Migration learning rolling bearing fault diagnosis method based on MSCNN-ACGAN

Through the MSCNN-ACGAN transfer learning method, combining multi-scale convolutional neural network and auxiliary classifier to generate an adversarial network, the problem of insufficient sample number and inconsistent data distribution in rolling bearing fault diagnosis under multiple operating conditions is solved, and the accuracy of fault diagnosis is improved and recognition capabilities.

CN118981683AActive Publication Date: 2024-11-19CHONGQING UNIV OF POSTS & TELECOMM

Patent Information

Application Number
CN202411098198.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-12
Publication Date
2024-11-19
Estimated Expiration
2044-08-12

AI Technical Summary

Technical Problem

In multiple operating conditions, in rolling bearing fault diagnosis, the problem of insufficient sample number and inconsistent data distribution in different operating conditions lead to poor fault diagnosis performance of traditional methods.

Method used

Using the transfer learning method based on MSCNN-ACGAN, the vibration signal characteristics are extracted through multi-scale convolutional neural networks, and the auxiliary classifier generates adversarial network to generate supplementary data. Combined with subdomain adaptive technology, the data distribution of source domain and target domain are aligned to improve the Accuracy of fault diagnosis.

Benefits of technology

It effectively solves the problems of insufficient sample size and inconsistent data distribution, improves the accuracy of fault diagnosis in multiple operating conditions, and improves the classification and identification capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118981683A_ABST
    Figure CN118981683A_ABST
Patent Text Reader

Abstract

The invention provides a migration learning rolling bearing fault diagnosis method based on MSCNN-ACGAN, and belongs to the technical field of bearing fault diagnosis. The rolling bearing fault diagnosis method based on the combination of MSCNN-ACGAN and transfer learning is provided for solving the problems that rolling bearing data collection is difficult, fault label samples are lacked, enough diagnosis knowledge is difficult to obtain, and the data distribution difference between a training task and a target task is large under multiple working conditions. Vibration signal samples obtained through sampling are subjected to non-overlapping sliding cutting according to different window sizes and then serve as input, a multi-scale convolutional neural network (MSCNN) is trained to extract signal multi-scale features, a classifier generative adversarial network (ACGAN) is assisted to carry out data enhancement, when training reaches Nash equilibrium, a transfer learning network model based on the MSCNN-ACGAN is trained, and the transfer learning effect is improved. And the fault diagnosis precision of the target domains with different data distribution is improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of rolling bearing fault diagnosis, and specifically relates to a rolling bearing fault diagnosis method based on migration learning of MSCNN-ACGAN, and a rolling bearing fault diagnosis method under multiple working conditions based on a multi-scale convolutional neural network (MSCNN), an auxiliary classifier generative adversarial network (ACGAN) and subdomain adaptation (DA). Background Art

[0002] With the rapid development of deep learning, deep learning is increasingly being used in machine health detection and fault diagnosis, especially in bearing fault diagnosis. However, in actual industrial applications, due to working in complex and changing environments, training data and test data may come from different feature spaces and have different distributions, and it is difficult to collect enough samples from a single distribution. The emergence of these situations has greatly affected the performance of traditional data-driven fault diagnosis methods, which means that the carefully trained fault diagnosis model does not generalize well enough in fault diagnosis tasks under multiple working conditions. Therefore, it is a feasible method to use multi-scale convolutional neural network (MSCNN) and auxiliary classifier generative adversarial network (ACGAN) to extract vibration signal data features and generate supplementary data sets, and perform subdomain adaptive fault diagnosis on the target task.

[0003] Multiscale convolutional neural network (MSCNN) plays an important role in the multi-scale characteristics of rolling bearing vibration signals. Multiscale convolutional neural network (MSCNN) has become an important tool in the field of vibration signal analysis because it can effectively capture the characteristics of rolling bearing vibration signals at different time scales. Using MSCNN, multiple scale characteristics from local subtle vibrations to overall vibration modes can be considered simultaneously, effectively improving the accuracy and reliability of fault diagnosis.

[0004] Auxiliary Classifier Generative Adversarial Network is a variant of Generative Adversarial Network (GAN), which introduces an auxiliary classifier to enhance the adversarial learning between the generator and the discriminator. In ACGAN, the generator not only attempts to generate realistic data samples, but also predicts the category labels of the generated samples through the auxiliary classifier, while the discriminator not only judges the authenticity of the generated samples, but also predicts the correctness of their category labels. This structure enables ACGAN to generate samples with specific categories, thereby expanding the application capabilities of traditional GAN ​​in generating multi-category datasets.

[0005] After searching, the application publication number CN112254964A, a rolling bearing fault diagnosis method based on a fast multi-scale convolutional neural network, belongs to the technical field of rolling bearing fault diagnosis methods based on a fast multi-scale convolutional neural network; the technical problem to be solved is: to provide an improvement of a rolling bearing fault diagnosis method based on a fast multi-scale convolutional neural network; the technical solution adopted to solve the above technical problem is: comprising the following steps: collecting a one-dimensional vibration signal from the driving end of the rolling bearing; standardizing each monitoring vibration signal collected in step one; building a fast multi-scale convolutional neural network model; training the fast multi-scale convolutional neural network model; collecting the acceleration vibration signal of the current rolling bearing and processing it to obtain a sample test set; inputting the sample test set in step five into the fast multi-scale convolutional neural network model trained in step four, thereby outputting the rolling bearing fault diagnosis result; the present invention is applied to bearing fault diagnosis.

[0006] In the above invention, although a multi-scale convolutional neural network is used to effectively extract multi-scale time information, it is unable to solve the problems of insufficient sample quantity and inconsistent data distribution between different working condition data. In the present invention, corresponding solutions are proposed for these two problems. First, to address the problem of insufficient sample quantity, after effectively extracting data features through a multi-scale convolutional neural network (MSCNN), ACGAN is used to generate effective fault feature data, thereby enriching samples and increasing the number of samples; secondly, to address the problem of inconsistent data distribution between different working condition data, a subdomain adaptive method is used for migration, and the distribution difference between the source domain and target domain data is reduced by aligning the data distribution of the source domain and the target domain. Summary of the invention

[0007] The present invention aims to solve the problems of lack of fault label samples of rolling bearings under multiple working conditions, difficulty in obtaining sufficient diagnostic knowledge, and large differences in data distribution between training tasks and target tasks. A rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN is proposed. The technical solution of the present invention is as follows:

[0008] A rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN, comprising the following steps:

[0009] S1. Signal processing: Use sensors to collect vibration signals of rolling bearing faults under different conditions, obtain signal samples, and perform preprocessing including non-overlapping moving average cutting and standardization;

[0010] S2. Network construction: Build a multi-scale convolutional neural network MSCNN, an auxiliary classifier generative adversarial network ACGAN, and a subdomain adaptive network. The MSCNN network includes a feature extractor, and the ACGAN network includes a generator and a discriminator.

[0011] S3, network training: input the source domain signal and its corresponding label into the network built in step S2 for training, the input of the feature extractor is the source domain signal, the input of the generator is the label information and random noise, and the input of the discriminator is the label information, the real signal, the generated signal output by the generator and the corresponding generated label information;

[0012] S4. Transfer learning: After the training of the multi-scale convolutional neural network MSCNN and the auxiliary classifier generative adversarial network ACGAN is completed, the subdomain adaptive network is used to perform rolling bearing fault diagnosis under multiple working conditions.

[0013] Furthermore, the step S1 specifically includes the following steps:

[0014] S1.1, using sensors to collect rolling bearing fault vibration signals under different states, obtaining original signal samples, and dividing them into several data sets according to different states;

[0015] S1.2, unify the original vibration signals in each data set into a length of N, arrange and combine them according to the requirements, and perform non-overlapping moving average cutting on the original vibration signals according to three different window sizes M;

[0016] S1.3, the mathematical model of Min-Max standardization is described as:

[0017]

[0018] Among them, x′ represents the normalized value, x represents the original data value, and x max is the maximum value in the data set, x min is the minimum value in the data set;

[0019] S1.4. According to the needs of network training, the dataset is divided into source domain and target domain for subdomain adaptive training. The source domain and target domain data are divided into training set and test set respectively in a 1:1 ratio.

[0020] Furthermore, the step S2 builds a multi-scale convolutional neural network MSCNN, including three groups of different CNN structures, specifically:

[0021] The first set of convolutional modules includes a 32×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 16×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer;

[0022] The second group of CNN structures includes a 64×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer.

[0023] The third group of CNN structures includes a 64×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, an 8×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer.

[0024] Furthermore, the step S2 builds an auxiliary classifier generation adversarial network ACGAN, and the specific structures of the generator and the discriminator are:

[0025] Generator: Fully connected layer with input (input length + 1, (input length + 1) * batch size), 16×1 convolution layer, ReLU activation function and BN algorithm, 3×1 convolution layer, ReLU activation function and BN algorithm, 1×1 convolution layer, ReLU activation function and BN algorithm, 1×1 convolution layer, ReLU activation function and BN algorithm, linear layer, Tanh activation function;

[0026] Discriminator: 5×1 convolution layer, LeakyReLU activation function and BN algorithm, 3×1 convolution layer, LeakyReLU activation function and BN algorithm, 1×1 convolution layer, LeakyReLU activation function and BN algorithm, 1×1 convolution layer, LeakyReLU activation function and BN algorithm, Flatten layer, Linear layer, LeakyReLU layer, Linear layer.

[0027] Furthermore, in step S2, the multi-scale convolutional neural network and the auxiliary classifier generative adversarial network are both composed of convolutional modules stacked on top of each other;

[0028] The mathematical model of one-dimensional convolution is described as the process of sliding the convolution kernel on the input data to calculate the feature map:

[0029]

[0030] Where x is the input sequence with a length of n, w is the convolution kernel with a length of k, y(t) is the value of the output feature map, t is the index of the output sequence, and * represents the convolution operation.

[0031] Furthermore, the step S3, network training: inputting the source domain signal and its corresponding label into the network constructed in step S2 for training, the input of the feature extractor is the source domain signal, the input of the generator is the label information and random noise, the input of the discriminator is the label information, the real signal and the false signal generated by the generator, specifically includes:

[0032] The loss function of the ACGAN model is:

[0033]

[0034] Among them, L S is the discrimination loss, L C is the classification loss;

[0035] In the training of ACGAN, the optimization goal is to train the discriminator to make L S +L C Maximize, train the generator to make L S -L C minimize;

[0036] ACGAN training uses the Adam optimization algorithm, the generator learning rate is set to 0.03, the discriminator learning rate is set to 0.001, the Softmax classifier is used for category distinction, the binary classification loss function uses the BCEWithLogitsLoss function, and the classification loss function uses the CrossEntropyLoss function.

[0037] Furthermore, in step S4, the subdomain adaptation network applies a subdomain distance loss function CORAL, calculates the covariance matrix of the source domain and the target domain data, and adjusts the source domain features so that the covariance matrix of the source domain features is as similar as possible to the covariance matrix of the target domain features;

[0038] Given the source domain feature matrix X s and the target domain feature matrix X t , their covariance matrices are C s and C t , the CORAL loss is

[0039]

[0040] Among them, ||·|| F The dimension of the feature representing the Frobenius norm, d.

[0041] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, a rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN as described in any one of the above is implemented.

[0042] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN as described in any one of the above.

[0043] The advantages and beneficial effects of the present invention are as follows:

[0044] 1) The present invention firstly utilizes a multi-scale convolutional neural network to directly extract features from the original vibration signal, thereby avoiding the problem of disappearance of signal features during the signal conversion process and reducing the signal processing time; 2) In view of the problem of insufficient signal features, the present invention uses an auxiliary classifier to generate adversarial networks to generate data for the signal features extracted by the multi-scale convolutional neural network MSCNN, thereby solving the problem of insufficient signal features; 3) In view of the scarcity of labeled samples of rolling bearings under multiple working conditions and the large difference in data distribution characteristics between training tasks and target tasks, the subdomain adaptive network uses the learned transferable knowledge to reduce the corresponding subdomain distance between source domain and target domain data after feature transformation, thereby reducing the data distribution difference; 4) In view of the problem of low accuracy of rolling bearings under multiple working conditions, the present invention integrates a multi-scale convolutional neural network and an auxiliary classifier to generate adversarial networks, applies a simple convolution structure, adjusts the convolution kernel size for the original cutting signal, improves the network's ability to extract and generate features, and comprehensively applies subdomain adaptive technology (main innovation point) to improve the network's ability to adapt to multiple working conditions and improve classification and recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of a rolling bearing fault diagnosis process according to a preferred embodiment of the present invention;

[0046] Figure 2 It is a signal segmentation schematic diagram of the present invention;

[0047] Figure 3 It is a schematic diagram of a multi-scale convolutional neural network;

[0048] Figure 4 (a) is a schematic diagram of the auxiliary classifier generative adversarial network generator network;

[0049] Figure 4 (b) is a schematic diagram of the auxiliary classifier generative adversarial network discriminator network;

[0050] Figure 5is the discriminator classification confusion matrix diagram; DETAILED DESCRIPTION

[0051] The following will describe the technical solutions in the embodiments of the present invention in detail in conjunction with the accompanying drawings in the embodiments of the present invention. The described embodiments are only part of the embodiments of the present invention.

[0052] The technical solution of the present invention to solve the above technical problems is:

[0053] See also Figure 1 , the flowchart of the research on the rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN proposed by the present invention, the specific steps of this embodiment are as follows:

[0054] S1. Signal processing: Use sensors to collect rolling bearing fault vibration signals under different states to obtain signal samples. Use three different window sizes to perform non-overlapping moving average cutting on the signal samples to obtain three different groups of original signals, divide the data, and set them as source domain and target domain respectively. Use Min-Max standardization to standardize the collected signal data to facilitate subsequent model training and optimization;

[0055] S1.1. The sampling method is to use the rolling bearing fault vibration signal under different states of the sensor to obtain the original signal samples, which are divided into several data sets according to different states.

[0056] S1.2. The original vibration signals in each data set are unified into a length of N, arranged and combined as required, and non-overlapping moving average cutting is performed on the original vibration signals according to three different window sizes M.

[0057] S1.3, the mathematical model of Min-Max standardization is described as:

[0058]

[0059] Among them, x′ represents the normalized value, x represents the original data value, and x max is the maximum value in the data set, x min is the minimum value in the data set.

[0060] The original vibration signal with a length of N is subjected to sliding cutting of the sample signal using three different window sizes to obtain three different sets of original cutting signals, such as Figure 2As shown. The experimental data used in this embodiment is the SKF6205 deep groove ball bearing data set in the rolling bearing fault vibration signal opened by Case Western Reserve University (CWRU) in the United States. In the experiment, an acceleration sensor installed at the motor drive end is used to collect vibration signals. Three types of faults are introduced by electrospark machining of the bearing: outer ring fault (ORF), ball fault (BF) and inner ring fault (IRF). The degree of fault for each fault type is different (fault diameter: 0.007 inches, 0.014 inches and 0.021 inches). The vibration signals under three motor speed conditions (1772rpm, 1750rpm, 1730rpm) are selected as experimental data, and the workload is 0HP, 1HP, 2HP and 3HP. The experimental bearing model is SKF6205. The data sampling frequency is 12000Hz. The experimental data representation method is shown in Table 1.

[0061] Table 1 Experimental data representation method

[0062]

[0063] S1.4. According to the needs of network training, the data set is divided into source domain and target domain for subdomain adaptive training. The division results are shown in Table 2. The source domain and target domain data are divided into training set and test set respectively in a ratio of 1:1.

[0064] Table 2 Division of source and target domains in subdomain adaptation training

[0065]

[0066]

[0067] S2. Network construction: Build a multi-scale convolutional neural network (MSCNN), an auxiliary classifier generative adversarial network (ACGAN), and a subdomain adaptive network. The network includes a feature extractor, a generator, and a discriminator, all of which are composed of stacked convolutional modules.

[0068] First, we build a multi-scale convolutional neural network (MSCNN), which includes three different CNN structures, such as Figure 3 shown.

[0069] The first group of convolution modules includes a 32×1 convolution layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 16×1 convolution layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolution layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer.

[0070] The second group of CNN structures includes a 64×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer.

[0071] The third group of CNN structures includes a 64×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, an 8×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer.

[0072] The specific parameters of the multi-scale convolutional neural network are shown in Table 3.

[0073] Table 3 Multi-scale convolutional neural network parameters

[0074]

[0075]

[0076] Secondly, we build an auxiliary classifier generative adversarial network (ACGAN). The specific structures of the generator and discriminator are as follows:

[0077] Generator: fully connected layer with input of (input length + 1, (input length + 1) * batch size), 16×1 convolution layer, ReLU activation function and BN algorithm, 3×1 convolution layer, ReLU activation function and BN algorithm, 1×1 convolution layer, ReLU activation function and BN algorithm, 1×1 convolution layer, ReLU activation function and BN algorithm, linear layer, Tanh activation function.

[0078] Discriminator: 5×1 convolution layer, LeakyReLU activation function and BN algorithm, 3×1 convolution layer, LeakyReLU activation function and BN algorithm, 1×1 convolution layer, LeakyReLU activation function and BN algorithm, 1×1 convolution layer, LeakyReLU activation function and BN algorithm, Flatten layer, Linear layer, LeakyReLU layer, Linear layer.

[0079] S3, network training: The source domain signal and its corresponding label are input into the network built in step S2 for training. The input of the feature extractor is the source domain signal, the input of the generator is the label information and random noise, and the input of the discriminator is the label information, the real signal and the false signal generated by the generator.

[0080] The loss function of the ACGAN model is:

[0081]

[0082] Among them, L S is the discrimination loss, L C is the classification loss;

[0083] In the training of ACGAN, the optimization goal is to train the discriminator to make L S +L C Maximize, train the generator to make L S -L C minimize.

[0084] ACGAN training uses the Adam optimization algorithm, the generator learning rate is set to 0.03, the discriminator learning rate is set to 0.001, the Softmax classifier is used for category distinction, the binary classification loss function uses the BCEWithLogitsLoss function, and the classification loss function uses the CrossEntropyLoss function.

[0085] To verify the effectiveness of the above method, the present invention is verified on the Case Western Reserve University (CWRU) data set. The data used in the experiment comes from vibration data under three different working conditions. The data used in the experiment includes normal operation signals and 9 fault signals. Figure 5 This is the confusion matrix diagram after the discriminator classification. It can be seen that the discriminator has a strong classification ability.

[0086] S4. Transfer learning: After the training of the multi-scale convolutional neural network (MSCNN) and the auxiliary classifier generative adversarial network (ACGAN) is completed, the subdomain adaptive network is used to perform rolling bearing fault diagnosis under multiple working conditions.

[0087] The subdomain adaptation network applies the subdomain distance loss function CORAL to calculate the covariance matrix of the source domain and target domain data, and adjusts the source domain features so that the covariance matrix of the source domain features is as similar as possible to the covariance matrix of the target domain features.

[0088] Given the source domain feature matrix X s and the target domain feature matrix X t , their covariance matrices are C s and Ct , the CORAL loss is

[0089]

[0090] Among them, ||·|| F The dimension of the feature representing the Frobenius norm, d.

[0091] This method combines the advantages of auxiliary classifier generative adversarial networks and multi-scale convolutional neural networks. It can supplement similar data when the amount of data is insufficient, improve the accuracy of system fault diagnosis, and has broad application prospects.

[0092] The systems, devices, modules or units described in the above embodiments may be implemented by computer chips or entities, or by products with certain functions.

[0093] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0094] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0095] The above embodiments should be understood to be only used to illustrate the present invention and not to limit the protection scope of the present invention. After reading the contents of the present invention, technicians can make various changes or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN, characterized in that: The following steps are involved: S1. Signal processing: Use sensors to collect vibration signals of rolling bearing faults under different conditions, obtain signal samples, and perform preprocessing including non-overlapping moving average cutting and standardization; S2. Network construction: Build a multi-scale convolutional neural network MSCNN, an auxiliary classifier generative adversarial network ACGAN, and a subdomain adaptive network. The MSCNN network includes a feature extractor, and the ACGAN network includes a generator and a discriminator. S3, network training: input the source domain signal and its corresponding label into the network built in step S2 for training, the input of the feature extractor is the source domain signal, the input of the generator is the label information and random noise, and the input of the discriminator is the label information, the real signal, the generated signal output by the generator and the corresponding generated label information; S4. Transfer learning: After the training of the multi-scale convolutional neural network MSCNN and the auxiliary classifier generative adversarial network ACGAN is completed, the subdomain adaptive network is used to perform rolling bearing fault diagnosis under multiple working conditions.

2. According to claim 1, a rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN is characterized in that: The step S1 specifically includes the following steps: S1.1, using sensors to collect rolling bearing fault vibration signals under different states, obtaining original signal samples, and dividing them into several data sets according to different states; S1.2, unify the original vibration signals in each data set into a length of N, arrange and combine them according to the requirements, and perform non-overlapping moving average cutting on the original vibration signals according to three different window sizes M; S1.3, the mathematical model of Min-Max standardization is described as: Among them, x ′ represents the normalized value, x represents the original data value, and x max is the maximum value in the data set, x min is the minimum value in the data set; S1.

4. According to the needs of network training, the dataset is divided into source domain and target domain for subdomain adaptive training. The source domain and target domain data are divided into training set and test set respectively in a 1:1 ratio.

3. The rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN according to claim 1 is characterized in that: The step S2 builds a multi-scale convolutional neural network MSCNN, including three groups of different CNN structures, specifically: The first set of convolutional modules includes a 32×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 16×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer; The second group of CNN structures includes a 64×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer. The third group of CNN structures includes a 64×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, an 8×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, a 3×1 convolutional layer processed with the BN algorithm, ReLU activation function, and a maximum pooling layer with a convolution kernel of 2, and finally a 128-channel fully connected layer.

4. The rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN according to claim 1 is characterized in that: The step S2 builds an auxiliary classifier to generate an adversarial network ACGAN, and the specific structures of the generator and the discriminator are: Generator: Fully connected layer with input (input length + 1, (input length + 1) * batch size), 16×1 convolution layer, ReLU activation function and BN algorithm, 3×1 convolution layer, ReLU activation function and BN algorithm, 1×1 convolution layer, ReLU activation function and BN algorithm, 1×1 convolution layer, ReLU activation function and BN algorithm, linear layer, Tanh activation function; Discriminator: 5×1 convolution layer, LeakyReLU activation function and BN algorithm, 3×1 convolution layer, LeakyReLU activation function and BN algorithm, 1×1 convolution layer, LeakyReLU activation function and BN algorithm, 1×1 convolution layer, LeakyReLU activation function and BN algorithm, Flatten layer, Linear layer, LeakyReLU layer, Linear layer.

5. A rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN according to claim 3 or 4, characterized in that: In step S2, the multi-scale convolutional neural network and the auxiliary classifier generative adversarial network are both composed of convolutional modules stacked on top of each other; The mathematical model of one-dimensional convolution is described as the process of sliding the convolution kernel on the input data to calculate the feature map: Where x is the input sequence with a length of n, w is the convolution kernel with a length of k, y(t) is the value of the output feature map, t is the index of the output sequence, and * represents the convolution operation.

6. The rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN according to claim 1 is characterized in that: The step S3, network training: inputting the source domain signal and its corresponding label into the network constructed in step S2 for training, the input of the feature extractor is the source domain signal, the input of the generator is the label information and random noise, the input of the discriminator is the label information, the real signal and the false signal generated by the generator, specifically including: The loss function of the ACGAN model is: Among them, L S is the discrimination loss, L C is the classification loss; In the training of ACGAN, the optimization goal is to train the discriminator to make L S +L C Maximize, train the generator to make L S -L C minimize; ACGAN training uses the Adam optimization algorithm, the generator learning rate is set to 0.03, the discriminator learning rate is set to 0.001, the Softmax classifier is used for category distinction, the binary classification loss function uses the BCEWithLogitsLoss function, and the classification loss function uses the CrossEntropyLoss function.

7. The rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN according to claim 1 is characterized in that: In step S4, the subdomain adaptation network applies the subdomain distance loss function CORAL, calculates the covariance matrix of the source domain and the target domain data, and adjusts the source domain features so that the covariance matrix of the source domain features is as similar as possible to the covariance matrix of the target domain features; Given the source domain feature matrix X s and the target domain feature matrix X t , their covariance matrices are C s and C t , the CORAL loss is Among them, ||·|| F The dimension of the feature representing the Frobenius norm, d.

8. An electronic device, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN as described in any one of claims 1 to 7 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the rolling bearing fault diagnosis method based on transfer learning of MSCNN-ACGAN as described in any one of claims 1 to 7 is implemented.

Citation Information

Patent Citations

  • Rolling bearing fault diagnosis method based on rapid multi-scale convolutional neural network

    CN112254964A

  • Bearing fault intelligent diagnosis method and diagnosis system, computer equipment and medium

    CN112633339A

  • Rolling bearing fault diagnosis method based on subdomain self-adaption

    CN115563536A

  • SAR image ship target generation and classification method

    CN116895016A

  • Rotary machinery fault diagnosis method based on improved auxiliary classification generative adversarial network

    CN117290771A

Cited By

  • Cross-domain bearing fault diagnosis method based on prototype contrast learning

    CN119760458B

  • Intelligent acquisition and analysis monitoring method and system for urban tube well safety

    CN119760513A

  • A smart data acquisition and analysis monitoring method and system for urban manhole safety

    CN119760513B