JP7666203B2 - Anomaly detection device, anomaly detection method, and anomaly detection program - Google Patents

Description

The present invention relates to an anomaly detection device, an anomaly detection method, and an anomaly detection program.

Conventionally, anomaly detection technology has been proposed that predicts failures based on waveform data from sensors installed in production equipment. Generally, anomaly detection does not collect enough abnormal data, and detecting unknown anomalies is an issue. Therefore, conventional technology has proposed machine learning methods that learn only normal data and detect anomalies.

Generally, production equipment can operate under multiple operating conditions even in a normal state. Therefore, Patent Document 1 proposes a method for detecting anomalies that takes multiple operating conditions into account. Meanwhile, Patent Document 2 proposes a method for determining anomalies by inputting a measured waveform into a model that has been trained on only normal data, outputting a generated waveform, and comparing the measured waveform with the generated waveform.

JP 2021-51698 A JP 2021-9441 A

However, the technology of Patent Document 1 has problems with detection accuracy because it judges an anomaly to be present when all inference results of each learning model prepared for each operating condition exceed all thresholds. In addition, although the method extracts features of the generated waveform to detect an anomaly, it is not easy to design effective features when there is a lack of abnormal data. In contrast, the technology of Patent Document 2 does not require feature design because it compares waveforms. However, the input to the autoencoder proposed in Patent Document 2 is only a single-variable explanatory variable, and multivariate cases in which many parameters are used are not assumed. Furthermore, both Patent Documents 1 and 2 only notify the degree of anomaly of a single variable as an indicator of an anomaly, which causes a problem of poor explanatory power for anomaly judgment.

The present invention has been made to solve the above problems, and aims to provide an anomaly detection device, an anomaly detection method, and an anomaly detection program that does not require designing the features of the waveforms to be handled, and that is highly interpretable in response to waveforms of multiple parameters.

The present invention comprises an acquisition unit that acquires a measured waveform, which is time-series data relating to multiple parameters of a target machine; a waveform generation unit that generates a generated waveform, which is time-series data relating to the multiple parameters, by inputting the measured waveform; a deviation calculation unit that calculates a deviation by comparing the measured waveform with the generated waveform; and a determination unit that determines that the measured waveform is abnormal if the deviation exceeds a predetermined threshold value, and the waveform generation unit has a trained learning model that has been learned based on the measured waveform when the target machine is normal.

In the above anomaly detection device, the learning model can be learned based on the measurement waveforms obtained under normal operating conditions of the target machine.

In the above anomaly detection device, the learning model can be generated by unsupervised learning.

In the above anomaly detection device, the learning model can be constructed using a neural network.

In the above anomaly detection device, the neural network includes an encoding unit and a generating unit, the encoding unit receives the measured waveform as an input and outputs a latent variable, the generating unit receives the latent variable as an input and outputs the generated waveform, and the encoding unit and the generator can be configured using a convolutional neural network.

In the above anomaly detection device, the learning model can be configured, for example, using an Efficient GAN.

In the above anomaly detection device, the deviation can be a single scalar value.

In the above anomaly detection device, the deviation calculation unit can be configured to calculate an individual deviation of the time series for each of the parameters.

The anomaly detection device may further include a display unit that displays at least one of the measured waveform, the generated waveform, the deviation, and the individual deviation of the time series for each parameter.

The anomaly determination method according to the present invention includes the steps of acquiring a measured waveform, which is time-series data relating to a plurality of parameters of a target machine; generating a generated waveform, which is time-series data relating to the plurality of parameters, by inputting the measured waveform; calculating a deviation by comparing the measured waveform with the generated waveform; and determining that the measured waveform is abnormal if the deviation exceeds a specified threshold value. The step of generating the generated waveform is performed by a trained learning model that has been trained based on the measured waveform of the target machine when it is normal.

The abnormality determination program of the present invention causes a computer to execute the steps of acquiring a measured waveform, which is time-series data on multiple parameters of a target machine, generating a generated waveform, which is time-series data on the multiple parameters, by inputting the measured waveform, calculating a deviation by comparing the measured waveform with the generated waveform, and determining that the measured waveform is abnormal if the deviation exceeds a specified threshold value. The step of generating the generated waveform is performed by a trained learning model that has been learned based on the measured waveform of the target machine when it is normal.

The present invention provides an anomaly detection device, an anomaly detection method, and an anomaly detection program that does not require the design of features of the waveforms being handled and is highly interpretable in response to waveforms of multiple parameters.

1 is a configuration example of an embodiment of a system including an anomaly detection device of the present invention. FIG. 2 is a block diagram showing a hardware configuration of the anomaly detection device. FIG. 2 is a block diagram showing the hardware configuration of the learning device. FIG. 2 is a block diagram showing the software configuration of the learning device. 13 is a graph showing an overlay of measured and generated waveforms under different operating conditions. FIG. 1 is a diagram illustrating the correspondence between an encoding unit, a generating unit, and a CNN. FIG. 2 is a block diagram showing a software configuration of the anomaly detection device. 13 is a graph showing a graph in which measured waveforms and generated waveforms under normal conditions are superimposed for three parameters, and a graph showing deviation time series data for each parameter. 13 is a graph showing a graph in which measured waveforms and generated waveforms are superimposed upon each other when an abnormality occurs, and a graph showing deviation time-series data of each parameter for three parameters. FIG. 13 is a diagram showing rankings of deviation degrees. FIG. 13 is a diagram illustrating a determination of normality or abnormality based on the relationship between the deviation and a threshold value. 4 is a flowchart showing a learning process of the learning device. 4 is a flowchart showing an anomaly detection method in the anomaly detection device.

One embodiment of the anomaly detection device according to the present invention will be described below with reference to the drawings. As an example, an anomaly detection device that detects anomalies in a ball screw mechanism of production equipment will be described below. Figure 1 shows an example of the configuration of a system including this anomaly detection device, which includes an anomaly detection device 1, a learning device 2 that generates a learning model used in this anomaly detection device 1, and a conveying device 3 that includes a ball screw mechanism. Each component will be described in detail below.

<1. Hardware configuration>
<1-1. Conveyor device>
1, the transport device 3 includes a ball screw 31, a nut 32 screwed onto the ball screw 31, and a stage 33 for a workpiece W provided on the nut 32. A servo motor 34 is connected to the ball screw 31, and the ball screw 31 is rotated by the servo motor 34. A servo driver 35 that controls the servo motor 34 is also connected to the servo motor 34. In this embodiment, the rotation speed, output torque, and stage position of the servo motor 34, which are controlled by the servo driver 35, are used as parameters for abnormality detection, and abnormalities are judged using the time-series waveforms of these.

<1-2. Anomaly detection device>
Next, an example of the hardware configuration of the anomaly detection device 1 according to the present embodiment will be described. Fig. 2 shows an example of the hardware configuration of the anomaly detection device according to the present embodiment.

This anomaly detection device 1 is a computer in which a control unit 11, a memory unit 12, an external interface 13, and a communication interface 14 are electrically connected, and can be configured, for example, as a PLC. Note that in FIG. 2, the external interface 13 and the communication interface 14 are written as "external I/F" and "communication I/F." This also applies to the various devices described below.

The control unit 11 includes a CPU, RAM, ROM, etc., and is configured to execute various information processing based on programs and data. The storage unit 12 is configured, for example, to be configured as an auxiliary storage device such as an HDD or SSD, and stores an anomaly detection program 121, learning result data 122, and various data for determining anomalies and for driving the device. The anomaly detection program 121 is a program for causing the anomaly detection device 1 to determine whether or not there is an anomaly in the measured waveform. The control unit 11 is configured to execute the processing of each step described below by interpreting and executing this anomaly detection program 121.

The external interface 13 is an interface for connecting to an external device, and is configured appropriately according to the external device to be connected. In this embodiment, the external interface 13 is connected to a display (display unit) 4. This display 4 is used, for example, to display the results of anomaly determinations. The display 4 is not particularly limited, and a known liquid crystal display, touch panel display, etc. can be used. In addition, input devices, output devices, etc. can be connected to the external interface 13 as appropriate.

The communication interface 14 is, for example, a wired LAN (Local Area Network) module, a wireless LAN module, etc., and is an interface for performing wired or wireless communication. In other words, the communication interface 14 is an example of a communication unit configured to communicate with other devices. In this embodiment, it is connected to the above-mentioned learning device 2 and transport device 3 via a network.

Regarding the specific hardware configuration of the anomaly detection device 1, components can be omitted, replaced, or added as appropriate depending on the embodiment. For example, the control unit 11 may include multiple processors. Furthermore, the control unit 11 may be configured by an FPGA. The storage unit 12 may be configured by RAM and ROM included in the control unit 11. In addition to a PLC, the anomaly detection device 1 can use an information processing device designed specifically for the service to be provided. Furthermore, depending on the target for which an anomaly is to be determined, it may be a general-purpose desktop PC, tablet PC, etc. The above points are the same for the learning device 2 described later.

<1-3. Learning device>
Next, an example of the hardware configuration of the learning device 2 according to this embodiment will be described. Fig. 3 shows an example of the hardware configuration of the learning device according to this embodiment.

As shown in FIG. 3, the learning device 2 is a computer to which a control unit 21, a memory unit 22, an external I/F 23, and a communication I/F 24 are electrically connected, similar to the anomaly detection device 1.

The configuration of the control unit 21 and the memory unit 22 is the same as that of the anomaly detection device. The memory unit stores a learning program 221, a learning model 222, first learning result data 223, second learning result data 224, and various data for driving the learning device. The learning program 221 is a program for learning a learning model 222 for outputting a generated waveform from a measured waveform. The control unit 21 is configured to interpret and execute this learning program 221, thereby executing the processing of each step described below.

The rest of the hardware configuration is the same as that of the anomaly detection device 1, so it will not be described here.

2. Software configuration
<2-1. Learning device>

The learning device 2 learns the learning model 222 using measured waveforms, which are time-series data relating to the rotational speed, output torque, and stage position of the servo motor 34 transmitted from the servo driver 35 of the transport device 3. In particular, this learning model 222 is learned using normal measured waveforms when the transport device 3 is operating normally. The software configuration of the learning device 2 will be described below with reference to Figure 4. Figure 4 is a block diagram showing the software configuration of the learning device.

<2-1-1. Learning model>
As shown in Figure 4, when the control unit 21 of the learning device 2 expands the learning program 221 stored in the memory unit 22 into RAM, the control unit 21 interprets and executes the learning program 221 using the CPU, and functions as a computer equipped with an acquisition unit 211, an encoding unit 212, a generation unit 213, a discrimination unit 214, and a storage processing unit 215.

As described above, the acquisition unit 211 acquires a measurement waveform, which is time-series data relating to the rotation speed, output torque, and stage position of the servo motor 34. This measurement waveform is transmitted after preprocessing by the anomaly detection device 1, as described below. However, it may also be transmitted directly from the servo driver 35, and preprocessing may be performed by the acquisition unit 211, as described below.

The acquisition unit 211 acquires measurement waveforms not only under the same operating conditions, but also under different operating conditions when the conveying device 3 is operated. For example, a measurement waveform when a light workpiece is conveyed (operating condition 1) and a measurement waveform when a heavy workpiece is conveyed (operating condition 2) are input. Examples of measurement waveforms under each operating condition are shown in FIG. 5. In this example, it can be seen that when a heavy workpiece W is conveyed, the speed decreases and the torque increases. In addition to the above, there are various other different operating conditions, such as when different processed parts are used (for example, when the stage is replaced due to a change in setup), and when the device operates at different environmental temperatures (for example, when the temperature changes between morning and afternoon).

Next, the encoding unit 212, the generation unit 213, and the discrimination unit 214 will be described. The encoding unit 212, the generation unit 213, and the discrimination unit 214 form a learning model using an Efficient Generative Adversarial Network (Efficient GAN). The encoding unit 212 and the generation unit 213 are formed, for example, of a convolutional neural network (CNN). This will be described later. The discrimination unit 214 is also formed of a similar CNN. However, the configuration of the discrimination unit 214 may be different from that of the encoding unit 212 and the generation unit 213.

When learning is performed, a first training step for training the discrimination unit 214 and a second training step for training the encoding unit 212 and the generation unit 213 are alternately performed. In the first training step, the discrimination unit 214 is trained to discriminate whether a generated waveform input to the discrimination unit 214 is a generated waveform generated by the generation unit 213 or a measured waveform transmitted from the acquisition unit 211. In other words, the discrimination unit 214 is trained to discriminate whether the input waveform originates from the acquisition unit 211 or the generation unit 213. In addition, in the second training step, the encoding unit 212 and the generation unit 213 are trained to generate a waveform that causes the discrimination unit 214 to make an erroneous discrimination. Note that here, originating from the acquisition unit 211 is expressed as "true", and originating from the generation unit 213 is expressed as "false". However, the method of expressing each origin is not limited to such an example, and may be appropriately selected according to the embodiment.

To be more specific, in the first training step, a measured waveform is input to the encoding unit 212. The encoding unit 212 outputs a latent variable z, and when this latent variable z is input to the generation unit 213, the generation unit 213 outputs a generated waveform that is approximate to the measured waveform input to the encoding unit 212.

This generated waveform is input to the discrimination unit 214, where arithmetic processing is performed. As a result, an output value corresponding to the result of discriminating whether the generated waveform originates from the generation unit 213 or the acquisition unit 211 is output. In this case, since the generated waveform input to the discrimination unit 214 is the generated waveform generated by the generation unit 213, the discrimination unit 214's correct answer is "false". Then, the error between the output value and this correct answer is calculated. In this way, the error between the output value and this correct answer is calculated for each generated waveform.

Similarly, the measured waveform is input to the discrimination unit 214, and the discrimination unit 214 executes calculation processing. As a result, the discrimination unit 214 outputs an output value corresponding to the result of determining whether the input waveform originates from the generation unit 213 or the acquisition unit 211. In this case, since the input waveform is a measured waveform, the discrimination unit 214 correctly determines that it is "true." In this way, the error between the output value and this correct answer is calculated for each waveform.

Then, the parameter values of the CNN in the discrimination unit 214 are adjusted so that the sum of the calculated errors becomes smaller. The adjustment of the parameter values of the discrimination unit 214 is repeated by the above series of processes until the sum of the errors between the output value and the correct answer of true or false becomes equal to or less than the threshold value. As a result, in the first training step, the discrimination unit 214 is trained by backpropagation or the like so as to discriminate whether the waveform is a generated waveform generated by the generation unit 213 or a measured waveform.

On the other hand, in the second training step, the measured waveform is input to the encoding unit 212. The encoding unit 212 outputs a latent variable z, and when this latent variable z is input to the generation unit 213, the generation unit 213 outputs a generated waveform that is approximate to the measured waveform input to the encoding unit 212.

The generated generated waveform is then input to the discrimination unit 214, which executes calculation processing. As a result, an output value corresponding to the result of determining whether the input generated waveform originates from the generation unit 213 or the acquisition unit 211 is output. In this training of the generation unit 213, the correct answer is to cause the discrimination result by the discrimination unit 214 to be erroneous. In other words, it is sufficient that the generated waveform generated by the generation unit 213 is close enough to be indistinguishable from the measured waveform, and the correct answer is that the obtained output value corresponds to the "true". For the combination of the measured waveform and the generated waveform, the error between the output value obtained by the series of processes and this correct answer (i.e., the "true") is calculated.

Then, the parameter values of the CNN in the encoding unit 212 and the generating unit 213 are adjusted so that the sum of the calculated errors becomes smaller. The parameter values of the encoding unit 212 and the generating unit 213 are repeatedly adjusted by the above series of processes until the sum of the errors between the output value obtained by the series of processes and the "true" for each combination of measured waveform and generated waveform becomes equal to or less than a threshold value. As a result, in the second training step, the generating unit 213 (and the encoding unit 212) is trained by backpropagation or the like so as to generate a generated waveform that will cause the discrimination unit 214 to make an incorrect discrimination.

In this way, by alternately performing the first training step and the second training step, the accuracy of the discrimination unit 214 and the generation unit 213 (and the encoding unit 212) is alternately improved. This allows the generation unit 213 to acquire the ability to appropriately generate a generated waveform that is approximately the same as the measured waveform.

After this learning is completed, the storage processing unit 215 generates first learning result data 223 and second learning result data 224, which respectively indicate the configurations of the constructed encoding unit 212 and generating unit 213 and the discrimination unit 214 (e.g., the number of layers of the neural network, the number of neurons in each layer, the connection relationships between neurons, and the transfer function of each neuron) and the calculation parameters (e.g., the weights of the connections between each neuron, the threshold value of each neuron). Then, the storage processing unit 215 stores the generated first and second learning result data 223, 224 in the memory unit 22.

In this way, when a measured waveform is input to the learned encoding unit 212 and generation unit 213, a generated waveform that approximates the measured waveform can be generated.

<2-1-2. CNN>
Next, an example of a CNN constituting the encoding unit 212 and the generating unit 213 will be described with reference to Fig. 6. This CNN is composed of a one-dimensional CNN, specifically, the encoding unit 212 is composed of a one-dimensional CNN, and the generating unit 213 is composed of a one-dimensional transposed CNN. Fig. 6 is a diagram showing the correspondence between the encoding unit and the generating unit and the CNN. The measurement waveform input to the encoding unit 212 is generated by the pre-processing unit 112 of the anomaly detection device 1.

This point will be explained first. The preprocessing unit 112 performs preprocessing on the collected measurement waveform. As preprocessing, the following three processes can be performed, for example. First, in the transport device 3, the waveform during the period when the stage 33 moves from the start point to the end point through the section (frame) is extracted and used for judgment. In addition, the value of the measurement waveform is normalized so that it is a numerical value between 0 and 1. Furthermore, a linear conversion is performed so that the frame length of the time series is 64.

As shown in FIG. 6, in this CNN, data represented by (C, T) with the frame length converted to 64 for the three parameters mentioned above is input to the encoding unit 212. Here, C is the number of dimensions of the variable, and T is the frame length of the data. The data input to the encoding unit 212 is (3, 64) because the number of parameters is 3 and the frame length is 64. In other words, 3*64=192 pieces of data are input. Therefore, as shown in FIG. 6, the data for each layer of the CNN is represented by (C, T).

In addition, in this CNN, the encoding unit 212 performs calculations to reduce the data frame length of the data in four layers to (32, 32), (64, 16), and (128, 8), and generates 10 pieces of data corresponding to the latent variable z in the fully connected layer. In other words, this latent variable z reduces the dimension of the measured waveform and shows the characteristics of the measured waveform in low dimensions. By expressing the measured waveform in low dimensions in this way, learning can be easily performed. After that, the data corresponding to this latent variable z is input to the generation unit 213, and calculations are performed to expand the reduced data frame length of the data in five layers to (512, 4), (256, 8), and (128, 16), and finally generates data of (3, 64) corresponding to the measured waveform, that is, the generated waveform.

Note that this CNN is just one example, and a normal neural network can also be used.

<2-2. Anomaly detection device>
Next, a software configuration of the anomaly detection device 1 will be described. Fig. 7 is a block diagram showing the software configuration of the anomaly detection device. As shown in Fig. 7, when the control unit 11 of the anomaly detection device 1 loads the anomaly detection program 121 stored in the storage unit 12 into the RAM, the control unit 11 interprets and executes the anomaly detection program 121 by the CPU, and functions as a computer including a collection unit 111, a preprocessing unit 112, a waveform generation unit 113, a deviation calculation unit 114, a determination unit 115, and an output unit 116.

The collection unit 111 collects measurement waveforms, which are time-series data relating to the rotation speed, output torque, and stage position of the servo motor 34 transmitted from the servo driver 35 of the transport device 3. The collected measurement waveforms are stored in the memory unit 12 or RAM.

The preprocessing unit 112 performs preprocessing on the collected measurement waveform. This processing is as described above. The preprocessed measurement waveform is then transmitted to the learning device 2 and used for the learning described above.

As described above, the waveform generating unit 113 is composed of a trained encoding unit 212 and a generating unit 213 to which the first learning result data 223 generated by the learning device 2 has been introduced.

The deviation calculation unit 114 compares the measured waveform with the generated waveform and calculates the deviation. As deviations, an individual deviation and an overall deviation are calculated. The calculation of the individual deviations is performed as follows. First, for each of the above three parameters, deviation time series data is calculated by calculating the difference between the measured waveform and the generated waveform shown in the time series data in a time series. For example, the difference is calculated for each of 64 data pieces, which is the frame length, and deviation time series data consisting of the data of the 64 differences is calculated. Next, the 64 data pieces are summed to calculate the individual deviation. In other words, the deviation for each of the three parameters can be expressed.

Figure 8 shows a graph of the measured waveforms and generated waveforms superimposed under normal conditions for three parameters, and a graph showing the deviation time series data for each parameter. Meanwhile, Figure 9 shows a graph of the measured waveforms and generated waveforms superimposed under abnormal conditions for three parameters, and a graph showing the deviation time series data for each parameter. In the deviation time series data under normal conditions shown in Figure 8, the deviation is roughly 0 for all parameters at any time period. Note that in the graph of the measured waveforms and generated waveforms superimposed in Figure 8, the two are nearly the same, so they appear as a single line. Meanwhile, in the deviation time series data under abnormal conditions shown in Figure 9, it can be seen that the deviation changes. For example, it can be seen that the deviation time series data for torque changes significantly at the start of transport (circled area).

Also, based on the individual deviations, a ranking of deviations as shown in FIG. 10 can be generated. On the other hand, the overall deviation is a scalar value obtained by summing up the individual deviations. In this way, the individual deviations and the overall deviations are calculated for each combination of measured waveforms and generated waveforms.

The judgment unit 115 judges that there is an abnormality when the overall deviation exceeds a specified threshold, and judges that there is a normality when it is equal to or less than the threshold. The waveform generation unit 113 includes a learning model that is trained based on the measured waveform under normal conditions, so when a measured waveform under normal conditions is input, a generated waveform that is close to the measured waveform is generated. On the other hand, when a measured waveform under abnormal conditions is input, a generated waveform different from the generated waveform under normal conditions is generated, resulting in a large deviation. Therefore, when the overall deviation exceeds a specified threshold, it is judged to be an abnormality. FIG. 11 is an example. It can be seen that when a normal measured waveform is input, an overall deviation lower than the threshold is calculated, whereas when an abnormal measured waveform is input, an overall deviation higher than the threshold is calculated.

The output unit 116 outputs at least one of the measured waveform, the generated waveform, the deviation time series data, the individual deviation, and the overall deviation, and displays it on the display 4. In addition, when an abnormality occurs, an alert can be displayed on the display 4, a sound can be generated, or other output can be performed.

3. Learning process in the learning device
Next, the learning process of the learning device configured as above will be described with reference to the flowchart of FIG.

As shown in FIG. 12, first, the conveying device 3 is operated (step S11), and normal measurement waveforms are collected for the above-mentioned three parameters (step S12). Next, preprocessing is performed on the measurement waveform (step S13). Next, the preprocessed measurement waveform is input to the learning model 222 to generate the above-mentioned first and second learning result data 223, 224 (step S14). Following this, a threshold is set (step S15). The setting of the threshold is not particularly limited, but for example, the normal distribution of the above-mentioned overall deviation degree is calculated, and the threshold can be set to 3σ or more, for example, 3σ, 4σ, etc. Note that the threshold can be set automatically in the learning device 2, or can be set in the anomaly detection device 1. Alternatively, it can be arbitrarily set by the user. After that, the first learning result data 223 is transmitted to the anomaly detection device 1 and incorporated into the waveform generation unit 113 (step S16). The first learning result data 223 is created periodically and can be incorporated into the waveform generating unit 113 each time. This completes the learning process in the learning device 2.

4. Anomaly detection method using an anomaly detection device
Next, an anomaly detection method in the anomaly detection device configured as above will be described with reference to the flowchart of FIG.

As shown in FIG. 13, first, while the transport device 3 is in operation, the collection unit 111 collects measured waveforms for the three parameters described above (step S21). Next, the preprocessing unit 112 preprocesses the measured waveform (step S22). Next, the preprocessed measured waveform is input to the waveform generation unit 113 incorporating the first learning result data 223, and a generated waveform is generated (step S23). Following this, the deviation calculation unit 114 compares the calculated generated waveform with the measured waveform and calculates the overall deviation (step S24). Finally, if the judgment unit 115 determines that the overall deviation is less than the threshold (YES in step S25), the operation of the transport device is continued as normal. On the other hand, if the overall deviation is equal to or greater than the threshold (NO in step S25), an abnormality is determined to have occurred, and an abnormality alert is displayed on the display 4 (step S26).

<5. Features>
As described above, according to this embodiment, the waveform generator 113 generates a generated waveform, which is time series data relating to a plurality of parameters, from a measured waveform during normal operation, which is time series data relating to the plurality of parameters measured by the transport device 3, and the measured waveform is compared with the generated waveform to calculate a deviation, and if this deviation exceeds a predetermined threshold value, it is determined that an abnormality has occurred. Therefore, the following effects can be obtained.

(1) It is possible to detect abnormalities with high accuracy in response to a plurality of operating conditions.
(2) Since the deviation time series data and individual deviations can be calculated, it is possible to check, for example, at what time the deviation becomes large for each parameter. This improves the interpretability of the abnormality determination.
(3) Since the measured waveform that has been subjected to simple preprocessing is input and the generated waveform is output, there is no need to design the waveform features.

6. Modifications
Although the embodiment of the present invention has been described above, the above description is merely an example of the present invention in every respect, and it goes without saying that various improvements and modifications can be made without departing from the scope of the present invention. For example, the following modifications are possible. In the following, the same reference numerals are used for the same components as in the above embodiment, and the description of the same points as in the above embodiment is omitted as appropriate. In addition, the following modifications can be combined as appropriate.

<6-1>
In the above embodiment, the learning model is based on GAN, but is not limited thereto, and may be based on neural networks such as autoencoder, GANomality, Skip GANomality, etc. In other words, any method may be used as long as it can learn multivariate and multidimensional data such as time-series data related to multiple parameters as described above.

<6-2>
The method of calculating the degree of deviation for performing an anomaly determination in deviation calculation section 114 is not particularly limited as long as it is possible to calculate the difference between the measured waveform and the generated waveform as the degree of deviation that can be evaluated by determination section 115.

<6-3>
In the above embodiment, the anomaly detection device 1 and the learning device 2 are configured as separate computers, but they may be configured as one computer. That is, a learning function may be added to the anomaly detection device 1.

<6-4>
The method of pre-processing the measured waveform is not particularly limited and can be changed as appropriate depending on the learning model used. Alternatively, learning can be performed or a generated waveform can be generated without pre-processing.

<6-5>
The abnormality detection device 1 of the above embodiment determines an abnormality in a conveying device having a ball screw mechanism, but is not limited to this and can be applied to determine abnormalities in various target machines.

REFERENCE SIGNS LIST 1 Anomaly detection device 11 Acquisition unit 13 Waveform generation unit 14 Deviation calculation unit 15 Determination unit 2 Learning device 212 Encoding unit 213 Generation unit 214 Discrimination unit 222 Learning model 3 Conveyor device (target machine)
4. Display (display unit)

Claims (11)

an acquisition unit that acquires a measurement waveform, which is a set of time series data relating to each of a plurality of parameters of a target machine;
a waveform generating unit that receives the measured waveform and generates a generated waveform, which is a set of time-series data relating to each of the plurality of parameters;
a deviation calculation unit that calculates a deviation by comparing the measured waveform with the generated waveform;
a determination unit that determines that the measured waveform is abnormal when the deviation exceeds a predetermined threshold value;
Equipped with
An anomaly detection device, wherein the waveform generation unit is equipped with a trained learning model that has been learned based on the measured waveform of the target machine under normal conditions. The anomaly detection device according to claim 1, wherein the learning model is trained based on the measured waveforms obtained under normal operating conditions of the target machine. The anomaly detection device according to claim 1 or 2, wherein the learning model is generated by unsupervised learning. The anomaly detection device according to claim 3, wherein the learning model is configured by a neural network. The neural network includes an encoding unit and a generating unit,
The encoding unit receives the measured waveform as an input and outputs a latent variable,
The generating unit receives the latent variables as an input and the generated waveform as an output,
The anomaly detection device according to claim 4 , wherein the encoding unit and the generating unit are configured with a convolutional neural network. The anomaly detection device according to claim 4 or 5, wherein the learning model is configured using an efficient GAN. The anomaly detection device according to any one of claims 1 to 6, wherein the deviation is a single scalar value. The anomaly detection device according to any one of claims 1 to 7, wherein the deviation calculation unit is configured to calculate an individual deviation of the time series for each of the parameters. The anomaly detection device according to any one of claims 1 to 8, further comprising a display unit that displays at least one of the measured waveform, the generated waveform, the deviation, and an individual deviation of the time series for each parameter. acquiring a measurement waveform which is a set of time series data relating to each of a plurality of parameters of a target machine;
generating a generated waveform, which is a set of time-series data relating to each of the plurality of parameters, by inputting the measured waveform;
calculating a deviation by comparing the measured waveform with the generated waveform;
determining that the measured waveform is abnormal when the deviation exceeds a specified threshold value;
Equipped with
An anomaly detection method, in which the step of generating the generated waveform is performed by a trained learning model that has been learned based on the measured waveform of the target machine under normal conditions. On the computer,
acquiring a measurement waveform which is a set of time series data relating to each of a plurality of parameters of a target machine;
generating a generated waveform, which is a set of time-series data relating to each of the plurality of parameters, by inputting the measured waveform;
calculating a deviation by comparing the measured waveform with the generated waveform;
determining that the measured waveform is abnormal when the deviation exceeds a specified threshold value;
Run the command,
An anomaly detection program, wherein the step of generating the generated waveform is performed by a trained learning model that has been learned based on the measured waveform of the target machine under normal conditions.