Cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network

Abstract

The invention provides a cross-domain fault diagnosis method based on a self-learning convolution domain adversarial network, and the method comprises the following steps: constructing a self-adaptive feature extractor based on self-learning convolution, and achieving the dynamic perception of signal features through a convolution parameter self-learning mechanism; constructing a self-learning convolution domain adversarial network comprising a self-learning convolution-based self-adaptive feature extractor, a domain classifier with a gradient inversion layer and a fault classifier; and inputting the source domain data set and the target domain data set into the self-learning convolutional domain adversarial network by adopting a cross-domain joint training strategy, synchronously optimizing cross-domain feature distribution confusion loss and source domain classification loss, and realizing feature alignment of the source domain and the target domain. According to the method, the advantage that self-learning convolution can dynamically adjust convolution kernel parameters according to different domain data is fully utilized, so that the model better adapts to cross-domain data distribution with difference in the feature extraction process, invariant features of a source domain and a target domain can be effectively extracted and aligned, and the cross-domain fault diagnosis performance of the model is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of mechanical fault diagnosis, and in particular relates to a cross-domain fault diagnosis method based on a self-learning convolutional domain adversarial network. Background Art

[0002] Failures in key components of mechanical equipment can lead to equipment downtime or even accidents, making efficient fault diagnosis technology crucial. In the field of fault diagnosis, domain adaptation technology has become an important tool for addressing cross-operational and cross-device diagnostic challenges. Domain adversarial networks (DANNs), by aligning feature distributions in the source and target domains, mitigate the diagnostic challenges associated with data distribution differences across different operating conditions or devices. By constructing a domain discriminator, DANNs force the feature extractor to generate domain-invariant features, thereby enabling knowledge transfer across different operating conditions or devices.

[0003] However, traditional DANNs have significant limitations. Their feature extraction modules are primarily based on fixed-structure convolutional layers. In practical applications, mechanical vibration signals exhibit complex dynamic characteristics due to changes in operating conditions (such as load and speed). Fixed-structure convolutional layers lack flexibility and struggle to adaptively capture these changing fault characteristics under varying operating conditions. This results in an inability to fully exploit the effective information in the signal during key fault feature extraction, significantly limiting the accuracy of cross-domain diagnosis and making it difficult to meet the high-precision requirements for bearing fault diagnosis under complex and variable operating conditions in real-world engineering applications.

[0004] Therefore, how to break through the constraints of the traditional convolutional layer structure and design a new convolutional layer that can dynamically adjust the convolution kernel structure according to the working conditions to enhance the model's ability to extract mechanical fault features under variable working conditions has become a technical problem that needs to be solved in the current field of mechanical fault diagnosis. Summary of the Invention

[0005] To solve the above problems, the present invention proposes a cross-domain fault diagnosis method based on a self-learning convolutional domain adversarial network. Through a unique self-learning convolution, the structure of the convolution kernel is flexibly and adaptively changed according to the characteristic differences of the vibration signal of the mechanical equipment under different working conditions (such as load, speed change, etc.). This self-learning convolution significantly enhances the robustness of the model to cross-domain distribution differences through domain-sensitive adaptive reorganization of the convolution kernel's multi-sub-core structure. Compared with the traditional fixed-structure convolution layer, the dynamic offset characteristics of its sub-core spatial position can explicitly compensate for the time-frequency feature offset caused by working condition differences. Driven by adversarial training, it achieves alignment of the feature distribution of the source domain and the target domain, effectively extracts domain-invariant fault features, and improves the accuracy of cross-domain fault diagnosis.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] The cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network includes the following steps:

[0008] S1: Build an adaptive feature extractor based on self-learning convolution, and realize dynamic perception of signal features through the self-learning mechanism of convolution parameters;

[0009] S2: Construct a self-learning convolutional domain adversarial network including an adaptive feature extractor based on self-learning convolution, a domain classifier with a gradient reversal layer, and a fault classifier.

[0010] S3: A cross-domain joint training strategy is adopted to input the source domain and target domain datasets into the self-learning convolutional domain adversarial network, and the cross-domain feature distribution confusion loss and the source domain classification loss are simultaneously optimized to achieve the alignment of the source domain and target domain feature spaces, thereby realizing cross-domain fault diagnosis.

[0011] Furthermore, the parameter learning mechanism of the self-learning convolution in S1 is specifically as follows:

[0012] Perform a standard convolution operation on the input signal to determine the batch dimension b and sequence length l of the output feature map;

[0013] The original convolution kernel W∈R Ks , divided into n independently adjustable sub-cores, generating a sub-core group W sub , the sub-nuclear distance is the self-learning parameter:

[0014] W sub =Reshape(W,[n,K s / n]) (1)

[0015] Dynamic adjustment of sub-core spacing is achieved by constructing convolution position index:

[0016] Based on the step size s of the standard convolution, an s×l sequence is constructed and the dimension is expanded to form [b, K s , l] tensor, generating the position index p0 when the standard convolution operation is performed:

[0017]

[0018] Among them, p0∈Z b×Ks×l , the value of each element is determined only by index i, which is equal to s×i and has nothing to do with the dimensions of b and Ks;

[0019] Then construct K s dimensional sequence, and expanded to [b, K s , 1], generate the original convolution kernel position index p1:

[0020]

[0021] Where p1∈Zb×Ks×1 , the value of each element is determined only by index j, equal to j, and has nothing to do with the dimension b. The third dimension is always 0;

[0022] The two are superimposed to obtain the global position index p2 of the standard convolution operation:

[0023]

[0024] Among them, the third dimension of p1 is automatically expanded to l and added element-by-element with p0 to generate the domain-sensitive global position encoding p2;

[0025] A lightweight convolutional layer is further introduced to calculate the domain offset between sub-nuclei. After channel compression and nonlinear activation, n-1 sub-nuclei spacing parameters are extracted.

[0026] Finally, through differentiable sub-kernel position remapping, the dynamic spacing parameter is fused with the position encoding p2 to construct a self-learning convolution with domain invariance perception.

[0027] The core architecture of the adaptive feature extractor based on self-learning convolution in S1 is a multi-layer convolutional neural network:

[0028] The first layer uses self-learning convolution, and the input is the original vibration signal;

[0029] The subsequent layers stack m layers of standard convolutional layers, normalization layers, pooling layers, and activation function layers in sequence.

[0030] Furthermore, the construction of the adversarial domain adaptation network architecture in S2 includes the following core components:

[0031] (a) The adaptive feature extractor based on self-learning convolution according to claim 1;

[0032] (b) The domain classifier is connected to the feature extractor through a gradient reversal layer, and its forward propagation process performs:

[0033] Receive the feature vector output by the feature extractor and output the domain classification prediction result;

[0034] Back propagation process is performed:

[0035] The gradient reversal layer multiplies the input gradient by a negative coefficient, and the calculation formula is:

[0036]

[0037] Among them, x is the feature extractor output feature, L is the domain classification loss, is the gradient from the domain classifier, and λ is the gradient reversal strength (usually set to 1.0 or dynamically adjusted);

[0038] (c) Fault classifier, which is directly connected to the feature extractor and accepts domain-invariant features from the source domain for fault pattern recognition;

[0039] Among them, the domain discriminator and the fault classifier are connected in parallel to the output of the feature extractor. The fault classifier is optimized by minimizing the classification loss of the source domain samples, and the domain discriminator is optimized by minimizing the feature distribution confusion loss.

[0040] Optionally, the domain classifier connected by the gradient reversal layer is composed of multiple layers of fully connected layers and nonlinear activation functions, and outputs the predicted probability of the domain label. The feature distribution confusion loss adopts the binary cross entropy loss with class balance weight, and its expression is:

[0041]

[0042] Among them, L domain is the domain discrimination loss, N is the total number of samples, and the N samples are summed and averaged when calculating the loss, y d (i) is the domain belonging identifier of the i-th sample, which takes the value of 0 (source domain) or 1 (target domain). D(·) is the domain classifier, the input is the feature, and the output is the probability that the sample belongs to the domain label. f(x (i) ) is the feature extractor f for the i-th sample x (i) The extracted cross-domain features, α is the balance coefficient, and the calculation formula is:

[0043]

[0044] Among them, N target and N source are the sample sizes of the source domain and target domain respectively.

[0045] Optionally, the fault classifier is composed of a fully connected layer network, and the cross entropy classification loss is calculated based on the source domain labeled data, and its expression is as follows:

[0046]

[0047] Among them, N s is the number of source domain samples, K is the number of fault categories, y k (i) is the true fault category label of the source domain sample, is the predicted probability of the kth class by the fault classifier.

[0048] Furthermore, in S3, the cross-domain joint training strategy is:

[0049] Input the source domain dataset (with fault labels and domain labels) and the target domain dataset (with domain labels) into the feature extractor;

[0050] The self-learning convolutional layer performs:

[0051] Dynamically adjust the convolution kernel based on the sub-kernel spacing parameter self-learning mechanism and perform convolution operations on the input signal;

[0052] The source domain and target domain features are input to the domain classifier, which outputs the feature distribution confusion loss. The source domain features are input to the fault classifier alone, which outputs the source domain classification loss.

[0053] The feature distribution confusion loss and source domain classification loss are optimized simultaneously, and the total loss function is defined as:

[0054] L total =L domain +L class (9)

[0055] Among them, L domain is the feature distribution confusion loss, L class is the source domain classification loss.

[0056] Testing phase execution:

[0057] The target domain test data is input into the feature extractor, the extracted feature vector is input into the fault classifier, and the fault classifier outputs the fault mode prediction label.

[0058] The beneficial effects of the present invention are as follows:

[0059] The present invention proposes a cross-domain fault diagnosis method based on a self-learning convolutional domain adversarial network. Aiming at the problem of time-frequency feature distribution offset caused by working condition changes in mechanical fault diagnosis, the proposed self-learning convolution uses a parameterized sub-kernel structure reorganization mechanism. Compared with the feature extraction method of the traditional fixed-structure convolution kernel, it can dynamically adjust the convolution kernel structure according to the distribution characteristics of the input data, effectively capture the changes in fault characteristics under different working conditions, reduce the distribution differences between the source domain and the target domain, enhance the ability to extract domain-invariant features, and effectively improve the accuracy of mechanical fault diagnosis across working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention provides the following figure for explanation:

[0061] Figure 1 This is a diagram of the self-learning process of the sub-core spacing parameters of the self-learning convolution in the present invention.

[0062] Figure 2 This is a flowchart of the self-learning convolutional domain adversarial network in the present invention. DETAILED DESCRIPTION

[0063] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.

[0064] In one embodiment of the present invention, a cross-domain fault diagnosis method based on a self-learning convolutional domain adversarial network includes the following steps:

[0065] S1: Build an adaptive feature extractor based on self-learning convolution, and realize dynamic perception of signal features through the parameter learning mechanism of self-learning convolution;

[0066] The parameter learning mechanism of the self-learning convolution in S1 is as follows: perform a standard convolution operation on the input signal to determine the batch dimension b and sequence length l of the output feature map; convert the original convolution kernel W∈R Ks , divided into n independently adjustable sub-cores, generating a sub-core group W sub , the sub-nuclear distance is the self-learning parameter:

[0067] W sub =Reshape(W,[n,K s / n]) (10)

[0068] Dynamic adjustment of sub-core spacing is achieved by constructing convolution position index: Based on the step size s of standard convolution, an s×l sequence is constructed, and the dimension is expanded to form [b, K s , l] tensor, generating the position index p0 when the standard convolution operation is performed:

[0069]

[0070] Among them, p0∈Z b×Ks×l , the value of each element is determined only by index i, which is equal to s×i and has nothing to do with the dimensions of b and Ks; then construct K s dimensional sequence, and expanded to [b, K s , 1], generate the original convolution kernel position index p1:

[0071]

[0072] Where p1∈Z b×Ks×1 , the value of each element is determined only by index j, which is equal to j, and has nothing to do with the dimension b. The third dimension is always 0. The two are superimposed to obtain the global position index p2 of the standard convolution operation:

[0073]

[0074] Among them, the third dimension of p1 is automatically expanded to l and added element-by-element to p0 to generate a domain-sensitive global position encoding p2; a lightweight convolutional layer is further introduced to calculate the domain offset between sub-nuclei, and after channel compression and nonlinear activation, n-1 sub-nuclei spacing parameters are extracted; finally, through differentiable sub-nuclei position remapping, the dynamic spacing parameters are fused with the position encoding p2 to construct a self-learning convolution with domain invariance perception.

[0075] The core architecture of the adaptive feature extractor based on self-learning convolution in S1 is a multi-layer convolutional neural network: the first layer uses self-learning convolution, and the input is the original vibration signal; the subsequent layers stack multiple layers of standard convolution layers, normalization layers, pooling layers and activation function layers in sequence.

[0076] S2: Design an adversarial domain adaptation network architecture consisting of an adaptive feature extractor based on self-learning convolution, a domain classifier connected by a gradient reversal layer, and a fault classifier;

[0077] The construction of the adversarial domain adaptation network architecture in S2 includes the following core components:

[0078] (a) The adaptive feature extractor based on self-learning convolution according to claim 1;

[0079] (b) The domain classifier is connected to the feature extractor through a gradient reversal layer. Its forward propagation process performs: receiving the feature vector output by the feature extractor and outputting the domain classification prediction result; the backward propagation process performs: the gradient reversal layer multiplies the input gradient by a negative coefficient, and the calculation formula is:

[0080]

[0081] Among them, x is the feature extractor output feature, L is the domain classification loss, is the gradient from the domain classifier, and λ is the gradient reversal strength (usually set to 1.0 or dynamically adjusted);

[0082] (c) The fault classifier is directly connected to the feature extractor and accepts the domain-invariant features of the source domain for fault pattern recognition. The domain discriminator and the fault classifier are connected in parallel to the output of the feature extractor. The fault classifier is optimized by minimizing the classification loss of the source domain samples, and the domain discriminator is optimized by minimizing the feature distribution confusion loss.

[0083] The domain classifier connected by the gradient reversal layer is composed of multiple layers of fully connected layers and nonlinear activation functions, and outputs the predicted probability of the domain label. The feature distribution confusion loss adopts the binary cross entropy loss with class balance weight, and its expression is:

[0084]

[0085] Among them, L domain is the domain discrimination loss, N is the total number of samples, and the N samples are summed and averaged when calculating the loss; y d (i) is the domain belonging identifier of the i-th sample, which takes the value of 0 (source domain) or 1 (target domain). D(·) is the domain classifier, the input is the feature, and the output is the probability that the sample belongs to the domain label. f(x (i) ) is the feature extractor f for the i-th sample x (i) The extracted cross-domain features; α is the balance coefficient, and the calculation formula is:

[0086]

[0087] Among them, N target and N source are the sample sizes of the source domain and target domain respectively.

[0088] The fault classifier consists of a fully connected layer network and calculates the cross entropy classification loss based on the source domain labeled data. Its expression is as follows:

[0089]

[0090] Among them, N s is the number of source domain samples, K is the number of fault categories, y k (i) is the true fault category label of the source domain sample, is the predicted probability of the kth class by the fault classifier.

[0091] S3: A cross-domain joint training strategy is adopted to input the source domain and target domain datasets into the self-learning convolutional domain adversarial network, and the feature distribution confusion loss and the source domain classification loss are simultaneously optimized to achieve adversarial alignment of the source domain and target domain feature spaces, thereby realizing cross-domain fault diagnosis.

[0092] In S3, the cross-domain joint training strategy is as follows: the source domain dataset (with fault labels and domain labels) and the target domain dataset (with domain labels) are input into the feature extractor, and the self-learning convolution layer performs: dynamically adjusting the convolution kernel based on the sub-kernel spacing parameter self-learning mechanism, and performing a convolution operation on the input vibration signal; the source domain and target domain features are input into the domain classifier, and the feature distribution confusion loss is output; the source domain features are input into the fault classifier alone, and the source domain classification loss is output; the feature distribution confusion loss and the source domain classification loss are simultaneously optimized, and the total loss function is defined as:

[0093] L total =L domain +L class (18)

[0094] Among them, Ldomain is the feature distribution confusion loss, L class The test phase is performed by inputting the target domain test data into the feature extractor, and the extracted feature vector is input into the fault classifier, which outputs the fault mode prediction label.

[0095] Transfer training under different load conditions was conducted on the bearing dataset of Hanoi University of Science and Technology. The cross-condition fault diagnosis accuracy of the proposed method was compared with that of the traditional DANN method. As shown in Table 1, the dataset contains four fault categories: normal, inner ring, outer ring, and sphere. There are three working conditions: 0W, 200W, and 400W, which are defined as [0], [1], and [2] respectively. The sampling frequency is 51.2k. When constructing the dataset, the sliding window is set to 6400 to include information about multiple rotations of the bearing. The collected samples are downsampled to 1024 length to match the model input size. There are 600 samples for each fault category in the three working conditions, including 500 single-class training samples and 100 test samples. A total of 6 cross-condition tasks were constructed with one of the working conditions as the source domain and the other two working conditions as the target domain. The experiment was repeated 5 times, and the average classification accuracy was taken as the final classification performance indicator of the algorithm. In the experiment, the optimizer used was SGD, the initial learning rate was 0.05, the batch size was set to 500, the epochs were set to 100 rounds, and the training was iterated until the loss of the training set tended to be stable. After training, the model was applied to the test data of the target domain to evaluate its fault diagnosis performance under different working conditions.

[0096] Table 1. Comparison of accuracy on the Hanoi University of Science and Technology dataset

[0097]

[0098] The results show that the diagnostic accuracy of the method proposed in the present invention in the target domain is significantly improved, with an average accuracy of 95.75%. In addition, the diagnostic accuracy in different cross-operating tasks is higher than that of the traditional DANN method, which shows that the method proposed in the present invention has obvious advantages and higher accuracy in cross-domain fault diagnosis, further verifying its effectiveness.

Claims

1. A cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network, characterized in that: The steps include: S1: Build an adaptive feature extractor based on self-learning convolution, and realize dynamic perception of signal features through the self-learning mechanism of convolution parameters; S2: Construction includes: An adaptive feature extractor based on self-learning convolution, a domain classifier with a gradient reversal layer, and a self-learning convolutional domain adversarial network for fault classifiers; S3: A cross-domain joint training strategy is adopted to input the source domain and target domain datasets into the self-learning convolutional domain adversarial network, and the cross-domain feature distribution confusion loss and the source domain classification loss are simultaneously optimized to achieve the alignment of the source domain and target domain feature spaces, thereby realizing cross-domain fault diagnosis.

2. A cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network according to claim 1, characterized in that: The parameter learning mechanism of the self-learning convolution in S1 is as follows: perform a standard convolution operation on the input signal to determine the batch dimension b and sequence length l of the output feature map; convert the original convolution kernel W∈R Ks , divided into n independently adjustable sub-cores, generating a sub-core group W sub , the sub-nuclear distance is the self-learning parameter: W sub =Reshape(W,[n,K s / n]) (1) Dynamic adjustment of sub-core spacing is achieved by constructing convolution position index: Based on the step size s of standard convolution, an s×l sequence is constructed, and the dimension is expanded to form [b, K s , l] tensor, generating the position index p0 when the standard convolution operation is performed: Among them, p0∈Z b×Ks×l , the value of each element is determined only by index i, which is equal to s×i and has nothing to do with the dimensions of b and Ks; then construct K s dimensional sequence, and expanded to [b, K s , 1], generate the original convolution kernel position index p1: Where p1∈Z b×Ks×1 , the value of each element is determined only by index j, which is equal to j, and has nothing to do with the dimension b. The third dimension is always 0. The two are superimposed to obtain the global position index p2 of the standard convolution operation: Among them, the third dimension of p1 is automatically expanded to l and added element-by-element to p0 to generate a domain-sensitive global position encoding p2; a lightweight convolutional layer is further introduced to calculate the domain offset between sub-nuclei, and after channel compression and nonlinear activation, n-1 sub-nuclei spacing parameters are extracted; finally, through differentiable sub-nuclei position remapping, the dynamic spacing parameters are fused with the position encoding p2 to construct a self-learning convolution with domain invariance perception.

3. The cross-domain fault diagnosis method based on the self-learning convolutional domain adversarial network according to claim 1 is characterized in that: The core architecture of the adaptive feature extractor based on self-learning convolution in S1 is a multi-layer convolutional neural network: the first layer adopts self-learning convolution, and the input is the original state monitoring signal; the subsequent layers stack m layers of standard convolution layers, one normalization layer, a pooling layer and an activation function layer in sequence.

4. The cross-domain fault diagnosis method based on the self-learning convolutional domain adversarial network according to claim 1 is characterized in that: The construction of the adversarial domain adaptation network architecture in S2 includes the following core components: (a) The adaptive feature extractor based on self-learning convolution according to claim 1; (b) The domain classifier is connected to the feature extractor through a gradient reversal layer. Its forward propagation process performs: receiving the feature vector output by the feature extractor and outputting the domain classification prediction result; the backward propagation process performs: the gradient reversal layer multiplies the input gradient by a negative coefficient, and the calculation formula is: Among them, x is the feature extractor output feature, L is the domain classification loss, is the gradient from the domain classifier, λ is the gradient reversal strength; (c) The fault classifier is directly connected to the feature extractor and accepts the domain-invariant features of the source domain for fault pattern recognition. The domain discriminator and the fault classifier are connected in parallel to the output of the feature extractor. The fault classifier is optimized by minimizing the classification loss of the source domain samples, and the domain discriminator is optimized by minimizing the feature distribution confusion loss.

5. The cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network according to claim 4 is characterized in that: The domain classifier connected by the gradient reversal layer is composed of multiple layers of fully connected layers and nonlinear activation functions, and outputs the predicted probability of the domain label. The feature distribution confusion loss adopts the binary cross entropy loss with class balance weight, and its expression is: Among them, L domain is the domain discrimination loss, N is the total number of samples, and the N samples are summed and averaged when calculating the loss; y d (i) is the domain belonging identifier of the i-th sample, which takes the value of 0 (source domain) or 1 (target domain). D(·) is the domain classifier, the input is the feature, and the output is the probability that the sample belongs to the domain label. f(x (i) ) is the feature extractor f for the i-th sample x (i) The extracted cross-domain features; α is the balance coefficient, and the calculation formula is: Among them, N target and N source are the sample sizes of the source domain and target domain respectively.

6. The cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network according to claim 4 is characterized in that: The fault classifier consists of a fully connected layer network and calculates the cross entropy classification loss based on the source domain labeled data. Its expression is as follows: Among them, N s is the number of source domain samples, K is the number of fault categories, y k (i) is the true fault category label of the source domain sample, is the predicted probability of the kth class by the fault classifier.

7. The cross-domain fault diagnosis method based on self-learning convolutional domain adversarial network according to claim 1 is characterized in that: In S3, the cross-domain joint training strategy is as follows: the source domain dataset with fault labels and the target domain dataset are input into the feature extractor. The self-learning convolution layer performs: dynamically adjusts the convolution kernel based on the sub-kernel spacing parameter self-learning mechanism and performs a convolution operation on the input signal; the source domain and target domain features are input into the domain classifier, which outputs the feature distribution confusion loss; the source domain features are input into the fault classifier alone, which outputs the source domain classification loss; the feature distribution confusion loss and the source domain classification loss are simultaneously optimized, and the total loss function is defined as: L total =L domain +L class (9) Among them, L domain is the feature distribution confusion loss, L class is the source domain classification loss; in the testing phase, the target domain test data is input into the feature extractor, the extracted feature vector is input into the fault classifier, and the fault classifier outputs the fault mode prediction label.

Images