An intelligent fault diagnosis method for bearings with small sample class imbalance

Abstract

The present invention discloses a small-sample class-imbalanced intelligent fault diagnosis method for bearings. First, a vibration acceleration sensor is used to collect bearing vibration signals; secondly, the collected original vibration signals are subjected to fast Fourier transform to be converted into frequency-domain signals, and then the transformed signals are subjected to 1-norm regularization processing to scale each sample to unit norm and retain the distribution of the original data; thirdly, a generative adversarial network AD2GAN network is constructed by combining an autoencoder network and a dual discriminator; fourthly, the AD2GAN model is continuously iteratively trained according to the set loss objective function, and the generated data is saved after the training is stable; then the saved samples are supplemented to the original imbalanced samples as required to make them reach class balance; finally, a deep convolutional neural network is used as a fault diagnosis model for fault diagnosis. The method proposed by the present invention has a substantial improvement in the aspect of fault diagnosis effect, with high diagnostic accuracy, good stability and strong generalization ability.

Description

Technical Field

[0001] The present invention relates to the field of intelligent fault diagnosis of mechanical rotating components, and particularly to an intelligent fault diagnosis method for bearings that solves small-sample class imbalance. Background Art

[0002] During the entire life cycle of a mechanical device from the start of operation to complete scrapping, faults will more or less occur. Once a fault occurs in an important component of a mechanical device during operation, if it is not discovered and the machine is not stopped in time, its long-term operation in a faulty state may cause the entire mechanical device to malfunction and may even cause certain casualties. According to relevant investigations, 45% - 55% of mechanical faults are caused by bearing faults. Therefore, fault monitoring and diagnosis of bearings play an extremely important role in ensuring the safe and reliable operation of mechanical devices.

[0003] Bearing fault monitoring and intelligent fault diagnosis are data-driven. During the entire life cycle of a machine, the machine operates in a normal state most of the time. Therefore, most of the collected data is healthy data (majority class), and the fault data is often less (minority class), resulting in a sample imbalance between healthy data and fault data. If existing mature fault diagnosis algorithms directly use a small-sample class-imbalanced data set for fault diagnosis, the minority class will have a high probability of misclassification, which limits the diagnostic performance and generalization ability of the intelligent fault diagnosis algorithm. Therefore, how to effectively identify minority-class samples under a small-sample class-imbalanced data set and improve the diagnostic ability and generalization ability of the intelligent fault diagnosis algorithm is an urgent problem to be solved.

[0004] To solve the small-sample class imbalance problem, starting from the intelligent fault diagnosis process, data and algorithms are used as the entry points. In terms of data, the minority-class samples can be replicated in large quantities through oversampling techniques, so that their quantity is balanced with that of the majority-class samples, and then the balanced data set is used to train the fault diagnosis model. In terms of fault diagnosis algorithms, the structure of the algorithm or the final loss function is modified to directly process imbalanced data. However, the specific algorithm design has certain limitations, which limit the generalization ability of the fault diagnosis algorithm and affect its diagnostic performance to a certain extent. Therefore, in order to fundamentally solve the problems faced by small-sample class-imbalanced fault diagnosis, the data oversampling technique has become the most effective way.

[0005] Currently, the commonly used data oversampling techniques mainly include the Synthetic Minority Oversampling Technique (SMOTE) and the Generative Adversarial Network (GAN). Although both can solve the problem of class imbalance in fault diagnosis to a certain extent, their existing drawbacks also limit their application in class imbalance fault diagnosis. The data generated by SMOTE has an edge distribution problem, which may increase the difficulty of fault classification and reduce the diagnostic effect of the fault diagnosis model. Due to its own structural limitations, GAN has poor stability, and even mode collapse may occur during training. The quality and diversity of the generated samples cannot be well guaranteed, resulting in a reduction in the diagnostic effect of the fault diagnosis model. Summary of the Invention

[0006] In order to overcome the drawbacks of the existing technology, the purpose of the present invention is to propose a new generator and discriminator structure to handle the problem of small sample class imbalance. Among them, the discriminator adopts a dual discriminator structure, aiming to overcome the mode collapse problem and improve the quality and diversity of the generated samples. The generator combines an autoencoder network, constructs the loss between the deep features extracted by the generator and the decoded features of the autoencoder network, and improves the similarity between the samples generated by the generator and the original real samples by minimizing this loss. Furthermore, it improves the ability of the generator to deceive the two discriminators at the same time, and assists the dual discriminators to reach the Nash equilibrium through positive mutual confrontation.

[0007] To achieve the above purpose, the present invention provides a method for intelligent fault diagnosis of bearings with small sample class imbalance. The specific implementation steps are as follows: A method for intelligent fault diagnosis of bearings with small sample class imbalance includes the following steps:

[0008] 7) Data acquisition: Use a vibration acceleration sensor to collect the original vibration signal of the bearing;

[0009] 8) Data preprocessing: Perform data preprocessing on the original vibration signal collected in step 1);

[0010] 9) Network model construction and training: Construct an Auxiliary Dual Discriminator Generative Adversarial Network (AD2GAN) network model and train the model;

[0011] 10) Data generation: Save the fake samples generated by the model according to the input signal;

[0012] 11) Class balance: Balance the number of various fault samples and normal samples;

[0013] 12) Fault diagnosis: Use the class-balanced data set to train the fault diagnosis model, and then use the test set to test the trained fault diagnosis model to complete intelligent fault diagnosis.

[0014] Preferably, the implementation process of step 1) is as follows: Install the vibration acceleration sensor on the bearing housing of the mechanical equipment, configure the LMS data acquisition system, then run the mechanical equipment, and start the LMS data acquisition system to collect the time-series vibration acceleration signals of the bearing under various states as the original vibration signals of the bearing, and obtain the original real data set of N types of samples, including 1 type of normal sample and N-1 types of faulty samples.

[0015] Preferably, the implementation process of step 2) is to convert the collected original vibration signals of the bearing into frequency-domain signals through fast Fourier transform, and then perform 1-norm regularization processing on the frequency-domain signals to scale the frequency-domain signals of each sample to the unit norm and retain the distribution of the original vibration signals.

[0016] Preferably, the implementation process of step 3) is as follows: The AD2GAN network model includes a generator and a dual discriminator constructed by combining an autoencoder network; where:

[0017] The autoencoder network includes an encoding input layer, a first hidden layer of the encoder, a feature output layer, which is also the input layer of the decoder, a first hidden layer of the decoder, and a reconstructed output layer of the decoder; where the encoding input layer is used to input frequency-domain signals; the first hidden layer of the encoder and the feature output layer are both used for feature extraction; the first hidden layer of the decoder and the reconstructed output layer of the decoder are both used for feature reconstruction of the features output by the encoder;

[0018] The generator includes an input layer, multiple fully connected layers, and an adversarial output layer, where the third fully connected layer in the multiple fully connected layers is shared with the first hidden layer of the decoder, and the adversarial output layer is shared with the reconstructed output layer of the decoder;

[0019] The dual discriminator includes discriminator D1 and discriminator D2, which have the same composition structure but do not share weights. Both discriminator D1 and discriminator D2 include an input layer, a first fully connected layer, a second fully connected layer, and a score evaluation output layer.

[0020] Preferably, the implementation process of step 4) is as follows: Select a small number of samples of K types of samples from the N-1 types of faulty samples obtained in step 1) as the templates for the samples generated by the generator, denoted as true samples X; Select random noise with the same number and dimension of 100 as the input of the generator for the true samples X of the K fault types, denoted as random noise Z, and the generator generates false samples of fault type K, denoted as G K (Z), and then input the generated false samples G K (Z) and the true samples X into the dual discriminator for score evaluation.

[0021] Preferably, the implementation process of the score evaluation is as follows: the discriminator D1 is set to give a high score when the input sample is a true sample X, that is, D1(X) obtains a high score; when the input sample is a false sample G K (Z), the discriminator D1 gives a low score, that is, D1(G K (Z)) obtains a low score;

[0022] The second discriminator D2 of the double discriminator is set to give a low score when the input sample is a true sample X, that is, D2(X) obtains a low score; when the input sample is a false sample G K (Z), the discriminator D2 gives a high score, that is, D2(G K (Z)) obtains a high score.

[0023] Preferably, the loss objective functions of the discriminator D1 and the discriminator D2 are defined as:

[0024]

[0025] The loss objective function of the generator is defined as:

[0026]

[0027] where X represents the true sample, G(Z) represents the generated sample, Z represents the random noise, D 1 (X) and D 2 (X) respectively represent the scores of the discriminator D1 and D2 on the real sample, D 1 (G(Z)) and D 2 (G(Z) respectively represent the scores of the discriminator D1 and D2 on the false sample generated by the generator, represents the expectation of the real data distribution, represents the expectation of the generated data distribution, X feature represents the feature obtained by the decoder after feature extraction of the true sample X and decoding in the first hidden layer of the decoder, G(Z) feature represents the feature obtained by the generator through feature extraction of G(Z) in the third fully connected layer, and α, β, and λ are hyperparameters.

[0028] Preferably, when the generator and the double discriminator are iteratively trained to a certain extent, that is, when the loss functions of the generator and the double discriminator are basically stabilized to a fixed value during training, store the false samples generated by the generator during each iteration until the training ends.

[0029] Preferably, the implementation process of step 6) is as follows: Train the N - 1 types of fault samples in turn, and save the fault fake samples corresponding to each type generated by the generator during the training process of each type of fault sample and when the dual discriminators are iteratively trained to a certain extent. Stack the saved fault fake samples of each type in order. After stacking, the dataset dimension of each type of fault fake sample is [a, b]. Then embed this dataset into the corresponding class labels, and the dataset dimension of the class label is [a, c], where a is the number of each type of fault fake sample, b is the number of data points of a single fault fake sample, and c is the length of a single label. After the dataset of each type of fault fake sample is stacked and embedded with the corresponding class labels, supplement the same - type samples in the corresponding N - 1 types of fault samples according to the imbalance ratio until the total number of each type of fault sample after supplementation is the same as the total number of normal samples, and all types of samples reach balance. The imbalance ratio refers to the ratio of the number of each type of fault sample to the number of normal samples.

[0030] Preferably, the fault diagnosis model selects a deep convolutional neural network. The model training set used is the supplemented dataset of all types of samples including normal samples, and the test set is the dataset with the same sample categories and the same number of samples per category as those in the training set, which is re - collected by the LMS data acquisition system.

[0031] Beneficial effects:

[0032] (1) Using dual discriminators to form an adversarial neural network solves the problem of mode collapse in traditional generative adversarial networks.

[0033] (2) At the same time, embedding a stacked auto - encoder network into the adversarial network generation module, with the help of the input signal restoration ability of the auto - encoder network to assist the mutual adversarial process of the dual discriminators, making the generated samples closer and closer to the input samples, and then achieving the effect of confusing the two discriminators at the same time, so that the dual discriminators cannot give corresponding correct scores according to their respective criteria, and then achieving Nash equilibrium.

[0034] (3) The method proposed in the present invention can use a small number of samples to generate fake samples with a highly similar distribution, which highly supplements the original class - imbalanced dataset and provides an effective solution to the small - sample class - imbalance problem. Compared with directly using the small - sample class - imbalanced dataset for fault diagnosis, the proposed method has a substantial improvement in the fault diagnosis effect, with high diagnostic accuracy, good stability, and strong generalization ability. Description of the drawings

[0035] Figure 1 is the basic process of a small - sample class - imbalanced bearing intelligent fault diagnosis method proposed by the present invention;

[0036] Figure 2It is the structural diagram of the AD2GAN data generation model and the DCNN fault diagnosis model described in the present invention;

[0037] Figure 3 (a) is the spectrogram of three types of fault true samples in the present invention;

[0038] Figure 3 (b) is the spectrogram of three types of fault generated samples in the present invention;

[0039] Figure 4 It is the classification accuracy change curve of the method proposed in the present invention and two other methods under 4 imbalance rates;

[0040] Figure 5 It is the diagnostic result confusion matrix of the method proposed in the present invention and one other method under 4 imbalance rates. Specific implementation manner

[0041] The following further explains the present invention with reference to the accompanying drawings.

[0042] The present invention provides a bearing intelligent fault diagnosis method for small sample class imbalance, and its specific implementation steps are as follows:

[0043] Step1: Data acquisition. Install a vibration acceleration sensor on the bearing seat of the mechanical equipment, configure the LMS data acquisition system as required, then run the mechanical equipment, and start the data acquisition system to collect the time-series vibration acceleration signals of the bearing in various health states to obtain the original real dataset. Among them, the original real dataset contains 1 type of normal sample and N - 1 types of fault samples.

[0044] Step2: Data preprocessing. Perform data preprocessing on the collected original vibration signals. Convert the collected original vibration signals into frequency domain signals through the Fast Fourier Transform (FFT), and then perform 1-norm regularization processing on the frequency domain signals to scale each sample to the unit norm and retain the distribution of the original data.

[0045] Step3: Build a data generation model. Build an AD2GAN network model. Combine an autoencoder network to construct a generator, and use a dual discriminator structure for the discriminator. Combine the generator and the discriminator to construct a data generation model.

[0046] The auto-encoding network mainly consists of an input layer, a hidden layer, and an output layer. The input layer is used to input the original data or input feature vectors; the hidden layer is used to perform feature transformation on the input data; the output layer is used to reconstruct the features transformed by the hidden layer. The process of feature transformation from the input layer to the hidden layer is called the encoding process; the process of feature reconstruction from the hidden layer to the output layer is called the decoding process. The encoder input layer includes 600 neurons, the first hidden layer of the encoder includes 512 neurons, and the feature output layer includes 256 neurons; the decoder input layer includes 256 neurons, the first hidden layer of the decoder includes 512 neurons, and the decoder reconstruction output layer includes 600 neurons.

[0047] The generator consists of multiple fully-connected layers. Among them, the input layer includes 100 neurons; the first fully-connected layer includes 128 neurons, the second fully-connected layer includes 256 neurons, the third fully-connected layer includes 512 neurons, which is shared with the first hidden layer of the decoder, and the adversarial output layer includes 600 neurons, which is shared with the decoder reconstruction output layer.

[0048] The discriminator also consists of multiple fully-connected layers. The two discriminators have the same composition structure but do not share weights. Among them, the input layer includes 600 neurons, the first fully-connected layer includes 512 neurons, the second fully-connected layer includes 256 neurons, and the score evaluation output layer includes 1 neuron.

[0049] Step4: Model training, training of the AD2GAN network model. Select a few true samples of fault type K as the template for the samples generated by the generator, denoted as X; select a random noise with the same number as the true samples of class K and a dimension of 100 as the input of the generator, denoted as Z. Through forward propagation, the generator generates false samples of fault type K, denoted as G K (Z). Then, the generated false samples and true samples are simultaneously input into the two discriminators for score evaluation.

[0050] The first discriminator D1 of the two discriminators is set such that when the sample input into it is a true sample X, the discriminator D1 gives a high score, that is, D1(X) obtains a high score; when the sample input into it is a false sample G K (Z), the discriminator D1 gives a low score, that is, D1(G K (Z)) obtains a low score.

[0051] The second discriminator D2 of the two discriminators is set such that when the sample input into it is a true sample X, the discriminator D2 gives a low score, that is, D2(X) obtains a low score; when the sample input into it is a false sample G K (Z), the discriminator D2 gives a high score, that is, D2(G K(Z)) to obtain a high score. This setting mode of the two discriminators is to solve the mode collapse problem faced by GAN and generate more diverse fake data more effectively to improve the effect of fault diagnosis.

[0052] The generator and the dual discriminators are continuously iteratively trained according to their respective set loss objective functions. When the loss function gradually converges to a certain value, the fake samples generated by the generator become more and more similar to the real samples. At this time, the generator can deceive the two discriminators at the same time, making the two discriminators unable to correctly identify the true and false samples and give correct evaluation scores, and the model reaches the Nash equilibrium, and the training ends. The specific definitions of the loss objective functions of the discriminator and the generator are as follows:

[0053]

[0054]

[0055] where X represents the real sample, G(Z) represents the generated sample, and Z represents the random noise, represents the expectation of the real data distribution, represents the expectation of the generated data distribution, X feature represents the feature obtained by the decoder after feature extraction of X and decoding in the first hidden layer of the decoder, G(Z) feature represents the feature obtained by the generator after feature extraction of G(Z) in the third fully connected layer. α, β, and λ are hyperparameters, which are all set to 0.2 here.

[0056] Step5: Data generation, save the fault samples generated by the generator based on the original input signal. It includes that when the generator and the discriminator are iteratively trained to a certain extent, that is, when the loss function of the generator and the loss functions of the dual discriminators are basically stable to a fixed value during the training process, store the fake samples generated by the generator in each iteration until the training ends.

[0057] Step6: Class balance, balance the number of various fault samples and normal samples. Rotate the training of N-1 types of fault samples, and save the fake samples corresponding to their categories generated by the generator during the training of each type of fault sample. Stack the saved fault samples of the same type in the order of saving. After stacking, the dataset dimension of each type of fault sample is [1000, 600]. Then embed the corresponding category label into this dataset. The sample label with the fault type of K is denoted as K, and the dataset dimension of the K-class label is [1000, 1].

[0058] Among them, 1000 refers to the number of samples of each type of fault, 600 refers to the number of data points of a single sample of each type of fault, and 1 refers to the length of a single label.

[0059] After the above-mentioned various types of fault data are stacked and the corresponding category labels are embedded, the false samples generated by different types are used to supplement the corresponding real fault samples according to the imbalance ratio. After supplementation, the total number of each type of fault sample is the same as that of the normal samples, and all types of samples reach balance.

[0060] Step 7: Fault diagnosis. Use the balanced dataset to train the fault diagnosis model, and then use the real test set to test the trained fault diagnosis model to complete intelligent fault diagnosis. The fault diagnosis model selects a relatively mature intelligent fault diagnosis network - Deep Convolutional Neural Network (DCNN).

[0061] To more clearly and intuitively illustrate the effectiveness of the method proposed in the present invention, a specific example will be used for specific analysis and explanation below.

[0062] The dataset of this example comes from the data of the self-customized test bench of this research group. The test object is a roller bearing, the bearing model is NU205EM, and the rated power of the motor is 0.75 kW. The dataset mainly includes 4 types of samples in different health states: 1 type of normal sample and 3 types of fault samples. Among them, the 3 types of fault samples are respectively: rolling element fault, inner race fault, and outer race fault. Each type of health state sample contains 400 groups. 200 are selected as the test set, and the remaining 200 groups are used as the real training set. The length of each sample is 1200 data points, and after fast Fourier transform, it is 600 data points. Then, the 1-norm regularization process is performed on the frequency domain signal to scale each sample to the unit norm and retain the distribution of the original data. The real training set refers to the dataset without the supplementation of the generated false samples.

[0063] To verify the proposed small-sample class imbalance intelligent fault diagnosis method based on the mutual confrontation of auxiliary double discriminators in the present invention, 4 imbalance ratios are set, namely 1:100, 1:50, 1:20, and 1:10. Among them, the imbalance ratio refers to the ratio of the number of each type of fault sample to the number of normal samples. For example, for the imbalance ratio of 1:100, that is, there are 200 groups of normal samples and 2 groups of each type of fault sample. The other imbalance ratios are calculated in the same way. The number of each type of fault sample is randomly selected from the corresponding real training dataset. The specific description of the roller bearing is shown in Table 1, and the number of training samples with different imbalance ratios under different health conditions is shown in Table 2.

[0064] Table 1 Specific description of roller bearing

[0065]

[0066] Table 2 Number of training samples with different imbalance ratios under different health conditions

[0067]

[0068] Randomly select 20 groups of original signals with fault type K to train the AD2GAN model. Similarly, set 20 groups of random noises for generating fake samples. During the training process, the Batchsize is set to 1, the number of iteration steps is set to 20,000 steps, and the learning rate is set to 0.0001. When the number of iteration steps is greater than 15,000, the generated fake samples with category K are saved every 5 steps until the iteration is completed. When the model training ends, a total of 1,000 groups of generated samples with category K are saved. Train each type of fault sample in turn as above. When the model training ends, each type of fault sample also saves 1,000 groups of generated samples corresponding to its category. The saved generated samples are mainly used to supplement the imbalanced real samples of the same category.

[0069] To evaluate the quality of the generated samples, first, use the Pearson Correlation Coefficient (PCC) and Cosine Similarities (CS) evaluation metrics to evaluate the generated fake samples. The average evaluation results of 1,000 groups of generated fault samples for each category are shown in Table 3. It can be seen from Table 3 that the PCC of each fault type is greater than 0.7, and the CS is greater than 0.8, indicating that the generated fake samples have a high similarity with the original real samples.

[0070] Table 3 PCC and CS between the generated samples and the original samples corresponding to their categories under different fault types

[0071]

[0072] To more intuitively display the effect of the model-generated samples, the spectrograms of one group of real samples and the corresponding generated fake samples are respectively plotted, as Figure 3 shown. It can be more intuitively seen from the Figure 3 spectrogram comparison diagrams of real samples and generated samples of various faults that, after ignoring some small deviations, the generated fake samples are highly similar to the original real samples of the same category. In summary, the method proposed in the present invention can better learn the distribution according to the distribution of the original real samples and generate high-quality fake samples corresponding to their categories.

[0073] To further prove the fault diagnosis performance of the fault samples generated by the proposed method under the condition of small sample class imbalance, a deep convolutional neural network fault diagnosis algorithm is used for fault diagnosis. The training set includes three forms under each imbalance probability, namely, an unbalanced training set (without using the data generation method), a supplementary balanced training set (the method proposed in the present invention), and a full real training set (all using real original data); the test set includes one form, namely, a full real test set. Through various forms of training and testing under each imbalance rate, the average fault diagnosis results of the test set obtained after 10 consecutive trainings are as follows Figure 4 shown.

[0074] As can be seen Figure 4 from, after supplementing and balancing the original unbalanced training data set using the method proposed in the present invention, the diagnostic effect of the model has been substantially improved. Compared with the unbalanced data set generated without using any data generation method, the diagnostic accuracy of the balanced data set after data supplementation under 4 kinds of imbalance states has increased by 70%, 65%, 27%, and 23% in turn, and the diagnostic accuracy is generally maintained at about 97%. In addition, the diagnostic accuracy of the method proposed in the present invention is only about 2% lower than that of the full real data set using all real original data. To sum up, the method proposed in the present invention can well handle the small sample class imbalance problem, greatly improve the fault diagnosis accuracy of small sample class imbalance, and has a better diagnostic effect.

[0075] To more intuitively illustrate the fault diagnosis effect of using the method proposed in the present invention compared with not using any data generation method, the confusion matrices of the above two methods under 4 kinds of imbalance rates are respectively drawn, as shown in Figure 5 shown. It can be seen Figure 5 from that the number of misclassified samples of each class in the confusion matrix after supplementing and balancing using the method proposed in the present invention is relatively small compared with the original unbalanced confusion matrix, especially when the imbalance ratio is small, this situation is more obvious. In addition, through calculation, it can be obtained that the diagnostic accuracy of the method proposed in the present invention under each imbalance condition has a small difference, and the general error is maintained at 1%-3%. Therefore, this also further illustrates that the method proposed in the present invention can well handle the small sample class imbalance problem, has a high diagnostic accuracy, and strong stability.

[0076] The series of detailed descriptions listed above are only the preferred embodiments of the present invention, but it should be noted that they do not limit the protection scope of the present invention. Any addition, deletion, and replacement made on the core basis of the present invention shall be included in the protection scope of the present invention.

Claims

1. An intelligent fault diagnosis method for bearings with small sample class imbalance, characterized in that, it includes the following steps: 1) Data acquisition: Use a vibration acceleration sensor to collect the original vibration signal of the bearing; 2) Data preprocessing: Perform data preprocessing on the original vibration signal collected in step 1); 3) Network model construction and training: Construct an auxiliary dual discriminator generative adversarial network AD2GAN network model and train the model; 4) Data generation: Save the fake samples generated by the model according to the input signal; 5) Class balance: Balance the number of various fault samples and normal samples; 6) Fault diagnosis: Use the class-balanced data set to train the fault diagnosis model, and then use the test set to test the trained fault diagnosis model to complete intelligent fault diagnosis; The implementation process of step 3) is as follows: The AD2GAN network model includes a generator constructed by combining an autoencoder network and a dual discriminator; where: The autoencoder network includes an encoding input layer, a first hidden layer of the encoder, a feature output layer, which is also the decoder input layer, a first hidden layer of the decoder, and a decoder reconstruction output layer; where the encoding input layer is used to input frequency domain signals; the first hidden layer of the encoder and the feature output layer are both used for feature extraction; the first hidden layer of the decoder and the decoder reconstruction output layer are both used for feature reconstruction of the features output by the encoder; The generator includes an input layer, multiple fully connected layers, and an adversarial output layer, where the third fully connected layer in the multiple fully connected layers is shared with the first hidden layer of the decoder, and the adversarial output layer is shared with the decoder reconstruction output layer; The dual discriminator includes discriminator D1 and discriminator D2, which have the same composition structure but do not share weights. Both discriminator D1 and discriminator D2 include an input layer, a first fully connected layer, a second fully connected layer, and a score evaluation output layer; The implementation process of step 4) is as follows: Among the N - 1 types of fault samples obtained in step 1), a small number of samples of K types of samples are selected as the templates for the samples generated by the generator, denoted as true samples X; Random noise with the same number as the true samples X of the K fault types and a dimension of 100 is selected as the input of the generator, denoted as random noise Z. The generator generates false samples of fault type K, denoted as G K (Z), and then the generated false samples G K (Z) and the true samples X are simultaneously input into the dual discriminator for score evaluation; Define the loss objective functions of discriminator D1 and discriminator D2 as: Define the loss objective function of the generator as: Among them, X represents the real sample, G(Z) represents the generated sample, Z represents the random noise, D 1 (X) and D 2 (X) respectively represent the scores of the discriminators D1 and D2 on the real sample, D 1 (G(Z)) and D 2 (G(Z) respectively represent the scores of the discriminators D1 and D2 on the fake samples generated by the generator, represents the expectation of the real data distribution, represents the expectation of the generated data distribution, X feature represents the feature obtained by decoding at the first hidden layer of the decoder after the decoder extracts features from the real sample X, G(Z) feature represents the feature obtained by the generator extracting features from G(Z) at the third fully connected layer, and α, β, and λ are hyperparameters.

2. An intelligent fault diagnosis method for bearings with small sample class imbalance as described in claim 1, characterized in that, The implementation process of step 1) is as follows: Install the vibration acceleration sensor on the bearing seat of the mechanical equipment, configure the LMS data acquisition system, then run the mechanical equipment, start the LMS data acquisition system to collect the time-series vibration acceleration signals of the bearing in various states as the original vibration signal of the bearing, and obtain the original real data set of N types of samples, where there is 1 type of normal sample and N - 1 types of fault samples.

3. An intelligent fault diagnosis method for bearings with small sample class imbalance as described in claim 2, characterized in that, The implementation process of step 2) is to convert the collected original vibration signal of the bearing into a frequency domain signal through fast Fourier transform, and then perform 1-norm regularization processing on the frequency domain signal to scale the frequency domain signal of each sample to the unit norm, while preserving the distribution of the original vibration signal.

4. An intelligent fault diagnosis method for bearings with small sample class imbalance as described in claim 1, characterized in that, The implementation process of the score evaluation is as follows: the discriminator D1 is set to give a high score when the input sample is a true sample X, that is, D1(X) obtains a high score; when the input sample is a false sample G K (Z), the discriminator D1 gives a low score, that is, D1(G K (Z)) obtains a low score; The second discriminator D2 of the double discriminator is set such that when the sample input into it is a true sample X, the discriminator D2 gives a low score, that is, D2(X) obtains a low score; when the sample input into it is a false sample G K (Z), the discriminator D2 gives a high score, that is, D2(G K (Z)) obtains a high score.

5. An intelligent fault diagnosis method for bearings with small sample class imbalance as described in claim 1, characterized in that, When the generator and the double discriminator are iteratively trained to a certain extent, that is, when the loss functions of the generator and the double discriminator are basically stabilized to a fixed value during the training process, the fake samples generated by the generator in each iteration are stored until the training ends.

6. A small-sample class-imbalanced bearing intelligent fault diagnosis method according to claim 5, characterized in that the implementation process of step 6) is as follows: alternately train the N-1 types of fault samples, and save the fault fake samples corresponding to the category generated by the generator when the generator and the double discriminator are iteratively trained to a certain extent during the training process of each type of fault sample. Stack the saved fault fake samples of each type in order. After stacking, the dataset dimension of each type of fault fake sample is [a, b]. Then embed this dataset into the corresponding category label, and the dataset dimension of the category label is [a, c], where a is the number of each type of fault fake sample, b is the number of data points of a single fault fake sample, and c is the length of a single label. After the dataset of each type of fault fake sample is stacked and embedded with the corresponding category label, each type of fault fake sample is used to supplement the same type of samples in the corresponding N-1 types of fault samples according to the imbalance ratio until the total number of each type of fault sample after supplementation is the same as the total number of normal samples, and all types of samples reach balance. The imbalance ratio refers to the ratio of the number of each type of fault sample to the number of normal samples.

7. A small-sample class-imbalanced bearing intelligent fault diagnosis method according to claim 6, characterized in that the fault diagnosis model selects a deep convolutional neural network, and the model training set used is the dataset of all types of samples including normal samples after supplementation, and the test set is the dataset with the same sample categories and the same number of samples in each category re-collected by the LMS data acquisition system.

Images