Generative adversarial network-based part anomaly detection method and apparatus therefor

Abstract

A generative adversarial network (GAN)-based part anomaly detection method, comprising: by means of a denoising auto-encoder, deriving a first anomaly score of input data, comprising: extracting manifold coordinates of the input data, and calculating a distance between the manifold coordinates and average coordinates of normal data so as to use same as the first anomaly score; by means of a discriminator, deriving a second anomaly score of the input data; and adding the first anomaly score and the second anomaly score up, so as to calculate an anomaly degree of the input data. The present disclosure further relates to an apparatus used for part anomaly detection. (FIG. 5)

Description

Parts anomaly detection method and device based on generative adversarial network Technical Field

[0001] The present disclosure relates to a method and device for detecting anomalies, and in particular to a method and device for detecting anomalies in parts based on a generative adversarial network (GAN). Background Art

[0002] In the process of detecting whether machine parts are abnormal, the following problems are encountered:

[0003] (1) Since most of the data that can be collected in industrial applications are generated under normal operation, the amount of positive / abnormal data is uneven.

[0004] (2) Due to changes in production factors such as work orders, molds, and materials, a continuous production process will include multiple working conditions. Different working conditions will make it difficult for the AI ​​model to distinguish between normal and abnormal conditions.

[0005] (3) The training data for supervised algorithms requires clear labels (e.g., the type of factor causing the anomaly). In industrial applications, many anomalies rarely occur or have never occurred. Unsupervised algorithms are more adaptable, but some unsupervised algorithms cannot adapt to various working conditions.

[0006] When there is an imbalance in the amount of positive / abnormal data, the AI ​​model training performance will be poor due to the lack of available abnormal data. In addition, because machine parts have various working conditions, most algorithms are unable to effectively identify actual abnormal conditions.

[0007] Therefore, a method and device for detecting abnormal parts based on a generative adversarial network are needed to improve the above problems.

[0008] Summary of the Invention

[0009] Therefore, the main purpose of the present invention is to provide a method and device for detecting part anomalies based on a generative adversarial network.

[0010] The present disclosure provides a part anomaly detection method for use in an industrial automation system, comprising: inferring a first anomaly score for input data using a denoising autoencoder, including: extracting a manifold coordinate of the input data and calculating a distance between the manifold coordinate and an average coordinate of normal data as the first anomaly score; generating a second anomaly score for the input data using a discriminator; and adding the first anomaly score and the second anomaly score to calculate an anomaly level of the input data.

[0011] The present disclosure provides a part anomaly detection device, comprising: a processor; and a computer storage medium coupled to the processor and configured to store computer-readable instructions for instructing the processor to execute the part anomaly detection method.

[0012] Based on the Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), this paper replaces the WGAN-GP generator with a denoising autoencoder while retaining the WGAN-GP discriminator to form an AI model architecture. The disclosed method and device for detecting abnormal parts based on a GAN have the following advantages:

[0013] (1) The denoising autoencoder can extract the manifold coordinates of the input data. The higher the dimensionality and the more diverse the input data, the lower the probability of overlap of the manifold coordinates, resulting in higher discrimination. Therefore, in industrial automation applications, when machine parts face complex operating conditions, the disclosed method and apparatus for detecting anomalies in parts based on a generative adversarial network exhibits high discrimination.

[0014] (2) The AI ​​model architecture disclosed in this paper uses both denoising autoencoders and discriminators to perform inference, which can improve the discrimination. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] FIG1 is a schematic diagram of a parts anomaly detection system according to an embodiment of the present disclosure.

[0016] FIG2 is a schematic diagram of a model training system according to an embodiment of the present disclosure.

[0017] FIG3 is a schematic diagram of an inference system according to an embodiment of the present disclosure.

[0018] FIG4 is a flowchart of a method for training an AI model according to an embodiment of the present disclosure.

[0019] FIG5 is a flowchart of a method for detecting anomalies according to an embodiment of the present disclosure.

[0020] FIG6 is a data characteristic coordinate diagram of manifold coordinates according to an embodiment of the present disclosure.

[0021] FIG. 7 is a schematic diagram of an electronic device according to an embodiment of the present disclosure.

[0022] Explanation of the reference numerals: 10: Part anomaly detection system 110: Belt 112: Drive wheel 12: Variable frequency drive 13: Controller 20: Model training system 21: Data preprocessing module 22: Model training module 23: Noise reduction autoencoder 24: Discriminator 25: Noise generator 26: First objective function 27: Second objective function 28: Data collection module 30: Inference system 31: Data preprocessing module 32: AI model 33: Noise reduction autoencoder 34: Discriminator 35: Numerical converter 36: Subtractor 37: Adder 40: Training AI model method S41-S48: Step 50: Part anomaly detection method S51, S52, S53, S11, S12, S13, S14: Step 60: Data characteristic coordinate diagram 70: Electronic device 71: Memory 72: Processor 73: Display element 74: I / O port 75: Peripheral element 76: Power supply 77: Bus 78: Program code DETAILED DESCRIPTION

[0023] The present disclosure provides a method and apparatus for detecting abnormal parts based on a generative adversarial network (GAN). In industrial automation applications where abnormal data is difficult to obtain, a GAN-based AI model architecture is trained using normal training data to address the previously difficult problem of distinguishing normal from abnormal states in various operating situations.

[0024] 1 is a schematic diagram of a part anomaly detection system 10 according to an embodiment of the present disclosure. The part anomaly detection system 10 includes a belt 110 , a plurality of drive pulleys 112 , a variable frequency drive (VFD) 12 , and a controller 13 , wherein the controller 13 may be equipped with an inference system 30 .

[0025] As shown in Figure 1, a belt 110 is wound around a plurality of drive pulleys 112. Drive pulleys 112 are connected to and controlled by a motor (not shown), which in turn is connected to and controlled by a variable frequency drive 12. As the motor continuously rotates and drives the belt 110, the controller 13 collects component parameters (e.g., motor current, motor torque, motor speed, and operating cycle) through the variable frequency drive 12 to generate a corresponding data set. The inference system 30 is configured to perform online inference based on the input real-time data to predict the degree of component abnormality.

[0026] Generally speaking, the inference system 30 must be trained before it can infer the degree of abnormality of a part. In this embodiment, the inference system 30 is trained by the controller 13, wherein the inference system 30 receives part parameters to generate a data set and performs training based on the data set. In one embodiment, the controller 13 outputs the data set to an external device, and the inference system 30 is trained by the external device (such as the electronic device 70 in Figure 7), and then the external device outputs the trained inference system 30 to the controller 13 for online inference. In another embodiment, the controller 13 outputs the data set to an external device, and the inference system 30 is trained by the external device, and online inference is performed in the external device. The advantage of training the inference system 30 by an external device is that a higher performance computer or processor can be selected to speed up the training process. The inference system 30 can be stored in a built-in or external memory of the controller 13 or the external device.

[0027] In one embodiment, the controller 13 may be, for example, a central processing unit (CPU), a micro control unit (MCU), a system on chip (SoC), or a programmable logic controller (PLC), but is not limited thereto.

[0028] FIG2 is a schematic diagram of a model training system 20 according to an embodiment of the present disclosure. Model training system 20 includes a data collection module 28, a data preprocessing module 21, and a model training module 22. Model training system 20 can be installed in controller 13 or an external training device (e.g., electronic device 70 in FIG7 ).

[0029] Before model training begins, data collection module 28 must collect relevant historical data for model training. For example, for motor operation, motor current, motor torque, motor speed, and operating cycles can be collected as historical data. In one embodiment, the relevant historical data is numerical data or images. A portion of the historical data can be used as training data, while another portion can be used as test data.

[0030] The data preprocessing module 21 is configured to perform statistical analysis on the training and test data to obtain statistical values. Next, the data preprocessing module 21 is configured to select multiple influential features from the statistical values ​​to generate a feature set for subsequent model training and testing. Specifically, during model training and online inference, the data preprocessing module 21 is configured to delete parameters and data that do not match the feature set and retain parameters and data that do match the feature set, allowing this data to serve as input to the model training module 22 and the inference system 30.

[0031] The model training module 22 includes a denoising auto-encoder (DAE) 23, a discriminator 24, a noise generator 25, a first objective function 26, and a second objective function 27, wherein the denoising auto-encoder 23 and the discriminator 24 conform to a deep neural network (DNN) model architecture.

[0032] In operation, the noise generator 25 adds noise to normal data to generate noisy data, and then outputs the noisy data to the denoising autoencoder 23. The denoising autoencoder 23 is an initial denoising autoencoder configured to convert the noisy data into generated data, and input the generated data to the discriminator 24.

[0033] Discriminator 24 is an initial discriminator, configured to generate a discriminant value Dg for generated data and a discriminant value Dn for real data, respectively, which are fed back to the denoising autoencoder 23. The discriminant values ​​Dg and Dn are arbitrary real numbers representing the degree to which the input data is real. Our experimental results show that these real numbers range between 1 and -1. The closer the real number is to 1, the more likely the input data is to be considered real data; the closer the real number is to -1, the more likely the input data is to be considered generated data. Based on the discriminant values ​​Dg and Dn, discriminator 24 calculates a first expected value and adjusts its parameters to maximize the first objective function 26. Through repeated training, the performance of discriminator 24 can be gradually optimized. Based on the fed-back discriminant values ​​Dg and Dn, denoising autoencoder 23 calculates a second expected value and adjusts its parameters, then continues training to minimize the second objective function 27. Through repeated training, the performance of denoising autoencoder 23 can be gradually optimized.

[0034] Furthermore, the generated data and the real data (i.e., normal data) are input to the discriminator 24 to generate the discriminant values ​​Dg and Dn, and the first and second expected values ​​are calculated, so that the denoising autoencoder 23 and the discriminator 24 adjust the parameters according to their respective objective functions. The formula of the first objective function 26 of the discriminator 24 is as follows:

[0035] The formula of the second objective function 27 of the denoising autoencoder 23 is as follows:

[0036] Pdata and PDAE are the probability distributions of real data (i.e. normal data) and generated data, respectively. are the expected values ​​of the real data and the generated data, respectively. The first objective function 26 aims to maximize the difference between the discriminant values ​​Dg and Dn, which means that the discriminator 24 is better able to distinguish between real and synthetic data. The second objective function 27 aims to minimize the difference between the discriminant values ​​Dg and Dn, which means that the generated data is closer to the real data, resulting in the discriminator 24 being unable to effectively distinguish. Furthermore, a penalty term is added to address the problem of GANs often failing to converge during training.

[0037] Specifically, the discriminator 24 refers to the first objective function max D To adjust the parameters, after multiple parameter iterations, the discriminator 24 can be gradually optimized; the denoising autoencoder 23 refers to the second objective function min DAE To adjust the parameters, the denoising autoencoder 23 can be gradually optimized after multiple parameter iterations. The parameter adjustment process will continuously repeat the following steps, including: (1) fixing the parameters of the denoising autoencoder 23 and adjusting the parameters of the discriminator 24 until the discrimination gap is maximized; (2) fixing the parameters of the discriminator 24 and adjusting the parameters of the denoising autoencoder 23 until the discrimination gap is minimized.

[0038] FIG3 is a schematic diagram of an inference system 30 according to an embodiment of the present disclosure. Inference system 30 includes a data preprocessing module 31 and an AI model 32, wherein AI model 32 includes a denoising autoencoder 33, a discriminator 34, a digital converter 35, a subtractor 36, and an adder 37. Inference system 30 can be installed in controller 13 or an external detection device (e.g., electronic device 70 in FIG7 ).

[0039] In operation, the inference system 30 obtains real-time data on the operation of machine parts through the controller 13. Taking motor operation as an example, motor current, motor torque, motor speed, and operating cycle can be collected as real-time data. The data preprocessing module 31 preprocesses the real-time data 310 to generate input data for the denoising autoencoder 33 and the discriminator 34. Both the denoising autoencoder 33 and the discriminator 34 are trained as shown in Figure 2. The denoising autoencoder 33 is configured to extract the manifold coordinates of the input data. The subtractor 36 then calculates the distance between the manifold coordinates and the average coordinates of the normal data as a first anomaly score. In one embodiment, the dimension of the manifold coordinates corresponds to the number of nodes in the hidden layer of the denoising autoencoder. The discriminator 34 infers a discriminant value D(x) for the real-time data, where the discriminant value D(x) is an arbitrary real number. A larger discriminant value D(x) indicates a greater similarity between the input data and the normal data; a smaller discriminant value D(x) indicates a lower similarity between the input data and the normal data. The identification value D(x) can be expressed by the following formula: D(x) = W x x + b

[0040] Where D(x) represents the judgment result of the discriminator 34 on the input data x, W refers to the weight matrix of the last layer of the neural network, and b is the bias. The bias introduces an additional parameter into the discriminator 34, allowing the discriminator 34 to better fit the input data and improve performance. The bias b is typically initialized to a small constant. The discriminator 34 finally transforms the judgment result through an activation function to obtain a discriminant value, where the activation function can be a linear function.

[0041] The numerical converter 35 is configured to convert the identification value D(x) into a second anomaly score. To ensure that the magnitude of the identification value D(x) can intuitively represent the similarity between the input data and normal data, in one embodiment, the numerical converter 35 multiplies the identification value D(x) by -1 to convert it into the second anomaly score. In another embodiment, the numerical converter 35 subtracts the identification value D(x) from a predetermined value to convert it into the second anomaly score. In another embodiment, when the identification value D(x) is greater than zero and less than 1, the numerical converter 35 takes the inverse of the identification value D(x) as the second anomaly score. Finally, the adder 37 adds the first anomaly score and the second anomaly score to calculate the degree of anomaly of the input data. A larger value for the degree of anomaly indicates a greater degree of anomaly between the input data and normal data; conversely, a smaller value for the degree of anomaly indicates a smaller degree of anomaly between the input data and normal data.

[0042] It is worth noting that although the examples of abnormality detection in Figures 1-3 use motors as an example, the present disclosure should not be limited to this application. For example, the present disclosure can also be applied to computer numerical control (CNC) milling machine processes. Values ​​such as tool conditions, feed rates, and clamping pressure collected from CNC machines can be used to determine whether the wear state of a part is normal or abnormal.

[0043] FIG4 is a flowchart of a method 40 for training an AI model according to an embodiment of the present disclosure. The method 40 for training an AI model may be executed by the controller 13 of FIG1 or the electronic device 70 of FIG7 and includes the following steps.

[0044] Step S41: inferring identification values ​​of multiple sets of input data through the discriminator, wherein the multiple sets of input data include real data and generated data.

[0045] Step S42: Utilize the identification values ​​of the real data and the generated data to obtain a first expected value and calculate a first objective function.

[0046] Step S43: Determine whether the first objective function is maximized. If so, proceed to step S45; if not, proceed to step S44.

[0047] Step S44: Fix the denoising autoencoder parameters and adjust the discriminator parameters. Return to step S41.

[0048] Step S45: The discriminator is used to infer the discriminant values ​​of the real data and the generated data, and the discriminant values ​​are fed back to the denoising autoencoder to calculate the second expected value.

[0049] Step S46: Obtain a second expected value and calculate a second objective function.

[0050] Step S47: Determine whether the second objective function is minimized. If so, end; if not, proceed to step S48.

[0051] Step S48: Fix the discriminator parameters and adjust the parameters of the denoising autoencoder. Return to step S45.

[0052] Steps S41, S42, S43, S44, and S45 can be performed by the discriminator 24, while steps S45, S46, S47, and S48 can be performed by the denoising autoencoder 23. Steps S42 and S46 can be performed by the first and second objective functions. In one embodiment, after step S48, the controller 13 or processor can determine whether to return to step S41 for training again. By repeatedly executing the AI ​​model training method 40, the performance of the denoising autoencoder 23 and discriminator 24 in Figure 2 can be gradually optimized for application to the denoising autoencoder 33 and discriminator 34 in the AI ​​model 32 in Figure 3.

[0053] FIG5 is a flow chart of a method 50 for detecting anomalies in a part based on a generative adversarial network according to an embodiment of the present disclosure. The method 50 can be executed by the controller 13 of FIG1 or the electronic device 70 of FIG7 and includes the following steps.

[0054] Step S51: Inferring a first anomaly score for the input data using a denoising autoencoder. In this embodiment, step S51 includes step S11: extracting manifold coordinates of the input data; and step S12: calculating the distance between the manifold coordinates and the average coordinates of normal data as the first anomaly score.

[0055] Step S52: Inferring a second anomaly score of the input data through the discriminator. In this embodiment, step S52 includes step S13: Inferring an identification value of the input data through the discriminator; and step S14: Converting the identification value into a second anomaly score.

[0056] Step S53: Add the first anomaly score and the second anomaly score to calculate the anomaly level of the input data. In one embodiment, when the anomaly level is higher than a custom threshold, the controller 13 or the electronic device 70 determines that the input data is abnormal and issues an alarm signal to notify the user.

[0057] Steps S51 and S11 may be executed by the denoising autoencoder 33 , step S12 may be executed by the subtractor 36 , steps S52 and S13 may be executed by the discriminator 34 , step S14 may be executed by the digital converter 35 , and step S53 may be executed by the adder 37 .

[0058] Before the part anomaly detection method 50 begins, the controller 13 may first obtain real-time data and perform preprocessing on the real-time data (including statistical analysis, deletion of non-matching feature set data, and retention of matching feature set data) to generate input data. By executing the part anomaly detection method 50, the controller 13 or the electronic device 70 can perform online inference to determine the degree of part anomaly and thereby determine whether the input data is abnormal.

[0059] Figure 6 shows a data characteristic coordinate diagram 60 of manifold coordinates according to an embodiment of the present disclosure. The historical data, training data, test data, and input data used in the present disclosure can be numerical data or images. Figure 6 uses images as an example to illustrate that different types of images are clustered in low-dimensional manifold spaces within a high-dimensional space, and that the manifold coordinates of different types of images rarely overlap. In other words, the higher the dissimilarity and the higher the dimensionality, the lower the probability of overlap, resulting in higher discrimination.

[0060] As described above, the disclosed method and apparatus for detecting abnormal parts only need to train the denoising autoencoder and the discriminator with normal training data to solve the problem that the amount of abnormal data is too small and cannot be distinguished under various working conditions.

[0061] It should be noted that the denoising autoencoders 23, 33 and the discriminators 24, 34 in Figures 2 and 3 are configured to be implemented in hardware, software, firmware, or any combination thereof. For example, the denoising autoencoders 23, 33 and the discriminators 24, 34 may be implemented as computer program code configured to be executed on one or more processors. Alternatively, the denoising autoencoders 23, 33 and the discriminators 24, 34 may be implemented as hardware logic / circuitry.

[0062] FIG7 is a schematic diagram of an electronic device 70 according to an embodiment of the present invention.

[0063] The present invention may be implemented in the form of computer program code or machine-usable instructions. The instructions may be computer-executable instructions in the form of program modules, which are executed by a computer or other machine, such as a personal digital assistant or other portable device. Generally speaking, program modules include routines, programs, objects, components, data structures, etc. A program module is a program code that performs a specific task or implements a specific abstract data type. The present invention may be implemented in a variety of system configurations, including portable devices, consumer electronics, general-purpose computers, and more specialized computing devices. The present invention may also be implemented in a distributed computing environment, processing devices connected by a communication network.

[0064] 7 . Electronic device 70 includes a bus 77 that directly or indirectly couples the following devices: memory 71 , one or more processors 72 , one or more display elements 73 , input / output (I / O) ports 74 , peripheral components 75 , and an illustrative power supply 76 . Bus 77 represents an element that can be one or more buses (e.g., an address bus, a data bus, or a combination thereof).

[0065] The memory 71 includes but is not limited to random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical disc storage device, magnetic disk, magnetic disk, magnetic disk storage device or other magnetic storage device, or any other media that can be used to store the required information and can be accessed by the electronic device 70.

[0066] The electronic device 70 includes a processor that reads data from various entities such as a memory 71 or peripheral components 75. The display component 73 displays information or instructions to a user or other devices, such as a display device, a speaker, a printing component, a vibration component, etc.

[0067] The I / O ports 74 allow the electronic device 70 to be logically connected to other devices including peripheral components 75, some of which are built-in devices. Exemplary components include microphones, joysticks, game consoles, satellite dishes, scanners, printers, wireless devices, etc. The peripheral components 75 can provide a user interface for processing gestures, sounds, or other physiological inputs generated by the user. In some examples, these inputs can be transmitted to a suitable network element for further processing. The electronic device 70 can be equipped with a depth camera, such as a stereo camera system, an infrared camera system, an RGB camera system, and combinations of these systems, to detect and identify objects.

[0068] In addition, the processor 72 in the electronic device 70 may also execute the program code 78 and instructions in the memory 71 to present the actions and steps of the above embodiments, or other descriptions in the specification.

[0069] The present disclosure has been disclosed as above by way of implementation examples. Anyone skilled in the art may make slight changes and modifications without departing from the concept and scope of the present disclosure. Therefore, the scope of protection of the present disclosure shall be determined by the claims.

Claims

1. A method for detecting abnormal parts based on a generative adversarial network, comprising the steps of: 1) Inferring a first anomaly score of an input data through a denoising autoencoder, comprising: 11) extracting a manifold coordinate of the above input data; as well as 12) calculating a distance between the manifold coordinates and an average coordinate of normal data as the first anomaly score; 2) inferring a second anomaly score of the input data through a discriminator; and 3) Adding the first anomaly score and the second anomaly score to calculate an anomaly level of the input data.

2. The method for detecting abnormal parts according to claim 1, wherein step 2) comprises: 21) inferring a first identification value of the input data through the identification device; and 22) Converting the first identification value into the second anomaly score.

3. The method for detecting abnormal parts according to claim 2, wherein step 22) comprises one of the following steps: Multiplying the first identification value by -1 to convert it into the second anomaly score, wherein the identification value is an arbitrary real number; Subtracting the first identification value from a preset value to convert it into the second anomaly score; and When the first identification value is greater than zero and less than 1, the reciprocal of the first identification value is taken as the second abnormality score.

4. The method for detecting abnormal parts according to claim 1, wherein before step 1), it further comprises: Preprocessing real-time data to generate processed data; Deleting the processed data that does not match a feature set; as well as The processed data matching the feature set is retained as the input data.

5. The method for detecting abnormal parts according to claim 1, wherein before step 1), a training process is further performed, comprising: A1) inferring second identification values ​​of a plurality of sets of input data through an initial identification device, wherein the plurality of sets of input data include at least one real data and at least one generated data; A2) calculating a first expected value according to the second identification value by the initial identification device; A3) inputting the first expected value into a first objective function to determine whether the first objective function is maximized; A4) when the first objective function has not been maximized, fixing the parameters of the initial denoising autoencoder, and adjusting the parameters of the initial discriminator according to the first expected value; and A5) Repeat steps A1) to A4) to train the initial discriminator until the first objective function is maximized, thereby generating the discriminator.

6. The method for detecting abnormal parts according to claim 5, wherein the training process further comprises the steps of: A6) inferring a third identification value between the generated data and the at least one real data through the initial identification device, and feeding it back to the denoising autoencoder; A7) calculating a second expected value according to the third identification value by the denoising autoencoder; A8) inputting the second expected value into a second objective function to determine whether the second objective function is minimized; A9) when the second objective function has not been minimized, fixing the parameters of the discriminator, and adjusting the parameters of the initial denoising autoencoder according to the second expected value; and A10) Repeat steps A6) to A9) to train the initial denoising autoencoder until the second objective function is minimized, thereby generating the denoising autoencoder.

7. The method for detecting abnormal parts according to claim 5, wherein step A1) comprises: adding noise to the normal data to generate noisy data; and The noisy data is converted into the generated data through the initial discriminator, and the generated data is input into the discriminator.

8. The method for detecting abnormal parts according to claim 5, wherein before the training process, a pre-processing process is further performed, including: B1) selecting a plurality of influential features from the statistical values ​​of a training data set by a feature selection method to generate a feature set; as well as B2) Selecting data and parameters matching the above feature set as a training data set and providing it to the model training stage.

9. The method for detecting abnormal parts according to claim 1, further comprising: 4) When the abnormality level is higher than a custom threshold, the input data is determined to be abnormal.

10. A device for detecting abnormality of a part, comprising: A processor; as well as A computer storage medium is coupled to the processor and is configured to store computer-readable instructions for instructing the processor to execute the part abnormality detection method as claimed in claim 1.

Images