Abnormality degree calculation system and method

Abstract

Provided is an abnormality degree calculation system and method that more accurately calculates the abnormality degree of a device. The abnormality degree calculation system includes a concept category assignment unit (11) that assigns a prescribed concept category (C1) based on the identification number of a target device (M); a feature vector extraction unit (12) that extracts a feature vector based on sensor data (D2) from a sensor corresponding to the target device; a likelihood calculation unit (13) that calculates the likelihood of the feature vector using a machine learning model obtained from a learning database (DB2); a loss calculation unit (14) that calculates a loss using a loss function defined as a function of the likelihood; a model update unit (15) that updates the model using the loss and the model; a relearning necessity determination unit (16) that determines whether relearning is necessary based on the calculated likelihood; and an abnormality degree calculation unit (17) that calculates the abnormality degree in the case where it is determined that relearning is not necessary.

Description

Technical Field

[0001] This invention relates to an anomaly calculation system and method. Background Technology

[0002] Typically, in anomaly detection, since it is difficult to obtain sufficient anomalous data for all types of anomalies, the method of inferring normal distribution using only normal data and using it for anomaly determination is often adopted.

[0003] On the other hand, in methods that infer normal distributions using only normal data, there are also cases where normal data cannot be collected sufficiently, or where the normal distribution changes during the period of collecting a sufficient amount of normal data. Therefore, there is a need for methods that can obtain sufficient detection accuracy with a small amount of normal data.

[0004] Therefore, a technique was proposed that uses normal data of machines similar to the object of detection in addition to the normal data of the machine being detected to improve the inference accuracy of the normal distribution, so as to achieve sufficient detection accuracy even when the normal data of the object of detection is small (Patent Document 1).

[0005] Patent Document 1 describes the following: "The processing unit 11 calculates a first value, which indicates the degree to which the derived state value of the object being observed deviates from a first distribution area, where the first distribution area represents the distribution of state values ​​for each device of the same type as the object being observed. The processing unit 11 calculates a second value, which indicates the degree to which the derived state value of the object being observed deviates from a second distribution area, where the second distribution area represents the distribution of past state values ​​for each object being observed. Based on the first and second values, the processing unit 11 determines whether the object being observed is abnormal."

[0006] Existing technical documents

[0007] Patent documents

[0008] Patent Document 1: Japanese Patent Application Publication No. 2019-008354 Summary of the Invention

[0009] The problem that the invention aims to solve

[0010] The invention described in Patent Document 1 assumes that the sensor data of the device detecting the anomaly and the sensor data of devices of the same type as the device being detected follow the same normal distribution. Therefore, the technology of Patent Document 1 can only be used when the device being detected and the devices of the same type as the device being detected are the same in terms of a certain concept, such as model, and further, in terms of a concept subordinate to the above concept, such as model number; that is, only when both the model and model number can be considered the same.

[0011] Therefore, in Patent Document 1, there is no assumption that although the device to be tested and the device of the same kind as the device to be tested have the same superordinate concept, such as machine type, but different subordinate concepts, such as model, and the normal distribution of the two are different.

[0012] In cases where the normal distribution of sensor data for the target device differs from the normal distribution of sensor data for the same type of device, the technology in Patent Document 1 learns the non-normally distributed areas of the target device as normal distributions, resulting in decreased detection accuracy compared to using only the sensor data of the target device.

[0013] Therefore, the purpose of this invention is to provide an anomaly calculation system and method that can calculate the anomaly degree of a device with higher accuracy.

[0014] Methods used to solve problems

[0015] To address the aforementioned issues, an anomaly calculation system according to the present invention is provided. This system calculates the anomaly of an object device and comprises: a concept category assignment unit that assigns a predetermined concept category based on the object device's identification number; a feature vector extraction unit that extracts feature vectors based on sensor data obtained from a sensor corresponding to the object device; a likelihood calculation unit that calculates the likelihood of the feature vectors using a machine learning model obtained from a learning database; a loss calculation unit that calculates a loss using a loss function defined as a function of the likelihood calculated by the likelihood calculation unit; a model update unit that updates the machine learning model using the loss calculated by the loss calculation unit and the learned machine learning model; a relearning necessity determination unit that determines whether relearning is necessary based on the likelihood calculated by the likelihood calculation unit; and an anomaly calculation unit that calculates the anomaly if the relearning necessity determination unit determines that relearning is not necessary.

[0016] Invention Effects

[0017] According to the present invention, when an anomaly of the target device is detected, the need for relearning is determined based on the calculated likelihood. If it is determined that relearning is not needed, the anomaly degree is calculated, thus improving the accuracy of the anomaly degree calculation. Attached Figure Description

[0018] Figure 1 This is an explanatory diagram showing the overall outline of this embodiment.

[0019] Figure 2 This is an illustration of the process of extracting concept categories from input.

[0020] Figure 3 This is an illustration of the process of calculating loss using data from the training database.

[0021] Figure 4 This is an illustration of the relationship between negative log-likelihood and the number of model updates.

[0022] Figure 5 This is an illustration of the method for determining whether relearning is necessary.

[0023] Figure 6 This is a diagram showing the hardware and software configuration of the anomaly calculation device.

[0024] Figure 7 It is a block diagram of the feature vector extraction part.

[0025] Figure 8 This is a block diagram for learning.

[0026] Figure 9 This is the processing flow of the anomaly calculation system during learning.

[0027] Figure 10 This is the block composition diagram used for anomaly calculation.

[0028] Figure 11 This is the processing flow of the anomaly calculation system during anomaly calculation.

[0029] Figure 12 This is an explanatory diagram illustrating the differences between the present invention and conventional methods.

[0030] Figure 13 Regarding the second embodiment, it is a block diagram for anomaly calculation.

[0031] Figure 14 This is the processing flow of the anomaly calculation system during anomaly calculation.

[0032] Figure 15 Regarding the third embodiment, it is a block diagram during learning.

[0033] Figure 16 This is the processing flow of the anomaly calculation system during learning.

[0034] Figure 17 This is the block composition diagram used for anomaly calculation.

[0035] Figure 18 This is the processing flow of the anomaly calculation system during anomaly calculation.

[0036] Label Explanation

[0037] 1: Anomaly calculation system; 11: Concept category assignment unit; 12: Feature vector extraction unit; 13: Likelihood calculation unit; 14: Loss calculation unit; 15: Model update unit; 16: Relearning necessity determination unit; 17: Anomaly calculation unit; 21: Likelihood calculation unit for learning model; 22: Likelihood calculation unit for learned model; 31: Corresponding concept category assignment unit; M: Device. Detailed Implementation

[0038] Hereinafter, embodiments of the present invention will be described based on the accompanying drawings. In an environment where there are both target devices and other devices (corresponding concept devices, also called other devices) with the same overarching concept but different subordinate concepts, the anomaly calculation system of this embodiment improves the accuracy of anomaly detection even when the sensor data of the target device and the sensor data of the corresponding concept devices follow different normal distributions, by using the sensor data of both. Therefore, in the anomaly calculation system of this embodiment, sufficient detection accuracy can be obtained using a relatively small amount of normal data from the target device.

[0039] The anomaly calculation system of this embodiment obtains input D0 from n devices M that are the same type, such as the target device, with respect to a higher-level concept. Input D0 includes model number D1, which is an example of a lower-level concept of the type, and sensor data D2.

[0040] The anomaly calculation system of this embodiment includes a concept category assignment unit 11, which outputs a concept category D3 based on the model number D1, indicating whether each device is a device that is a detection target (target device) or a device with the same concept as the detection target device (other device).

[0041] Furthermore, the anomaly calculation system of this embodiment includes: a feature vector extraction unit 12 that outputs a feature quantity D4 based on sensor data D2; a likelihood calculation unit 13 that outputs a likelihood D5 based on the feature quantity D4 using a machine learning model; a loss calculation unit 14 that calculates a loss value D6 using a loss function defined by a function of likelihood D5, which increases the likelihood of data from the detection target device and decreases the likelihood of data from the device with the same concept; a model update unit 15 that updates the parameters of the machine learning model of the likelihood calculation unit 13 to minimize the loss value D6; a relearning necessity determination unit 16 that determines whether relearning is necessary based on the likelihood D5 and the concept category D3; and an anomaly calculation unit 17 that, if it is determined that relearning is not necessary, sets the likelihood D7 of the detection target as an anomaly.

[0042] In this embodiment, by learning by treating normal data of the target device as normal data with increased likelihood and abnormal data of devices with similar normal distributions as anomalous data with decreased likelihood, a model can be obtained that assigns a high likelihood to the data of the target device and treats it as normal. Therefore, compared to conventional techniques that learn using only the data of the target device, the normal distribution of the target device's data can be correctly inferred, and sufficient detection accuracy can be obtained even when the amount of data from the target device is small.

[0043] [Example 1]

[0044] use Figures 1-12 The first embodiment is described below. Figure 1 This is an explanatory diagram showing the overall outline of this embodiment.

[0045] Anomaly Calculation System 1 (Reference) Figure 6 The calculation represents the degree of anomaly of multiple devices M, which are physical changes measurable according to their operation. These physical changes, measurable according to their operation, include, for example, the sound, vibration (or the sound and vibration combined, referred to as the vibration-accompanied signal), temperature, and color associated with the operation of the device, such as electric motors, hydraulic motors, pneumatic cylinders, hydraulic cylinders, solenoids, and linear actuators. Here, sound-based anomaly detection is illustrated as an example, but anomalies can also be detected based on physical changes other than sound.

[0046] Figure 1 The anomaly calculation system shown calculates the anomaly degree of a selected device among multiple devices M1 to Mn belonging to a common superordinate concept C2. Each device M1 to Mn has a different subordinate concept C1 under its superordinate concept C2. For example, the superordinate concept C2 is the machine type, and the subordinate concept C1 is different models belonging to the same machine type. To indicate that the subordinate concept C1 differs for each device, in... Figure 1 The subcategories are represented as C11 to C13. As long as the model number is common, the difference in serial number is not a problem.

[0047] The superordinate concept C2 and the subordinate concept C1 can also be represented as follows. For example, the smallest unit for classifying device M is the subordinate concept C1, and the concept immediately above the subordinate concept C1 is the superordinate concept C2. Here, the device that is being detected is designated as device M1, and the other devices M2 to Mn are called corresponding concept devices. Devices M1 to Mn are referred to as device M or device unless otherwise distinguished.

[0048] Input D0 includes sensor data D2 obtained from n devices M of the same model as the example of the higher-level concept, and model number D1 as the example of the lower-level concept. Hereinafter, model number D1 will be simply referred to as model D1.

[0049] The concept category assignment unit 11 transforms the model number D1 into the concept category D3. The feature vector extraction unit 12 outputs the feature vector D4 from the sensor data D2. Hereinafter, the feature vector D4 may be simply referred to as feature quantity D4. The training database DB1 stores the sensor data D2 and the concept category D3. The training database DB1 may also be called the training digital input signal database DB1.

[0050] The likelihood calculation unit 13 uses the machine learning model given from the learning database DB2 to output the likelihood D5 of the sensor data of each device M based on the feature D4.

[0051] The loss calculation unit 14 uses concept category D3 and likelihood D5 to calculate a loss function such that the likelihood of device M1, which is the detection object, is higher, and the likelihood of devices M2 to Mn, which are co-concepts, is lower, and outputs the loss D6.

[0052] The model update unit 15 uses loss D6 to update the parameters of the machine learning model stored in the learning database DB2. Hereinafter, the machine learning model will be referred to simply as the model.

[0053] During anomaly detection, the relearning requirement determination unit 16 uses likelihood D5 and concept category D3 to determine whether relearning is necessary.

[0054] If the anomaly calculation unit 17 determines that no further learning is needed, it uses the likelihood D7 of the detected object in the likelihood D5 to calculate the anomaly of the detected object.

[0055] use Figure 2 This explains the relationship between input D0, model number D1, sensor data D2, and concept category D3. Input D0 obtained from each device M includes model number D1 and sensor data D2. Model number D1 is transformed into concept category D3 by concept category assignment unit 11. Concept category D3 is set to "1" for the detection target device M1, and set to "0" for other devices M2 to Mn with the same concept as the detection target.

[0056] exist Figure 3 The diagram illustrates a method for calculating likelihood and loss using input from the training database DB1. First, a batch of N sensor data points D2 and concept categories D3 are obtained from the training database DB1.

[0057] Here, we assume that we obtain, for example, N / 2 combinations of sensor data D2 from the target device and concept category D3, and, for example, N / 2 combinations from devices with the same concept as the target device. This is because, in the case of multiple other devices with the same concept as the target device, if we randomly obtain a batch of N data points from the set of data from the target device and data from other devices with the same concept, the number of data points from the target device will be less than the number of data points from other devices with the same concept, thus slowing down the learning process of the machine learning model.

[0058] Next, the feature vector extraction unit extracts feature vectors from the sensor data D2 to obtain N feature quantities D4. The likelihood calculation unit 13 calculates the likelihood using a machine learning model and obtains N likelihoods D5 based on the feature quantities D4.

[0059] Any machine learning model can update its parameters using a loss defined by a function of likelihood. For example, a Normalizing Flow (NF), a Gaussian Mixture Model (GMM), or Mahalanobis distance can be used.

[0060] For the loss calculation unit 14, N likelihoods D5 and the concept category D3 corresponding to each likelihood D5 are input. The loss calculation unit 14 calculates the loss and outputs it using a loss function. The loss function is given, for example, by the following equation 1.

[0061] [Formula 1]

[0062]

[0063] N target N / 2 is the number of combinations of likelihood D5 and concept category D3 from the detection object device included in a batch. outlier NLL(x) is the number of samples in the batch that satisfy the conditions within the indicator function I, from the number of combinations of likelihood D5 and concept category D3 from other devices with the same concept as the detection target. NLL(x) is the negative log-likelihood of feature D4. K is a hyperparameter of 0 ≤ k ≤ 1. By adjusting k, the ratio of increasing the likelihood of the detection target device data in the first term of Equation 1 to decreasing the likelihood of the detection target device data in the second term can be adjusted. Decreasing k increases the likelihood of the detection target device data in the first term.

[0064] In the case of using, for example, a GMM as a machine learning model, it is necessary to make k smaller than 1 (k < 1). The indicator function I[NLL(x) < c] takes 1 when NLL(x) < c and 0 when NLL(x) ≥ c. c is a threshold value, which is determined by the method shown below. The feature quantity from the detection target device and the feature quantity from the device of the same concept as the detection target are given to the likelihood calculation unit, and a loss function is given as shown in the following mathematical formula 2.

[0065] [Mathematical formula 2]

[0066]

[0067] The value of NLL(x) when the loss function is minimized by performing model updates multiple times is used as the threshold value c.

[0068] Figure 4 is a graph showing the transition of the negative log-likelihood of each data in the case of updating the model using Mathematical formula 1. Figure 4 The vertical axis of represents NLL(x), Figure 4 The horizontal axis of represents the number of model updates. Above the graph of, the change in the negative log-likelihood of the data of other devices of the same concept as the detection target is shown. Below the graph of, the change in the negative log-likelihood of the data of the detection target device is shown. Figure 4 Figure 4

[0069] By using Mathematical formula 1, a constraint can be imposed such that NLL(x) of the data of the detection target device is smaller than NLL(x) of the data of other devices of the same concept as the detection target. In the abnormality calculation unit 17, since the abnormality is calculated with the negative log-likelihood as the abnormality, this constraint means that the model learns in a way that reduces the abnormality of the data of the detection target device and increases the abnormality of the data of other devices of the same concept as the detection target.

[0070] Furthermore, by using the indicator function I, divergence of NLL(x) of the data of other devices of the same concept as the detection target device can be prevented.

[0071] Using Figure 5 , the method for the relearning necessity determination unit 16 to determine the necessity of relearning is described. Figure 5 The vertical axis of represents NLL(x), Figure 5 The horizontal axis of represents time. Above the graph of, the negative log-likelihood of the data of other devices of the same concept as the detection target device is shown. Below the graph of, the negative log-likelihood of the data of the detection target device is shown. Figure 5 Figure 5

[0072] If an anomaly detection system is applied using a model made during model learning, the distribution of data from the target device and data from other devices with similar concepts changes over time from the time of learning.

[0073] Therefore, as Figure 5 As shown on the left side of the "relearning execution" time point in the curve graph, it can be inferred that the negative log-likelihood of the data of the detection target device increases, while the negative log-likelihood of the data of other devices with the same concept as the detection target decreases.

[0074] If the negative log-likelihood of the data from the target device increases, an anomaly may have occurred in the target device. Therefore, relearning is only performed when the negative log-likelihood of the data from other devices with the same concept as the target device decreases. For example, the user sets a relearning decision threshold Thrc, and relearning is performed when the negative log-likelihood of the data from other devices with the same concept as the target device is lower than the relearning decision threshold C.

[0075] exist Figure 6 The diagram illustrates the hardware and software configuration of the anomaly calculation system. By executing a prescribed computer program, the computer can be used as an anomaly calculation device 100. Figure 6 The example shown is an anomaly calculation device 100 consisting of a single computer, but it is also possible for multiple computers to form one or more anomaly calculation devices 100. The computer can also be a virtual computer.

[0076] The anomaly calculation device 100 is connected to n sensor terminals T, which are installed on n devices respectively, via a communication network CN.

[0077] The sensor terminal T is configured, for example, as a portable recording terminal.

[0078] The anomaly calculation device 100 is, for example, a computer that includes an arithmetic unit 1001, a main storage device 1002, an auxiliary storage device 1003, an input unit 1004, an output unit 1005, and a communication unit 1006.

[0079] The arithmetic unit 1001 includes one or more microprocessors, which read a predetermined computer program stored in the auxiliary storage device 1003 into the main storage device 1002 and execute it to achieve... Figure 1 The functions described in the text include the concept classification assignment unit 11, feature vector extraction unit 12, likelihood calculation unit 13, loss calculation unit 14, model update unit 15, relearning need determination unit 16, and anomaly calculation unit 17.

[0080] The input unit 1004 may include, for example, a keyboard, a touch panel, a pointer device, etc., to accept input from a user using the anomaly calculation device 100. The output unit 1005 may include, for example, a monitor, a speaker, a printer, etc., to provide information to the user.

[0081] The communication unit 1006 communicates with the sensor terminal T via the communication network CN. The communication unit 1006 can also communicate with other computers not shown.

[0082] The storage medium MM is, for example, a flash memory or a hard disk, which transfers computer programs or data to and stores them in the anomaly calculation device 100, or reads computer programs or data from and stores them in the anomaly calculation device 100. The storage medium MM can be directly connected to the anomaly calculation device 100, or it can be connected to the anomaly calculation device 100 via a communication network CN.

[0083] The structure of the sensor terminal T is described below. The sensor terminal T includes, for example, a sensor unit 210, a control unit 220, a storage unit 230, and a sensor terminal communication unit 240.

[0084] The sensor unit 210 is a device that acquires sensor data from the equipment, such as a microphone, accelerometer, magnetometer, or camera. The data acquired by the sensor unit 210 is stored in the storage unit 230. The control unit 220, which controls the sensor terminal T, sends the sound data stored in the storage unit 230 to the anomaly calculation device 100. Furthermore, the anomaly calculation device 100 can calculate the anomaly level based on the received sensor data.

[0085] Figure 7 This is a block diagram of the feature vector extraction unit 12. Figure 8 This is a block diagram of the anomaly calculation device 100 during learning. Figure 9 This is the processing flow within the anomaly calculation device 100 during learning.

[0086] The configuration of the feature vector extraction unit 12 will be explained. The case where the sensor unit 210 of the sensor terminal T is a microphone will be described, but it is not limited to a microphone; any device capable of acquiring sensor data from the device M, such as an accelerometer, magnetic sensor, or camera, can be used. When using a sensor other than a microphone, a well-known feature vector extraction method will be used. When the sensor unit 210 is another type of sensor, a well-known feature vector extraction method corresponding to the type of sensor will be used.

[0087] The input sound acquisition unit 101 converts the analog input signal from the microphone into a digital input signal using an A / D (analog-to-digital) converter (S101) and saves it to the training digital input signal database DB1 (S102).

[0088] The frame segmentation unit 102 divides the digital input signal retrieved from the training digital input signal database DB1 into segments according to a predetermined number of time points (hereinafter referred to as frame size) and outputs the frame signal (S104). Frames may also overlap.

[0089] The window function multiplication unit 103 multiplies the input frame signal by a window function and outputs the window function multiplication signal (S105). For example, a Hanning window is used for the window function.

[0090] The frequency domain signal calculation unit 104 performs a short-time Fourier transform on the input signal after multiplication by the window function, and outputs a frequency domain signal (S106). If the frame size is N, the frequency domain signal is a group of (N / 2+1) = M frequency intervals, each containing one complex number and its corresponding M complex numbers. The frequency domain signal calculation unit 104 can also use frequency transformation methods such as constant Q transform (CQT) instead of short-time Fourier transform.

[0091] The power spectrum calculation unit 105 outputs the power spectrum based on the input frequency domain signal (S107). The filter bank multiplication unit 106 outputs the Mel power spectrum by multiplying the input power spectrum by a Mel filter bank (S108). The filter bank multiplication unit 106 can also use a filter bank such as a 1 / 3 octave filter instead of a Mel filter bank.

[0092] The instantaneous characteristic calculation unit 107 outputs a logarithmic Mel power spectrum by applying a logarithm to the input Mel power spectrum (S109). Alternatively, the Mel frequency cepstral coefficients (MFCCs) can be calculated instead of the logarithmic Mel power spectrum. In this case, the logarithmic value of the power spectrum is calculated instead of the filter bank multiplication unit 106 and the logarithm calculation unit 107, multiplied by the filter bank, subjected to a discrete cosine transform, and the MFCCs are output.

[0093] The feature quantity time series calculation unit 108 outputs feature quantity D4 by linking adjacent L frames for the input log-Mel power spectrum or for MFCC (S110). Alternatively, instead of the log-Mel power spectrum or MFCC, the time series of their time difference or time derivative can be input, and the feature quantity D4 can be output by linking adjacent L frames.

[0094] Alternatively, by inputting the time difference or time derivative of the time series, or the time series of the time derivative (δ, δ), adjacent L frames can be linked to output the feature quantity D4. Furthermore, by selecting a combination of these and linking them along the feature axis, adjacent L frames can be linked to output the feature quantity D4. The feature quantity D4 generated by the feature vector extraction unit 12 is input to the likelihood calculation unit 13.

[0095] The likelihood calculation unit 13 calculates the likelihood D5 based on the feature D4. The calculated likelihood D5 is input to the loss calculation unit 14 (S111).

[0096] The loss calculation unit 14 calculates the value of a predetermined loss function D6, which is defined based on the likelihood of the data of the detection target device and the likelihood of the data of other devices with the same concept as the detection target device, according to the likelihood D5 and the concept category D3. The loss D6 calculated by the loss calculation unit 14 is input to the model update unit 15 (S112).

[0097] The model update unit 15 repeatedly learns the parameters of the machine learning model to minimize the value of the loss D6 (S113-S116). These machine learning model parameters are stored in the learning database DB2 (S117).

[0098] That is, the model update unit 15 determines whether the convergence condition is met, or whether the number of repetitions C1 of this process exceeds the upper limit ThC (S113).

[0099] If the convergence condition is not met or the number of iterations C1 is below the upper limit ThC, the model update unit 15 updates the parameters of the machine learning model (S114), calculates the convergence condition (S115), increments the number of iterations C1 by 1, and returns to step S112.

[0100] The model update unit 15 saves the parameters of the machine learning model to the learning database DB2 (S117).

[0101] Figure 10 This is a block diagram of the anomaly calculation device 100 used to calculate anomalies. Figure 11 This describes the processing flow within the anomaly calculation device 100 when calculating anomalies. (Refer to...) Figure 10 and Figure 11 Please provide an explanation.

[0102] Anomaly calculation device 100 replaces loss calculation unit 14 and model update unit 15, and has relearning need determination unit 16 and anomaly calculation unit 17.

[0103] When calculating the anomaly of feature quantity D4 extracted from sensor data D2, likelihood calculation unit 13 reads parameters from learning database DB2 (S201).

[0104] Steps S101 to S110 have already been described, so repeated explanations are omitted. In these steps S101 to S110, feature quantity D4 is generated regarding the sensor data D2 obtained from the sensor terminal T, and input into the machine learning model of the likelihood calculation unit 13. Repeated explanations are also omitted in the following description. In addition, in S102, the input sound is also saved to the training database DB1 during anomaly calculation, but this is to save training data for relearning.

[0105] The likelihood calculation unit 13 calculates the likelihood D5 based on the feature D4, and inputs the calculated likelihood D5 into the relearning need determination unit 16 (S111).

[0106] The relearning need determination unit 16 uses the likelihood D5 and concept category D3 to perform relearning if the negative log likelihood of other devices with the same concept as the detection target device is lower than a certain relearning need determination threshold Thrc. If it is higher, the likelihood D7 of the detection target device's data is input to the anomaly calculation unit 17 (S202).

[0107] Anomaly calculation unit 17 calculates the anomaly of sensor data of the detection target device based on the likelihood D7 of the data of the detection target device (S203).

[0108] Here, as a machine learning model, the configuration is disclosed, for example, when using a normalizing flow (NF).

[0109] The likelihood calculation unit 13 is composed of a multi-layer neural network. The input to the likelihood calculation unit 13 consists of nodes equal to the dimension of the feature vector, each node receiving elements of the feature vector. The input layer is connected to a predetermined transformation function, a multi-layer neural network with the same number of nodes in both the input and output layers, and a non-linear activation function (e.g., the ReLU function). The final layer of the multi-layer neural network has the same number of nodes as the input layer. The predetermined transformation function is set such that the transformation from the input layer to the output of the multi-layer neural network is a reversible transformation. The likelihood calculation unit 13 assumes that the distribution function of the output of the final layer of the multi-layer neural network follows a known distribution (e.g., a normal distribution), and calculates the likelihood based on the output of the final layer of the multi-layer neural network. Furthermore, using a well-known probability variable transformation formula, the likelihood D4 of the input feature vector is calculated.

[0110] The loss calculation unit 14 uses a loss function defined by the likelihood D5, such as Equation 1, to calculate the loss D6. At this time, the loss function is set, for example as Equation 1, to increase the likelihood D5 of the data of the target device and decrease the likelihood D5 of the data of other devices that are in the same category as the target device.

[0111] The model update unit 15 updates the parameters of the machine learning model to minimize the loss D6 calculated by the loss calculation unit 14. Minimization can be performed, for example, by well-known optimal algorithms such as SGD, Momentum SGD, AdaGrad, RMSprop, AdaDelta, Adam, etc.

[0112] According to this embodiment, learning can be performed by increasing the likelihood of normal data from the target device and treating it as normal data, and decreasing the likelihood of normal data from other devices that are in the same category as the target device and treating them as abnormal data.

[0113] Figure 12 This describes the effect of this embodiment. Figure 12 The upper side indicates the case where the machine learning model learns using only normal data from the device being tested. Figure 12 The area below indicates a situation where the machine learning model learns to treat normal data from devices with the same conceptual relationship as the detection target device as anomalous data. The shaded area with a fine grid represents the region where the machine learning model in this embodiment treats the data as normal.

[0114] like Figure 12 As shown in (1), when learning is performed using only the normal data of the detection target device, since the normal data of the detection target device and the normal data of other devices with the same concept as the detection target device have similar normal distributions, the learned model will also regard a portion of the normal data of other devices with the same concept as the detection target device as normal data, giving a high likelihood.

[0115] However, in this embodiment, the learned model treats normal data from other devices with the same conceptual relationship as the target device as anomalous data, giving a lower likelihood. Therefore, according to this embodiment, as Figure 12 As shown in (2), it can reduce the area that the machine learning model considers to be normal and more accurately infer the normal distribution of the data of the detection target device.

[0116] Therefore, according to this embodiment, by using normal data from other devices that are in the same category as the target device, the normal distribution of data of the target device can be inferred more accurately, and sufficient detection accuracy can be obtained even when the data of the target device is small.

[0117] [Example 2]

[0118] use Figure 13 , Figure 14 The second embodiment will be described. In the following embodiments, including this embodiment, the differences from the first embodiment will be the main focus.

[0119] Figure 13 It is a block configuration diagram in the case of performing sequential learning processing. Figure 14 It is the processing flow of sequential learning processing.

[0120] In this embodiment, first, the machine learning model is learned through batch learning that is the same as the learning method described in the first embodiment. Figure 7 , Figure 8 The difference between this embodiment and the first embodiment is that the anomaly degree is calculated using the learned model, and at the same time, the learning model is updated using the data obtained at runtime, and the anomaly degree is also calculated according to the updated model.

[0121] The input D0 from the sensor terminal T is divided into a model number D1 and sensor data D2. The model number D1 is transformed into a concept category D3 by the concept category assigning unit 11, and the sensor data D2 is transformed into a feature quantity D4 by the feature vector extraction unit 12. The sensor data D2 and the concept category D3 are stored in the data storage database DB3 for determining the cause of an anomaly when an anomaly occurs.

[0122] In this embodiment, there are two likelihood calculation units. One is the learning model likelihood calculation unit 21 that calculates the likelihood of the feature quantity D4 using the model obtained from the sequentially updated learning database. The other is the learned model likelihood calculation unit 22 that calculates the likelihood of the feature quantity D4 using the model obtained from the learned database DB4 that stores the learned model.

[0123] The loss calculation unit 14 calculates the loss value according to the likelihood calculated by the learning model likelihood calculation unit 21 using a specified loss function that is defined as a function of likelihood. The specified loss function is a loss function that gives a higher likelihood to the data of the detection target device and a lower likelihood to the data of other devices of the same concept as the detection target device.

[0124] In addition, in the sequential learning processing used in this embodiment, since the model is sequentially updated using the sequentially input sensor data, unlike in batch learning, batch learning cannot be used. Therefore, among the n sensor data input from n devices via the sensor terminal T, the data of one detection target device and the data of one other device of the same concept as the detection target device are extracted, and the loss is calculated based on their likelihoods. For example, when using Equation 1, Ntarget is 1, when NLL(x) < c, Noutlier is 1, and when NLL(x) ≧ c, Equation 1 becomes only the first term.

[0125] The model update unit 15 updates the parameters based on the loss value and saves the obtained parameters to the learning database DB2. In minimizing the loss function, well-known optimal algorithms for iterative parameter updates can be used, such as Online Gradient Descent (OGD).

[0126] The anomaly calculation unit 17 calculates the anomaly degree based on the likelihoods calculated by the learning model likelihood calculation unit 21 and the learned model likelihood calculation unit 22, respectively. When using only the likelihood from the gradually updated model, it is difficult to handle anomalies such as aging that occur over a long period of time. Therefore, the likelihood from the learned model at the start of operation is added to address these anomalies. As a method for calculating the anomaly degree using both the likelihood obtained from the learned model and the likelihood obtained from the gradually updated model, a function such as Equation 3 can be used, for example.

[0127] [Formula 3]

[0128]

[0129] A is the final anomaly score. A1 is the anomaly score calculated using the model obtained from the training database. A2 is the anomaly score calculated using the model obtained from the trained database. α is a parameter set at runtime.

[0130] This embodiment, thus configured, achieves the same effect as the first embodiment. Furthermore, in this embodiment, during anomaly calculation, the model is gradually learned while anomaly calculation is being performed, thus enabling the learning of a model that is more robust to changes in the normal distribution of each device.

[0131] In this embodiment, the learning database DB2 at the start of the operation is used as the learned database DB4. At the same time, the likelihood given by the model at the start of the operation is calculated and reflected in the final anomaly score, so that the detection of anomalies such as aging due to the long passage of time can also be addressed.

[0132] Furthermore, this embodiment can also accommodate changes in the input data of the same-position concept, such as when other devices with newly added same-position concepts are added.

[0133] [Example 3]

[0134] use Figures 15-18 The third embodiment is described below. In the first and second embodiments, all devices with the same upper-level concept (e.g., machine type) as the device being tested, but with a lower-level concept (e.g., model number) that is different from the device being tested, are treated as other devices with the same upper-level concept as the device being tested.

[0135] On the other hand, among other devices with the same concept, there are also multiple types of subordinate concepts, such as model number, and there are cases where the similarity to the data of the target device varies depending on the type. For example, consider devices that are the same type as the target device but different models, but with the same or different driving methods. In this embodiment, the aim is to further improve detection accuracy by utilizing such types of subordinate concepts.

[0136] use Figure 15 and Figure 16 The learning process in this embodiment is explained. The model number D1 in the input D0 from the sensor terminal T is transformed into a concept category D3 by the concept category assignment unit 11. Sensor data D2 is transformed into a feature quantity D4 by the feature vector extraction unit 12. Here, the concept category assignment unit 11 differs from that in the first and second embodiments, outputting both the concept category D3 and the model number D1.

[0137] Sensor data D2 is transformed into feature quantity D4 by feature vector extraction unit 12. When the set of other devices with the same concept as the detection target device includes multiple models, the corresponding concept classification unit 31 investigates the relationship between each model of the other devices with the same concept and the model of the detection target device, and assigns a weight w according to each model of the other devices with the same concept. This weight w can, for example, be set based on the commonality of elements related to the source of sound emitted from the detection target device, i.e., the sound source. For example, the sensor data differs depending on whether the device's drive method is an electric motor or a hydraulic motor. Therefore, in this embodiment, for models of other devices with the same drive type as the detection target device, a higher value is given as indicating a higher similarity to the detection target device, and for models of other devices with a different drive type than the detection target device, a lower value is given as indicating a lower similarity to the detection target device.

[0138] In the training database DB1, in addition to sensor data D2 and concept category D3, the weights w corresponding to each model of other devices with the same concept are also stored.

[0139] The likelihood calculation unit 13 uses the model obtained from the learning database DB2 and calculates the likelihood D5 based on the feature quantity D4 calculated by the feature vector extraction unit 12.

[0140] The loss calculation unit 14 uses a loss function defined as likelihood, which gives a higher likelihood for data from the target device and a lower likelihood for data from other devices in the same category as the target device. Based on the likelihood calculated by the likelihood calculation unit 13, the loss D6 is calculated. For example, the loss function used here can be Equation 4, which is a modification of Equation 1.

[0141] [Formula 4]

[0142]

[0143] Here, wi is the weight of the i-th model set in the corresponding concept category assignment section 31.

[0144] The model update unit 15 updates the parameters of the machine learning model to minimize the loss D6 calculated by the loss calculation unit 14. Minimization can be achieved, for example, by well-known optimal algorithms such as SGD, Momentum SGD, AdaGrad, RMSprop, AdaDelta, Adam, etc.

[0145] use Figure 17 and Figure 18 The anomaly calculation process in this embodiment is explained. For the model number D1 in the input D0 from the sensor terminal T, the concept category assignment unit 11 assigns it the concept category D3. Sensor data D2 is transformed into feature quantity D4 by the feature vector extraction unit 12.

[0146] In the case where the set of other devices with the same concept as the test object includes multiple models, the corresponding concept category assignment section 31 investigates the relationship between each model of the other devices with the same concept and the model of the test object, and assigns a weight w according to each model of the other devices with the same concept.

[0147] In the training database DB1, in order to perform relearning, in addition to sensor data D2 and concept category D3, the weights w corresponding to each model of other devices with the same concept are also stored.

[0148] The likelihood calculation unit 13 uses the model obtained from the learning database DB2 to calculate the likelihood D5 based on the feature quantity D4 calculated by the feature vector extraction unit 12.

[0149] The relearning need determination unit 16 uses likelihood D5 and concept category D3 to perform model relearning if the negative log likelihood of other devices with the same concept as the detection target device is lower than a certain relearning need determination threshold Thc. If the negative log likelihood is higher than the threshold Thc, the likelihood D5 of the detection target device data is input to the anomaly calculation unit 17 (S202).

[0150] Anomaly calculation unit 17 calculates the anomaly of sensor data of the detection target device based on the likelihood D5 of the data of the detection target device (S203).

[0151] This embodiment, thus configured, achieves the same effect as the first embodiment. Furthermore, in this embodiment, a model information database DB5 that maintains model information and a corresponding concept category assignment unit 31 that assigns weights based on the models of other devices with the same corresponding concept are added. Therefore, according to this embodiment, the impact of data from other devices with each corresponding concept on the value of loss D6 can be adjusted according to the model of the other devices with each corresponding concept. Thus, in this embodiment, for example, by assigning a larger weight to devices with the same driving method as the target device and a smaller weight to devices with different driving methods, the model can learn to reduce the likelihood of data more similar to the target device.

[0152] Furthermore, the present invention is not limited to the embodiments described above, but includes various modifications. For example, the embodiments described above have been explained in detail to facilitate understanding of the present invention, and are not limited to necessarily possessing all the described configurations. Moreover, a portion of the configuration of one embodiment can be replaced with the configuration of another embodiment; furthermore, the configuration of another embodiment can be added to the configuration of one embodiment. Furthermore, other configurations can be added, deleted, or replaced for portions of the configurations of each embodiment.

[0153] The constituent elements of this invention can be arbitrarily selected or omitted, and inventions with the selected or omitted components are also included in this invention. Furthermore, the components described in the claims can be combined in ways other than those explicitly stated in the claims.

Claims

1. An anomaly calculation system for calculating the anomaly degree of an object device, characterized in that, have: The concept category assignment section assigns a prescribed concept category based on the device's identification number. Other devices that are in the same concept as the aforementioned object device are assigned a different concept category than the aforementioned object device. The aforementioned other devices in the same concept refer to devices that have the same higher concept as the aforementioned object device but a different lower concept. The feature vector extraction unit extracts feature vectors based on sensor data obtained from the sensor corresponding to the aforementioned device. The likelihood calculation unit uses a machine learning model obtained from the learning database to calculate the likelihood of the above feature vectors. The loss calculation unit calculates the loss using a loss function, which is defined as a function of the likelihood calculated by the likelihood calculation unit. The loss function is set to increase the likelihood of sensor data from the target device and decrease the likelihood of sensor data from the other devices in the same category. The model update unit updates the machine learning model using the loss calculated by the loss calculation unit and the learned machine learning model in a way that minimizes the loss calculated by the loss calculation unit. The relearning need determination unit determines whether the machine learning model needs to be relearned based on the likelihood of the sensor data of the other devices of the above-mentioned co-existing concept calculated by the likelihood calculation unit. If the negative log-likelihood of the sensor data of the other devices of the above-mentioned co-existing concept is lower than the relearning determination threshold, it is determined that relearning is needed. as well as The anomaly calculation unit calculates the anomaly degree when the relearning necessity determination unit determines that relearning is not required.

2. The anomaly calculation system as described in claim 1, characterized in that, The aforementioned object equipment and other equipment with the same concept share the commonality of machine type as the smallest unit for classifying equipment, but the models included in the aforementioned machine type are different.

3. The anomaly calculation system as described in claim 1, characterized in that, The identification number mentioned above is information assigned to a specific model.

4. The anomaly calculation system as described in claim 1, characterized in that, Set the above concept category to "1" for the above object device and "0" for the above other devices.

5. The anomaly calculation system as described in claim 1, characterized in that, Multiple devices belonging to the same type but different models are pre-grouped, and these multiple devices demonstrate the ability to measure physical changes based on the work. Select one of the aforementioned target devices from the grouped devices; Among the grouped devices, at least one device other than the target device is selected as the other device.

6. The anomaly calculation system as described in claim 1, characterized in that, It also has: The model likelihood calculation unit calculates the likelihood of the model being learned, which is used for model learning not only during learning but also during runtime; and The learned model likelihood calculation unit calculates the likelihood of the learned model.

7. The anomaly calculation system as described in claim 6, characterized in that, The aforementioned model likelihood calculation unit calculates the likelihood by reading the model from the learning database, which is stored in the model and updated gradually during runtime. The learned model likelihood calculation unit mentioned above calculates the likelihood by reading the learned model from the learned database of the model saved at the start of the run.

8. The anomaly calculation system as described in claim 1, characterized in that, It also has a corresponding concept category assignment unit that assigns weights to the aforementioned other devices based on the relationship between the subordinate concepts of the aforementioned object device and the subordinate concepts of each of the aforementioned other devices.

9. The anomaly calculation system as described in claim 8, characterized in that, The aforementioned equivalent concept category assignment unit assigns weights to each of the aforementioned other devices, and by using the aforementioned weights within the loss function, adjusts the impact of the likelihood values ​​of each of the aforementioned other devices on the loss value.

10. The anomaly calculation system as described in claim 1, characterized in that, The aforementioned equipment is a device that produces sound or vibration based on its operation.

11. A method for calculating anomaly degree, characterized in that, In the anomaly calculation system where a computer calculates the anomaly degree of an object device, the following steps are performed by the aforementioned computer: Based on the device's identification number, a specified conceptual category is assigned. Other devices that are of the same conceptual level as the aforementioned target device are assigned a different conceptual category than the aforementioned target device. The aforementioned other devices of the same conceptual level refer to devices that have the same higher-level concept as the target device but a different lower-level concept. Based on sensor data obtained from the sensors corresponding to the aforementioned devices, feature vectors are extracted; The likelihood of the above feature vectors is calculated using a machine learning model obtained from the learning database. The loss is calculated using a loss function defined as a function of the likelihood calculated above, wherein the loss function is set to increase the likelihood of sensor data from the aforementioned object device and decrease the likelihood of sensor data from the aforementioned other devices of the aforementioned co-existing concept. Using the calculated loss and the learned machine learning model, the machine learning model is updated in a way that minimizes the calculated loss. When an anomaly is detected in the aforementioned target device, the likelihood of sensor data from other devices of the aforementioned co-existing concept is calculated to determine whether the aforementioned machine learning model needs to be relearned. If the negative log-likelihood of sensor data from the aforementioned other devices of the aforementioned co-existing concept is lower than the relearning determination threshold, it is determined that relearning is required. If it is determined that no further learning is needed, calculate the anomaly degree.

Images