A Cross-Domain Fault Diagnosis Method for Rotating Machinery Based on a Multi-Source Subdomain Adaptation Network

Abstract

The present invention discloses a cross-domain fault diagnosis method for rotating machinery based on a multi-source sub-domain adaptation network, including: acquiring vibration signals of a faulty bearing; performing wavelet packet transform on the vibration signals; introducing a deep convolutional generative adversarial network to augment target domain data; obtaining shared features of multi-source domains through a pre-trained deep residual network; applying a multi-branch parallel structure and using local maximum mean discrepancy to align sub-domains of each source domain and the target domain; setting a weighted module according to the magnitude of the local mean discrepancy metric loss; adopting a selection criterion of minimum loss and closest distance to achieve joint diagnosis of multi-source domains. The present invention is applicable to fault diagnosis of rotating machinery with few samples under multiple working conditions, and can effectively improve the stability and reliability of small-sample bearing fault diagnosis under different working conditions as a whole.

Description

Technical Field

[0001] The present invention belongs to the field of small-sample vibration signal processing and cross-domain fault diagnosis of rotating machinery, and particularly relates to a cross-domain fault diagnosis method for rotating machinery based on a multi-source subdomain adaptation network. Background Art

[0002] Rotating machinery is widely used in power devices such as engines, fans, and diesel engines. Due to the harsh operating conditions of the power devices (low-speed heavy-load conditions, overloaded operation, alternating shocks, and loads), rotating machinery is extremely prone to failure. Research shows that faults caused by rotating components such as bearings and gears account for more than 40% of the total mechanical faults, indicating that rotating machinery is more prone to failure than other components. At the same time, due to the complex working conditions of rotating machinery, it is difficult to sample. Therefore, conducting research on fault diagnosis of small-sample vibration signals of rotating machinery can effectively improve the limitations of traditional fault diagnosis relying on manual experience.

[0003] Vibration signals often accompany faults, and the signals contain rich fault features. In recent years, due to the rapid development of deep learning algorithms, more and more deep learning methods have also been applied to the field of fault diagnosis. However, in actual equipment working application scenarios, it is difficult and expensive to obtain labeled samples. Therefore, it is difficult to obtain sufficient labeled data for deep learning training. This makes deep learning methods based on large amounts of data no longer applicable. Therefore, there is an urgent need to develop a fault diagnosis method for rotating machinery based on scarce samples to achieve feature extraction and fault diagnosis of vibration signals of rotating machinery under the condition of scarce samples. Summary of the Invention

[0004] Object of the Invention: To overcome the deficiencies in the prior art, a cross-domain fault diagnosis method for rotating machinery based on a multi-source subdomain adaptation network is provided, which is applicable to cross-domain fault diagnosis of rotating machinery under the condition of scarce samples.

[0005] Technical Solution: To achieve the above object, the present invention provides a cross-domain fault diagnosis method for rotating machinery based on a multi-source subdomain adaptation network, including the following steps:

[0006] S1: Obtain vibration signals of the rotating machinery under different rotational speed conditions;

[0007] S2: Introduce a deep convolutional generative adversarial network as a sample generator to perform data augmentation on the small-sample vibration signals obtained in step S1;

[0008] S3: Use the trained deep residual network as a shared feature extractor to perform shared feature extraction on the vibration signals augmented in step S2 and multiple source domain vibration signal datasets;

[0009] S4: Apply a multi-branch parallel structure as the domain-specific feature extractor and domain-specific feature classifier, and use the local maximum mean discrepancy to align the sub-domains of each source domain and target domain;

[0010] S5: Set a weighting module according to the magnitude of the loss measured by the local mean discrepancy;

[0011] S6: Take the minimum loss and the closest distance as the classifier selection criterion to achieve multi-source domain joint diagnosis and determine the fault type.

[0012] Further, the acquisition method of the vibration signal in step S1 is as follows:

[0013] Use an acceleration sensor to collect the vibration signals x1(t), x2(t), x3(t), x4(t) of the rotating machinery;

[0014] Further, the process of deep convolutional generative adversarial network sample augmentation in step S2 is as follows:

[0015] A1: Input the generated signal and the real signal into the discriminator, and let the discriminator distinguish between the real and generated signals, that is, maximize the objective function V(D, G);

[0016] A2: Input the generated signal with a real label into the training discriminator, and require the generated data to deceive the discriminator, that is, minimize the objective function V(D, G). The rise and fall of V(D, G) form an adversarial relationship, and the network finds a balance between the two to generate more similar signal data;

[0017] A3: The objective function to be optimized is:

[0018]

[0019] In the formula: D(x) is the probability that the discriminator judges whether the real data is real; D[G(Z)] is the probability that the discriminator judges whether the generated signal is real; P data is the real sample distribution; P G is the prior distribution of the vector Z.

[0020] Further, the method of shared feature extraction in step S3 is as follows:

[0021] Since the transferability of shallow features is relatively good, they can be frozen after pre-training and then fine-tuned. Concentrate the main computational resources on the extraction of domain-specific features, thereby improving the training efficiency. For this part, VggNet and ResNet can be used as the backbone networks, and the input and output layers can be adjusted.

[0022] The feature extraction of the deep residual network mainly includes three parts: local perception, parameter sharing, and pooling;

[0023] Local perception means that the network is partially connected. Each neuron is only connected to some neurons in the previous layer, perceiving only locally rather than the entire signal. Local perception is achieved through a sliding window. The adjacent parts of the signal are closely related, while the correlation of the more distant parts is weak. Therefore, only local perception is needed, and the global information can be obtained by integrating the local information at a higher layer. Weight sharing means that the information learned from a local area is applied to other parts of the signal. That is, the entire signal is convolved with an identical convolution kernel, which is equivalent to filtering the signal. Different features are achieved by multiple different convolution kernels. Pooling, such as max pooling, takes the maximum value of a region. Therefore, when the signal undergoes small changes such as translation and scaling, it is still very likely to obtain the maximum value at the same position, with the same response as before the change, thus achieving affine invariance. The same applies to mean pooling. After a small affine change occurs, the mean value may remain unchanged.

[0024] Further, the specific steps of step S4 are as follows:

[0025] B1: Set a specific feature space for each source domain; considering only a single network branch, the loss measured by LMMD is:

[0026]

[0027] B2: Weight the losses of each specific feature space. The source domain and target domain losses measured by LMMD are expressed as follows:

[0028]

[0029] where ω j is the weight corresponding to each source domain, G j (·) is the source-specific feature extractor. By weighting each source domain, the actual distance between the reorganized source domain and the target domain can be obtained;

[0030] B3: The domain-specific classifier receives the output features of the domain-specific feature extractor and outputs its probability distribution. Its classification loss is:

[0031]

[0032] The source domain dataset is:

[0033] X S ={Z S1 ,Z S2 ,Z S3 ,...Z SL}

[0034] where Z Sl is a subclass of X S , and L is the number of subcategories;

[0035] B4: Distance metric after mapping the same subclasses of the source domain and the target domain:

[0036]

[0037] B5: Recombining the dataset of the source domain and the classification loss is:

[0038]

[0039] Furthermore, in step S5, a weighted module is set according to the size of the loss measured by the local mean difference, and its formula is:

[0040] The weight allocator ω(j) is essentially a binary classifier that determines the source of the subdomain based on the distribution distance between each source domain and the target domain:

[0041]

[0042] Among them, the source domain serial number j ∈ {1, 2, 3... N}, and the subdomain serial number l ∈ {1, 2, 3... L}.

[0043] Furthermore, in step S6, the criterion for selecting the classifier is to minimize the loss and make the distance the closest, so as to realize the multi-source domain joint diagnosis. The fault type judgment formula is:

[0044]

[0045] Furthermore, an objective function optimization module is set, and the parameters of the shared feature extractor, domain-specific feature extractor, and domain-specific classifier are continuously adjusted to minimize the classification loss and the domain adaptation loss.

[0046] The method for setting the objective function optimization module is:

[0047] The parameters of each source domain for the domain-specific feature extractor and the domain-specific classifier are independent. The weight allocator and the task classifier are responsible for the result diagnosis after weight allocation. The total loss of the network is the LMMD domain adaptation loss and the classification loss:

[0048] L total = L cls + λL LMMD

[0049] The goal of the function optimization module is to minimize L cls and L LMMD , and the overall loss is minimized by continuously adjusting the parameters of the shared feature extractor, domain-specific feature extractor, and domain-specific classifier. At the same time, it is obviously unrealistic to optimize the parameters of all three. We need to fix two of the parameters unchanged to optimize the remaining one. The optimization method is as follows:

[0050]

[0051] Among them, θ F , θ G , θ C are respectively the optimizable parameters of the shared feature extractor, the domain-specific feature extractor, and the domain-specific classifier, and

[0052] In the present invention, the SGD iteration method is used to train θ F , θ G , θ C , and gradually approaches the optimal value through random gradients, so as to obtain

[0053] In the present invention, the deep convolutional generative adversarial network expands the small-sample target domain dataset; secondly, multi-source domain shared features are obtained through the network branch structure; thirdly, the local maximum mean discrepancy is used to align each sub-domain of the source domain and the target domain; finally, a weighted module is adopted to minimize the global loss, realizing multi-source domain joint diagnosis.

[0054] Aiming at the problems of few samples of rotating machinery fault signals, large differences between the source domain and the target domain, and small distances between sub-domains of the target domain, the present invention mainly includes four parts: data augmentation by the deep convolutional generative adversarial network, obtaining multi-source domain shared features through the network branch structure, aligning each sub-domain of the source domain and the target domain by the local maximum mean discrepancy, and minimizing the global loss by adopting a weighted module to realize multi-source domain joint diagnosis. The deep convolutional generative adversarial network is used as a sample generator to expand the target domain samples; the network branch structure is used as a shared feature extractor to obtain multi-source domain shared features, while reducing the complexity of subsequent network training and saving computing resources; the domain-specific feature extractor and the domain-specific classifier combine the local maximum mean discrepancy to align each sub-domain of the source domain and the target domain, thereby improving the integrity of the transferred information; a weighted module is adopted to minimize the global loss and realize multi-source domain joint diagnosis. Therefore, the present invention proposes a fault diagnosis method for multi-source sub-domain transfer learning to realize cross-domain fault diagnosis of rotating machinery with small-sample vibration signals.

[0055] Beneficial effects: Compared with the prior art, the present invention is applicable to the fault diagnosis of rotating machinery for small-sample cross-domain vibration signals, which is mainly divided into four parts: data augmentation by deep convolutional generative adversarial network, obtaining multi-source domain shared features by network branch structure, alignment of sub-domains in the source domain and the target domain, and implementing joint diagnosis of multi-source domains by using a weighted module. Data augmentation by deep convolutional generative adversarial network increases the data volume of small-sample vibration signals, which can effectively prevent the overfitting phenomenon caused by a small sample size during model training; the network branch structure obtains the shared features of multi-source domains, and by screening out the shared features of samples, the computational complexity in subsequent migration processes is reduced; local maximum mean discrepancy aligns each sub-domain in the source domain and the target domain, and by increasing the number of source domains, the expansion of transferable features in the source domain is achieved; the weighted module is used to minimize the global loss by constructing an appropriate loss function. As a result, the present invention can effectively improve the stability and reliability of small-sample rotating machinery cross-domain fault diagnosis as a whole. Description of the Drawings

[0056] Figure 1 It is a schematic diagram of the framework of the present invention;

[0057] Figure 2 It is a diagram of a deep convolutional generative adversarial network;

[0058] Figure 3 It is a structural diagram of the VALENIAN-PT500 rotating machinery test bench;

[0059] Figure 4 It is a diagram comparing the accuracies of the present invention and various algorithms on the CWRU dataset;

[0060] Figure 5 It is a diagram comparing the accuracies of the present invention and various algorithms on the VP500 dataset;

[0061] Figure 6 It is a confusion matrix diagram of the recognition accuracies of various categories of the present invention on the CWRU dataset;

[0062] Figure 7 It is a confusion matrix diagram of the recognition accuracies of various categories of the present invention on the VP500 dataset;

[0063] Figure 8 It is a t-sne visualization diagram of the feature extraction of the present invention for the CWRU dataset;

[0064] Figure 9 It is a t-sne visualization diagram of the feature extraction of the present invention for the VP500 dataset;

[0065] Figure 10 It is a diagram comparing the accuracies of the present invention and various algorithms on the CWRU dataset when taking different proportions of target domain samples. Detailed Embodiments

[0066] The present invention will be further clarified below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, various equivalent forms of modification by those skilled in the art fall within the scope defined by the appended claims of this application.

[0067] The present invention provides a cross-domain fault diagnosis method for rotating machinery based on a multi-source sub-domain adaptation network, as Figure 1 shown, including the following steps:

[0068] S1: Obtain small-sample vibration signals of the rotating machinery under different rotational speed conditions;

[0069] S2: Introduce a deep convolutional generative adversarial network as a sample generator to perform data augmentation on the small-sample vibration signals obtained in step S1;

[0070] S3: Use the trained deep residual network as a shared feature extractor to extract shared features from the vibration signals augmented in step S2 and multiple source-domain vibration signal datasets;

[0071] S4: Apply a multi-branch parallel structure as a domain-specific feature extractor and a domain-specific feature classifier, and use the local maximum mean discrepancy to align the sub-domains of each source domain and the target domain;

[0072] S5: Set a weighting module according to the magnitude of the local mean discrepancy metric loss;

[0073] S6: Adopt the criterion of the smallest loss and the closest distance as the classifier selection criterion to achieve multi-source domain joint diagnosis and determine the fault type;

[0074] S7: Set an objective function optimization module to minimize the classification loss and the domain adaptation loss by continuously adjusting the parameters of the shared feature extractor, the domain-specific feature extractor, and the domain-specific classifier.

[0075] Based on the above solution, in this embodiment, the above rotating machinery fault diagnosis method is applied as an example, specifically as follows:

[0076] The experimental data of this example comes from the CWRU public dataset and the VALENIAN-PT500 rotating machinery test bench. The test bench is as Figure 3As shown. The module numbers of the experimental gears 1 - 7 are all 25. Therefore, the test can be carried out by easily sliding and meshing any gear between the good gear and the faulty gear. The transmission ratio of the gearbox is 4.97:1 and it consists of two - stage gears. The number of teeth of each stage of gears is 58:25 and 60:28 respectively. By sliding the gears along the intermediate shaft, the normal gear on the intermediate shaft can mesh with the gear on the input shaft, and the other gear on the intermediate shaft can also mesh with the gear on the output shaft. The experimental bearing is the conventional model UC206 of rotating machinery, and the bearing structure parameters are shown in Table 1.

[0077] Table 1 Structural parameters of UC206 rotating machinery

[0078]

[0079] During the experiment, the speed of the motor was adjusted by the frequency converter to make the gearbox operate at variable speeds. During the experiment, the fault signals of the gears and bearings were collected by vibration acceleration sensors. Dewesoft was used for vibration signal acquisition. The signal sampling frequency was 10,000 Hz and the sampling duration was 3 s.

[0080] The simulated fault types adopted in this embodiment are the local fault of the inner - ring crack of the rotating machinery and the gear wear fault.

[0081] The overall framework of the network is as Figure 1 shown, and the specific diagnosis process is as follows:

[0082] 1) The acquisition method of vibration signals is:

[0083] Vibration signals x1(t), x2(t), x3(t), x4(t) of the rotating machinery are collected by acceleration sensors;

[0084] 2) Sample expansion of the deep convolutional generative adversarial network, as Figure 2 shown, and its process is:

[0085] A1: The generated signal and the real signal are input into the discriminator, and the discriminator is required to distinguish the real and the generated signals as much as possible, that is, to maximize the objective function V(D, G);

[0086] A2: The generated signal is labeled as real and input into the trained discriminator. The generated data is required to deceive the discriminator, that is, to minimize the objective function V(D, G). The rise and fall of V(D, G) form an adversarial relationship, and the network finds the balance between the two to generate more similar signal data;

[0087] A3: The objective function to be optimized is:

[0088]

[0089] Where: D(x) is the probability that the discriminator determines whether the real data is real; D[G(Z)] is the probability that the discriminator determines whether the generated signal is real; P data is the real sample distribution; P G is the prior distribution of the vector Z.

[0090] 3) Use the trained deep residual network as a shared feature extractor to extract shared features from the denoised multi-source vibration signals:

[0091] The feature extraction of the deep residual network mainly includes three parts: local perception, parameter sharing, and pooling;

[0092] Local perception means that the network is partially connected, and each neuron is only connected to some neurons in the previous layer, only perceiving the local part rather than the entire signal; local perception is achieved through a sliding window. The adjacent parts of the signal are closely related, and the correlation of the farther parts is weak. Therefore, only local perception is needed, and the global information can be obtained by integrating the local information at a higher layer. Weight sharing is the information learned from a local area and applied to other parts of the signal. That is, convolving the entire signal with the same convolution kernel is equivalent to filtering the signal. Different features are achieved by multiple different convolution kernels. Pooling, such as max pooling, is to take the maximum value of a region. Therefore, when the signal undergoes small changes such as translation and scaling, it is still very likely to obtain the maximum value at the same position, and the response is the same as before the change, thus achieving affine invariance. The same is true for mean pooling. After a small affine change, the mean may still remain unchanged

[0093] Since the transferability of shallow features is relatively good, they can be frozen after pre-training and then fine-tuned. Concentrate the main computing resources on the extraction of domain-specific features, thereby improving the training efficiency. For this part, VggNet and ResNet can be used as the backbone networks, and the input and output layers can be adjusted.

[0094] 4) Apply a multi-branch parallel structure as the domain-specific feature extractor and domain-specific feature classifier:

[0095] B1: It is necessary to set specific feature spaces for each source domain. Considering only a single network branch, the loss measured by LMMD is:

[0096]

[0097] B2: Weight the losses of each specific feature space. The source domain and target domain losses measured by LMMD are expressed as follows:

[0098]

[0099] Among them, ω j is the weight corresponding to each source domain, Gj (·) is a source-specific feature extractor. By weighting each source domain, the actual distance between the reorganized source domain and the target domain can be obtained;

[0100] B3: The domain-specific classifier receives the output features of the domain-specific feature extractor and outputs its probability distribution. Its classification loss is:

[0101]

[0102] The source domain dataset is:

[0103] X S ={Z S1 ,Z S2 ,Z S3 ,...Z SL}

[0104] where Z Sl is a subclass of X S , and L is the number of subcategories;

[0105] B4: Distance metric after mapping the same subclasses of the source domain and the target domain:

[0106]

[0107] B5: The dataset and classification loss of the reorganized source domain are:

[0108]

[0109] 5) Set the weighting module according to the size of the loss measured by the local mean difference:

[0110] The weight allocator ω(j) is essentially a binary classifier, which determines the source of the subdomain based on the distribution distance between each source domain and the target domain:

[0111]

[0112] where the source domain serial number j ∈ {1, 2, 3... N}, and the subdomain serial number l ∈ {1, 2, 3... L}.

[0113] 6) Adopt the criterion of minimizing the loss and the closest distance as the classifier selection criterion to achieve multi-source domain joint diagnosis:

[0114]

[0115] 7) Set the objective function optimization module:

[0116] Each source domain is independent of the parameters of the domain-specific feature extractor and the domain-specific classifier. The weight allocator and the task classifier are responsible for the result diagnosis after weight allocation. The total loss of the network is the LMMD domain adaptation loss and the classification loss:

[0117] L total = L cls + λL LMMD

[0118] The goal of the function optimization module is to minimize L cls and L LMMD , and the overall loss is minimized by continuously adjusting the parameters of the shared feature extractor, the domain-specific feature extractor, and the domain-specific classifier. However, it is obviously unrealistic to optimize the parameters of all three at the same time. It is necessary to fix two of the parameters and optimize the remaining one. The optimization method is as follows:

[0119]

[0120] θ F ,θ G ,θ C are the optimizable parameters of the shared feature extractor, the domain-specific feature extractor, and the domain-specific classifier respectively, is the finally obtained optimal parameter. In this embodiment, the SGD iteration method is used to train θ F ,θ G ,θ C , and gradually approaches the optimal value through the stochastic gradient to obtain

[0121] To verify the effectiveness of the multi-source sub-domain adaptation network proposed in the present invention, this embodiment is tested on the CWRU public dataset and the VP500 dataset measured on the built test platform, and the accuracy is compared with various domain adaptation methods. The test accuracies on the CWRU dataset and the VP500 dataset are shown in Figure 4 and Figure 5 . It can be seen from the figure that the accuracy of the method proposed in the present invention is higher than that of the current single-source domain test network on the test accuracies of the two datasets; the specific accuracies of each category on the two datasets are shown in Figure 6 and Figure 7 . To clearly show the feature representation of the network feature mapping, the t-distributed stochastic neighbor embedding (TSNE) visualization analysis is performed on the final output features of the network, as shown in Figure 8 and Figure 9 . Among them, the projection coincidence degree of the target domain and the source domain is high, the boundary is clear, and the recognition degree is extremely high, which proves the excellent generalization ability of the method proposed in the present invention on different datasets. From Figure 10It can be seen that when the number of samples of other domain adaptation methods is reduced to a certain extent, the recognition accuracy drops significantly, indicating the occurrence of overfitting. It can be seen that when the data is insufficient, the proposed method can make the most of the limited labeled data.

Claims

1. A cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network, characterized in that, It includes the following steps: S1: Obtain the vibration signals of the rotating machinery under different rotational speed conditions; S2: Introduce a deep convolutional generative adversarial network as a sample generator to perform data augmentation on the small-sample vibration signals obtained in step S1; S3: Use the trained deep residual network as a shared feature extractor to extract shared features from the vibration signals augmented in step S2 and multiple source domain vibration signal datasets; S4: Apply a multi-branch parallel structure as a domain-specific feature extractor and a domain-specific feature classifier, and use local maximum mean discrepancy to align each sub-domain of the source domain and the target domain; S5: Set a weighted module according to the magnitude of the loss measured by the local mean discrepancy; S6: Adopt the criterion of minimum loss and closest distance as the classifier selection criterion to achieve multi-source domain joint diagnosis and determine the fault type; Step S4 is specifically as follows: B1: Set a specific feature space for each source domain; only considering a single network branch, the loss measured by LMMD is: B2: Weight the losses of each specific feature space, and the losses of the source domain and the target domain measured by LMMD are expressed as follows: Among them, ω j is the weight corresponding to each source domain, and G j (·) is the source-specific feature extractor. By weighting each source domain, the actual distance between the recombined source domain and the target domain can be obtained; B3: The domain-specific classifier receives the output features of the domain-specific feature extractor and outputs its probability distribution, and its classification loss is: The source domain dataset is: X S = {Z S1 , Z S2 , Z S3 ,... Z SL} Among them, Z Sl is a subclass of X S , and L is the number of subcategories; B4: Measure the distance after mapping the same sub-classes of the source domain and the target domain; B5: Reorganize the dataset and classification loss of the source domain as: In step S5, a weighted module is set according to the magnitude of the loss measured by the local mean discrepancy, and its formula is: The weight allocator ω(j) is essentially a binary classification discriminator, which determines the source of the sub-domain based on the distribution distance between each source domain and the target domain: Among them, the source domain serial number j ∈ {1, 2, 3... N}, and the sub-domain serial number l ∈ {1, 2, 3... L}.

2. The cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 1, characterized in that, The acquisition method of the vibration signal in step S1 is: Use an acceleration sensor to collect the vibration signals x1(t), x2(t), x3(t), x4(t) of the rotating machinery.

3. The cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 1, characterized in that, The process of deep convolutional generative adversarial network sample augmentation in step S2 is: A1: Input the generated signal and the real signal into the discriminator, and the discriminator distinguishes the real and the generated signals, that is, maximize the objective function V(D, G); A2: Input the generated signal with a real label into the training discriminator, and require the generated data to deceive the discriminator, that is, minimize the objective function V(D, G). The rise and fall of V(D, G) form an adversarial situation, and the network finds the balance between the two to generate more similar signal data; A3: The objective function to be optimized is: Where: D(x) is the probability that the discriminator judges whether the real data is real; D[G(Z)] is the probability that the discriminator judges whether the generated signal is real; P data is the real sample distribution; P G is the prior distribution of the vector Z.

4. A cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 1, characterized in that, The method of shared feature extraction in step S3 is: The feature extraction by the deep residual network includes three parts: local perception, parameter sharing, and pooling.

5. A cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 1, characterized in that, In step S6, the criterion of minimum loss and closest distance is adopted as the classifier selection criterion to achieve multi-source domain joint diagnosis, and its fault type judgment formula is:

6. A cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 1, characterized in that, Set an objective function optimization module, and minimize the classification loss and the domain adaptation loss by continuously adjusting the parameters of the shared feature extractor, the domain-specific feature extractor, and the domain-specific classifier.

7. A cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 6, characterized in that, The method of setting the objective function optimization module is: The total loss of the network is the LMMD domain adaptation loss and the classification loss: L total = L cls + λL LMMD The goal of the function optimization module is to minimize L cls and L LMMD , and the overall loss is minimized by continuously adjusting the parameters of the shared feature extractor, domain-specific feature extractor, and domain-specific classifier. The optimization method is as follows: Among them, θ F , θ G , θ C are the optimizable parameters of the shared feature extractor, the domain-specific feature extractor, and the domain-specific classifier respectively, is the finally obtained optimal parameter.

8. A cross - domain fault diagnosis method for rotating machinery based on a multi - source sub - domain adaptation network according to claim 7, characterized in that, The acquisition method is: training θ using the SGD iterative method F , θ G , θ C , gradually approaching the optimal value through the stochastic gradient, so as to obtain

Images