Anomaly Detection System

Abstract

To provide an abnormality detection system and method capable of performing a more stable abnormality detection.SOLUTION: An abnormality detection system 1 for detecting an abnormality in a target machine 2 using a computer comprises: a communication unit that acquires first data D0 from a first sensor 11 attached to the target machine and acquires second data D1 from a second sensor 12 attached to the target machine; a calculation unit, and a storage unit, wherein the calculation unit comprises: a coding unit 13 that is learned to generate a latent expression including a predetermined latent expression for estimating the second data based on the first data; a decoding unit 14 that is learned to restore the first data from the latent expression; and an abnormality detection unit 16 for detecting an abnormality in the target machine based on a restoration error between the first data and the first data restored by the decoding unit.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to anomaly detection systems and methods. [Background technology]

[0002] When a factory machine malfunctions and stops for a long time, the company will suffer losses. Therefore, during round inspections, skilled workers detect early abnormalities in the machine based on their five senses. However, in recent years, there has been a shortage of skilled workers, so mechanization and labor saving are desired. In order to detect abnormalities in the target machine at an early stage, it is necessary to attach many types of sensors to the target machine and collect and analyze data. However, due to the physical constraints or cost of the target machine, it is difficult to attach all the necessary sensors to the target machine. Therefore, a technology has been proposed that uses a regression model to estimate sensor data from other sensor data (Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2017-207904 A Summary of the Invention [Problem to be solved by the invention]

[0004] In the method described in Patent Document 1, all variations in the sensor data input to the regression model are explained as variations in the estimation target. Therefore, in Patent Document 1, when variations unrelated to the sensor data of the estimation target are added to the input sensor data, the unrelated variations are reflected in the value of the sensor data of the estimation target.

[0005] Therefore, an object of the present disclosure is to provide an anomaly detection system and method that are capable of performing more stable anomaly detection. [Means for solving the problem]

[0006] In order to solve the above-mentioned problems, an anomaly detection system according to one aspect of the present invention is an anomaly detection system that detects an anomaly in a target machine using a computer, the computer including a communication unit that acquires first data from a first sensor attached to the target machine and acquires second data from a second sensor attached to the target machine, a calculation unit and a memory unit, the calculation unit including an encoding unit trained to generate latent expressions including a predetermined latent expression that estimates the second data based on the first data, a decoding unit trained to restore the first data from the latent expressions, and an anomaly detection unit that detects an anomaly in the target machine based on a restoration error between the first data and the first data restored by the decoding unit. Effect of the Invention

[0007] According to the present invention, the second data output from the second sensor is estimated using a predetermined latent expression that is a part of the latent expression. Therefore, when the first data fluctuates due to a factor other than the second sensor, the fluctuation can be explained using a latent expression other than the predetermined latent expression, and the second data can be estimated more robustly than conventional methods. [Brief description of the drawings]

[0008] [Figure 1] FIG. 1 is a block diagram of an anomaly detection system. [Diagram 2] FIG. 1 is a hardware configuration diagram of an anomaly detection system. [Diagram 3] FIG. 1 is an explanatory diagram for learning a model. [Figure 4] FIG. 4 is an explanatory diagram showing the effect of the present embodiment. [Diagram 5] 13 is a flowchart of a learning process. [Figure 6] 13 is a flowchart of an abnormality detection process (during operation). [Figure 7] FIG. 11 is a block diagram of an anomaly detection system according to a second embodiment. [Figure 8] FIG. 11 is an explanatory diagram of a method for selecting a first sensor. [Figure 9] FIG. 11 is a block diagram of an anomaly detection system according to a modified example of the second embodiment. [Figure 10] FIG. 11 is an explanatory diagram of a method for selecting a first sensor. [Figure 11] FIG. 11 is a block diagram of an anomaly detection system according to a third embodiment. [Figure 12] FIG. 13 is an explanatory diagram of a method for detecting an abnormality. [Figure 13] FIG. 13 is a block diagram of an anomaly detection system according to a fourth embodiment. [Figure 14] An illustration of a pre-trained model. [Figure 15] FIG. 13 is a block diagram of an anomaly detection system according to a fifth embodiment. [Figure 16] FIG. [Figure 17] FIG. 13 is a block diagram of an anomaly detection system according to a sixth embodiment. [Figure 18] FIG. 1 is a hardware configuration diagram of an anomaly detection system. [Figure 19] FIG. 11 is an explanatory diagram showing a method of setting hyper-parameters in a loss calculation unit. [Figure 20] FIG. 13 is an explanatory diagram showing the relationship between the degree of anomaly and a probability distribution function. [Figure 21] 13 is a flowchart of a learning process. [Figure 22] 13 is a flowchart of an abnormality detection process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0009] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. In this embodiment, the state of a machine that is the subject of anomaly detection (target machine) is determined using as few sensors as possible. The target machine 2 includes various industrial machines such as press machines, injection molding machines, heating furnaces, NC machine tools, 3D printers, electric discharge machines, welding machines, machining centers, polishing machines, industrial sewing machines, and industrial robots. The target machine 2 is not limited to industrial machines, and can also include electrical equipment such as air conditioners, freezers, and fans.

[0010] A single target machine 2 may be composed of multiple devices, such as fans, pumps, sliders, valves, and motors. Physical quantities, such as temperature, pressure, speed, voltage, current, load weight, load torque, vibration, and operating noise, are measured by sensors. The anomaly detection system 1 determines the state of the target machine 2 and the presence or absence of an anomaly based on the data output by the sensors.

[0011] The anomaly detection system 1 according to this embodiment estimates second data D1 output from a second sensor 12 different from the first sensor 11, based on first data D0 output from the first sensor 11.

[0012] The first sensor 11 is attached to the machine 2 that is the target of anomaly detection both during training of the machine learning model and during operation of the machine learning model, and outputs the first data D0. The second sensor 12 may be used only during training of the machine learning model and output the second data D1, or may be used also during operation of the machine learning model and output the second data D1.

[0013] In one embodiment, when the number of sensors that can be attached to the target machine 2 is limited, a first sensor 11 that can estimate all sensors necessary for monitoring the condition of the target machine 2 is selected. The number of first sensors 11 may be one or more. In the following explanation, a case where one first sensor 11 is used will be mainly explained.

[0014] The anomaly detection system 1 of the above-mentioned one embodiment estimates second data D1 output by a second sensor 12 that is necessary for monitoring the state of the target machine 2 but is not installed on the target machine 2, based on first data D0 output by a first sensor 11. The anomaly detection system 1 monitors the state of the target machine 2, detects anomalies, and identifies the cause of the anomaly, based on the first data D0 and the estimated second data D1. Note that there may be one or more second sensors 12 whose output is estimated based on the first data D0 measured by the first sensor 11.

[0015] According to this embodiment, a machine learning model capable of estimating the second data D1 of the second sensor 12 that is not installed from the first data D0 of the installed first sensor 11 is generated during learning, so that condition monitoring and abnormality detection can be performed even if multiple sensors cannot be installed on the target machine 2 during operation.

[0016] In another example of this embodiment, when a physical quantity corresponding to the state of the target machine 2 is measured by a sensor, a set of hyperparameters to be used when calculating the difference between the observed value and the restored value is selected according to the relationship between the physical quantity and the type of the target machine. The set of hyperparameters can be prepared in advance according to, for example, the physical properties of the target machine 2 and the required robustness. This makes it possible to appropriately detect an abnormality in the target machine 2 according to the physical properties of the target machine 2 and the required robustness. EXAMPLES

[0017] A first embodiment will be described with reference to Figs. 1 to 6. Fig. 1 is a block diagram of an anomaly detection system 1 according to this embodiment. In this embodiment, the state of a target machine 2, such as a motor, a fan, a pump, a valve, or a slider, is measured with as few first sensors 11 as possible, and the state of the target machine 2 is monitored and the degree of anomaly is calculated. The power source of the target machine 2 may be any of electricity, hydraulics, and pneumatics.

[0018] The anomaly detection system 1 includes, for example, a first sensor 11, a second sensor 12, an encoding unit 13, a decoding unit 14, a learning unit 15, an anomaly detection unit 16, a state grasping unit 17, a training database DB1, and a learning model database DB2.

[0019] The anomaly detection system 1 estimates second data D1 output from a second sensor 12 that is not installed during operation, from first data D0 output from a first sensor 11 that can be installed during operation of the target machine 2 (during operation of the machine learning model). For this reason, the encoding unit 13 acquires a latent representation for estimating the second data D1 from the first data D0 during learning of the machine learning model.

[0020] The decoding unit 14 is configured to restore the first data D0 output by the first sensor 11 using all of the latent representations obtained by the encoding unit 13.

[0021] During learning of the machine learning model (sometimes abbreviated as model learning or learning mode), learning is performed by acquiring the first data D0 and the second data D1 from the machine 2 equipped with all the sensors 11, 12 required for grasping the state of the target machine 2. During learning, the encoding unit 13 is trained to obtain a latent representation representing the data D1 of the second sensor 12 that is not installed during operation, from the data D0 of the first sensor 11 that can be installed during operation.

[0022] During operation, data D0 from a first sensor 11 installed on the target machine 2 is used. The anomaly detection system 1 uses an encoding unit 13 to estimate data D1 from a second sensor 12 that is not installed on the target machine 2.

[0023] The state grasping unit 17 of the anomaly detection system 1 grasps the state of the target machine 2 from the first data D0 of the first sensor installed on the target machine 2 and the second data (estimated value) of the second sensor 12 not installed on the target machine 2.

[0024] The state grasping unit 17 outputs the observed value or the estimated value of the first sensor 11 or the second sensor 12 as data D5. During learning, the state grasping unit 17 outputs the observed value of the first sensor 11 and the observed value of the second sensor 12. As will be described in an embodiment later, the state grasping unit 17 can also output the estimated value of the second sensor 12 during learning. During operation, the state grasping unit 17 outputs the observed value of the first sensor 11 and the estimated value of the second sensor 12 as data D5.

[0025] Here, the observed value of the first sensor 11 is the first data D0 that is actually measured by the first sensor 11. The estimated value of the second sensor 12 is the second data D1 that is estimated from the first data.

[0026] The anomaly detection unit 16 of the anomaly detection system 1 calculates and outputs an anomaly degree D6 indicating whether an anomaly has occurred in the target machine 2 based on the difference between the first data D0 observed by the first sensor 11 and the first data D0 restored by the encoding unit 13 and the decoding unit 14 that configure the machine learning model. The anomaly degree D6 can be used in other systems such as a production management system (not shown).

[0027] The learning unit 15 learns the parameters used in the encoding unit 13 and the parameters used in the decoding unit 14 based on the training data stored in the training database DB1, the first data D0 of the first sensor 11, and the second data D1 of the second sensor 12. Furthermore, the learning unit 15 updates the parameters used in the encoding unit 13 and the parameters used in the decoding unit 14 so as to minimize a predetermined loss function, and stores them in the learning database DB2.

[0028] 2 shows a hardware configuration of the anomaly detection system 1. The anomaly detection system 1 includes, for example, an anomaly detection device 100, a first sensor 11, and a second sensor 12.

[0029] The anomaly detection device 100 is configured from a computer, and by causing the computer to execute a specific computer program, the computer can be used as the anomaly detection device 100. Although Fig. 4 shows an example in which the anomaly detection device 100 is configured from one computer, instead, one or more anomaly detection devices 100 may be formed from multiple computers. The computer may be a virtual computer.

[0030] The anomaly detection device 100 is connected to sensors 11 and 12 via a communication network CN. The sensors 11 and 12 are, for example, microphones, vibration sensors, temperature sensors, current sensors, voltage sensors, weight sensors, torque sensors, etc. Here, an example will be described in which the first sensor 11 is a microphone (sound sensor) and the second sensor 12 is a sensor that measures a physical quantity other than sound.

[0031] The anomaly detection device 100 includes, for example, a calculation unit 101, a main memory device 102, an auxiliary memory device 103, an input unit 104, an output unit 105, and a communication unit 106.

[0032] The calculation unit 101 includes one or more microprocessors, and realizes predetermined functions of the anomaly detection device 100 by reading a predetermined computer program stored in the auxiliary storage device 103 into the main storage device 102 and executing it. The predetermined functions are, for example, an encoding unit 13, a decoding unit 14, a learning unit 15, an anomaly detection unit 16, and a state grasping unit 17.

[0033] The input unit 104 may include, for example, a keyboard, a touch panel, a pointing device, etc., and accepts input from a user who uses the anomaly detection device 100. The output unit 105 may include, for example, a monitor display, a speaker, a printer, etc., and provides information to the user.

[0034] The communication unit 106 communicates with the sensors 11 and 12 via the communication network CN. The communication unit 106 can also communicate with other computers (not shown).

[0035] The storage medium MM is, for example, a flash memory or a hard disk, and transfers and stores computer programs or data to the anomaly detection device 100, and reads and stores computer programs or data from the anomaly detection device 100. The storage medium MM may be directly connected to the anomaly detection device 100, or may be connected to the anomaly detection device 100 via the communications network CN.

[0036] The configurations of the sensors 11 and 12 will be described. The first sensor 11 will be described as an example. The second sensor 12 can be configured in a similar manner. The first sensor 11 includes, for example, a sensor unit 111, a control unit 112, a storage unit 113, and a communication unit 114.

[0037] The sensor unit 111 is a microphone that detects sound from the target machine 2. Therefore, hereinafter, the sensor unit 111 may be referred to as the microphone 111. Sound data detected by the sensor unit 111 is stored in the storage unit 113. The control unit 112 that controls the sensor 11 transmits the sound data stored in the storage unit 113 to the anomaly detection device 100.

[0038] Fig. 3 shows a learning method of a machine learning model. As shown in the upper part of Fig. 3, first data D0 output from a first sensor 11 is input to an encoding unit 13. The encoding unit 13 calculates a feature amount of the input first data D0 and outputs encoded data D2. The encoded data D2 has latent expressions (which may also be called latent variables) LV12(1) to LV12(3) and LV0 that restore the first data from the input first data D0.

[0039] Here, four latent expressions are illustrated for the purpose of explanation, but in reality, more latent expressions are included. Among the encoded data D2, the predetermined latent expressions LV12(1) to LV12(3) are used to estimate the output of the second sensor 12 (second data D1). The first predetermined latent expression LV12(1) is used to estimate the output of the second sensor 12(1), such as a temperature sensor. The second predetermined latent expression LV12(2) is used to estimate the output of the second sensor 12(2), such as a voltage sensor. The third predetermined latent expression LV12(3) is used to estimate the output of the second sensor 12(3), such as a load weight. The temperature sensor, voltage sensor, and load weight sensor are examples for explanation.

[0040] The other latent representation LV0 is not used to estimate the output of the second sensor 12, but is used only to restore the output of the first sensor 11. That is, only a portion of the predetermined data LV12(1) to LV12(3) of the encoded data D2 output by the encoding unit 13 is used to estimate the output of the second sensor 12, and the remaining data LV0 is used to estimate the output of the first sensor 11 together with the predetermined data LV12(1) to LV12(3).

[0041] As shown in the center of Figure 3, when training the machine learning model, not only the first sensor 11 but also the second sensors 12(1) to 12(3) are attached to the target machine 2, and data RV12(1) to RV12(3) are output as observation values.

[0042] Learning unit 15 compares data estimated from predetermined latent expressions LV12(1) to LV12(3) with actual observation data RV12(1) to RV12(3) measured by second sensors 12(1) to 12(3), and calculates an estimation error, which is the difference between the two.

[0043] On the other hand, the decoding unit 14 restores the first data D0 of the first sensor 12 from the encoded data D2 output by the encoding unit 13, and outputs the restored first data D0A. For ease of understanding, the following description will be given by switching the code of the observed first data to "D0" and the code of the restored first data to "D0A."

[0044] The learning unit 15 compares the first data D0 output by the first sensor 11 with the first data D0A reconstructed by the decoding unit 14, calculates a reconstruction error (reconstruction error) which is the difference between the two, and adjusts the parameters of the encoding unit 13 and the parameters of the decoding unit 14 so as to minimize the loss obtained by summing up the estimated error and the reconstruction error.

[0045] Fig. 4 is an explanatory diagram showing the effect of this embodiment. Fig. 4(1) shows a comparative example. In the comparative example, the second data D1 output by the second sensor 12 is estimated using all of the latent expressions LV12(1) to LV12(3) obtained from the first data D0 output by the first sensor 11. Therefore, even if a variation unrelated to the second data D1 is added to the first data D0, the variation is reflected in the estimation of the second data D1.

[0046] For example, a case will be described in which the first sensor 11 measures the operation sound of the target machine 2, and the operation speed of the target machine is estimated from the operation sound. The operation sound of another device adjacent to the target machine 2 is not a measurement target of the first sensor 11, and therefore becomes noise. If the first sensor 11 detects the operation sound of the other device as being included in the operation sound of the target machine 2, the noise component will be reflected in the estimated value of the operation speed.

[0047] In contrast, in the method of the present embodiment shown in Fig. 4(2), only a part of the latent representations generated by the encoding unit 13 is used to estimate the second data D1 of the second sensor 12, and the other latent representations LV0 are not used to estimate the second data. Therefore, if the input data D0 to the encoding unit 13 fluctuates due to a cause unrelated to the data D1 to be estimated, the fluctuation is aggregated to the other latent representations LV0 that are not used for estimation. Therefore, it can be confirmed that the fluctuation is unrelated to the estimation of the second data D1.

[0048] 5 is a flowchart of the learning process S1. The anomaly detection system 1 acquires the first data D0 from the first sensor 11 (S101), and stores the acquired first data D0 in the auxiliary storage device 103 (S102).

[0049] The anomaly detection system 1 loads the first data D0 stored in the auxiliary storage device 103 into the main storage device 102 (S103), and causes the encoding unit 13 to encode the first data D0 (S104). The anomaly detection system 1 estimates the second data D1 of the second sensor 12 from a predetermined latent expression among the latent expressions generated by the encoding unit 13 (S105).

[0050] The anomaly detection system 1 causes the decoding unit 14 to reconstruct the first data D0 using all the latent representations generated by the encoding unit 13 (S106). The anomaly detection system 1 calculates, as losses, a reconstruction error between the first data D0 input to the encoding unit 13 and the first data D0A reconstructed by the decoding unit 14, and an estimation error between the second data observed by the second sensor 12 and the second data estimated from a predetermined latent representation (S107).

[0051] The anomaly detection system 1 repeatedly learns the parameters of the machine learning model so as to minimize the calculated loss value (S109 to S111). The parameters of the machine learning model are stored in the learning database DB2 (S112).

[0052] That is, the anomaly detection system 1 determines whether a predetermined convergence condition is satisfied or whether the number of iterations C1 of this process exceeds the upper limit value ThC (S108). If the convergence condition is not satisfied or if the number of iterations C1 is equal to or less than the upper limit value ThC, the model update unit 15 updates the parameters of the machine learning model (S109), calculates the convergence condition (S110), increments the number of iterations C1 by one, and returns to step S108.

[0053] When a predetermined convergence condition is satisfied (S108: YES), the anomaly detection system 1 stores the parameters of the machine learning model in the learning database DB2 (S112).

[0054] 6 is a flowchart showing a process S2 for detecting an abnormality in the target machine 2. The abnormality detection process S2 corresponds to a process during operation when the machine learning model is operated. During operation in this embodiment, only the first sensor 11 is provided in the target machine 2, and the second sensor 12 is not provided in the target machine.

[0055] The anomaly detection system 1 reads parameters of a machine learning model from the learning model database DB2 and sets them in the encoding unit 13 and the decoding unit 14 (S201). The anomaly detection system 1 acquires the first data D0 from the first sensor 11 (S202) and stores it in the auxiliary storage device 103 (S203).

[0056] The anomaly detection system 1 reads the first data D0 stored in the auxiliary storage device 103 into the main storage device 102 (S204), and generates a latent representation by causing the encoding unit 13 to encode the first data D0 (S205).

[0057] The anomaly detection system 1 estimates the second data D1 of the second sensor 12 from a predetermined latent expression among the latent expressions generated by the encoding unit 13 (S206).

[0058] On the other hand, the decoding unit 14 reconstructs the first data D0 (D0A in FIG. 3) by using all of the latent representations generated by the encoding unit 13 (S207).

[0059] In the anomaly detection step S208, the anomaly detection system 1 treats the recovery error of the first data D0 as an anomaly degree, and judges whether the target machine 2 is in an abnormal state or a normal state based on the magnitude of the anomaly degree. Furthermore, the anomaly detection system 1 grasps the state of the target machine 2 based on the first data which is an observed value and the second data D1 which is an estimated value, and estimates the cause when it is judged that the target machine 2 is in an abnormal state.

[0060] For example, when the anomaly detection system 1 determines that the target machine 2 is in an abnormal state based on the operating sound of the target machine 2, it can estimate the cause of the anomaly of the target machine 2 from the estimated values ​​of the second data D1 output by the second sensor 12, such as the operating speed, temperature, and voltage. For example, when it is determined that the target machine 2 is abnormal based on the operating sound, if the temperature of the target machine 2 is higher than a predetermined temperature, the anomaly detection system 1 can estimate that the anomaly of the target machine 2 is caused by the temperature.

[0061] According to this embodiment configured as described above, the second data D1 of the second sensor 12 that is not used during operation can be estimated based on the first data D0 of the first sensor 11 that is used during operation. Therefore, even if multiple sensors cannot be attached to the target machine 2 due to sensor attachment restrictions, it is possible to detect abnormalities in the target machine 2 and monitor the condition from the data D0 and D1 of the multiple sensors.

[0062] In this embodiment, it is only necessary to attach a portion of the sensor group required to monitor the target machine 2, namely, sensor 11, to the target machine 2, thereby reducing the cost of purchasing and installing the sensors, and as a result, reducing the cost of the anomaly detection system 1.

[0063] In this embodiment, the second data D1 of the second sensor 12 is estimated from a predetermined latent representation that is part of the latent representations possessed by the machine learning model. Therefore, even if noise unrelated to the second sensor 12 is introduced into the first data D0 of the first sensor 11, the effect of the noise on the estimation of the second data D1 can be reduced, thereby improving reliability. EXAMPLES

[0064] The second embodiment will be described with reference to Figures 7 to 10. In the following embodiments including this embodiment, the differences from the first embodiment will be mainly described. In this embodiment, a sensor to be selected as the first sensor 11 is selected from among a plurality of first sensor candidates using a physical causal database DB3 that defines the causal relationship of physical quantities.

[0065] Fig. 7 is a block diagram of an anomaly detection system 1a according to this embodiment. The anomaly detection system 1A has all the components of the anomaly detection system 1 described in Fig. 1, and further includes a first sensor selection unit 18 and a physical causal database DB3.

[0066] When the first sensor selection unit 18 receives the sensor list L0 that specifies the sensors necessary for grasping the state of the target machine 2, it refers to the physical causal database DB3 and selects the first sensor 11 from the group of sensors listed in the sensor list L0. The sensors not selected as the first sensor 11 from the group of sensors listed in the sensor list L0 become second sensors.

[0067] 8 shows a method for selecting the first sensor 11. The first sensor selection unit 18 compares the group of sensors listed in the sensor list L0 with the physical causal database DB3, and selects the first sensor 11 and the second sensor 12 that can most efficiently grasp the state of the target machine 2. Grasping the state of the target machine 2 most efficiently means grasping the state of the target machine 2 using the fewest number of sensors.

[0068] An example will be described below. Assume that the sensor list L0 lists a sensor for detecting the motor rotation speed, a temperature sensor, a vibration sensor, a sensor for detecting the motor acceleration, and a sensor for detecting the operating sound as the types of sensors required to grasp the state of the target machine 2. In other words, the sensor list L0 specifies the physical quantities required to grasp the state of the target machine 2.

[0069] The physical causal database DB3 indicates causal relationships between a plurality of physical quantities. For example, the motor rotation speed of a machine can affect the operating sound, but is not affected by the operating sound. By using such physical causal relationships, it is possible to select the first sensor 11 that can efficiently grasp the state of the target machine 2. The physical causal relationships may be manually set in advance by an administrator of the anomaly detection system 1A, or may be created semi-automatically or automatically by simulation software, etc.

[0070] The present embodiment thus configured also achieves the same effects as those of the first embodiment. Furthermore, in this embodiment, a minimum number of sensors can be determined based on a sensor list L0 indicating a group of sensors (physical quantities requiring measured values) required for grasping the state of the target machine 2. Therefore, the convenience of the user of the anomaly detection system 1 is improved.

[0071] A modified example of this embodiment will be described with reference to Figures 9 and 10. In this modified example, the first sensor 11 is selected from the group of sensors listed in the sensor list L0 without using the physical causal database DB3. As in the above, the sensor not selected as the first sensor 11 becomes the second sensor 12.

[0072] 9 is a block diagram of an anomaly detection system 1A1 according to this modified example. The anomaly detection system 1A1 includes a first sensor selection unit 18 but does not include a physical cause and effect database DB3.

[0073] 10 shows a method for the first sensor selection unit 18 to select the first sensor 11. The first sensor selection unit 18 selects one sensor from the group of sensors listed in the sensor list L0 as the first sensor candidate 11A, and sets the unselected sensor as the second sensor candidate 12A. Then, the anomaly detection system 1A1 calculates a restoration error between the observed value D0 and the restored value D3 of the first sensor candidate 11A, and an estimation error between the observed value of the second sensor candidate 12A and an estimated value estimated from a predetermined latent expression generated by the encoding unit 13.

[0074] The anomaly detection system 1A1 calculates the loss during learning while changing the first sensor candidate 11A, and selects as the first sensor 11 the first sensor candidate 11A that minimizes the loss.

[0075] This modified example configured in this manner also achieves the same effects as the present embodiment. Furthermore, since this modified example does not use the physical causal database DB3, it is not necessary to prepare the physical causal database DB3 in advance. Therefore, this modified example further improves the convenience for the user of the anomaly detection system 1A1. EXAMPLES

[0076] A third embodiment will be described with reference to Fig. 11 and Fig. 12. In the anomaly detection system 1B according to this embodiment, when detecting an anomaly in the target machine 2, the degree of anomaly is calculated using an estimation error of the second data D2 in addition to a recovery error of the first data D0.

[0077] Fig. 11 is a block diagram of an anomaly detection system 1B according to this embodiment. Fig. 12 shows a method for detecting an anomaly. In this embodiment, a first sensor 11 and a second sensor 12 are provided in a target machine 2, and the respective data are output as observed values ​​not only during learning of a machine learning model but also during operation.

[0078] As shown in the upper part of Figure 12, when learning the machine learning model, the anomaly detection system 1B determines parameters of the machine learning model so that the second data of the second sensor 12 can be estimated based on a predetermined latent expression among the latent expressions based on the first data D0 of the first sensor 11 (S301).

[0079] When the machine learning model is in operation, the target machine 2 is provided with a first sensor 11 and a second sensor 12, the first sensor 11 outputs first data D0 as an observation value, and the second sensor 12 outputs second data D1 as an observation value.

[0080] During operation of the machine learning model, the anomaly detection system 1B compares the second data D1 observed by the second sensor 12 with the second data D2 estimated from the first data D0, and calculates an estimation error (S302).

[0081] The anomaly detection system 1B calculates a restoration error between the first data D0 and data D0A restored from the first data D0 (S303), and determines the degree of anomaly of the target machine 2 from the estimated error and the restoration error (S304).

[0082] The present embodiment thus configured also achieves the same effects as the first embodiment. Furthermore, in this embodiment, since the degree of anomaly is calculated by including the estimated error in addition to the restoration error, the abnormality of the target machine 2 can be detected with higher accuracy than a method of calculating the degree of anomaly using only the restoration error. In this embodiment, the abnormality of the target machine 2 can be detected from the restoration error of the first data D0, and further, the cause of the abnormality can be estimated from the first data D0 (observed value) and the second data D1 (observed value). Furthermore, in this embodiment, the estimated error between the second data D1 (observed value) and the second data D1 (estimated value) can be used to perform highly accurate abnormality detection. EXAMPLES

[0083] A fourth embodiment will be described with reference to Figures 13 and 14. In this embodiment, when there is little data on the target machine 2, a learning model created from data on a machine of the same type as the target machine 2 is used as an initial model, and the initial model is fine-tuned using data obtained from the target machine 2.

[0084] 13 is a block diagram of an anomaly detection system 1C. The anomaly detection system 1C of this embodiment includes a pre-trained model database DB4. The pre-trained model database DB4 stores a pre-trained model M0 (see FIG. 14).

[0085] The pre-trained model M0 will be described with reference to Figure 14. The pre-trained model M0 is a learning model generated for a homogeneous machine, i.e., a machine that belongs to the same machine genre as the target machine 2 but has a different individual identification number.

[0086] The machine genre can be classified based on information including the machine type, vendor name, drive system, etc., such as "hydraulic press machine made by company A" and "electric die casting system made by company B." Alternatively, the machine genre may be classified based on the machine type and drive system. Furthermore, for example, in the case of a machine that is highly standardized in the industry, the machine genre may be classified based only on the machine type.

[0087] In the example of Fig. 14, two machine genres MG1 and MG2 are shown. For example, the machine genres MG1 and MG2 are products of the same vendor, but their drive systems are different, such as hydraulic and electric. Assume that the factory of a user who uses the anomaly detection system 1 is equipped with machines belonging to the machine genres MG1 and MG2.

[0088] The machine genre MG1 includes two identification numbers IN1 and IN2. Meanwhile, another machine genre MG2 includes two identification numbers IN3 and IN4. The target machine 2 is a machine with the identification number IN1 of the machine genre MG1.

[0089] The anomaly detection system 1 trains the learning model in the learning model database DB2 as described above based on the data measured by the sensors 11, 12 for the target machine 2 (MG1, IN1) (S401).

[0090] Here, if the target machine 2 has only recently been introduced and there is little measured data on the target machine 2, it takes time to generate a machine learning model with good accuracy. Therefore, the anomaly detection system 1 uses a machine learning model generated for a similar machine (MG1, IN2) that has a different identification number IN from the target machine 2 but belongs to the same machine genre MG1 as the target machine 2, as the pre-trained model M0 (S402).

[0091] The anomaly detection system 1 acquires the pre-trained model M0 stored in the pre-trained model database DB4, and uses the pre-trained model M0 as the initial value of the machine learning model to be used for monitoring the target machine 2 (S403).

[0092] Then, the anomaly detection system 1 fine-tunes the machine learning model (=initial value is the pre-trained model M0) using data measured on the target machine 2 by the first sensor 11 and the second sensor 12 (S404).

[0093] The present embodiment thus configured also achieves the same effects as those of the first embodiment. Furthermore, in this embodiment, when the amount of measured data D0 and D1 for the target machine 2 is small, a pre-trained model created for a machine of the same type as the target machine 2 is used as the initial value of the machine learning model, and fine tuning is performed using the measured data for the target machine 2, so that an appropriate machine learning model can be created in a shorter time. EXAMPLES

[0094] A fifth embodiment will be described with reference to Figures 15 and 16. In this embodiment, when a physical quantity measured for the target machine 2 changes continuously depending on the state of the machine 2, data between observed data is interpolated.

[0095] FIG. 15 is a block diagram of an anomaly detection system 1D of this embodiment. The anomaly detection system 1D includes a data extension unit 19 for extending the first data D0 detected by the first sensor 11. In this embodiment, both the first sensor 11 and the second sensor 12 can be used during learning of a machine learning model used to grasp the state of the target machine 2. During operation of the machine learning model, the first sensor 11 is usable, but the second sensor 12 may or may not be usable. That is, during operation, the second data of the second sensor 12 may be estimated based on the first data.

[0096] Fig. 16 shows a data expansion method. For example, a case will be described in which the first sensor 11 is a microphone that measures the operation sound of the target machine 2, and the operation sound of the target machine 2 changes according to continuously changing conditions such as the temperature, operation speed, and pressure of the target machine 2.

[0097] In this case, it is difficult to measure all of the operation sounds corresponding to the continuous changes of the target machine 2. Therefore, the anomaly detection system 1D generates extended data (interpolated data) D0E by using a data extension method such as Mix-up based on operation sound data D0 (operation sound at operating speed v100) measured in a certain state of the target machine 2 and operation sound data D0 (operation sound at operating speed v300) measured in another state of the target machine 2.

[0098] In this example, unmeasured data such as operation sound data at operating speed v101, operation sound data at operating speed v102, operation sound data at operating speed v103, ... operation sound data at operating speed v299 can be obtained in advance.

[0099] The present embodiment thus configured also achieves the same effects as those of the first embodiment. Furthermore, in this embodiment, even if the first data D0 outputted from the first sensor 11 changes in accordance with the continuous state change of the target machine 2, the data D0 of the first sensor 11 can be expanded and acquired in advance, and a machine learning model can be created at an early stage. EXAMPLES

[0100] A sixth embodiment will be described with reference to Fig. 17 to Fig. 22. In this embodiment, a loss function used in a machine learning model is optimized in accordance with the characteristics of a target machine 2, thereby improving the accuracy of anomaly detection and satisfying the required robustness.

[0101] 17 is a block diagram of an anomaly detection system 1E during learning. The anomaly detection system 1E can include, for example, a sensor 51, a feature extraction unit 52, a machine learning unit 53, a loss calculation unit 54, an anomaly determination unit 55, a waveform determination unit 56, a hyperparameter setting unit 57, a training data database DB1E, a learning model database DB2E, and a hyperparameter set database DB5.

[0102] Here, a sensor such as a microphone that measures the operation sound of the target machine 2 will be described as an example of the sensor 51. When measuring vibration instead of the operation sound, a vibration sensor may be used as the sensor 51.

[0103] The feature extraction unit 51 extracts a feature vector (hereinafter, feature) D51 from the data D50 measured by the sensor 11, and stores the extracted feature D51 in the training data database DB1E.

[0104] The machine learning unit 53 trains a machine learning model stored in the learning model database DB2E using the features D51 extracted by the feature extraction unit 52.

[0105] The loss calculation unit 54 calculates the loss between the observed value of the sensor 51 and the restored value restored by the machine learning model. The magnitude of this loss makes it possible to determine the degree to which the target machine 2 is abnormal, so it can also be called an abnormality degree calculation unit 54.

[0106] The abnormality determination unit 55 determines whether the target machine 2 is abnormal, based on the data D53 (loss or abnormality degree) output from the loss calculation unit .

[0107] The waveform determination unit 56 determines whether the waveform pattern of the operation sound corresponds to any of the pre-registered waveform patterns based on the extracted feature amount D51. The hyper parameter setting unit 57 selects one hyper parameter set by referring to the hyper parameter set database DB5 based on the determination result of the waveform determination unit 56, and sets the selected hyper parameter set D52 in the loss calculation unit 54.

[0108] FIG. 18 is a hardware configuration diagram when the anomaly detection system 1E is configured using a computer 50. As shown in FIG.

[0109] The computer 50 includes, for example, a control unit 501, a memory unit 502, a storage 503, and an input / output device (I / O device) 504, and is connected to a communication network CN. A storage medium MM can also be connected to the computer 50, and computer programs or data can also be transmitted and received between the storage medium MM and the memory 502 or storage 503.

[0110] When a predetermined computer program stored in storage 503 is loaded into memory 502 and executed by a control unit (calculation unit) 501, a feature extraction unit 52, a machine learning unit 53, a loss calculation unit 54, and an abnormality determination unit 55 are realized. A part or all of the predetermined computer program can be stored in a storage medium MM.

[0111] FIG. 19 shows a method for setting a hyper-parameter set in the loss calculation unit 54 according to the characteristics of the target machine 2 and the required robustness.

[0112] The waveform determination unit 56 selects a waveform pattern or the like corresponding to the machine type based on a table T1 indicating a combination of the machine type and the waveform pattern or the like. The waveform pattern or the like indicates the tendency of the waveform pattern and noise for each target device 2. For example, a fan outputs a steady operating sound and hardly generates an unexpected sound (noise). A pump outputs a steady operating sound, but an unexpected sound may occur from time to time. A slider generates different sounds when it is operating and when it is stopped, so steady operating sounds and non-steady operating sounds are mixed. A valve operates from time to time, so its operating sound is non-steady. The waveform determination unit 56 may determine which machine type the target machine 2 corresponds to from the feature amount D51 of the operating sound of the target machine 2, or may obtain a signal indicating the type of the target machine 2 and determine it. For example, when the anomaly detection system 1E is installed in a factory, the type of the target machine 2 may be manually set.

[0113] The hyper parameter setting unit 57 refers to the hyper parameter set database DB5 based on the determination result by the waveform determining unit 56 (which waveform pattern or the like it corresponds to) to identify the hyper parameter set to be set in the loss calculation unit 54.

[0114] The hyper parameter set database DB5 manages, for example, a set number, a waveform pattern or the like, a required robustness, a hyper parameter, the number of samples, and a weight of a loss function in association with each other.

[0115] The required robustness indicates the robustness required for anomaly judgment of the target machine 2. Examples of the required robustness include "high", "low", "medium low", and "medium high". As shown in Fig. 20, the hyperparameters include, for example, the standard deviation δ and the mean value μ of the probability distribution function used by the loss calculation unit 54.

[0116] The hyper-parameter setting unit 57 selects a hyper-parameter set corresponding to the determined waveform pattern, etc., and sets the hyper-parameters and weights of the loss function included in the selected hyper-parameter set in the loss calculation unit 54.

[0117] 20, when the operating sound of a fan or the like is stable and there is little possibility of noise contamination, the required robustness is not required to be high, so the required robustness is set low. Therefore, in this case, a hyper-parameter set (Set1) is selected so that the range determined as normal by the loss calculation unit 54 is narrowed.

[0118] In contrast, in the case of a machine that operates irregularly and outputs a non-stationary operating sound, such as a valve, the required robustness is high. This is because the operating sound of a valve is not a stable pattern, and so lowering the required robustness increases the possibility of erroneously determining that the valve is abnormal. In this case, a hyper-parameter set is selected (Set4) so ​​that the range in which the loss calculation unit 54 determines that something is normal is broadened. The hyper-parameter set used when the target machine 2 is a pump or slider is an intermediate value between the range in which the loss calculation unit 54 determines that something is normal between the range in which the loss calculation unit 54 determines that something is normal for a fan and the range in which the loss calculation unit 54 determines that something is normal for a valve.

[0119] 21 is a flowchart showing processing S11 during learning. The anomaly detection system 1E acquires data D50 measured by the sensor 51 (S1101), and stores the acquired data D50 in the storage 503 (S1102). The anomaly detection system 1E reads the data D50 from the storage 503 into the memory 502 (S1103), and causes the feature extraction unit 52 to extract features D51 (S1104).

[0120] The abnormality detection system 1E causes the waveform determination unit 56 to determine the waveform pattern, etc. of the feature amount D51 (S1105), and causes the hyper parameter setting unit 57 to select a hyper parameter set according to the waveform pattern, etc. (S1106).

[0121] The loss calculation unit 54 calculates the loss of the machine learning model using the hyper-parameter set set by the hyper-parameter setting unit 57 (S1108).

[0122] The anomaly detection system 1E repeatedly learns the parameters of the machine learning model so that the calculated loss value is minimized (S1109 to S1111). The parameters of the machine learning model are stored in the learning database DB2E (S1112).

[0123] That is, the anomaly detection system 1E determines whether a predetermined convergence condition is satisfied or whether the number of iterations C1 of this process exceeds the upper limit value ThC (S1108). If the convergence condition is not satisfied or if the number of iterations C1 is equal to or less than the upper limit value ThC, the machine learning unit 53 updates the parameters of the machine learning model (S1109), calculates the convergence condition (S110), increments the number of iterations C1 by one, and returns to step S1108.

[0124] When a predetermined convergence condition is satisfied (S1108: YES), the anomaly detection system 1E stores the parameters of the machine learning model in the learning database DB2E (S112).

[0125] 22 is a flowchart showing a process S21 for detecting an abnormality in the target machine 2. The abnormality detection system 1E reads and sets parameters of a machine learning model from the learning model database DBE2 (S2101).

[0126] The anomaly detection system 1E acquires data D50 from the sensor 51 (S2102) and stores the data in the storage 503 (S2103).

[0127] The anomaly detection system 1E reads the data D50 stored in the storage 503 into the memory 502 (S2104), and causes the feature extraction unit 52 to extract the feature D51 (S2105).

[0128] The abnormality detection system 1E causes the waveform determination unit 56 to determine the waveform pattern, etc. (S2106), and causes the hyper parameter setting unit 57 to set a hyper parameter set according to the determination result in the loss calculation unit 54 (S2107).

[0129] The anomaly detection system 1E causes the loss calculation unit 54 to calculate the degree of anomaly using a method according to the properties of the target machine 2 (S2108). The anomaly detection system 1E causes the anomaly determination unit 55 to determine whether the calculated degree of anomaly is greater than a predetermined threshold value (S2109).

[0130] If the calculated abnormality degree is greater than the predetermined threshold (S2109: YES), the abnormality determination unit 54 determines that the target machine 2 is in an abnormal state (S2110). On the other hand, if the calculated abnormality degree is equal to or less than the predetermined threshold (S2109: NO), the abnormality determination unit 54 determines that the target machine 2 is in a normal state (S2111).

[0131] According to this embodiment configured as above, a hyper-parameter set selected according to the type of the target machine 2 can be set in the loss calculation unit 54, so that abnormalities in the target machine 2 can be detected more appropriately.

[0132] In this embodiment, robustness is determined according to the degree of change in the physical quantity measured for the target machine 2 (whether it is steady or non-steady, whether noise is present, etc.), so that abnormalities can be detected according to the characteristics of the target machine 2.

[0133] This embodiment may be combined with the first to fifth embodiments. For example, the first and second sensors may be used in this embodiment as in the first embodiment. Then, a part of the latent expression obtained from the first sensor may be used to estimate the value of the second sensor.

[0134] The present invention is not limited to the above-described embodiments, and includes various modified examples. For example, the above-described embodiments have been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to those including all of the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration.

[0135] The components of the present invention may be selected arbitrarily, and the invention having the selected components is also included in the present invention. Furthermore, the components described in the claims may be combined in combinations other than those explicitly stated in the claims. [Explanation of symbols]

[0136] 1: Anomaly detection system, 11: Concept classification 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: Learning model likelihood calculation unit, 22: Learned model likelihood calculation unit, 31: Peer concept classification unit, M: Equipment

Claims

1. An anomaly detection system that detects an anomaly of a target machine by a computer, the computer includes a communication unit that acquires first data from a first sensor attached to the target machine and acquires second data from a second sensor attached to the target machine, a calculation unit, and a storage unit; The calculation unit is an encoding unit trained to generate latent representations including a predetermined latent representation that estimates the second data based on the first data; a decoder trained to recover the first data from the latent representation; an anomaly detection unit that detects an anomaly of the target machine based on a restoration error between the first data and the first data restored by the decoding unit; Equipped with the storage unit stores a sensor correspondence relationship management unit in which the second sensor that outputs the second data that can be estimated from the first data is previously associated with each of the first sensor candidates; The calculation unit further includes a first sensor selection unit that selects one of the first sensor candidates based on the sensor correspondence relationship management unit. Anomaly detection system.

2. the second sensor is temporarily attached to the target machine in a learning mode to obtain the predetermined latent representation; The calculation unit further includes a state monitoring unit that monitors a state of the target machine based on the first data and the second data estimated from the predetermined latent expression. The anomaly detection system according to claim 1 .

3. The anomaly detection unit detects an anomaly of the target machine based on a restoration error between the first data observed by the first sensor and the first data restored by the decoding unit, and an estimation error between the second data observed by the second sensor and the second data estimated from the predetermined latent representation. The anomaly detection system according to claim 1 .

4. The storage unit stores a pre-learned model that is trained in advance based on first data from the first sensor and second data from the second sensor attached to a device of the same type as the target machine but having a different individual identification number, The calculation unit uses the pre-learning model as an initial model of a learning model for adjusting parameters of the encoding unit and parameters of the decoding unit. The anomaly detection system according to claim 1 .

5. The calculation unit further includes a data complementation unit that interpolates the first data from the first sensor. The anomaly detection system according to claim 1 .

Images