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

Abstract

A method for fault diagnosis of rotating machinery based on an improved auxiliary classification generative adversarial network, step 1, data collection: collecting vibration signals of key transmission components of rotating machinery in various health states; step 2, data preprocessing: using short-time Fourier transform to convert the collected one-dimensional time domain signal into a two-dimensional time-frequency image; step 3, model building: constructing an auxiliary classification generative adversarial network, improving the network structure and loss function; step 4, model training: using a small amount of labeled data and a large amount of unlabeled data to perform adversarial training of the generator and the discriminator; step 5, data generation and fault diagnosis: using the trained generator to generate multi-mode samples, and using the trained discriminator to identify faults. The present invention realizes multi-mode sample generation and fault diagnosis under limited label data. Under different label ratios, the method can not only generate high-quality multi-mode samples, but also has high fault diagnosis accuracy and stability.

Description

Technical Field

[0001] The present invention belongs to the technical field of rotating machinery fault diagnosis, and specifically is a rotating machinery fault diagnosis method based on an improved auxiliary classification generative adversarial network. Background Art

[0002] With the vigorous development of modern industrial systems, rotating machinery plays an irreplaceable role in intelligent equipment, and its safety has received extensive attention from academia and industry. Key transmission components of rotating machinery (such as bearings and gears) will inevitably suffer from wear, corrosion, deformation, cracks and other faults under complex operating environments such as high speed and heavy load. The failure of transmission components directly affects the operating reliability of rotating machinery, which may cause major accidents, economic losses and even casualties. Therefore, fault diagnosis and predictive maintenance of rotating machinery have important research value.

[0003] In recent years, fault diagnosis methods based on deep learning have gradually become a research hotspot due to their powerful fault feature learning capabilities and end-to-end diagnostic characteristics, including autoencoders (AE), deep belief networks (DBN), convolutional neural networks (CNN), and recurrent neural networks (RNN). Although the above models have shown effectiveness in the field of fault diagnosis, they rely on a large amount of labeled data for supervised learning. However, it is difficult to obtain enough labeled samples in actual engineering, resulting in serious overfitting problems in supervised models. Therefore, it is worth studying how to use limited labeled samples to train accurate and reliable deep learning-based fault diagnosis models.

[0004] Generative adversarial network (GAN), as one of the key technologies of unsupervised generative models, is created based on zero-sum game theory and mainly consists of generator and discriminator. GAN can generate new samples similar to existing samples using only unlabeled samples. Semi-supervised generative adversarial network (SGAN) is a semi-supervised generative model improved from GAN. It extends GAN to semi-supervised scenarios by forcing the discriminator to output category labels. The discriminator of traditional GAN ​​uses Sigmoid function as output, while SGAN uses Softmax function as output. For a data set with N categories, the discriminator expands the output to N+1 dimensions, including 1-dimensional discriminant output and N-dimensional classification output. Auxiliary classification generative adversarial network (ACGAN) is a supervised generative model improved from GAN. It embeds category labels as auxiliary information into random noise input to guide the generator to generate multi-modal samples. The discriminator of ACGAN uses both Sigmoid function and Softmax function as output to achieve discrimination and classification of input samples. In recent years, ACGAN has been used to solve the problems of multi-modal sample generation and fault diagnosis.

[0005] The above research shows that SGAN is a semi-supervised learning mechanism, but it cannot generate multimodal samples due to the lack of category information; ACGAN can generate multimodal samples by using category labels as auxiliary information, but its supervised learning mechanism requires the participation of a large amount of labeled data.

[0006] Compared with the prior art, this application has the following advantages:

[0007] Technical comparison with patent CN114861789A "A method and system for fault diagnosis of rotating machinery based on generative adversarial mechanism"

[0008] Patent CN114861789 A belongs to the field of rotating machinery fault diagnosis. Although this patent also belongs to the field of rotating machinery fault diagnosis, the difference between them is that the problems they solve are different. The problem that patent CN114861789 A aims to solve is only the fault diagnosis of rotating machinery, while the problem that this patent aims to solve includes two parts: multi-mode sample generation and fault identification of rotating machinery.

[0009] The deep learning model used in patent CN114861789 A is an autoencoder, which introduces the generative adversarial mechanism into the autoencoder to construct an autoencoder deep learning network based on the generative adversarial mechanism; while the deep learning model used in this patent is an auxiliary classification generative adversarial network, which introduces the semi-supervised learning mechanism into the auxiliary classification generative adversarial network to construct a semi-supervised auxiliary classification generative adversarial network. There is an essential difference between the two.

[0010] Technical comparison with patent CN112308038 B "Mechanical equipment fault signal identification method based on classroom-style generative adversarial network model"

[0011] Patent CN112308038 B belongs to the field of mechanical fault diagnosis, and this patent is aimed at fault diagnosis of rotating machinery and belongs to the field of mechanical fault diagnosis. The difference is that patent CN112308038 B is to solve the problem of low accuracy of existing mechanical equipment fault signal recognition methods, while this patent is to solve the problem of multi-mode sample generation and fault diagnosis of rotating machinery under limited label data.

[0012] The deep learning model used in patent CN112308038 B is a generative adversarial network, which constructs a classroom-based generative adversarial network model containing a generator and multiple discriminators; while the deep learning model used in this patent is an auxiliary classification generative adversarial network, which introduces a semi-supervised learning mechanism into the auxiliary classification generative adversarial network and constructs a semi-supervised auxiliary classification generative adversarial network. Although both belong to generative adversarial networks, there are essential differences in their principles and structures. Summary of the invention

[0013] In order to solve the above technical problems, the present invention proposes a rotating machinery fault diagnosis method based on an improved auxiliary classification generative adversarial network to achieve multi-modal sample generation and fault diagnosis under limited label data.

[0014] To achieve the above object, the technical solution adopted by the present invention is:

[0015] The rotating machinery fault diagnosis method based on the improved auxiliary classification generative adversarial network has the following specific steps:

[0016] Step 1: Data collection:

[0017] Collect vibration signals of key transmission components of rotating machinery under various health conditions;

[0018] Step 2: Data preprocessing:

[0019] The one-dimensional time domain signal samples of various health states are converted into two-dimensional time-frequency images using short-time Fourier transform;

[0020] The short-time Fourier transform is a function of time and frequency, as shown in the following formula:

[0021]

[0022] Where STFT(·) represents short-time Fourier transform, x(t) represents a one-dimensional time domain signal, t and w represent time and frequency, and g(st) represents a window function centered at time t.

[0023] The time and frequency resolution of the two-dimensional time-frequency image obtained by short-time Fourier transform depends on the length of the window function, as shown in the following formula:

[0024]

[0025]

[0026]

[0027] Among them, N x Indicates the length of the signal to be processed, N w Indicates the length of the window function, N o represents the overlapping length during the window function translation process, [·] represents rounding down, and X represents the time-frequency matrix with a dimension of F×T obtained by short-time Fourier transform;

[0028] Step 3: Model building:

[0029] Construct auxiliary classification generative adversarial network and improve network structure and loss function;

[0030] Step 4: Model training:

[0031] Use a small amount of labeled data and a large amount of unlabeled data to perform adversarial training between the generator and the discriminator;

[0032] Step 5: Data generation and fault diagnosis:

[0033] The trained generator is used to generate multi-modal samples, and the trained discriminator is used to identify faults.

[0034] As a further improvement of the present invention, in step one, the collected vibration signals include vibration signals in K healthy states, i.e., normal states, and K-1 fault states, and one-dimensional time domain signal samples are obtained by sliding window sampling.

[0035] As a further improvement of the present invention, in step 2, the one-dimensional time domain signal samples of various health states are converted into two-dimensional time-frequency images using short-time Fourier transform, and all samples are normalized to [-1,1], and then the sample set is split into a training set and a test set, where the training set is divided into labeled training samples and unlabeled training samples according to proportion.

[0036] As a further improvement of the present invention, in step three, the constructed auxiliary classification generative adversarial network includes a generator G and a discriminator D. The generator G is composed of an embedding layer and a deconvolution block, and its input is a noise vector and a label vector, and its output is a generated sample; the discriminator D is composed of a convolution block and a fully connected layer, and its input is labeled samples, unlabeled samples and generated samples, and its output is a discrimination result and a classification result.

[0037] As a further improvement of the present invention, the input of the generator G is a 200-dimensional Gaussian random noise vector z and a category label vector c, the label vector c is embedded in the noise vector z through an embedding layer, and a sample is generated after five deconvolution blocks are output, and each deconvolution block is connected in sequence to include: a deconvolution layer Deconv, a batch normalization layer BN, and an activation function layer;

[0038] The convolution kernel size of the first four deconvolution blocks is 3×3, the stride is 2×2, and the activation function is ReLU; the convolution kernel size of the last deconvolution block is 4×4, the stride is 2×2, and the activation function is Tanh. The number of channels of the five deconvolution blocks are 200, 64, 32, 16, and 1, respectively.

[0039] As a further improvement of the present invention, the input of the discriminator D is a real sample or a generated sample, and a 1-dimensional discriminant vector and a K-dimensional classification vector are output after three convolution blocks and two fully connected layers. Each convolution block is connected in sequence to include: a convolution layer Conv, a batch normalization layer BN, an activation function layer, and a discard layer Dropout. The convolution kernel size of the first convolution block is 5×5, the stride is 4×4, and the padding is 2×2; the convolution kernel size of the last two convolution blocks is 5×5, the stride is 2×2, and the padding is 2×2. The activation functions of the three convolution blocks are all Leaky ReLU, the number of channels is 32, 64, and 128 respectively, and the two fully connected layers FC use Sigmoid and Softmax functions as outputs respectively.

[0040] As a further improvement of the present invention, the loss function of the auxiliary classification generative adversarial network includes the loss function of the generator G and the loss function of the discriminator D, as shown in the following formula:

[0041]

[0042]

[0043]

[0044]

[0045] Among them, L G , L D Represent the loss functions of the generator G and the discriminator D respectively; Respectively represent the supervised loss and unsupervised loss of the discriminator D; λ1 and λ2 are scaling factors; P r(x) , P r(x,y) They represent the distribution of unlabeled true samples and labeled true samples respectively, c is the label of the labeled true sample, P(c=c|x) and P(c=c|G(z,c)) represent the conditional probability distribution of the category labels of the labeled true samples and the generated samples respectively.

[0046] As a further improvement of the present invention, in step 4, the generator G and the discriminator D are trained against each other using labeled training samples and unlabeled training samples to simultaneously improve the multi-modal sample generation capability of the generator G and the recognition capability of the discriminator D. The training process is as follows:

[0047] (1) Input the noise vector z and the label vector c into the generator G to obtain the generated sample;

[0048] (2) Input the labeled training samples, unlabeled training samples and generated samples into the discriminator D to obtain a 1-dimensional discriminant vector and a K-dimensional classification vector;

[0049] (3) Calculate the loss of the generator G and the discriminator D;

[0050] (4) Fix the parameters of the generator G and optimize the parameters θ of the discriminator D D ;

[0051] (5) Fix the parameters of the discriminator D and optimize the parameters θ of the generator G G ;

[0052] (6) Repeat steps 1-5 until the number of iterations is reached;

[0053] (7) Save the trained generator G and discriminator D.

[0054] As a further improvement of the present invention, the adversarial training of the generator G and the discriminator D uses the Adam optimization algorithm to update the model parameters. The Adam algorithm provides independent adaptive learning rates for different model parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, as shown in the following formula:

[0055]

[0056]

[0057]

[0058] Among them, θ G ,θ D They represent the model parameters of the generator G and the discriminator D, respectively. α, β1, and β2 represent the learning rate of the Adam optimizer, the exponential decay rate of the first-order moment estimate, and the exponential decay rate of the second-order moment estimate, respectively. These are the optimal points of the model found by the error back propagation derivative method.

[0059] Compared with the prior art, the advantages of the present invention are:

[0060] (1) The present invention uses short-time Fourier transform to convert a one-dimensional time-domain vibration signal into a two-dimensional time-frequency image. Short-time Fourier transform is a joint time-frequency analysis method for time-varying and non-stationary signals. It uses a fixed-length window function to translate on the time-domain signal, intercept the signal and perform Fourier transform to obtain a local spectrum set for each time period. Therefore, the two-dimensional time-frequency image matrix obtained by short-time Fourier transform contains both time domain information and frequency domain information. It carries richer features than the original time domain signal, represents a more complex structural distribution, and is helpful for the training of fault diagnosis models based on deep learning.

[0061] (2) The present invention combines the characteristics of the semi-supervised generative adversarial network SGAN and the auxiliary classification generative adversarial network ACGAN to construct a semi-supervised auxiliary classification generative adversarial network SACGAN, and improves the network structure and loss function. By using the semi-supervised learning mechanism of SGAN, the generation, discrimination and classification capabilities are improved through adversarial training of a small amount of labeled data and a large amount of unlabeled data, and ACGAN is transformed from supervised learning to semi-supervised learning, realizing multi-modal sample generation and fault diagnosis under limited labeled data.

[0062] (3) The present invention uses a small amount of labeled data and a large amount of unlabeled data to perform adversarial training between the generator and the discriminator, thereby simultaneously improving the multi-modal sample generation capability of the generator and the recognition capability of the discriminator. The method can not only generate high-quality multi-modal samples, but also has high fault diagnosis accuracy and stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 is a flow chart of the method of the present invention;

[0064] Figure 2 The time domain vibration signals of various health states of the bearing and the converted time-frequency images;

[0065] Figure 3 A schematic diagram of the structure of the improved auxiliary classification generative adversarial network of the present invention;

[0066] Figure 4 It is a structural schematic diagram of the generator of the present invention;

[0067] Figure 5 It is a schematic diagram of the structure of the discriminator of the present invention;

[0068] Figure 6 A training flow chart for generating an adversarial network for improved auxiliary classification of the present invention;

[0069] Figure 7 Schematic diagram of the generation of real samples of various health conditions of bearings into multi-mode generated samples;

[0070] Figure 8 SSIM quality assessment plots of multi-mode generated samples of various bearing health states;

[0071] Fig. 9 Recognition results of test samples under different label ratios of bearing data. DETAILED DESCRIPTION

[0072] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] In this embodiment, if Figure 1 As shown, a rotating machinery fault diagnosis method based on an improved auxiliary classification generative adversarial network mainly includes: step one, data acquisition: collecting vibration signals of key transmission components of rotating machinery (such as bearings and gears) in various health states; step two, data preprocessing: using short-time Fourier transform to convert the collected one-dimensional time domain signal into a two-dimensional time-frequency image; step three, model building: constructing an auxiliary classification generative adversarial network, and improving the network structure and loss function; step four, model training: using a small amount of labeled data and a large amount of unlabeled data to perform adversarial training of the generator and the discriminator; step five, data generation and fault diagnosis: using the trained generator to generate multi-modal samples, and using the trained discriminator to identify faults.

[0074] In this embodiment, step 1, the vibration data of 10 health states of the bearing are collected, including normal state and 9 fault states (3 fault locations and 3 fault sizes). The vibration data of various health states are sampled by sliding window (window length is 1024) to obtain 300 one-dimensional time domain signals as available samples. The health state and category label of the bearing are shown in Table 1 below:

[0075]

[0076] In step 2, short-time Fourier transform is used to convert the one-dimensional time-domain signal samples of various health states into two-dimensional time-frequency images.

[0077] The short-time Fourier transform is a function of time and frequency, as shown in the following equation (1):

[0078]

[0079] Where STFT(·) represents short-time Fourier transform, x(t) represents a one-dimensional time domain signal, t and w represent time and frequency, and g(st) represents a window function centered at time t.

[0080] The time and frequency resolution of the two-dimensional time-frequency image obtained by short-time Fourier transform depends on the length of the window function, as shown in the following equations (2), (3), and (4):

[0081]

[0082]

[0083]

[0084] Among them, N x Indicates the length of the signal to be processed, N w Indicates the length of the window function, N o represents the overlapping length during the window function translation process, [·] represents rounding down, and X represents the time-frequency matrix with a dimension of F×T obtained by short-time Fourier transform.

[0085] In this embodiment, the window function of the short-time Fourier transform is a "hann" window, and the window function length N w With overlap length N o 256 and 250 respectively.

[0086] In this embodiment, in step 2, the one-dimensional time domain signal samples (length 1024) of various health states of the bearing are converted into two-dimensional time-frequency images using short-time Fourier transform, and all samples are normalized to [-1, 1], and then geometrically processed into 64×64 image samples. Each health state of the bearing contains 300 time-frequency image samples, of which 100 samples are taken as test samples, and the remaining samples are divided into labeled training samples and unlabeled training samples according to different proportions.

[0087] In this embodiment, if Figure 2 As shown in the figure, the one-dimensional time domain vibration signals of the bearing in various health states are converted into two-dimensional time-frequency images. It can be seen that the characteristic differences between different health states are more significant than the original time domain signals.

[0088] In this embodiment, if Figure 3 As shown in Figure 1, in step 3, the constructed semi-supervised auxiliary classification generative adversarial network SACGAN includes a generator G and a discriminator D. The generator G consists of an embedding layer and a deconvolution block, with the input being a noise vector and a label vector, and the output being a generated sample; the discriminator D consists of a convolution block and a fully connected layer, with the input being labeled samples, unlabeled samples, and generated samples, and the output being the discrimination results and classification results.

[0089] In this embodiment, if Figure 4As shown, the input of the generator G is a 200-dimensional Gaussian random noise vector z and a category label vector c. The label vector c is embedded into the noise vector z through an embedding layer, and the generated sample is output through five deconvolution blocks. Each deconvolution block is connected in sequence: a deconvolution layer Deconv, a batch normalization layer BN, and an activation function layer. The convolution kernel size of the first four deconvolution blocks is 3×3, the stride is 2×2, and the activation function is ReLU; the convolution kernel size of the last deconvolution block is 4×4, the stride is 2×2, and the activation function is Tanh. The number of channels of the five deconvolution blocks is 200, 64, 32, 16, and 1, respectively.

[0090] In this embodiment, if Figure 5 As shown, the input of the discriminator D is a real sample or a generated sample, and it outputs a 1-dimensional discriminant vector and a K-dimensional classification vector after three convolution blocks and two fully connected layers. Each convolution block is connected in sequence to include: a convolution layer Conv, a batch normalization layer BN, an activation function layer, and a dropout layer Dropout. The convolution kernel size of the first convolution block is 5×5, the stride is 4×4, and the padding is 2×2; the convolution kernel size of the last two convolution blocks is 5×5, the stride is 2×2, and the padding is 2×2. The activation function of the three convolution blocks is Leaky ReLU, and the number of channels is 32, 64, and 128 respectively. The two fully connected layers FC use Sigmoid and Softmax functions as outputs respectively.

[0091] The loss function of the SACGAN includes the loss function of the generator G and the loss function of the discriminator D, as shown in the following equations (5), (6), (7), and (8):

[0092]

[0093]

[0094]

[0095]

[0096] Among them, L G , L D Represent the loss functions of the generator G and the discriminator D respectively; Respectively represent the supervised loss and unsupervised loss of the discriminator D; λ1 and λ2 are scaling factors; P r(x) , P r(x,y) They represent the distribution of unlabeled true samples and labeled true samples respectively, c is the label of the labeled true sample, P(c=c|x) and P(c=c|G(z,c)) represent the conditional probability distribution of the category labels of the labeled true samples and the generated samples respectively.

[0097] In this embodiment, the loss function L of the discriminator D is D The proportional factors λ1 and λ2 are both 0.5.

[0098] In this embodiment, if Figure 6 As shown in Figure 1, in step 4, the generator G and the discriminator D are trained against each other using labeled training samples and unlabeled training samples to simultaneously improve the multi-modal sample generation capability of the generator G and the recognition capability of the discriminator D. The training process is as follows:

[0099] (1) Input the noise vector z and the label vector c into the generator G to obtain the generated samples.

[0100] (2) Input the labeled training samples, unlabeled training samples and generated samples into the discriminator D to obtain a 1-dimensional discriminant vector and a K-dimensional classification vector.

[0101] (3) Calculate the loss of the generator G and the discriminator D.

[0102] (4) Fix the parameters of the generator G and optimize the parameters θ of the discriminator D D .

[0103] (5) Fix the parameters of the discriminator D and optimize the parameters θ of the generator G G .

[0104] (6) Repeat steps 1 to 5 until the number of iterations is reached.

[0105] (7) Save the trained generator G and discriminator D.

[0106] The adversarial training of the generator G and the discriminator D uses the Adam optimization algorithm to update the model parameters. The Adam algorithm provides independent adaptive learning rates for different model parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, as shown in the following equations (9), (10), and (11):

[0107]

[0108]

[0109]

[0110] Among them, θ G ,θ D They represent the model parameters of the generator G and the discriminator D, respectively. α, β1, and β2 represent the learning rate of the Adam optimizer, the exponential decay rate of the first-order moment estimate, and the exponential decay rate of the second-order moment estimate, respectively. These are the optimal points of the model found by the error back propagation derivative method.

[0111] In this embodiment, the hyperparameter configuration of the SACGAN training is as follows: the learning rate α of the Adam algorithm is 0.0005, the exponential decay rate β1 of the first-order moment estimation is 0.5, the exponential decay rate β2 of the second-order moment estimation is 0.999, the batch size is K×10, and the number of iterations is 200.

[0112] In this embodiment, SACGAN is trained using training samples with a label ratio of 0.2 (40 labeled samples of each type and 160 unlabeled samples). The sample generation capability of the generator G is usually verified by evaluating the quality of the generated samples, such as Figure 7 As shown in the figure, it is a schematic diagram of the generation of real samples of various health conditions of bearings into multi-modal generated samples. It can be seen that the generated samples are highly similar to the real samples.

[0113] In this embodiment, in order to more objectively evaluate the sample generation ability of the generator G, the structural similarity (SSIM) is used to quantitatively evaluate the quality of the generated samples. SSIM aims to measure the similarity of two images in terms of brightness, contrast and structure. A larger SSIM value indicates that the two images have a higher similarity. For comparison, ACGAN was trained under the same conditions. Five pairs of real samples and generated samples were randomly selected, such as Figure 8 As shown in Figure 2, the average SSIM values ​​of ACGAN and SACGAN are compared. It can be seen that the SSIM values ​​of all types of SACGAN are greater than those of ACGAN, which indicates that the generated samples of SACGAN are closer to the real samples.

[0114] In this embodiment, in order to verify the fault identification ability of the discriminator D, various training samples are divided into labeled samples and unlabeled samples according to different proportions to train SACGAN. For comparison, ACGAN and the CNN of the discriminator D are trained under the same conditions. Each group of experiments is repeated ten times and the results are averaged. Fig. 9 As shown in the figure, the recognition results of test samples under different label ratios. The results show that all three models have high recognition accuracy, which illustrates the effectiveness of the discriminator D structure; ACGAN and SACGAN effectively improve the recognition ability of the discriminator D through adversarial learning; limiting the number of labeled training samples has a greater impact on the fault recognition ability of CNN and ACGAN, while SACGAN can weaken the impact of the reduction in label ratio and shows the highest recognition accuracy under different label ratios.

[0115] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A rotating machinery fault diagnosis method based on an improved auxiliary classification generative adversarial network, characterized in that: The specific steps are as follows: Step 1: Data collection: Collect vibration signals of key transmission components of rotating machinery under various health conditions; Step 2: Data preprocessing: The one-dimensional time domain signal samples of various health states are converted into two-dimensional time-frequency images using short-time Fourier transform; The short-time Fourier transform is a function of time and frequency, as shown in the following formula: Where STFT(·) represents short-time Fourier transform, x(t) represents a one-dimensional time domain signal, t and w represent time and frequency, and g(st) represents a window function centered at time t. The time and frequency resolution of the two-dimensional time-frequency image obtained by short-time Fourier transform depends on the length of the window function, as shown in the following formula: Among them, N x Indicates the length of the signal to be processed, N w Indicates the length of the window function, N o represents the overlapping length during the window function translation process, [·] represents rounding down, and X represents the time-frequency matrix with a dimension of F×T obtained by short-time Fourier transform; Step 3: Model building: Construct auxiliary classification generative adversarial network and improve network structure and loss function; In step 3, the constructed auxiliary classification generative adversarial network includes a generator G and a discriminator D. The generator G consists of an embedding layer and a deconvolution block, the input is a noise vector and a label vector, and the output is a generated sample; the discriminator D consists of a convolution block and a fully connected layer, the input is a labeled sample, an unlabeled sample, and a generated sample, and the output is a discrimination result and a classification result; the loss function of the auxiliary classification generative adversarial network includes the loss function of the generator G and the loss function of the discriminator D, as shown in the following formula: Among them, L G , L D Represent the loss functions of the generator G and the discriminator D respectively; Respectively represent the supervised loss and unsupervised loss of the discriminator D; λ1 and λ2 are scaling factors; P r(x) , P r(x,y) They represent the distribution of unlabeled true samples and labeled true samples respectively, c is the label of the labeled true sample, P(c=c|x) and P(c=c|G(z,c)) represent the conditional probability distribution of the category labels of the labeled true samples and the generated samples respectively; Step 4: Model training: Use a small amount of labeled data and a large amount of unlabeled data to perform adversarial training between the generator and the discriminator; Step 5: Data generation and fault diagnosis: The trained generator is used to generate multi-modal samples, and the trained discriminator is used to identify faults.

2. The rotating machinery fault diagnosis method based on improved auxiliary classification generative adversarial network according to claim 1 is characterized in that: In step 1, the collected vibration signals include vibration signals in K healthy states, i.e., normal states, and K-1 fault states, and one-dimensional time domain signal samples are obtained by sliding window sampling.

3. The rotating machinery fault diagnosis method based on improved auxiliary classification generative adversarial network according to claim 1 is characterized in that: In step 2, short-time Fourier transform is used to convert the one-dimensional time domain signal samples of various health states into two-dimensional time-frequency images, and all samples are normalized to [-1,1]. Then the sample set is split into a training set and a test set, where the training set is divided into labeled training samples and unlabeled training samples in proportion.

4. The rotating machinery fault diagnosis method based on improved auxiliary classification generative adversarial network according to claim 1 is characterized in that: The input of the generator G is a 200-dimensional Gaussian random noise vector z and a category label vector c. The label vector c is embedded into the noise vector z through an embedding layer, and the generated samples are output through five deconvolution blocks. Each deconvolution block is connected in sequence and includes: A deconvolution layer Deconv, a batch normalization layer BN, and an activation function layer; The convolution kernel size of the first four deconvolution blocks is 3×3, the stride is 2×2, and the activation function is ReLU; the convolution kernel size of the last deconvolution block is 4×4, the stride is 2×2, and the activation function is Tanh. The number of channels of the five deconvolution blocks are 200, 64, 32, 16, and 1, respectively.

5. The rotating machinery fault diagnosis method based on improved auxiliary classification generative adversarial network according to claim 1 is characterized in that: The input of the discriminator D is a real sample or a generated sample, and a 1-dimensional discriminant vector and a K-dimensional classification vector are output after three convolution blocks and two fully connected layers. Each convolution block is connected in sequence to include: a convolution layer Conv, a batch normalization layer BN, an activation function layer, and a discard layer Dropout. The convolution kernel size of the first convolution block is 5×5, the stride is 4×4, and the padding is 2×2; the convolution kernel size of the last two convolution blocks is 5×5, the stride is 2×2, and the padding is 2×2. The activation functions of the three convolution blocks are all LeakyReLU, the number of channels is 32, 64, and 128 respectively, and the two fully connected layers FC use Sigmoid and Softmax functions as outputs respectively.

6. The rotating machinery fault diagnosis method based on improved auxiliary classification generative adversarial network according to claim 1 is characterized in that: In step 4, the generator G and the discriminator D are trained adversarially using labeled training samples and unlabeled training samples to simultaneously improve the multi-modal sample generation capability of the generator G and the recognition capability of the discriminator D. The training process is as follows: (1) Input the noise vector z and the label vector c into the generator G to obtain the generated sample; (2) Input the labeled training samples, unlabeled training samples and generated samples into the discriminator D to obtain a 1-dimensional discriminant vector and a K-dimensional classification vector; (3) Calculate the loss of the generator G and the discriminator D; (4) Fix the parameters of the generator G and optimize the parameters θ of the discriminator D D ; (5) Fix the parameters of the discriminator D and optimize the parameters θ of the generator G G ; (6) Repeat steps 1-5 until the number of iterations is reached; (7) Save the trained generator G and discriminator D.

7. The rotating machinery fault diagnosis method based on improved auxiliary classification generative adversarial network according to claim 1 is characterized in that: The adversarial training of the generator G and the discriminator D uses the Adam optimization algorithm to update the model parameters. The Adam algorithm provides independent adaptive learning rates for different model parameters by calculating the first-order moment estimate and the second-order moment estimate of the gradient, as shown in the following formula: i G ←Adam(▽ θG L G ,i G ,a,b1,b2) i D ←Adam(▽ θD L D ,i D ,a,b1,b2) Among them, θ G ,θ D They represent the model parameters of the generator G and the discriminator D, respectively. α, β1, and β2 represent the learning rate of the Adam optimizer, the exponential decay rate of the first-order moment estimate, and the exponential decay rate of the second-order moment estimate, respectively. These are the optimal points of the model found by the error back propagation derivative method.

Images