Anomaly Severity Calculation System and Method

Abstract

To more accurately calculate the degree of abnormality of a device.SOLUTION: An abnormality degree calculation system includes: a concept type providing part 11 that provides a predetermined concept type C1 based on an identification number of a target device M; a feature amount vector extraction part 12 that extracts a feature amount vector based on sensor data D2 of a sensor for the target device; a likelihood calculation part 13 that calculates the likelihood of the feature amount vector by using a machine learning model obtained from a learning database DB2; a loss calculation part 14 that calculates loss by using a loss function defined as a function of the likelihood; a model update part 15 that updates the model by using the loss and the model; a relearning necessity determination part 16 that determines the necessity of relearning from the calculated likelihood; and an abnormality degree calculation part 17 that calculates the degree of abnormality upon determination that relearning is unnecessary.SELECTED DRAWING: Figure 1

Description

Technical Field

[0006] , , , , , , ,

[0001] The present invention relates to an abnormality degree calculation system and method.

Background Art

[0002] Generally, in abnormality detection, since it is difficult to obtain a sufficient amount of abnormality data for all types of abnormalities, a method is often adopted in which a normal distribution is estimated using only normal data and abnormality determination is performed using the estimated normal distribution.

[0003] On the other hand, even in a method of estimating a normal distribution using only normal data, there are cases where normal data cannot be sufficiently collected or where the normal distribution changes while a sufficient amount of normal data is being collected. Therefore, there is a need for a method that can obtain sufficient detection accuracy with a small amount of normal data.

[0004] Therefore, a technique has been proposed (Patent Document 1) that improves the estimation accuracy of the normal distribution by using normal data of a machine similar to the detection target in addition to the normal data of the detection target machine, and achieves sufficient detection accuracy even when the normal data of the detection target is small.

[0005] <00​​​​​​​​​​​​​​​[Overview of the Initiative] [Problems that the invention aims to solve]

[0007] The invention described in Patent Document 1 assumes that the sensor data of the device to be detected for abnormalities and the sensor data of a device of the same type as the device to be detected follow the same normal distribution. Therefore, the technology of Patent Document 1 can only be used when the device to be detected and a device of the same type as the device to be detected are identical in a certain concept, for example, with respect to the model, and further identical in a lower-level concept, for example, the type, that is, when both the model and type can be considered to be the same.

[0008] Therefore, Patent Document 1 does not consider the case where the device to be detected and a device of the same type as the device to be detected are identical in terms of higher-level concepts, such as model, but differ in terms of lower-level concepts, such as type, and their normal distributions are different.

[0009] If the normal distribution of sensor data for the device being detected differs from the normal distribution of sensor data for similar devices, the technology described in Patent Document 1 will learn the region that is not part of the normal distribution for the device being detected as the normal distribution. As a result, the detection accuracy will decrease compared to using only the sensor data of the device being detected.

[0010] Therefore, the purpose of this disclosure is to provide an abnormality calculation system and method that can calculate the abnormality level of equipment with higher accuracy. [Means for solving the problem]

[0011] To solve the above problems, an anomaly calculation system according to one aspect of the present invention is an anomaly calculation system for calculating the anomaly of a target device, comprising: a concept type assignment unit that assigns a predetermined concept type based on the identification number of the target device; a feature vector extraction unit that extracts a feature vector based on sensor data obtained from a sensor corresponding to the target device; a likelihood calculation unit that calculates the likelihood of the feature vector using a machine learning model obtained from a training database; a loss calculation unit that calculates the 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 trained machine learning model; a retraining necessity determination unit that determines whether retraining is necessary from the likelihood calculated by the likelihood calculation unit; and an anomaly calculation unit that calculates the anomaly when the retraining necessity determination unit determines that retraining is unnecessary. [Effects of the Invention]

[0012] According to the present invention, when an abnormality is detected in the target device, the necessity of relearning is determined from the calculated likelihood, and if it is determined that relearning is unnecessary, the abnormality score is calculated, thereby improving the accuracy of calculating the abnormality score. [Brief explanation of the drawing]

[0013] [Figure 1] An explanatory diagram showing the overall overview of this embodiment. [Figure 2] A diagram illustrating the process of extracting concept types from input. [Figure 3] A diagram illustrating the process of calculating loss using data from a training database. [Figure 4] An explanatory diagram showing the relationship between the negative log-likelihood and the number of model updates. [Figure 5] A diagram illustrating the method for determining whether relearning is necessary. [Figure 6] Hardware and software configuration diagram of the abnormality severity calculation device. [Figure 7] Block diagram of the feature vector extraction unit. [Figure 8] Block diagram for learning. [Figure 9]Processing flow of the abnormality degree calculation system during learning. [Figure 10] Block diagram during abnormality degree calculation. [Figure 11] Processing flow of the abnormality degree calculation system during abnormality degree calculation. [Figure 12] Explanatory diagram showing the difference between the present invention and the conventional method. [Figure 13] Regarding the second embodiment, block diagram during abnormality degree calculation. [Figure 14] Processing flow of the abnormality degree calculation system during abnormality degree calculation. [Figure 15] Regarding the third embodiment, block diagram during learning. [Figure 16] Processing flow of the abnormality degree calculation system during learning. [Figure 17] Block diagram during abnormality degree calculation. [Figure 18] Processing flow of the abnormality degree calculation system during abnormality degree calculation.

Mode for Carrying Out the Invention

[0014] Hereinafter, embodiments of the present invention will be described based on the drawings. The abnormality degree calculation system according to the present embodiment is in an environment where a detection target device and devices (also called co-concept devices and other devices) having the same upper concept and different lower concepts are mixed. Even when the sensor data of the detection target device and the sensor data of the co-concept devices follow different normal distributions, by using the sensor data of both, the accuracy of detecting abnormalities is improved. As a result, in the abnormality degree calculation system of the present embodiment, sufficient detection accuracy can be obtained by using a relatively small amount of normal data of the detection target device.

[0015] The abnormality degree calculation system of the present embodiment is n devices M including the target device, and acquires an input D0 from the devices M that are the same with respect to the upper concept, for example, the model type. The input D0 includes a model number D1, which is an example of the lower concept of the model type, and sensor data D2.

[0016] The abnormality calculation system of this embodiment includes a concept type assignment unit 11 that outputs a concept type D3 indicating whether each device is a device to be detected (target device) or a device with the same concept as the device to be detected (other device) based on the model number D1.

[0017] Furthermore, the anomaly calculation system of this embodiment includes a feature vector extraction unit 12 that outputs feature quantities D4 from sensor data D2, a likelihood calculation unit 13 that outputs likelihood D5 from feature quantities D4 using a machine learning model, a loss calculation unit 14 that calculates a loss value D6 using a loss function defined as a function of likelihood D5 that increases the likelihood of data from the device to be detected and decreases the likelihood of data from devices of the same concept, a learning unit 15 that updates the parameters of the machine learning model of the likelihood calculation unit 13 to minimize the loss value D6, a retraining necessity determination unit 16 that determines whether retraining is necessary from likelihood D5 and concept type D3, and an anomaly calculation unit 17 that calculates the anomaly from the likelihood D7 of the detected target if it is determined that retraining is unnecessary.

[0018] In this embodiment, by learning to assign a high likelihood to normal data of the device to be detected and to consider it as normal data, and to assign a low likelihood to normal data of devices with similar normal distributions to the device to be detected and to consider it as abnormal data, a model can be obtained that assigns a high likelihood only to the data of the device to be detected and considers it as normal. Therefore, compared to conventional techniques that learn using only the data of the device to be detected, it is possible to accurately estimate the normal distribution of the data of the device to be detected and to obtain sufficient detection accuracy even when the amount of data of the device to be detected is small. [Examples]

[0019] The first embodiment will be explained using Figures 1 to 12. Figure 1 is an explanatory diagram showing the overall overview of this embodiment.

[0020] The abnormality calculation system 1 (see Figure 6) calculates the abnormality of multiple devices M that exhibit measurable physical changes in response to their operation. Measurable physical changes in response to operation include, for example, sound, vibration (sound and vibration can be collectively called vibration-related signals), temperature, and color associated with the operation of devices such as electric motors, hydraulic motors, pneumatic cylinders, hydraulic cylinders, solenoids, and linear actuators. Here, we will explain abnormality detection using sound as an example, but abnormalities can also be detected from physical changes other than sound.

[0021] The abnormality calculation system shown in Figure 1 calculates the abnormality level for selected devices from among multiple devices M1 to Mn belonging to a common higher-level concept C2. Each device M1 to Mn has a different lower-level concept C1 belonging to the higher-level concept C2. For example, the higher-level concept C2 is the model, and the lower-level concept C1 are different model numbers belonging to the same model. To show that the lower-level concept C1 is different for each device, it is shown as C11 to C13 in Figure 1. If the model number is the same, differences in serial numbers are not considered.

[0022] The superordinate concept C2 and subordinate concept C1 can also be expressed as follows. For example, the smallest unit for classifying device M is subordinate concept C1, and the concept directly above subordinate concept C1 is the superordinate concept C2. Here, the device to be detected is referred to as device M1, and the other devices M2 to Mn are called equivalent devices. When devices M1 to Mn are not specifically distinguished, they are referred to as device M or simply device.

[0023] Input D0 includes sensor data D2 obtained from n identical devices M with respect to the same model, which is an example of a higher-level concept, and model number D1, which is an example of a lower-level concept. Hereafter, model number D1 may be abbreviated as model D1.

[0024] The concept type assignment unit 11 converts the model number D1 into concept type D3. The feature vector extraction unit 12 outputs a feature vector D4 from the sensor data D2. Hereafter, the feature vector D4 may be abbreviated as feature D4. The training database DB1 stores the sensor data D2 and concept type D3. The training database DB1 can also be called the training digital input signal database DB1.

[0025] The likelihood calculation unit 13 uses a machine learning model provided from the training database DB2 to output the likelihood D5 of the sensor data for each device M from the feature quantities D4.

[0026] The loss calculation unit 14 uses the concept type D3 and likelihood D5 to calculate a loss function such that the likelihood of the device M1, which is the detection target, is high, and the likelihood of the devices M2 to Mn, which are equivalent concepts, is low, and outputs the loss D6.

[0027] The model update unit 15 updates the parameters of the machine learning model stored in the training database DB2 using the loss D6. Hereafter, the machine learning model may be abbreviated as "model".

[0028] The retraining necessity determination unit 16 determines whether retraining is necessary when an anomaly is detected, using the likelihood D5 and the concept type D3.

[0029] If the anomaly score calculation unit 17 determines that retraining is not necessary, it calculates the anomaly score of the detected object using the likelihood D7 of the detected object from the likelihood D5.

[0030] The relationship between input D0, model number D1, sensor data D2, and concept type D3 is explained using Figure 2. The input D0 obtained from each device M includes the model number D1 and sensor data D2. The model number D1 is converted to concept type D3 by the concept type assignment unit 11. For example, concept type D3 is set to "1" for the detection target device M1, and to "0" for other devices M2 to Mn that have the same concept as the detection target.

[0031] Figure 3 shows how to calculate likelihood and loss using input from the training database DB1. First, combinations of sensor data D2 and concept type D3 in batches of N are obtained from the training database DB1.

[0032] Here, for example, N / 2 combinations of sensor data D2 and concept type D3 from the device to be detected are acquired, and for example, N / 2 combinations from devices with the same concept as the device to be detected are acquired. This is because, if there are multiple other devices with the same concept as the device to be detected, and data of batch size N is randomly acquired from the set of data from the device to be detected and the data from other devices with the same concept, the amount of data from the device to be detected will be less than the amount of data from other devices with the same concept, which will slow down the training of the machine learning model.

[0033] Next, the feature vector extraction unit extracts feature vectors from the sensor data D2 to obtain N features D4. The likelihood calculation unit 13 calculates the likelihood using a machine learning model to obtain N likelihoods D5 from the features D4.

[0034] A machine learning model can be any model that allows for parameter updates using a loss function defined by the likelihood function. Examples include Normalizing Flow (NF), Gaussian Mixture Model (GMM), and Mahalanobis.

[0035] The loss calculation unit 14 receives N likelihoods D5 and a concept type D3 corresponding to each likelihood D5 as input. The loss calculation unit 14 calculates and outputs the loss using a loss function. The loss function is given, for example, by the following equation 1.

[0036]

number

[0037] Ntarget is the number N / 2 of combinations of likelihood D5 and concept type D3 from the device to be detected included in the batch. Noutlier is the number of samples that satisfy the conditions within the indicator function I among the number N / 2 of combinations of likelihood D5 and concept type D3 from other devices of the same concept as the detection target included in the batch. NLL(x) is the negative log-likelihood of feature D4. k is a hyperparameter where 0 ≦ k ≦ 1, and by adjusting k, the ratio of the effect of increasing the likelihood of the data of the device to be detected in the first term of Equation 1 and the effect of decreasing the likelihood of the data of the device to be detected in the second term can be adjusted. By decreasing k, the effect of increasing the likelihood of the data of the device to be detected in the first term can be enhanced.

[0038] When using, for example, GMM as the 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 and is determined, for example, by the method shown below. Both the feature from the device to be detected and the feature from the device of the same concept as the detection target are given to the likelihood calculation unit, and the loss function is given as in Equation 2 below.

[0039]

Equation

[0040] The value of NLL(x) when the model update is performed several times to minimize this loss function is used as the threshold value c.

[0041] Figure 4 is a graph showing the transition of the negative log-likelihood of each data when the model is updated using Equation 1. The vertical axis of Figure 4 represents NLL(x), and the horizontal axis represents the number of model updates. Above the graph of Figure 4, 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 Figure 4, the change in the negative log-likelihood of the data of the device to be detected is shown.

[0042] By using equation 1, a constraint can be added to make the NLL(x) of the data of the detected device smaller than the NLL(x) of the data of other devices that are of the same concept as the detected device. In the anomaly calculation unit 17, the anomaly is calculated using the negative log-likelihood as the anomaly, so this constraint corresponds to the model learning to lower the anomaly of the data of the detected device and increase the anomaly of the data of other devices that are of the same concept as the detected device.

[0043] Furthermore, by using the indicator function I, it is possible to prevent the NLL(x) of data from other devices of the same concept as the detected device from diverging.

[0044] The method by which the retraining necessity determination unit 16 determines whether retraining is necessary will be explained using Figure 5. In Figure 5, the vertical axis represents NLL(x), and the horizontal axis represents time. The upper part of the graph in Figure 5 shows the negative log-likelihood of data for other devices of the same concept as the detected device. The lower part of the graph in Figure 5 shows the negative log-likelihood of data for the detected device.

[0045] When an anomaly detection system is operated using the model created during model training, the distribution of data from the target device and data from other devices with the same conceptual level as the target device changes over time from the time of training.

[0046] Therefore, as shown on the left side of the graph in Figure 5 at the "Retraining Execution" point, the negative log-likelihood of the data for the detected device will increase, while the negative log-likelihood of the data for other devices with the same concept as the detected device will decrease.

[0047] If the negative log-likelihood of the data for the detected device is high, there is a possibility that the detected device is malfunctioning. Therefore, retraining is performed only when the negative log-likelihood of the data for other devices with the same conceptual identity as the detected device decreases. For example, the user sets a retraining threshold Thc, and retraining is performed when the negative log-likelihood of the data for other devices with the same conceptual identity as the detected device falls below the retraining threshold c.

[0048] Figure 6 shows the hardware and software configuration of the anomaly severity calculation system. By having a computer execute a predetermined computer program, that computer can be used as an anomaly severity calculation device 100. Figure 6 shows an example in which an anomaly severity calculation device 100 is configured from a single computer, but instead, one or more anomaly severity calculation devices 100 may be formed from multiple computers. The computer may also be a virtual computer.

[0049] The abnormality calculation device 100 is connected to n sensor terminals T, each of which is attached to n devices, via a communication network CN.

[0050] The sensor terminal T is configured, for example, as a portable recording terminal.

[0051] The abnormality calculation device 100 is a computer that includes, for example, a calculation unit 1001, a main memory 1002, an auxiliary memory 1003, an input unit 1004, an output unit 1005, and a communication unit 1006.

[0052] The arithmetic unit 1001 includes one or more microprocessors and reads a predetermined computer program stored in the auxiliary storage device 1003 into the main storage device 1002 and executes it, thereby realizing functions such as the concept classification unit 11, feature vector extraction unit 12, likelihood calculation unit 13, loss calculation unit 14, model update unit 15, retraining necessity determination unit 16, and anomaly score calculation unit 17 as described in Figure 1.

[0053] The input unit 1004 may include, for example, a keyboard, touch panel, or pointing device, and receives input from a user using the abnormality calculation device 100. The output unit 1005 may include, for example, a monitor display, speaker, or printer, and provides information to the user.

[0054] 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).

[0055] The storage medium MM is, for example, a storage medium such as flash memory or a hard disk, and is used to transfer and store computer programs or data to the abnormality calculation device 100, or to read computer programs or data from the abnormality calculation device 100 and store them. The storage medium MM may be directly connected to the abnormality calculation device 100, or it may be connected to the abnormality calculation device 100 via a communication network CN.

[0056] The configuration of the sensor terminal T will now be described. The sensor terminal T includes, for example, a sensor unit 21, a control unit 22, a storage unit 23, and a communication unit 24.

[0057] The sensor unit 21 is a device that acquires sensor data from equipment such as a microphone, acceleration sensor, magnetic sensor, and camera. The data acquired by the sensor unit 21 is stored in the storage unit 23. The control unit 22, which controls the sensor terminal T, transmits the sound data stored in the storage unit 23 to the abnormality calculation device 100. The abnormality calculation device 100 can then calculate the abnormality level based on the received sensor data.

[0058] Figure 7 is a block diagram of the feature vector extraction unit 12. Figure 8 is a block diagram of the anomaly score calculation device 100 during training, and Figure 9 is a processing flow within the anomaly score calculation device 100 during training.

[0059] The configuration of the feature vector extraction unit 12 will be explained. The explanation will describe the case where the sensor unit 21 of the sensor terminal T is a microphone, but it is not limited to a microphone; any device capable of acquiring sensor data from equipment M, such as an acceleration sensor, magnetic sensor, or camera, is acceptable. If a sensor other than a microphone is used, a known feature extraction method shall be used. If the sensor unit 21 is any other type of sensor, a known feature vector extraction method appropriate to the type of sensor shall be used.

[0060] The input sound acquisition unit 101 converts the analog input signal received from the microphone into a digital input signal using an A / D (analog / digital) converter (S101) and stores it in the training digital input signal database DB1 (S102).

[0061] The frame division unit 102 divides the digital input signal taken from the training digital input signal database DB1 into a specified number of time points (hereinafter referred to as frame size) and outputs a frame signal (S104). Frames may overlap.

[0062] The window function multiplication unit 103 outputs a window function multiplied signal by multiplying the input frame signal by a window function (S105). For example, a Hanning window is used as the window function.

[0063] The frequency domain signal calculation unit 104 outputs a frequency domain signal (S106) by applying a short-time Fourier transform to the input window function multiplied signal. The frequency domain signal is a set of M complex numbers, where (N / 2+1)=M frequency bins, each corresponding to one complex number, if the frame size is N. The frequency domain signal calculation unit 104 can also use frequency transformation methods such as the constant Q transform (CQT) instead of the short-time Fourier transform.

[0064] The power spectrogram calculation unit 105 outputs a power spectrogram based on the input frequency domain signal (S107). The filter bank multiplication unit 106 outputs a Mel power spectrogram by multiplying the input power spectrogram by the Mel filter bank (S108). The filter bank multiplication unit 106 may use a filter bank such as a 1 / 3 octave band filter instead of the Mel filter bank.

[0065] The instantaneous feature calculation unit 107 outputs a logarithmic mel-power spectrogram by applying a logarithm to the input mel-power spectrogram (S109). Alternatively, mel-frequency cepstrum coefficients (MFCCs) may be calculated instead of the logarithmic mel-power spectrogram. In that case, instead of the filter bank multiplication unit 106 and the logarithmic calculation unit 107, the logarithmic value of the power spectrogram is calculated, multiplied by the filter bank, a discrete cosine transform is applied, and the MFCCs are output.

[0066] The feature time series calculation unit 108 outputs feature D4 by concatenating adjacent L frames for the input log-MelPower spectrogram or MFCC (S110). Alternatively, instead of a log-MelPower spectrogram or MFCC, a time series (delta) of their time difference or time derivative may be input, and feature D4 may be output by concatenating adjacent L frames.

[0067] The feature vector D4 may be output by concatenating adjacent L frames by inputting a time difference or time derivative time series (delta-delta). Furthermore, the feature vector D4 may be output by concatenating adjacent L frames to any combination of these in the feature vector axis direction. The feature vector D4 generated by the feature vector extraction unit 12 is input to the likelihood calculation unit 13.

[0068] 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).

[0069] The loss calculation unit 14 calculates a predetermined loss function value D6, which is defined as the likelihood of the data of the detected device and the likelihood of the data of other devices with the same concept as the detected device, based on the likelihood D5 and the concept type D3. The loss D6 calculated by the loss calculation unit 14 is input to the model update unit 15 (S112).

[0070] The model update unit 15 repeatedly trains the parameters of the machine learning model so that the value of the loss D6 is minimized (S113~S116). These machine learning model parameters are stored in the training database DB2 (S117).

[0071] In other words, the model update unit 15 determines whether the convergence condition is met or whether the number of iterations C1 of this process exceeds the upper limit ThC (S113).

[0072] If the convergence conditions are not met, or if the number of iterations C1 is less than or equal to the upper limit ThC, the model update unit 15 updates the parameters of the machine learning model (S114), calculates the convergence conditions (S115), increments the number of iterations C1 by one, and returns to step S112.

[0073] The model update unit 15 saves the parameters of the machine learning model to the training database DB2 (S117).

[0074] Figure 10 is a block diagram of the abnormality calculation device 100 used to calculate the abnormality level. Figure 11 shows the processing flow within the abnormality calculation device 100 used to calculate the abnormality level. The explanation will be given with reference to Figures 10 and 11.

[0075] The abnormality calculation device 100 has a retraining necessity determination unit 16 and an abnormality calculation unit 17 instead of a loss calculation unit 14 and a model update unit 15.

[0076] When calculating the anomaly score of feature quantity D4 extracted from sensor data D2, the likelihood calculation unit 13 reads parameters from the training database DB2 (S201).

[0077] The contents of steps S101 to S110 have already been described, so redundant explanations will be omitted. In these steps S101 to S110, feature quantities D4 are generated for the sensor data D2 acquired from the sensor terminal T, and these are input into the machine learning model of the likelihood calculation unit 13. Redundant explanations will also be omitted in the following description. Note that in S102, the input sound is saved to the training database DB1 even when anomaly calculation is performed, in order to save training data for when retraining is performed.

[0078] The likelihood calculation unit 13 calculates the likelihood D5 from the feature quantity D4 and inputs the calculated likelihood D5 to the retraining necessity determination unit 16 (S111).

[0079] The retraining necessity determination unit 16 uses the likelihood D5 and concept type D3 to perform retraining if the negative log-likelihood of other devices with the same concept as the detected device exceeds a certain retraining necessity determination threshold Thc. If it falls below that threshold, it inputs the likelihood D7 of the detected device's data to the abnormality calculation unit 17 (S202).

[0080] The abnormality calculation unit 17 calculates the abnormality of the sensor data of the device to be detected based on the likelihood D7 of the data of the device to be detected (S203).

[0081] Here, we disclose the configuration when using a machine learning model, for example, Normalizing Flow (NF).

[0082] The likelihood calculation unit 13 is composed of a multilayer neural network. The input to the likelihood calculation unit 13 consists of an element numbered by the dimensions of the feature vector, with each element receiving each element of the feature vector. The input layer is connected to a predetermined transformation function, a multilayer neural network having the same number of elements in its input and output, and a nonlinear activation function (e.g., ReLU function), with the final layer of the multilayer neural network having the same number of elements as the input layer. In this case, the predetermined transformation function is defined such that the transformation from the input layer to the output of the multilayer neural network is an inversely transformable transformation. The likelihood calculation unit 13 calculates the likelihood from the output of the final layer of the multilayer neural network, assuming that the distribution function follows a known distribution (e.g., a normal distribution). Furthermore, it calculates the likelihood D4 of the input feature vector using a known random variable transformation formula.

[0083] The loss calculation unit 14 calculates the loss D6 using a loss function defined as a function of likelihood D5, for example, Equation 1. In this case, the loss function is defined such that the likelihood D5 of the data of the detected device is high, and the likelihood D5 of the data of other devices that are of the same concept as the detected device is low, as in Equation 1.

[0084] 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 using known optimization algorithms such as SGD, Momentum SGD, AdaGrad, RMSprop, AdaDelta, and Adam.

[0085] According to this embodiment, it is possible to learn to recognize normal data of the detected device as normal data by increasing the likelihood of such data, and to recognize normal data of other devices with the same conceptual level as the detected device as abnormal data by decreasing the likelihood of such data.

[0086] Figure 12 illustrates the effects of this embodiment. The upper part of Figure 12 shows the case where the machine learning model is trained using only normal data from the device being detected. The lower part of Figure 12 shows the case where the machine learning model is trained by considering normal data from a device of the same concept as the device being detected as abnormal. The thin shaded area indicates the region that the machine learning model in this embodiment considers to be normal data.

[0087] As shown in Figure 12(1), when training is performed using only normal data of the device to be detected, the trained model will give a high likelihood because the normal data of the device to be detected and the normal data of other devices with the same concept as the device to be detected have similar normal distributions.

[0088] However, in this embodiment, the trained model considers normal data from other devices with the same conceptual identity as the device being detected as abnormal data, giving it a low likelihood. Therefore, according to this embodiment, as shown in Figure 12(2), the region that the machine learning model considers to be normal can be reduced, and the normal distribution of data from the device being detected can be estimated more accurately.

[0089] Therefore, according to this embodiment, by using normal data from other devices with the same conceptual background as the device to be detected, the normal distribution of data for the device to be detected can be estimated more accurately, and sufficient detection accuracy can be obtained even when there is little data for the device to be detected. [Examples]

[0090] The second embodiment will be explained using Figures 13 and 14. In each of the following embodiments, including this embodiment, the differences from the first embodiment will be described in detail.

[0091] Figure 13 is a block diagram for sequential learning processing. Figure 14 is the processing flow for sequential learning processing.

[0092] In this embodiment, first, a machine learning model is trained using batch learning, similar to the learning method described in Figures 7 and 8 in the first embodiment. The difference between this embodiment and the first embodiment is that, at the same time as calculating the anomaly score using the trained model, the trained model is updated using data obtained during operation, and the anomaly score is also calculated from the updated model.

[0093] The input D0 from the sensor terminal T is divided into a model number D1 and sensor data D2. The model number D1 is converted to a concept type D3 by the concept type assignment unit 11, and the sensor data D2 is converted to a feature vector D4 using the feature vector extraction unit 12. The sensor data D2 and concept type D3 are stored in the data storage database DB3 for purposes such as identifying the cause of an anomaly when one occurs.

[0094] In this embodiment, there are two types of likelihood calculation units. One is a training model likelihood calculation unit 21 that calculates the likelihood of feature D4 using a model obtained from a training database that is updated sequentially. The other is a trained model likelihood calculation unit 22 that calculates the likelihood of feature D4 using a model obtained from a trained database DB4 that stores trained models.

[0095] The loss calculation unit 14 calculates the loss value from the likelihood calculated by the learning model likelihood calculation unit 21 using a predetermined loss function defined as a function of likelihood. The predetermined loss function is a loss function that gives a high likelihood to the data of the device to be detected and a low likelihood to the data of other devices that are of the same conceptual level as the device to be detected.

[0096] Note that in the sequential learning process used in this embodiment, since the sequential model is updated using sequentially input sensor data, unlike during batch learning, learning cannot be performed using batches. Therefore, out of the n sensor data inputs from n devices through the sensor terminal T, one piece of data of the detection target device and one piece of data of another device having the same concept as the detection target device are extracted, and the loss is calculated from their likelihoods. For example, when using Equation 1, Ntarget is 1, Noutlier becomes 1 when NLL(x) < c, and Equation 1 consists of only the first term when NLL(x) ≥ c.

[0097] The model update unit 15 updates the parameters based on the value of the loss and stores the obtained parameters in the learning database DB2. For minimizing the loss function, a known optimization algorithm for performing sequential parameter updates, such as Online Gradient Descent (OGD), can be used.

[0098] The abnormality degree calculation unit 17 calculates the abnormality degree from 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 sequentially updated model, it is difficult to handle abnormalities that occur over a long period of time, such as aging deterioration. Therefore, the likelihood from the learned model at the start of operation is added to handle the above-mentioned abnormalities. As a method of calculating the abnormality degree using both the likelihood obtained from the learned model and the likelihood obtained from the sequentially updated model, for example, a function such as Equation 3 can be used.

[0099]

Equation

[0100] A is the final abnormality degree. A1 is the abnormality degree calculated using the model obtained from the learning database. A2 is the abnormality degree calculated using the model obtained from the learned database. α is a parameter determined during operation.

[0101] This embodiment, configured in this way, also produces the same effects as the first embodiment. Furthermore, in this embodiment, since the model is trained sequentially while calculating the anomaly score, a model that is robust to changes in the normal distribution of each device can be trained.

[0102] In this embodiment, the training database DB2 at the start of operation is used as the trained database DB4, and the likelihood given by the model at the start of operation is also calculated and reflected in the final anomaly score, making it possible to detect anomalies that occur over long periods of time, such as aging.

[0103] Furthermore, this embodiment can also handle cases where the input data for the isotope concept changes, such as when other devices for the isotope concept are newly added. [Examples]

[0104] A third embodiment will be explained using Figures 15 to 18. In the first and second embodiments, all devices that are the same as the detected device in terms of a higher-level concept, for example, the model, but different in terms of a lower-level concept, for example, the type, were treated as other devices of the same concept as the detected device.

[0105] On the other hand, within the same concept of other devices, there may be multiple types of sub-concepts, such as model numbers, and the degree of similarity with the data of the device to be detected may differ for each type. For example, there may be devices that are the same model as the device to be detected, but have different model numbers, and have the same or different drive methods. In this embodiment, the aim is to further improve detection accuracy by utilizing these types of sub-concepts.

[0106] The learning process of this embodiment will be explained using Figures 15 and 16. Of the input D0 from the sensor terminal T, the model number D1 is converted to a concept type D3 by the concept type assignment unit 11. The sensor data D2 is converted to a feature vector D4 by the feature vector extraction unit 12. Here, unlike the first and second embodiments, the concept type assignment unit 11 outputs both the concept type D3 and the model number D1.

[0107] Sensor data D2 is converted into feature vector D4 by the feature vector extraction unit 12. The homogeneous concept classification unit 31, when the set of other devices with the same concept as the device to be detected includes multiple models, examines the relationship between each model of the other devices with the same concept and the model of the device to be detected, and assigns a weight w to each model of the other devices with the same concept. This weight w may be set, for example, based on the commonality of elements related to the sound source emitted from the device to be detected, i.e., the sound source. For example, the sensor data will differ depending on whether the device uses an electric motor or a hydraulic motor. Therefore, in this embodiment, a high value is given to the model of other devices that have the same drive type as the device to be detected, indicating a high similarity to the device to be detected, and a low value is given to the model of other devices that have a different drive type than the device to be detected, indicating a low similarity to the device to be detected.

[0108] The training database DB1 stores sensor data D2 and concept type D3, as well as weights w corresponding to each model of other devices of the same concept.

[0109] The likelihood calculation unit 13 uses the model obtained from the training database DB2 to calculate the likelihood D5 from the features D4 calculated by the feature vector extraction unit 12.

[0110] The loss calculation unit 14 calculates the loss D6 from the likelihood calculated by the likelihood calculation unit 13, using a loss function defined as a likelihood function that gives a high likelihood to the data of the detected device and a low likelihood to the data of other devices that are of the same concept as the detected device. The loss function used here can be, for example, Equation 4, which is a modified version of Equation 1.

[0111]

number

[0112] Here, wi is the weight of the i-th type as defined in the equivalence concept type assignment unit 31.

[0113] 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 using known optimization algorithms such as SGD, Momentum SGD, AdaGrad, RMSprop, AdaDelta, and Adam.

[0114] The process for calculating the anomaly in this embodiment will be explained using Figures 17 and 18. The conceptual type assignment unit 11 assigns conceptual type D3 to the type number D1 of the input D0 from the sensor terminal T. The sensor data D2 is converted into feature vector D4 by the feature vector extraction unit 12.

[0115] The same-concept type assignment unit 31, when the set of other devices with the same concept as the device to be detected includes multiple models, examines the relationship between each model of the other devices with the same concept and the device to be detected, and assigns a weight w to each model of the other devices with the same concept.

[0116] The training database DB1 stores sensor data D2 and concept type D3, as well as weights w corresponding to each model of other devices of the same concept, for use in retraining.

[0117] The likelihood calculation unit 13 uses the model obtained from the training database DB2 to calculate the likelihood D5 from the features D4 calculated by the feature vector extraction unit 12.

[0118] The retraining necessity determination unit 16 uses the likelihood D5 and concept type D3 to retrain the model if the negative log-likelihood of other devices with the same concept as the device to be detected exceeds a certain retraining necessity determination threshold Thc. If the negative log-likelihood falls below the threshold Thc, it inputs the likelihood D5 of the data for the device to be detected into the abnormality calculation unit 17 (S202).

[0119] The abnormality calculation unit 17 calculates the abnormality of the sensor data of the device to be detected based on the likelihood D5 of the data of the device to be detected (S203).

[0120] This embodiment, configured in this way, also produces the same effects as the first embodiment. Furthermore, this embodiment adds a type information database DB5 that holds type information and an equivalent concept type assignment unit 31 that assigns weights according to the type of other devices of the equivalent concept. As a result, according to this embodiment, the influence of the data of other devices of each equivalent concept on the value of loss D6 can be adjusted for each type of other device of each equivalent concept. Therefore, in this embodiment, for example, by assigning a larger weight to devices of the same type as the detected device and a smaller weight to devices of different types, the model can be trained to lower the likelihood of data that is more similar to the detected device.

[0121] It should be noted that the present invention is not limited to the embodiments described above, and includes various modifications. For example, the embodiments described above are described in detail to make the present invention easier to understand, and are not necessarily limited to those having all the configurations described. Furthermore, it is possible to replace parts of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add configurations from other embodiments to the configuration of one embodiment. In addition, it is possible to add, delete, or replace parts of the configuration of each embodiment with other configurations.

[0122] Each component of the present invention can be arbitrarily selected or omitted, and an invention comprising the selected components is also included in the present invention. Furthermore, the components described in the claims can be combined in combinations other than those explicitly stated in the claims. (Note 1) An abnormality calculation system for calculating the degree of abnormality of the target equipment, A conceptual type assignment unit that assigns a predetermined conceptual type based on the identification number of the target device, A feature vector extraction unit extracts feature vectors based on sensor data obtained from a sensor corresponding to the aforementioned target device, A likelihood calculation unit that calculates the likelihood of the feature vector using a machine learning model obtained from a training database, A loss calculation unit that calculates the loss using a loss function defined as a function of the likelihood calculated by the likelihood calculation unit, A model update unit updates the machine learning model using the loss calculated by the loss calculation unit and the learned machine learning model. A retraining necessity determination unit determines whether retraining is necessary based on the likelihood calculated by the likelihood calculation unit, If the retraining necessity determination unit determines that retraining is unnecessary, an abnormality calculation unit calculates the abnormality level. Equipped with Anomaly severity calculation system. (Note 2) The aforementioned predetermined conceptual type is a conceptual type obtained from the identification number, and is an equivalent concept belonging to a common higher-level concept with the target device. The loss calculation unit calculates a loss to be fed back to the machine learning model such that it considers the sensor data of the target device as normal and gives a high likelihood, and considers the sensor data from the target device and other devices of the same concept as abnormal and gives a low likelihood. The abnormality calculation system described in Appendix 1. (Note 3) The aforementioned target device and other devices of the aforementioned equivalent concept share the same model, which is the smallest unit for classifying devices, but the types included in each model are different. The abnormality calculation system described in Appendix 2. (Note 4) The aforementioned identification number is information assigned according to the model. The abnormality calculation system described in Appendix 1. (Note 5) The aforementioned conceptual type is set to "1" for the target device and to "0" for the other devices. The abnormality calculation system described in Appendix 2. (Note 6) Multiple devices that exhibit measurable physical changes in response to operation, which belong to a common model and are pre-grouped together, One of the target devices is selected from the grouped devices. Of the grouped devices, at least one device other than the target device is selected as another device. The abnormality calculation system described in Appendix 2. (Note 7) The aforementioned relearning necessity determination unit determines whether relearning is necessary during operation using the likelihood of data from other devices of the same concept. The abnormality calculation system described in Appendix 2. (Note 8) Furthermore, in order to perform model learning not only during training but also during operation, the system includes a training model likelihood calculation unit that calculates the likelihood of the training model, and a trained model likelihood calculation unit that calculates the likelihood of the trained model. The abnormality calculation system described in Appendix 1. (Note 9) The aforementioned training model likelihood calculation unit calculates the likelihood by reading the model from the training database which stores the model that is updated sequentially during operation. The aforementioned trained model likelihood calculation unit calculates the likelihood by reading the model from the trained database which stores the model at the start of operation. The abnormality calculation system described in Appendix 8. (Note 10) The system further includes an equivalence concept classification unit that assigns weights to the other devices according to the relationship between the sub-concepts of the target device and the sub-concepts of each of the other devices. The abnormality calculation system described in Appendix 2. (Note 11) The isotype classification unit assigns weights to each of the other devices and uses these weights in the loss function to adjust the influence of the likelihood values ​​of each of the other devices on the loss value. The abnormality calculation system described in Appendix 10. (Note 12) The aforementioned device is a device that generates sound or vibration in response to its operation. The abnormality calculation system described in Appendix 1. (Note 13) An abnormality calculation system that calculates the degree of abnormality of the target equipment using a computer, According to the aforementioned computer, A predetermined conceptual category is assigned based on the identification number of the aforementioned target device. Based on the sensor data obtained from the sensor corresponding to the aforementioned target device, a feature vector is extracted. Using a machine learning model obtained from a training database, the likelihood of the feature vector is calculated. The loss is calculated using the loss function defined as a function of the calculated likelihood. The machine learning model is updated using the calculated loss and the trained machine learning model. When an abnormality is detected in the target device, the system determines whether or not relearning is necessary based on the calculated likelihood. If it is determined that retraining is not necessary, calculate the anomaly score. Abnormality degree calculation method. [Explanation of Symbols]

[0123] 1: Anomaly Score Calculation System, 11: Concept Classification Unit, 12: Feature Vector Extraction Unit, 13: Likelihood Calculation Unit, 14: Loss Calculation Unit, 15: Model Update Unit, 16: Retraining Necessity Determination Unit, 17: Anomaly Calculation Unit, 21: Training Model Likelihood Calculation Unit, 22: Trained Model Likelihood Calculation Unit, 31: Isotropic Concept Classification Unit, M: Equipment

Claims

1. An abnormality calculation system for calculating the degree of abnormality of the target equipment, A conceptual classification unit that assigns a predetermined conceptual classification based on the identification number of the device, A feature vector extraction unit extracts feature vectors based on sensor data obtained from a sensor corresponding to the aforementioned device, A likelihood calculation unit that calculates the likelihood of the feature vector using a machine learning model obtained from a training database, A loss calculation unit that calculates the loss using a loss function defined as a function of the likelihood calculated by the likelihood calculation unit, A model update unit updates the machine learning model using the loss calculated by the loss calculation unit and the learned machine learning model. An anomaly calculation unit calculates the anomaly score based on the likelihood calculated by the likelihood calculation unit. Equipped with, The concept classification unit assigns different concept classifications to the device when assigning a concept classification based on the device's identification number, depending on whether the normal distribution of the device's sensor data is the same as or different from the normal distribution of the sensor data of the target device for which the degree of abnormality is calculated. The loss calculation unit calculates the loss value using a loss function that increases the loss value as the value of the negative log-likelihood based on the likelihood derived from the sensor data of the device assigned the same conceptual type as the target device is larger, and increases the loss value as the value of the negative log-likelihood based on the likelihood derived from the sensor data of the device assigned a different conceptual type from the target device is smaller. The model update unit updates the machine learning model in such a way that it reduces the loss value calculated by the loss calculation unit. Anomaly severity calculation system.

2. A single model may include multiple variations. The model of the aforementioned target device and the model of other devices that are of the same concept belonging to a common higher-level concept as the aforementioned target device, which are devices to which a different conceptual type from the conceptual type of the aforementioned target device has been assigned, are common. The model number of the aforementioned target device and the model numbers of other devices of the same concept are different from each other. The abnormality level calculation system according to claim 1.

3. The aforementioned identification number is information assigned according to the model. The abnormality level calculation system according to claim 1.

4. The aforementioned conceptual type is set to "1" for the target device, and to "0" for other devices of the same concept belonging to a common higher-level concept with the target device, which are devices to which a different conceptual type from the target device has been assigned. The abnormality level calculation system according to claim 1.

5. Multiple devices that exhibit measurable physical changes in response to operation, which belong to a common model and are pre-grouped together, One of the target devices is selected from the grouped devices. Of the grouped devices, at least one device other than the target device is selected as another device of an equivalent concept belonging to a common higher-level concept with the target device, which is a device to which a different conceptual type from the conceptual type of the target device has been assigned. The abnormality level calculation system according to claim 1.

6. The anomaly calculation system comprises a trained model likelihood calculation unit and a training model likelihood calculation unit, which is the likelihood calculation unit. The aforementioned learning model likelihood calculation unit calculates the likelihood of the feature vector using the machine learning model obtained from the learning database, not only during learning but also during operation. The aforementioned trained model likelihood calculation unit calculates the likelihood of the feature vector using the trained machine learning model obtained from the trained database during operation. The abnormality level calculation system according to claim 1.

7. The learning model likelihood calculation unit calculates the likelihood by reading the machine learning model from the learning database which stores the machine learning model that is updated sequentially during operation. The trained model likelihood calculation unit calculates the likelihood by reading the trained machine learning model from the trained database which stores the trained machine learning model at the start of operation. The abnormality level calculation system according to claim 6.

8. The abnormality calculation system further comprises an equivalence concept type assignment unit that sets weights for the other devices according to the similarity of the lower concepts of the target device and the lower concepts of other devices that are equivalence concepts belonging to a common higher concept with the target device, which are devices to which a different conceptual type from the conceptual type of the target device has been assigned. The loss calculation unit calculates the loss value using a loss function that includes a product obtained by multiplying the negative log-likelihood value, which is based on the likelihood derived from the sensor data of the other device, by the weight set for the other device. The abnormality level calculation system according to claim 1.

9. The aforementioned device is a device that generates sound or vibration in response to its operation. The abnormality level calculation system according to claim 1.

10. An abnormality calculation method in which a computer calculates the degree of abnormality of the target equipment, The aforementioned computer, A predetermined conceptual category is assigned based on the device's identification number. Based on the sensor data obtained from the sensor corresponding to the aforementioned device, a feature vector is extracted. Using a machine learning model obtained from a training database, the likelihood of the feature vector is calculated. The loss is calculated using the loss function defined as the calculated likelihood function, The machine learning model is updated using the calculated loss and the trained machine learning model. A method for calculating the degree of abnormality based on the calculated likelihood when detecting an abnormality in the aforementioned target equipment, When the computer assigns a conceptual type based on the identification number of the device, it assigns different conceptual types depending on whether the normal distribution of the sensor data of the device is the same as or different from the normal distribution of the sensor data of the target device for which the degree of abnormality is calculated. When the computer calculates the loss using the loss function, it uses a loss function that increases the value of the loss as the value of the negative log-likelihood based on the likelihood derived from the sensor data of the device assigned the same conceptual type as the target device is larger, and increases the value of the loss as the value of the negative log-likelihood based on the likelihood derived from the sensor data of the device assigned a different conceptual type from the target device is smaller. When the computer updates the machine learning model, it updates the machine learning model in such a way that it reduces the calculated loss value. Anomaly degree calculation method.

Images