time t24.

[0075] Furthermore, by referring to the second hidden layer error da4, it can be estimated that an abnormality has occurred in the input data 71 (and thus that damage has occurred to the inner ring of the motor) during the period from time t21 to time t24. Looking more closely, it can be estimated that the abnormality that occurred in the input data 71 has shifted from time t21 to time t24, that is, that the damage to the inner ring of the motor is progressing.

[0076] The anomaly detection unit 10 may be configured to recognize slight changes such as those described above (for example, the trend change occurring in the first hidden layer anomaly degree da2 during the period from time t23 to time t24) as an anomaly. The anomaly detection unit 10 may be configured to be changeable in terms of the setting that defines whether the detection result is abnormal or not.

[0077] If the anomaly detection unit 10 is configured not to recognize such slight changes as an anomaly, there is a risk that the difference in trend between the normal period T1 and the anomaly periods T2a and T2b will not be determined to be an anomaly simply by referring to the input / output error da1. This could lead to a determination that there is no anomaly in the input data 71 during the period from time t0 to time t23. The same applies when only the first hidden layer anomaly degree da2 is referred to.

[0078] However, in the machine learning device 6 according to this embodiment, the anomaly detection unit 10 refers to the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4, respectively, to detect differences in the trends of the respective data. Therefore, anomalies in the input data 71 can be found with higher accuracy than when only some of the calculation results (for example, the input / output error da1 and the first hidden layer anomaly degree da2) are calculated.

[0079] <Yet another example of anomaly detection> Next, using yet another data example, we will explain the detection of abnormal values ​​in the input data 71. Here, we will explain an example in which motor vibration is treated as input data 71, and damage to the inner ring of the motor is detected by detecting the displacement of the motor vibration.

[0080] Figure 10 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71 other than the input data 71 illustrated in Figures 5, 8, and 9 is input into the machine learning model 80.

[0081] In FIG. 10, the input data 71, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 are shown in a time series graph with the horizontal axis representing time and the vertical axis representing a predetermined value.

[0082] In Figure 10, the start point of the data is time t30 and the end point of the data is time t34. The period from time t30 to time t31 is called normal period T1. The period from time t31 to time t32 is called abnormal period T2d. The period from time t32 to time t33 is called abnormal period T2e. The period from time t33 to time t34 is called abnormal period T2f. Abnormal periods T2d to T2f are synonymous with the aforementioned abnormal period T2. More specifically, the abnormal values ​​occurring in each of abnormal periods T2d to T2f are slightly different, similar to the aforementioned abnormal periods T2a to T2c.

[0083] 10, the period from time t30 to time tc is defined as learning period T3. The period from time tb to time t34 is defined as inference period T4. As explained with reference to FIGS. 8 and 9, it is determined whether an abnormality has occurred in input data 71 during inference period T4.

[0084] As shown in Figure 10, during the inference period T4 and the normal period T1, the input / output error da1 remains stable at 0. During the same period, the first hidden layer anomaly degree da2 and the second hidden layer error da4 each undergo slight fluctuations.

[0085] Next, in the abnormal period T2d, the input / output error da1 remains flat, just as it did in the normal period T1. In the abnormal period T2d, the first hidden layer anomaly degree da2 shows an increase in data value compared to the normal period T1. In the abnormal period T2d, the second hidden layer error da4 shows a large increase or decrease in data value compared to the normal period T1.

[0086] Next, in the abnormal period T2e, the input / output error da1 fluctuates slightly compared to the normal period T1 and the abnormal period T2d. In the abnormal period T2e, the first hidden layer anomaly degree da2 increases slightly compared to the normal period T1 and the abnormal period T2d, and then remains at the increased value. In the abnormal period T2e, the second hidden layer error da4 fluctuates in data value compared to the normal period T1. Furthermore, the fluctuation of the second hidden layer error da4 in the abnormal period T2e is not significantly different from that in the abnormal period T2d. However, it can also be said that the amount of fluctuation of the second hidden layer error da4 is slightly smaller than that in the abnormal period T2d.

[0087] Next, in the abnormal period T2f, the input / output error da1 exhibits larger fluctuations in data values ​​compared to the normal period T1, the abnormal period T2d, and the abnormal period T2e. In the abnormal period T2f, the first hidden layer anomaly degree da2 is similar to that in the normal period T1, the abnormal period T2d, and the abnormal period T2e, and is not significantly different.

[0088] During the abnormal period T2f, the second hidden layer error da4 exhibits fluctuations in the data values ​​compared to the normal period T1. However, the fluctuations in the second hidden layer error da4 during the abnormal period T2f are not significantly different from those during the abnormal periods T2e and T2f. However, the amount of fluctuation in the second hidden layer error da4 during the abnormal period T2f is smaller than those during the abnormal periods T2e and T2f.

[0089] From the above, it can be said that by referring to the input / output error da1, it can be estimated that an abnormality has occurred in the input data 71 (and thus that damage has occurred to the inner ring of the motor) at least during the period from time t32 to time t34. Furthermore, by focusing on the details, it can be estimated that the abnormality that occurred in the input data 71 has shifted between time t32 and time t34, that is, that the damage to the inner ring of the motor is progressing.

[0090] Furthermore, by referring to the first hidden layer anomaly degree da2, it can be said that it is possible to estimate that an abnormality has occurred in the input data 71 (and thus that damage has occurred to the inner ring of the motor) during the period from time t31 to time t34. When focusing on the details, it can be estimated that the abnormality that has occurred in the input data 71 has shifted, that is, that the damage to the inner ring of the motor is progressing, for the period from time t31 to time t32, the period from time t32 to time t33, and the period from time t33 to time t34.

[0091] Furthermore, by referring to the second hidden layer error da4, it can be estimated that an abnormality has occurred in the input data 71 (and thus that damage has occurred to the inner ring of the motor) during the period from time t31 to time t34. Focusing on the details, it can be estimated that the abnormality that has occurred in the input data 71 has shifted, that is, that the damage to the inner ring of the motor is progressing, for the period from time t31 to time t32, the period from time t32 to time t33, and the period from time t33 to time t34.

[0092] <Example of anomaly detection performed on simulation data> Next, a description will be given of detection of abnormal values ​​in the input data 71 when predetermined simulation data is used as the input data 71. Here, the input data 71 is simulation data obtained by simulating damage to the inner ring of the motor as described above.

[0093] FIG. 11 is a graph showing the calculation results of the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 when input data 71, which is simulation data, is input to the machine learning model 80.

[0094] In FIG. 11, the input data 71, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 are shown in a time series graph with the horizontal axis representing time and the vertical axis representing a predetermined value.

[0095] 11, the start point of the data is time t40 and the end point of the data is time t43. The period from time t40 to time t41 is called normal period T1. The period from time t41 to time t42 is called abnormal period T2g. The period from time t42 to time t43 is called abnormal period T2h.

[0096] 11, the period from time t40 to time td is defined as learning period T3. The period from time td to time t43 is defined as inference period T4. As explained with reference to FIGS. 8 to 10, it is determined whether an abnormality has occurred in input data 71 during inference period T4.

[0097] As shown in FIG. 11, during the inference period T4 and the normal period T1, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4 remain stable.

[0098] Next, in the abnormal period T2g, both the input / output error da1 and the first hidden layer anomaly degree da2 remain stable, just as they did in the normal period T1. In the abnormal period T2g, the second hidden layer error da4 exhibits fluctuations in the data value compared to the normal period T1.

[0099] Next, in the abnormal period T2h, the input / output error da1 and the second hidden layer error da4 both fluctuate in data value compared to the normal period T1 and the abnormal period T2g. Also, when looking at the details, the first hidden layer anomaly degree da2 fluctuates slightly in data value compared to the normal period T1 and the abnormal period T2g.

[0100] From the above, by referring to the input / output error da1, it can be estimated that an abnormality has occurred in the input data 71 (and thus that the inner ring of the motor has been damaged) at least during the period from time t42 to time t44.

[0101] Furthermore, by looking at the first hidden layer anomaly degree da2, if we focus on the details, we can estimate that an abnormality has occurred in the input data 71 (and therefore that damage has occurred to the inner ring of the motor) at least during the period from time t42 to time t44.

[0102] Furthermore, by referring to the second hidden layer error da4, it can be estimated that an abnormality has occurred in the input data 71 (and thus that the inner ring of the motor has been damaged) during the period from time t41 to time t43. It can also be estimated that the abnormality that has occurred in the input data 71 has shifted between the period from time t41 to time t42 and the period from time t42 to time t43, that is, that the damage to the inner ring of the motor is progressing.

[0103] <About input / output error da1> Next, a method for calculating the input / output error da1 will be described. The input / output error da1 is the error between the second output data do2 and the input data 71, and is calculated using a loss function, which will be described later. The second output data do2 is data of the inference result obtained by inputting the first output data do1 into the machine learning model 80 and performing inference. The input / output error da1 is calculated using a loss function based on the input data 71 and the second output data do2. Specifically, it is as follows.

[0104] Each input value included in the input data 71 is defined as an input value x. Each output value (each value of the inference result) included in the second output data do2 is defined as an output value y. As a loss function for calculating the input / output error da1, for example, MAE (Mean Absolute Error) or MSE (Mean Squared Error) can be used. In this case, when the loss function is MAE, the loss function L is expressed as in the following equation (4).

number

number

[0105] <About the second hidden layer error da4> Next, we will explain how to calculate the second hidden layer error da4. As mentioned above, the second hidden layer error da4 is the error between the first hidden layer anomaly degree da2 and the second hidden layer anomaly degree db1, and is calculated using a loss function described below. First, we will explain the first hidden layer anomaly degree da2 and the second hidden layer anomaly degree db1, and then we will explain the second hidden layer error da4.

[0106] As described above, the first hidden layer anomaly degree da2 represents the normalized distance between the first hidden layer vector ha and the mean vector of the first hidden layer vector ha. The first hidden layer vector ha is expressed by the following equation (6).

number

[0107] The mean vector of the first hidden layer vector ha is expressed by the following equations (7) and (8).

number

number

[0108] The first hidden layer abnormality degree da2 is calculated by the following equation (9).

number

[0109] As described above, the second hidden layer anomaly degree db1 represents the normalized distance between the second hidden layer vector hb and the mean vector of the second hidden layer vector hb. The second hidden layer vector hb is expressed by the following equation (10).

number

[0110] The mean vector of the second hidden layer vector hb is expressed by the following equations (11) and (12).

number

number

[0111] The second hidden layer abnormality degree db1 is calculated by the following equation (13).

number

[0112] The second hidden layer error da4 is calculated by a loss function L that represents the error between the first hidden layer anomaly degree da2 and the second hidden layer anomaly degree db1. When the loss function is MAE, the loss function L is expressed by the following equation (14).

number

number

[0113] <Anomaly detection method> Next, an anomaly detection method using the machine learning device 6 will be described.

[0114] 12 is a flowchart of an anomaly detection method using a machine learning device 6. As shown in FIG. 12, first, a first calculation step is executed (step St1). In the first calculation step, input data 71 is input to a machine learning model 80 to perform inference and calculate a first calculation result 30. As described above, the first calculation result 30 includes the first output data do1, the input / output error da1, and the first hidden layer vector ha.

[0115] Next, a second calculation step is executed (step St2). In the second calculation step, the first output data do1 is input to the machine learning model 80 to perform inference and calculate the second calculation result 31. As described above, the second calculation result 31 includes the second output data do2 and the second hidden layer vector hb.

[0116] Next, the third calculation step is executed (step St3). In the third calculation step, the second hidden layer error da4 is calculated based on the first calculation result 30 and the second calculation result 31. The first calculation step (step St1) to the third calculation step (step St3) will be described in detail below.

[0117] Fig. 13 is a flowchart showing a detailed configuration of the first calculation step. As shown in Fig. 13, first, a user prepares input data 71 (step St11). Step St11 includes selection and extraction of input data, predetermined pre-processing (for example, statistical processing, FFT analysis, etc.), etc.

[0118] Next, the inference calculation unit 92 executes inference based on the input data 71 and the machine learning model 80, and generates first output data do1 as the inference result (step St12). Next, the anomaly degree calculation unit 93 calculates the input / output error da1 (step St13). The anomaly degree calculation unit 93 also acquires the first hidden layer vector ha based on the hidden layer 50B (step St14). Therefore, the first output data do1, the input / output error da1, and the first hidden layer vector ha are generated as the first calculation result 30 after the first calculation step (see FIG. 4). Then, the process proceeds to the second calculation step (step St2).

[0119] 14 is a flowchart showing a detailed configuration of the second calculation step. As shown in FIG. 14, in the second calculation step, first, the inference calculation unit 92 acquires first output data do1 from the first calculation result 30 (step St21). The first output data do1 is data included in the output layer 50C of the machine learning model 80 after the first calculation step. The acquisition of the first output data do1 may be performed by the inference calculation unit 92, another calculation unit, or a user may select the first output data do1 from the first calculation result 30.

[0120] Next, the inference calculation unit 92 inputs the acquired first output data do1 into the machine learning model 80 and executes inference again, generating second output data do2 as the inference result (step StSt22). Next, the anomaly degree calculation unit 93 acquires the second hidden layer vector hb based on the hidden layer 50B (step St23). Therefore, the second output data do2 and the second hidden layer vector hb are generated as the second calculation result 31 after the second calculation step (see FIG. 4). Then, the process proceeds to the third calculation step (step St3).

[0121] Fig. 15 is a flowchart showing a detailed configuration of the third calculation step. As shown in Fig. 15, in the third calculation step, the anomaly calculation unit 93 calculates a first hidden layer anomaly degree da2 based on the acquired first hidden layer vector ha (step st31). The anomaly calculation unit 93 also calculates a second hidden layer anomaly degree db1 based on the acquired second hidden layer vector hb (step St32). The anomaly calculation unit 93 then calculates a second hidden layer error da4 using the first hidden layer anomaly degree da2 and the second hidden layer anomaly degree db1 according to a loss function (step St33). The anomaly calculation unit 93 then transmits the calculation result AS to the display unit 100E (step St34).

[0122] <Machine Learning Device 6 According to the Second Embodiment> Next, a machine learning device 6 according to the second embodiment will be described. The machine learning device 6 of this embodiment basically has a configuration common to that of the first embodiment. Therefore, the common parts will be assigned the same reference numerals and their description will be omitted, and only the different parts will be described.

[0123] Figure 16 is a diagram schematically showing how the input / output error da1, the first hidden layer anomaly degree da2, the first hidden layer error da3, and the second hidden layer error da4 are calculated in the machine learning device 6 according to the second embodiment.

[0124] The machine learning device 6 of this embodiment includes a data storage unit 7, a model holding unit 8, a calculation unit 9, and an anomaly detection unit 10 similar to those described above (not shown). The anomaly degree calculation unit 93 calculates a first hidden layer error da3 based on the first hidden layer vector ha and the second hidden layer vector hb according to this embodiment (see FIG. 16). The anomaly degree calculation unit 93 then transmits the calculated first hidden layer error da3, together with the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4, to the anomaly detection unit 10 as a calculation result AS. The method for calculating the first hidden layer error da3 will be specifically described as follows.

[0125] <About the first hidden layer error da3> The first hidden layer error da3 is calculated by a loss function L that represents the error between the first hidden layer vector ha and the second hidden layer vector hb. When the loss function is MAE, the loss function L is expressed by the following equation (16).

number

number

[0126] Depending on the input data 71, even if an abnormal value is included, there may be cases where the input / output error da1 does not show a significant change in the trend as described above. Even in such cases, by referring to the calculated first hidden layer error da3, the user can easily determine whether the input data 71 includes an abnormal value.

[0127] As described above, the first hidden layer vector ha is a feature vector that indicates the characteristics of the hidden layer 50B when inference is performed using the input data 71. On the other hand, the second hidden layer vector hb indicates the characteristics of the hidden layer 50B when further inference is performed using the inference result (=first output data do1) based on the input data 71. Therefore, even if the input data 71 contains an abnormal value, the second hidden layer vector hb is a feature vector that indicates a feature in which the abnormal value is diluted compared to the first hidden layer vector ha.

[0128] The first hidden layer error da3 is calculated based on the features of these two hidden layers (first hidden layer vector ha, second hidden layer vector hb). In other words, this is equivalent to determining an abnormal value of the input data 71 using information from hidden layer 50B, which is a condensed version of the features of the input data 71. For this reason, even if only a slight trend change can be confirmed in the input / output error da1, a significant change may occur in the first hidden layer error da3. Therefore, by calculating the first hidden layer error da3 in addition to the input / output error da1, it becomes easier to detect an abnormal value of the input data 71.

[0129] <Modification> The present disclosure is not limited to the above-described embodiments, and various modifications are possible without departing from the spirit of the present disclosure. For example, in the configuration according to the first embodiment, the anomaly degree calculation unit 93 transmits the calculation result AS (more specifically, the input / output error da1, the first hidden layer anomaly degree da2, and the second hidden layer error da4) to the anomaly detection unit 10, but this is not limiting. For example, the anomaly degree calculation unit 93 may be configured to directly transmit the calculation result AS to the display unit 100E, or to output the calculation result externally by other means. In this case, the user may visually check the externally output calculation result and confirm the change in trend as described above, thereby determining whether or not there is an anomaly in the input data 71.

[0130] Furthermore, for example, in step St31 of the anomaly detection method using the machine learning device 6, the first hidden layer anomaly degree da2 is calculated, but the timing for calculating the first hidden layer anomaly degree da2 is not limited to this. The timing may be any time after step St13 in which the anomaly degree calculation unit 93 acquires the first hidden layer vector ha and before step St32 in which the second hidden layer error da4 is calculated.

[0131] <Additional Notes> The machine learning device (6) disclosed in the specification includes a model storage unit (8) configured to store a machine learning model (80) including an input layer (50A), an output layer (50C), and at least one intermediate layer (50B) disposed between the input layer (50A) and the output layer (50C), and a machine learning device (80) configured to input input data (71) to the machine learning model (80) to perform inference to calculate a first calculation result (30), and to use output data (do1) included in the output layer (50C) of the first calculation result (30) as a machine learning result. and a calculation unit (9) configured to input the data into a learning model (80) to perform inference to calculate a second calculation result (31), and to calculate an intermediate layer error (da4) using a loss function (L) based on a first intermediate layer anomaly degree (da2) calculated based on data (ha) included in the intermediate layer (50B) of the first calculation result (30) and a second intermediate layer anomaly degree (db1) calculated based on data (hb) included in the intermediate layer (50B) of the second calculation result (31) (first configuration).

[0132] In the machine learning device (6) according to the first configuration, the first hidden layer anomaly degree (da2) represents the first normalized distance between the first hidden layer vector (ha), which is a feature vector of the hidden layer (50B) obtained by inputting input data (71) into the machine learning model (80) and executing inference, and the mean vector of the first hidden layer vector (ha), and the second hidden layer anomaly degree (db1) represents the second normalized distance between the second hidden layer vector (hb), which is a feature vector of the hidden layer (50B) obtained by inputting output data (do1) into the machine learning model (80), and the mean vector of the second hidden layer vector (hb) (second configuration).

[0133] A machine learning device (6) according to a second configuration includes: The first hidden layer vector (ha) is

number

number

number

number

[0134] In the machine learning device (6) according to the third configuration, the first normalized distance is

number

number

[0135] The machine learning device (6) according to the fourth configuration calculates the first hidden layer anomaly degree (da2) as da2 2 When this is done, da2 2 The following formula holds:

number

number

[0136] The electronic device (100A) disclosed in the specification includes a machine learning device (6) according to any one of the first to fifth configurations (sixth configuration).

[0137] The machine learning program (P) disclosed in the specification is intended to function as a machine learning device (6) according to any one of the first to fifth configurations (seventh configuration).

[0138] The simulation device (100) disclosed in the specification calculates an intermediate layer error (da4) using a machine learning device (6) according to any one of the first to fifth configurations (eighth configuration).

[0139] The anomaly detection method disclosed in the specification is an anomaly detection method using a machine learning device (6) including: a model holding unit (8) configured to hold a machine learning model (80) including an input layer (50A), an output layer (50C), and at least one intermediate layer (50B) arranged between the input layer (50A) and the output layer (50C); and a calculation unit (9) configured to input predetermined input data (71) to the machine learning model (80) to perform inference and calculate calculation results (30, 31), and to calculate an intermediate layer error (da4) based on the multiple calculation results (30, 31), The ninth configuration includes a step of inputting first input data (71) as input data to a machine learning model (80) to execute inference and calculate a first calculation result (30) as the calculation result, a step of inputting output data (do1) included in an output layer (50C) of the first calculation result (30) to the machine learning model (80) to execute inference and calculate a second calculation result (31) as the calculation result, and a step of calculating an intermediate layer error (da4) based on data (ha) included in an intermediate layer (50B) of the first calculation result (30) and data (hb) included in the intermediate layer (50B) of the second calculation result (31) (ninth configuration). [Explanation of symbols]

[0140] 5 One-sided view 6 Machine Learning Devices 7 Data storage unit 8 Model holding section 9 Arithmetic section 10. Anomaly detection unit 30 First calculation result 31 Second operation result 50 Three-layer neural network 50A input layer 50B Hidden layer (intermediate layer) 50C output layer 71 Input Data 72 Initial Value Data 80 Machine Learning Models 91 Learning calculation unit 92 Inference calculation section 93 Abnormality calculation unit 100 computers 100A CPU 100B memory 100C auxiliary storage 100D Operation input section 100E Display section Ap Abnormal location As Calculation result P Simulation Program T1 normal period T2 Abnormal period T2a~T2h Abnormal period T3 Study Period T4 inference period da1 input / output error da2 1st hidden layer anomaly score (1st hidden layer anomaly score) da3 1st hidden layer error da4 Second hidden layer error (intermediate layer error) db1 Second hidden layer anomaly score (second intermediate layer anomaly score) do1 First output data do2 Second output data ha 1st hidden layer vector (1st intermediate layer vector) hb Second hidden layer vector (first intermediate layer vector) x input value y output value α weight β weight

Claims

1. a model storage unit configured to store a machine learning model including an input layer, an output layer, and at least one intermediate layer disposed between the input layer and the output layer; a calculation unit configured to input the input data to the machine learning model and perform inference to calculate a first calculation result, input output data included in the output layer of the first calculation result to the machine learning model and perform inference to calculate a second calculation result, and calculate an intermediate layer error using a loss function based on a first intermediate layer anomaly degree calculated based on data included in the intermediate layer of the first calculation result and a second intermediate layer anomaly degree calculated based on data included in the intermediate layer of the second calculation result; A machine learning device comprising:

2. the first hidden layer anomaly degree represents a first normalized distance between a first hidden layer vector, which is a feature vector of the hidden layer, obtained by inputting the input data into the machine learning model and executing inference, and an average vector of the first hidden layer vector; 2. The machine learning device according to claim 1, wherein the second intermediate layer anomaly degree represents a second normalized distance between a second intermediate layer vector, which is a feature vector of the intermediate layer resulting from inputting the output data into the machine learning model, and an average vector of the second intermediate layer vectors.

3. The first hidden layer vector is [0000] year, The mean vector of the first hidden layer is [Equation 25] year, The second hidden layer vector is [Equation 26] year, The mean vector of the second hidden layer is [0000] The machine learning device according to claim 2 .

4. The first normalized distance is [0000] is the distance normalized using the covariance matrix The second normalized distance is [0000] The machine learning device according to claim 3 , wherein the distance is normalized using a covariance matrix expressed as:

5. The first intermediate layer anomaly degree is da2 ２ When this is done, da2 ２ The following formula holds true: [Equation 30] The second intermediate layer anomaly degree is db1 ２ When this is done, db1 ２ The machine learning device according to claim 4 , wherein the following formula holds: [Equation 31]

6. An electronic device comprising the machine learning device according to claim 1 .

7. A machine learning program for causing the machine learning device according to any one of claims 1 to 5 to function.

8. A simulation device that calculates the intermediate layer error using the machine learning device according to claim 1 .

9. a model storage unit configured to store a machine learning model including an input layer, an output layer, and at least one intermediate layer disposed between the input layer and the output layer; a calculation unit configured to input predetermined input data into the machine learning model, execute inference, and calculate a calculation result, and to calculate an intermediate layer error based on a plurality of the calculation results; An anomaly detection method using a machine learning device comprising: inputting first input data as the input data into the machine learning model to perform inference, and calculating a first calculation result as the calculation result; inputting output data included in the output layer of the first operation result into the machine learning model to perform inference, and calculating a second operation result as the operation result; calculating the intermediate layer error based on data included in the intermediate layer of the first calculation result and data included in the intermediate layer of the second calculation result; An anomaly detection method comprising:

Images