A method and apparatus for coordinating attention mechanism and auxiliary classification adversarial generative network applied to bearing fault diagnosis

By integrating a coordinated attention mechanism with an adversarial generative network for auxiliary classification, a CA-ACGAN model was established. This model solves the problems of noise interference and data imbalance in traditional bearing fault diagnosis methods under real working conditions, and achieves efficient bearing fault identification and classification.

CN118673360BActive Publication Date: 2026-11-17HUAQIAO UNIVERSITY
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410683101.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-29
Publication Date
2026-11-17
Estimated Expiration
2044-05-29

AI Technical Summary

Technical Problem

Traditional bearing fault diagnosis methods face problems such as noise interference, unbalanced data categories, and inconsistent data distribution under real-world conditions, resulting in poor performance when the models are applied to actual situations.

Method used

A CA-ACGAN model is established by integrating a coordinated attention mechanism with an auxiliary classification adversarial generative network. Through the interaction between the improved coordinated attention mechanism and the generator and discriminator, bearing fault features are generated and identified.

Benefits of technology

It improves the accuracy of bearing fault diagnosis and noise immunity, and achieves efficient fault identification and classification.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118673360B_ABST
    Figure CN118673360B_ABST
Patent Text Reader

Abstract

This invention relates to a method and apparatus for bearing fault diagnosis using a coordinated attention mechanism and an auxiliary classification generative adversarial network (ACGAN). The method integrates the coordinated attention mechanism (CA) with the auxiliary classification generative adversarial network (ACGAN) to establish a CA-ACGAN model for bearing fault diagnosis. The process includes: acquiring raw one-dimensional bearing data; converting the raw one-dimensional bearing data into a two-dimensional time-frequency graph using continuous wavelet transform as the real data; improving the coordinated attention mechanism and integrating it with the generator; feeding random noise along with labels into the generator to generate fake data; feeding real and fake data together into a discriminator for training; determining the algorithm's termination condition; repeating the steps until Nash equilibrium is reached; extracting the trained discriminator, feeding the test set into the discriminator for testing, and outputting the accuracy and classification results. This invention not only improves upon the shortcomings of the original algorithm but also exhibits excellent noise resistance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of rolling bearing fault diagnosis, and in particular, it can be used to identify bearing fault types in industrial production by converting fault signals into two-dimensional time-domain graphs. Specifically, it relates to a coordinated attention mechanism and auxiliary classification adversarial generative network method and device for bearing fault diagnosis. Background Technology

[0002] With the rapid development of industrial big data and IoT technologies, mechanized production is becoming increasingly widespread. Against this backdrop, machinery and equipment are gradually evolving towards automation, intelligence, and complexity. In industrial production, predictive and health management (PHM) has become particularly important. Especially the operating status and fault prediction of rotating machinery has always been one of the core research directions in fault diagnosis of industrial production equipment. Among these, rolling bearings, as a key component, directly affect the stable operation of the equipment due to their health status.

[0003] In actual factory production operations, high-speed rotating machinery often faces harsh working environments, and rolling bearings bear multiple loads. Prolonged wear, impacts, and other disturbances can lead to various forms of defects and failures in rolling bearings, resulting in significant economic losses for industrial production. Statistics show that over half of the failures of large machinery originate from problems in the operation of their bearing components. Compared to other components, bearings are more fragile, more unstable during their operating cycle, and prone to wear and eventual breakage, thus requiring regular inspection and replacement. Therefore, research on rolling bearing fault diagnosis methods is of great significance for ensuring the stable and safe operation of rotating machinery.

[0004] Traditional bearing fault diagnosis methods typically combine signal processing techniques and machine learning. First, the bearing's vibration signals need to be acquired, then processed to extract features. Finally, machine learning techniques are used to classify and identify the extracted features. This process mainly includes feature extraction and fault identification. However, in real-world operating conditions, traditional machine learning models face a series of challenges. Unlike ideal laboratory environments, real-world data often contains various noise interferences, imbalanced data categories, and even inconsistent data sample distributions. Therefore, applying these models to practical situations may encounter various problems. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings and limitations of the commonly used Auxiliary Classifier Generative Adversarial Network (ACGAN), which, although a powerful generative network, still has some drawbacks, such as training instability, lack of visual consistency in generated samples, and blurry and unclear generated samples. Therefore, this invention provides a method and device for applying a coordinated attention mechanism and an Auxiliary Classifier Generative Adversarial Network to bearing fault diagnosis. It integrates the Coordinate Attention mechanism (CA) with ACGAN to establish a CA-ACGAN model, utilizing the complementary advantages of both to solve the bearing fault diagnosis problem. This invention not only improves the shortcomings of the original algorithm but also exhibits excellent performance in noise resistance.

[0006] To achieve the above objectives, the technical solution of the present invention is: a method for coordinating attention mechanism and auxiliary classification adversarial generative network applied to bearing fault diagnosis, which integrates the coordinating attention mechanism (CA) and the auxiliary classification adversarial generative network (ACGAN) to establish a CA-ACGAN model to realize bearing fault diagnosis.

[0007] In one embodiment of the present invention, the method includes the following steps:

[0008] S1. Collect raw one-dimensional bearing data;

[0009] S2. Convert the original one-dimensional bearing data into a two-dimensional time-frequency graph using continuous wavelet transform as the real data.

[0010] S3. Improve the coordination attention mechanism and integrate it with the generator;

[0011] S4. Random noise and labels are fed into the generator to generate fake data;

[0012] S5. Feed real data and fake data together into the discriminator for training;

[0013] S6. Determine the algorithm termination condition;

[0014] S7. Repeat steps S2-S5 until Nash equilibrium is reached;

[0015] S8. Extract the trained discriminator, feed the test set into the discriminator for testing, and output the accuracy and classification results.

[0016] In one embodiment of the present invention, in step S1, the original one-dimensional bearing data adopts the CWRU rolling bearing dataset, and the vibration signal of the drive end bearing under load 0 condition is collected at a sampling frequency of 12kHz.

[0017] In one embodiment of the present invention, in step S2, the vibration signal of the drive end bearing in the original one-dimensional bearing data is divided into 600 data points as a segment, with 300 overlapping points between adjacent segments, to generate a two-dimensional time-frequency diagram of the state of each drive end bearing.

[0018] In one embodiment of the present invention, step S3 is specifically implemented as follows:

[0019] (1) Coordinate information embedding

[0020] The global pooling factor specified in the following formula is decomposed into a one-dimensional feature encoding:

[0021]

[0022] The above equation z c The output of the c-th channel in the SE module is the result of compression. The input x comes directly from a convolutional layer with a fixed kernel size. Given the input x, each channel is encoded along the horizontal and vertical axes using two spatial ranges (H,1) or (1,W) of the pooling kernel, respectively. Therefore, the output of the c-th channel at height h is represented as:

[0023]

[0024] Similarly, the c-th channel output with width w is shown in the following equation:

[0025]

[0026] The two transformations aggregate features along two spatial directions respectively, producing a pair of direction-aware feature maps;

[0027] (2) Coordinating attention generation

[0028] Given feature maps embedded in the formula using coordinate information, we first concatenate them, then feed them into a shared 1×1 convolutional transform function F1, resulting in the following equation:

[0029] f=δ(F1([z h ,z w ]))

[0030] [z h ,z w ] represents a splicing operation along a spatial dimension, δ is a non-linear activation function, and f∈R C / r×(H×W) Intermediate feature maps encoding spatial information in the horizontal and vertical directions are generated; where r is the reduction ratio of the control block size in the SE module. Then, f is divided into two tensors along the spatial dimension. h ∈R C / r×H and f w ∈R C / r×WAdditionally, two 1×1 convolution transformations F are used. h and F w f respectively h and f w Converting it to a tensor with the same channel number as the input X, and then activating it with the sigmoid function, yields the output shown in the following equation:

[0031] g h =σ(F h (f h ))

[0032] g w =σ(F w (f w ))

[0033] In the formula, σ is the sigmoid function. To reduce the complexity of the model, the scale r is reduced to decrease the number of channels in f, and then the output g is... h and g w After being expanded and used as weights for attention, the final expression for the output y is shown below:

[0034]

[0035] An improved coordination attention mechanism is achieved by using two 1×1 convolutional transformations F. h and F w f respectively h and f w Before converting to a tensor with the same channel number as the input x, a Hard-Swish activation function is added, and the structure of the coordinated attention mechanism is improved to adapt to the ACGAN network structure, thereby increasing the generator network's ability to capture features.

[0036] In one embodiment of the present invention, the Hard-Swish activation function is an improved version of the Swish function, and the formula is as follows:

[0037]

[0038] In one embodiment of the present invention, in step S4, the two-dimensional time-frequency image generated in step S2 is divided into a 7:3 ratio, and the CA-ACGAN model is built and trained in the Python-based PyTorch deep learning framework. The number of training rounds is set to 1750, and the learning rate of both the generator and discriminator networks is set to 0.0002. Randomly generated noise and labels are fed into the generator to generate fake data.

[0039] In one embodiment of the present invention, in step S7, if the probability of identifying the input data as true or false is 50% for both cases, it indicates that Nash equilibrium has been reached and the model training is complete.

[0040] The present invention also provides a coordinated attention mechanism and an auxiliary classification adversarial generative network device for bearing fault diagnosis, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the steps of the method described above.

[0041] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0042] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a CA-ACGAN model that integrates the Coordinate Attention mechanism (CA) and ACGAN, and uses the complementary advantages of the two to solve the bearing fault diagnosis problem. The present invention not only improves the shortcomings of the original algorithm, but also shows excellent performance in noise resistance. Attached Figure Description

[0043] Figure 1 This is a flowchart of a hybrid algorithm based on the CA-ACGAN model algorithm diagnostic flowchart;

[0044] Figure 2 Timing diagram of the status of each part of the driver;

[0045] Figure 3 Time-frequency plot of the dataset;

[0046] Figure 4 Traditional attention coordination mechanism module structure diagram;

[0047] Figure 5 Improved structure diagram of the coordination attention mechanism module;

[0048] Figure 6 Discriminator accuracy curve;

[0049] Figure 7 Confusion matrix;

[0050] Figure 8 Classification accuracy curve;

[0051] Figure 9 Classification accuracy of each algorithm under different signal-to-noise ratios. Detailed Implementation

[0052] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0053] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0054] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0055] like Figure 1 As shown, this invention provides a coordinated attention mechanism and auxiliary classification adversarial generative network method for bearing fault diagnosis. It integrates the coordinated attention mechanism (CA) and the auxiliary classification adversarial generative network (ACGAN) to establish a CA-ACGAN model for bearing fault diagnosis; the method includes the following steps:

[0056] S1. Collect raw one-dimensional bearing data;

[0057] S2. Convert the original one-dimensional bearing data into a two-dimensional time-frequency graph using continuous wavelet transform as the real data.

[0058] S3. Improve the coordination attention mechanism and integrate it with the generator;

[0059] S4. Random noise and labels are fed into the generator to generate fake data;

[0060] S5. Feed real data and fake data together into the discriminator for training;

[0061] S6. Determine the algorithm termination condition;

[0062] S7. Repeat steps S2-S5 until Nash equilibrium is reached;

[0063] S8. Extract the trained discriminator, feed the test set into the discriminator for testing, and output the accuracy and classification results.

[0064] The following is a detailed implementation process of the present invention.

[0065] To accomplish the task of bearing fault diagnosis, the specific implementation steps of this invention are as follows:

[0066] Step 1: This chapter uses the Case Western Reserve University (CWRU) rolling bearing dataset as the experimental data. Taking bearing fault data at 12kHz as an example, vibration signal analysis is performed on bearing data under zero load. The time sequence diagrams of various parts of the drive end are shown below. Figure 2 As shown.

[0067] Step 2: Perform continuous wavelet transform on the one-dimensional signal and collect the vibration signal of the DE bearing using a sampling frequency of 12kHz. Based on the different sizes and locations of defects, the data from the drive-end bearing is divided into 10 modes. Under 0hp load, the vibration signal generated per rotation contains 401 sampling points (12000×60 / 1797), with a sampling time of approximately 0.0334 seconds (60 / 1797). To ensure that each sample contains at least one cycle of vibration data, the vibration signal is divided into segments of 600 data points, with 300 overlapping points between adjacent segments. The generated time-frequency diagrams of each bearing state are shown below. Figure 3 As shown.

[0068] Step 3, the coordinated attention module, involves two steps: coordinate information embedding and coordinated attention generation. This process precisely encodes the location information of channel relationships and long-range dependencies. The traditional coordinated attention mechanism module structure is as follows: Figure 4 As shown,

[0069] (1) Coordinate information embedding

[0070] The global pooling factor specified in the following formula is decomposed into a one-dimensional feature encoding:

[0071]

[0072] The above equation z c The output of the c-th channel in the SE module is the result of compression. The input x comes directly from a convolutional layer with a fixed kernel size. Given the input x, each channel is encoded along the horizontal and vertical axes using two spatial ranges (H,1) or (1,W) of the pooling kernel, respectively. Therefore, the output of the c-th channel at height h is represented as:

[0073]

[0074] Similarly, the c-th channel output with width w is shown in the following equation:

[0075]

[0076] The two transformations above aggregate features along two spatial directions respectively, producing a pair of direction-aware feature mappings.

[0077] (2) Coordinating attention generation

[0078] Specifically, given the feature maps embedded in the formula by coordinate information, we first concatenate them, then send them to a shared 1×1 convolutional transformation function F1, resulting in the following equation:

[0079] f=δ(F1([z h ,z w ]))

[0080] [z h ,z w ] represents a splicing operation along a spatial dimension, δ is a non-linear activation function, and f∈R C / r×(H×W) Intermediate feature maps are used to encode spatial information in the horizontal and vertical directions. Here, r is the reduction ratio of the control block size in the SE module. Then, f is divided into two tensors along the spatial dimension. h ∈R C / r×H and f w ∈R C / r×W Additionally, two 1×1 convolution transformations F are used. h and F w f respectively h and f w Converting it to a tensor with the same channel number as the input X, and then activating it with the sigmoid function, yields the output shown in the following equation:

[0081] g h =σ(F h (f h ))

[0082] g w =σ(F w (f w ))

[0083] In the formula, σ is the sigmoid function. To reduce the complexity of the model, the number of channels in f is usually reduced by an appropriate reduction ratio r, and then the output g is calculated. h and g w Each value is expanded and used as a weight for attention, ultimately resulting in an expression for the output y of the coordinated attention module.

[0084] The formula is as follows:

[0085]

[0086] The modified coordination attention mechanism block structure is as follows: Figure 5 As shown, the Hard-Swish activation function is an improved version of the Swish function, and its formula is as follows:

[0087]

[0088] Step 4: Divide the 600 time-frequency maps for each fault category into a 7:3 ratio. Build and train the CA-ACGAN algorithm proposed in this chapter using the PyTorch deep learning framework based on Python. Set the number of training epochs to 1750, and set the learning rate for both the generator and discriminator networks to 0.0002. Feed randomly generated noise along with the labels into the generator to generate fake data.

[0089] Step 5: Feed the real data and the fake data generated by the generator into the discriminator for identification, and train the discriminator's identification ability.

[0090] Steps 6 and 7 test the discriminator's ability to distinguish data. If the probability of identifying the input data as true or false is 50% for both cases, then Nash equilibrium has been reached, and the model training is complete. The discriminator's accuracy is as follows: Figure 6 As shown.

[0091] Step 8: To rule out overfitting and further verify the model's performance, the trained discriminator was extracted separately for validation on the test set. The test set contained 180 samples for each data type. The test results are as follows. Figure 7 The confusion matrix shown indicates that although some misclassified samples were incorrectly assigned to other categories, overall, the bearing states of each category can be correctly distinguished. The classification accuracy curve is as follows. Figure 8 As shown, when ITER is 400, it can distinguish different fault types of bearing data very well. The fault diagnosis accuracy of the CA-ACGAN model eventually stabilized at a level close to 0.98, and showed a small fluctuation within the normal range.

[0092] The present invention also provides a coordinated attention mechanism and an auxiliary classification adversarial generative network device for bearing fault diagnosis, including a memory, a processor, and computer program instructions stored in the memory and executable by the processor. When the processor executes the computer program instructions, it can implement the steps of the method described above.

[0093] The present invention also provides a computer-readable storage medium having stored thereon computer program instructions that can be executed by a processor, wherein when the processor executes the computer program instructions, it can implement the steps of the method described above.

[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A bearing fault diagnosis method based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification, characterized in that, A CA-ACGAN model is established by fusing the Coordinating Attention Mechanism (CA) with the Generative Adversarial Network (ACGAN) for auxiliary classification to achieve bearing fault diagnosis; the method includes the following steps: S1. Collect raw one-dimensional bearing data; S2. Convert the original one-dimensional bearing data into a two-dimensional time-frequency graph using continuous wavelet transform as the real data. S3. Improve the coordination attention mechanism and integrate it with the generator; S4. Random noise and labels are fed into the generator to generate fake data; S5. Feed real data and fake data together into the discriminator for training; S6. Determine the algorithm termination condition; S7. Repeat steps S2-S5 until Nash equilibrium is reached; S8. Extract the trained discriminator, feed the test set into the discriminator for testing, and output the accuracy and classification results. The specific implementation method of step S3 is as follows: (1) Coordinate information embedding The global pooling factor specified in the following formula is decomposed into a one-dimensional feature encoding: The above formula For the SE module, the first Output after channel compression and They are height and width, respectively. Let j be the starting point for the pixel height and width, respectively. Directly from convolutional layers with a fixed kernel size, given input Then, use the two spatial ranges of the pooling kernel. or Encode each channel along both the horizontal and vertical axes; Therefore, the first The passage at height The output is represented as: Similarly, the width is The The channel output is shown in the following formula: The two transformations aggregate features along two spatial directions respectively, producing a pair of direction-aware feature maps; (2) Coordinating attention generation Given feature maps embedded from coordinate information in a formula, they are first linked together, and then sent to a shared [system / platform]. Convolution Transform Function We get the following formula: This indicates a splicing operation along a spatial dimension. It is a non-linear activation function. Intermediate feature maps that encode spatial information in the horizontal and vertical directions; where It is the reduction ratio of the control block size in the SE module, and then along the spatial dimension. Divide into two tensors and Two more are used. Convolution Transformation and Each and Convert to input Tensors with the same channel number are then used Function activation yields the output shown in the following equation: In the formula yes Functions, to reduce model complexity, reduce the proportion. To reduce The number of channels, then the output and Each weight is expanded and used as an attention parameter, and finally, the output is... The expression is shown in the following formula: Improve the coordination of attention mechanisms, that is, use two Convolution Transformation and Each and Convert to input Before tensors with the same channel number, a Hard-Swish activation function is added, and the structure of the coordinated attention mechanism is improved to adapt to the ACGAN network structure, thereby increasing the generator network's ability to capture features.

2. The bearing fault diagnosis method based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification as described in claim 1, characterized in that, In step S1, the original one-dimensional bearing data uses the CWRU rolling bearing dataset, and the vibration signal of the drive end bearing under load 0 is collected at a sampling frequency of 12kHz.

3. The bearing fault diagnosis method based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification as described in claim 1, characterized in that, In step S2, the vibration signal of the drive end bearing in the original one-dimensional bearing data is divided into 600 data points as a segment, with 300 overlapping points between adjacent segments, to generate a two-dimensional time-frequency diagram of the state of each drive end bearing.

4. The bearing fault diagnosis method based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification as described in claim 1, characterized in that, The formula for the Hard-Swish activation function is as follows: 。 5. The bearing fault diagnosis method based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification as described in claim 1, characterized in that, In step S4, the two-dimensional time-frequency graph generated in step S2 is divided into a 7:3 ratio. The CA-ACGAN model is built and trained in the Python-based PyTorch deep learning framework. The number of training rounds is set to 1750, and the learning rate of both the generator and discriminator networks is set to 0.0002. Randomly generated noise and labels are fed into the generator to generate fake data.

6. The bearing fault diagnosis method based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification as described in claim 1, characterized in that, In step S7, if the probability of identifying the input data as true or false is 50% for both cases, it indicates that Nash equilibrium has been reached and the model training is complete.

7. A bearing fault diagnosis device based on a coordinated attention mechanism and an adversarial generative network for auxiliary classification, characterized in that, It includes a memory, a processor, and computer program instructions stored in the memory and executable by the processor, which, when executed by the processor, enable the implementation of the steps of the method as described in any one of claims 1-6.

8. A computer-readable storage medium having stored thereon computer program instructions executable by a processor, wherein when the processor executes the computer program instructions, it is able to implement the steps of the method as described in any one of claims 1-6.

Citation Information

Patent Citations

  • Motor bearing fault diagnosis method

    CN114913396A

  • Non-equilibrium sample bearing fault diagnosis method based on multi-scale feature fusion

    CN116541771A