Rotating machinery cross-domain fault diagnosis method based on dynamic evolution and wavelet double-path structure

Abstract

The application discloses a rotating machinery cross-domain fault diagnosis method based on dynamic evolution and wavelet double-path structure, and relates to the technical field of fault diagnosis. The method comprises the following steps: obtaining labeled vibration signals under a certain working condition of rotating machinery and a large number of unlabeled vibration signal samples under an actual variable-speed working condition to be measured. Firstly, batch normalization processing is performed on the input original vibration signals to reduce signal distribution differences caused by speed changes. Secondly, a double-path feature extraction structure is constructed based on wavelet packet transformation, low-frequency global and high-frequency detail features are fused, and the model cross-domain fault feature extraction capability is improved. Then, a domain self-adaptive method based on a dynamic evolution mechanism is adopted to construct a series of mixed domains evolving from a source domain to a target domain, the domain offset mutation problem in the migration process is relieved through gradual transition and gradual migration, and finally the trained model is saved to realize the rotating machinery cross-domain fault diagnosis by using a small amount of labeled samples in the source domain and a large amount of unlabeled samples in the target domain.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure. Background Technology

[0002] Rotating machinery, as a key power transmission device, is widely used in industries such as machinery manufacturing, energy, automotive, aerospace, and power due to its high efficiency and durability. Core components in this type of equipment, such as gears and bearings, often operate under conditions of dynamically changing speeds, which poses a significant challenge to deep learning-based fault diagnosis methods. Speed ​​fluctuations enhance the non-stationary characteristics and nonlinear effects of vibration signals, causing the signals to simultaneously contain fault features and speed change information, increasing the difficulty of feature extraction and state identification. Furthermore, in practical engineering applications, the sample data used to train fault diagnosis models often originates from specific operating conditions, differing from the actual data distribution of the target domain. This mismatch between domains leads to a decline in model generalization performance. Particularly noteworthy is that data-driven diagnostic methods typically assume that training and testing data conform to the same distribution; if this condition is not met, the model is prone to mismatch, and the recognition accuracy decreases significantly. Therefore, how to achieve cross-domain fault diagnosis of gearboxes using limited samples under variable speed conditions has become an important research direction in the field of industrial intelligent operation and maintenance.

[0003] Cross-domain fault diagnosis has always been a key challenge in intelligent operation and maintenance. Under variable speed conditions, the non-stationarity of vibration signals caused by speed changes further increases the difficulty of fault feature extraction and matching. Transfer learning, as an effective solution, aims to leverage diagnostic knowledge trained in an existing source domain (usually from easily accessible or labeled data) and align the feature distributions of the source and target domains through inter-domain distribution adaptation methods, thereby improving the model's transfer performance and diagnostic accuracy in the target domain. Mainstream cross-domain fault diagnosis methods primarily improve diagnostic performance by minimizing distance metrics, using adversarial domain training, or combining both. While some progress has been made, two limitations remain: First, most methods directly align the source and target domains, leading to abrupt domain shifts and reducing the model's diagnostic performance in the target domain. Second, most existing methods suffer from insufficient feature extraction capabilities, which also affects the model's generalization performance in the target domain. Therefore, more effective cross-domain fault diagnosis methods need to be researched to address these two issues. Summary of the Invention

[0004] To address the aforementioned problems and technical requirements, the inventors have proposed a cross-domain fault diagnosis method for rotating machinery based on dynamic evolution and a wavelet dual-path structure. The technical solution of this invention is as follows:

[0005] A method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure includes the following steps:

[0006] Step 1: Under the constant speed condition of the rotating machinery, collect a sufficient amount of labeled one-dimensional time series vibration signals and define the constant speed condition as the source domain; at the same time, under the actual dynamic variable speed condition of the rotating machinery to be measured, collect a large amount of unlabeled vibration signals and a small amount of labeled vibration signals and define the dynamic variable speed condition as the target domain.

[0007] Step 2: Extract fixed-length data segments from the original vibration signals in the source and target domains to construct a sample set, which is then divided into a training set and a test set. The training set contains source-target domain sample pairs for transfer learning, while the test set consists of labeled vibration signals from the target domain. The source-target domain sample pairs include samples in the source domain with fault category labels and samples in the target domain without labels. The labels of the source domain samples are set according to the fault category of their corresponding vibration signals. The category labels of the labeled samples in the source domain are set to 1,...,W, where W is the number of fault categories. The number of fault categories in the target domain is the same as that in the source domain. The training set is used for batch input into the fault diagnosis model constructed below for training.

[0008] Step 3: Construct a fault diagnosis network model based on a wavelet dual-path structure and initialize the model parameters; the model is implemented based on a Domain Adversarial Network (DANN) and includes three parts: a feature extractor Φ, a label classifier Ψ, and a neighborhood discriminator Γ.

[0009] Step 4: Construct a series of dynamic hybrid domains based on the dynamic evolution mechanism, which gradually and smoothly transition from a distribution close to the source domain to a distribution close to the true distribution of the target domain, in order to alleviate the model mismatch problem caused by abrupt changes in domain offset during the migration process;

[0010] The dynamic evolution mechanism aims to achieve gradual distribution alignment between the source and target domains. This mechanism introduces a mixing ratio parameter that dynamically adjusts with each training epoch, systematically fusing labeled source domain samples and unlabeled target domain samples to generate a series of temporally intermediate transition domains—the "mixed domains." These mixed domains are closer to the source domain distribution in the early stages of training, and then gradually approach the true distribution of the target domain, effectively avoiding feature distortion or negative transfer that may occur in traditional domain adaptation methods due to direct alignment of the two ends of the distribution. By controlling the gradual rate of inter-domain offset, this mechanism improves the model's generalization ability and diagnostic stability in the target domain.

[0011] Step 5: Based on the network model that integrates the dynamic evolution mechanism and the wavelet dual-path structure, the network model is trained using source domain-target domain sample pairs based on the transfer learning method. The model parameters are updated by minimizing the total loss function and performing gradient backpropagation.

[0012] The total loss function includes: classification loss for source and mixed domain samples, domain adversarial loss between the source and target domains, domain adversarial loss for the mixed domain, distribution difference loss between the source and target domains, and conditional consistency loss for pseudo-labels in the target domain. The classification loss uses the standard cross-entropy loss function; the domain adversarial loss uses the binary cross-entropy loss function to achieve inter-domain feature alignment; the distribution difference loss is measured using multi-kernel maximum mean difference (MK-MMD); and the conditional consistency loss for pseudo-labels in the target domain is achieved by calculating the positive Kullback-Leibler (KL) divergence to improve the model's inference stability and generalization ability in the target domain.

[0013] Step 6: Truncate the labeled samples in the test set to a certain length, preprocess them, and then input them into the fault diagnosis network model trained in Step 5 to obtain the cross-domain fault diagnosis results of rotating machinery.

[0014] The further technical solution is as follows: In step two:

[0015] The original vibration signal was truncated using a random sliding window sampling method, with a sliding window size of 2048. The sample set data was preprocessed: the sample set was divided into multiple batches, and batch normalization was performed on the samples in each batch, including:

[0016] For all input samples in the same batch (where m represents the number of samples in the batch, and d represents the number of features in each sample), calculate the mean μ of X. x and variance

[0017]

[0018] In terms of feature dimensions, the calculated mean and variance are used to normalize all input samples in the batch:

[0019]

[0020] in Let represent the k-th feature in the i-th sample, where ∈ is a very small constant.

[0021] The further technical solution is that, in step three, constructing the wavelet dual-path structure includes:

[0022] A new feature extractor, WPT-MCNN, is designed by incorporating wavelet packet transform for frequency domain decomposition at the front end of the existing DANN feature extractor Φ (ECA-MCNN). The wavelet packet transform decomposes training or test set samples into high-frequency detail features and low-frequency global features, simultaneously capturing both local transient features and global trend information while maintaining the signal length before and after decomposition. Subsequently, the two features are concatenated and fused along the channel dimension to form a dual-channel composite signal. This composite signal is then input to the feature extractor Φ, which performs a series of convolutions and downsampling operations on the input dual-channel composite signal to achieve multi-scale feature abstraction and fusion. This constructs a wavelet-enhanced deep feature extraction structure with dual-path processing capabilities, referred to as the wavelet dual-path structure.

[0023] The further technical solution is as follows: In step four:

[0024] First, a pseudo-label generation mechanism based on forward reasoning is used to generate pseudo-labels in the target domain. Used when constructing hybrid domains using dynamic evolution mechanisms; then let This represents a labeled source domain sample set. This represents an unlabeled target domain sample set. To mitigate the problem of abrupt changes in inter-domain distribution, a dynamic mixing strategy is adopted to combine source domain samples... and real labels and target domain samples and pseudo-tags The mixing is performed according to the mixing ratio parameter μ(k) that changes dynamically with each training round, generating a series of intermediate mixing domains that progressively transition from the source domain to the target domain. Where N s N t and N m These represent the batch sample sizes for the source domain, target domain, and mixed domain, respectively.

[0025] Among them, the pseudo-label generation mechanism based on forward reasoning generates pseudo-labels on the target domain, specifically including:

[0026] High-level feature representations are extracted from the target domain samples using the feature extractor Φ. Then, the model parameters are frozen, preventing backpropagation and gradient updates; the network is in a pure inference state. In this state, the label classifier Ψ outputs the predicted class probabilities for the high-level features, which are then converted into one-hot vectors as pseudo-labels for the target domain, represented as:

[0027]

[0028] in, This represents a pseudo-label refinement function that incorporates prediction confidence information. This represents the confidence estimator for the predictions of the label classifier Ψ. Φ represents the target domain features extracted by the feature extractor, and W is the number of fault categories.

[0029] The aforementioned intermediate hybrid domain Constructed based on the interpolation method of optimal transmission:

[0030]

[0031] in, The hybrid transformation operator represents the parameter μ(k), and δ is used to control noise. Strength to improve robustness, Σ represents the covariance matrix of the target domain feature estimation, pseudo-label The predicted class assignment representing the target domain is encoded as a one-hot vector; the mixing ratio parameter μ(k) controls the contribution weight of source domain information in the migration process, and is expressed as:

[0032]

[0033] Where k represents the current training iteration number, n is the total number of iterations, ζ is the weight controlling the exponential decay term, τ controls the exponential decay rate, υ adjusts the order of the exponential decay curve, β adjusts the order of the polynomial decay curve, and μ0 controls the weight controlling the polynomial decay term, with μ0 fixed at 1 to ensure that the mixed domain is aligned with the source domain in the early stages of training, achieving sufficient learning of the source domain information. After a series of experiments, the values ​​of parameters ζ and υ were finally determined to be 0, and the values ​​of parameters τ and β were determined to be 1.

[0034] In the early stages of training, the dynamic evolution mechanism prioritizes source domain alignment (μ(k)≈1) to avoid potential interference from low-reliability pseudo-labels in the target domain. As training progresses, when the model enters the later stages (μ(k)→0), the quality of pseudo-labels in the target domain has been improved through previous training. The hybrid domain gradually converges to the target domain distribution using refined pseudo-labels, effectively improving domain adaptation performance in cross-domain fault diagnosis. The dynamic evolution mechanism proposed in this application fuses source and target domain samples through a dynamically decaying hybridization coefficient μ(k), constructing a series of intermediate hybrid domains that smoothly transition from the source domain to the target domain. This strategy gradually reduces the distribution differences between domains, significantly enhancing the model's generalization ability and diagnostic stability in the target domain.

[0035] The further technical solution is as follows: In step five:

[0036] (1) The classification loss consists of two parts. The first part is the classification loss of labeled samples in the source domain, expressed as:

[0037]

[0038] Where, Ns W is the number of labeled samples in the source domain, and W is the number of fault categories. θ represents the source domain features extracted by the feature extractor Φ, Ψ(·) represents the class probability prediction output by the label classifier Ψ, and θ represents the class probability prediction. Ψ For the network parameters of Ψ; It is an indicator function; its value is 1 when the subscript and superscript are equal, and 0 otherwise.

[0039] The second part is the classification loss for mixed-domain samples, expressed as:

[0040]

[0041] Where, N t It is the number of unlabeled samples in the target domain. Let Φ represent the target domain features extracted by the feature extractor, and μ(k) be the mixing ratio parameter that dynamically changes with the training epochs. Also an indicator function, its value is 1 when the subscript and superscript are equal, and 0 otherwise;

[0042] The total classification loss is the sum of the two classification losses mentioned above, that is:

[0043]

[0044] (2) The domain adversarial loss consists of two parts. The first part is the domain adversarial loss between the source domain and the target domain, expressed as:

[0045]

[0046] Where Γ(·) represents the predicted output of the neighborhood discriminator Γ, and θ Γ Let Γ be the network parameter, and GRL(·) denote the gradient reversal backpropagation function.

[0047] The second part is the domain adversarial loss for the hybrid domain, expressed as:

[0048]

[0049] Where N m It is the number of samples in the mixed domain; These are the mixed domain features extracted by Φ.

[0050] The total domain adversarial loss is the sum of the two losses mentioned above, namely:

[0051]

[0052] (3) The distribution difference loss is the multi-kernel maximum mean difference loss (MK-MMD) between the feature vectors of the source domain samples and the feature vectors of the target domain samples, expressed as:

[0053]

[0054] Where H represents the Hilbert space, K(·) represents the kernel function that maps the probability distribution to the reproducing kernel Hilbert space, and multiple kernels k are used. o This is indicated to improve the model's expressive power; E[·] represents the expected value. Let d represent the source and target domain samples respectively, where d is the vector dimension; β u These are weighting parameters for different cores, where m is the number of cores and k is the number of cores. u Represents the kernel function, multi-core k o Different kernels are used to improve the mapping effect of MMD, ultimately achieving the optimal and reasonable kernel selection.

[0055] (4) Target domain pseudo-label conditional consistency loss is represented by the positive KL divergence between the predicted output of the label classifier Ψ in the target domain without freezing the model parameters and the pseudo-labels generated by Ψ in the target domain with freezing the model parameters. Its function is to improve the quality of pseudo-labels predicted by the model in the target domain by stabilizing the model's output in the target domain. The target domain pseudo-label conditional consistency loss is expressed as:

[0056]

[0057] in, and Let E[·] represent the predicted probability distribution of the j-th target domain sample before and after parameter freezing, respectively, and let D[·] represent the expected value of the variance of the probability distribution. KL (·) represents the positive KL divergence function.

[0058] Finally, during model training, the classification loss L is minimized by the cross-entropy between the source domain and the mixed domain samples. sup The loss due to the difference in distribution between the source and target domain samples. fdd Conditional consistency loss with target domain pseudo-labels L pcc And the domain adversarial loss L that maximizes the samples from the source domain, target domain, and mixed domain. adv The total loss function is then expressed as:

[0059] L total =L sup -L adv +αL fdd +L pcc

[0060] Where α is the balancing hyperparameter, which is reduced in the early stages of training. fdd The weights are assigned and the value of α is gradually increased with each training epoch of the model. α is represented as:

[0061]

[0062] Furthermore, this application employs the Adam optimizer to perform gradient backpropagation and parameter updates. The parameter iterative update mechanism of the Adam optimizer is as follows:

[0063]

[0064] In the formula, m k and n k The gradient of the objective function g k Let be the first-order and second-order matrices, k represent the current iteration batch, k-1 represent the previous iteration batch; β1 and β2 represent the matrix exponential decay rate. and It is for m k and n k Correction; θ∈θ Φ ,θ Ψ ,θ Γ ε represents the model parameters, η represents the learning rate, and ε is a constant.

[0065] By adopting a dynamic range constraint method for the learning rate, and by setting an upper and lower limit for the learning rate, gradient oscillations or convergence stagnation caused by excessively large or small learning rates in the later stages of training are avoided, thereby ensuring that the model optimization process smoothly tends to a stable solution.

[0066] The beneficial technical effects of this invention are:

[0067] 1) The rotating machinery cross-domain fault diagnosis method based on dynamic evolution and wavelet dual-path structure proposed in this application effectively combines the dynamic evolution mechanism and wavelet dual-path structure, which can better learn richer feature representations from the generated mixed domain samples and unlabeled samples in the target domain, thereby improving the overall cross-domain fault diagnosis performance of the fault diagnosis model.

[0068] 2) This application addresses the problems of abrupt domain shifts and insufficient feature extraction capabilities during model transfer by proposing a dynamic evolution mechanism. This mechanism constructs a series of evolving hybrid domains between the source and target domains to achieve gradual transfer and alleviate the abrupt domain shift problem. Furthermore, wavelet packet transform is added before the feature extractor to decompose the input signal into high-frequency and low-frequency components, constructing a wavelet dual-path structure to enhance the model's feature extraction capabilities.

[0069] 3) The rotating machinery cross-domain fault diagnosis network model constructed in this application alleviates the problem of abrupt changes in domain differences by introducing a dynamic evolution mechanism, enhances the feature extraction capability of the model by adding wavelet packet transform to the front end of the feature extractor, and improves the quality of pseudo-labels in the target domain by introducing pseudo-label conditional consistency loss. It can achieve high-precision rotating machinery cross-domain fault diagnosis when the target domain samples are unlabeled. Attached Figure Description

[0070] Figure 1 This is a flowchart of the cross-domain fault diagnosis method for rotating machinery provided in this application.

[0071] Figure 2 This is a flowchart of the dynamic evolution mechanism provided in this application.

[0072] Figure 3 This is a structural diagram of the cross-domain fault diagnosis network model for rotating machinery provided in this application. Detailed Implementation

[0073] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0074] One embodiment of this application provides a method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure. The specific flowchart of this method is as follows: Figure 1 As shown, the specific implementation method includes the following steps:

[0075] Step 1: Use a data acquisition system to acquire vibration signal samples of rotating machinery under different operating conditions (including constant speed and variable speed conditions) and various fault modes. Specifically, under the actual variable speed condition to be tested, only a small number of labeled vibration signals need to be collected as a test set; the remaining large amount of data can be collected without manual labeling.

[0076] Step 2: The collected vibration signal data is randomly windowed and repeatedly truncated to a sample set of length 2048. After batch normalization of the samples, they are divided into training and test sets in an 8:2 length ratio. This embodiment uses the JNU bearing dataset and the BJTU gearbox dataset. The JNU bearing dataset has four different fault types (W=4): inner race fault, normal, outer race fault, and rolling element fault, corresponding to labels 0, 1, 2, and 3, respectively. Three speed conditions from the dataset were selected for experiments: 600 r / min, 800 r / min, and 1000 r / min. The BJTU gearbox dataset has five different fault types (W=5): broken tooth fault, gear wear fault, missing tooth fault, normal, and root crack fault, corresponding to labels 0, 1, 2, 3, and 4, respectively. Three speed conditions from the dataset were selected for experiments: 20 Hz, 35 Hz, and 50 Hz. Both datasets were trained using one-hot encoding; for example, the one-hot encoding for the label "3" is

[00100] . Information about the JNU bearing dataset and the BJTU gearbox dataset is shown in Table 1.

[0077] Table 1. Detailed Information about the Dataset

[0078]

[0079] Step 3: A cross-domain fault diagnosis network model for rotating machinery is built using the PyTorch deep learning framework. This network model consists of three parts: a feature extractor Φ, a label classifier Ψ, and a domain discriminator Γ. The gradient update process during model training is shown in Table 2.

[0080] The algorithm process of the proposed model is shown in Table 2.

[0081]

[0082] Following the aforementioned algorithm and formulas, the code is first written by adding a wavelet packet transform before the standard feature extractor ECA-MCNN. This decomposes the input signal into high-frequency detail components and low-frequency global components, which are then concatenated along the channel dimension to form a dual-channel composite signal. This signal is then input into the feature extractor for further convolution and pooling, thus constructing a wavelet dual-path structure. Next, the model is trained for label classification and domain discrimination. Pseudo-labels are generated in the target domain by predicting the model under frozen gradient conditions, preparing for the dynamic evolution mechanism to generate a hybrid domain. Then, through the dynamic evolution mechanism, a series of hybrid domains evolving from the source domain to the target domain are generated as the training process progresses. Finally, the network parameters are updated through backpropagation of the loss function.

[0083] The specific working process of the dynamic evolution mechanism is as follows: Figure 2 As shown, in the early stages of training, the mixed domain generated by the interpolation method is initially aligned with the source domain. This ensures that the mixed domain fully inherits information from the source domain during the initial training phase and avoids interference from low-reliability pseudo-labels in the target domain at this stage. As training progresses, the proportion of target domain information in the mixed domain is gradually increased, allowing the mixed domain to gradually transition from the source domain to the target domain. At this point, parameters are updated through backpropagation of the loss function, and the conditional consistency loss of the target domain pseudo-labels enhances the model's inference stability in the target domain, improving the reliability of the model's predicted output in the target domain. In the later stages of training, the quality of the pseudo-labels predicted in the target domain improves through the earlier training. At this point, target domain information dominates in the mixed domain, enabling the model to learn more information from the target domain and enhancing its generalization ability.

[0084] The network is trained sequentially according to the proposed transfer learning training strategy, and the overall structure of the proposed model is as follows: Figure 3 As shown in Table 3, the Adam optimizer was used with a batch size of 20 and a learning rate of 0.001. After adjustments through experiments, the total number of training rounds was finally determined to be 100.

[0085] Table 3 shows the network parameters of the proposed model.

[0086]

[0087] Note: "Classes" indicates the number of fault categories.

[0088] Step 4: Based on the algorithm written in Step 3, build the network model, set the relevant hyperparameters, and start training for the transfer task.

[0089] Step 5: Input the labeled samples of the target domain of the test set into the trained rotating machinery fault diagnosis model for online fault diagnosis, obtain the diagnosis results, and test the fault diagnosis performance of the model.

[0090] Using the method provided in this application, labeled vibration signals under a certain operating condition of rotating machinery and a large number of unlabeled vibration signal samples under the actual variable speed operating condition are obtained. First, the original vibration signals are preprocessed by batch normalization to reduce the distribution difference between the source and target domains. Second, a wavelet dual-path structure based on wavelet packet transform is constructed to decompose the input signal into high-frequency detail components and low-frequency global components, improving the model's feature extraction capability. Then, the network is trained based on a dynamic evolution mechanism, progressively aligning the differences between the source and target domains by generating a series of evolving mixed domains from the source to the target domain, thereby improving the model's generalization ability. Finally, the trained model is saved, enabling cross-domain fault diagnosis of rotating machinery using a small number of labeled samples from the source domain and a large number of unlabeled samples from the target domain.

[0091] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure, characterized in that, The method includes: Step 1: Under the constant speed condition of the rotating machinery, collect a sufficient amount of labeled one-dimensional time series vibration signals and define the constant speed condition as the source domain; at the same time, under the actual dynamic variable speed condition of the rotating machinery to be measured, collect a large amount of unlabeled vibration signals and a small amount of labeled vibration signals and define the dynamic variable speed condition as the target domain. Step 2: Extract fixed-length data segments from the original vibration signals in the source and target domains to construct a sample set, which is then divided into a training set and a test set. The training set contains source-target domain sample pairs for transfer learning, and the test set consists of labeled vibration signals in the target domain. The source-target domain sample pairs include samples in the source domain with fault category labels and samples in the target domain without labels, and the labels of the source domain samples are set according to the fault category of their corresponding vibration signals. Step 3: Construct a fault diagnosis network model based on a wavelet dual-path structure and initialize the model parameters; the model is implemented based on a Domain Adversarial Network (DANN) and includes a feature extractor. Tag classifier and Domain Discriminator These three parts; The feature extractor in the DANN A wavelet packet transform is added to the front end. This wavelet packet transform decomposes the training or test set samples into high-frequency detail features and low-frequency global features, while maintaining the signal length before and after decomposition. The two features are then concatenated and fused along the channel dimension to construct a dual-channel composite signal, which is then input to the feature extractor. The feature extractor This is used to perform a series of convolution and downsampling operations on the input dual-channel composite signal, thereby constructing the wavelet dual-path structure; Step 4: Construct a series of dynamic hybrid domains based on a dynamic evolution mechanism, which gradually and smoothly transition from a distribution close to the source domain to a distribution approaching the true distribution of the target domain, in order to alleviate the model mismatch problem caused by abrupt changes in domain offset during the migration process; Step 4 includes: A pseudo-label generation mechanism based on forward reasoning generates pseudo-labels on the target domain. This is used when constructing hybrid domains using dynamic evolution mechanisms; make This represents a labeled source domain sample set. This represents the unlabeled target domain sample set, and the source domain samples... and real labels and target domain samples and pseudo-tags According to the mixing ratio parameter that changes dynamically with the training rounds The mixture is blended to generate a series of intermediate blended domains that progressively transition from the source domain to the target domain. ,in , and These represent the batch sample sizes for the source domain, target domain, and mixed domain, respectively. In the early stages of training, the dynamic evolution mechanism prioritizes source domain alignment to avoid interference from low-reliability pseudo-labels in the target domain at this time; as the model enters the later stages of training, the quality of pseudo-labels in the target domain has been improved through the previous training. Step 5: Based on the network model that integrates the dynamic evolution mechanism and the wavelet dual-path structure, the network model is trained using the source domain-target domain sample pairs based on the transfer learning method. The model parameters are updated by minimizing the total loss function and performing gradient backpropagation. Step 6: Input the labeled samples from the test set into the fault diagnosis network model trained in Step 5 to obtain the cross-domain fault diagnosis results of rotating machinery.

2. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to claim 1, characterized in that, The forward-reasoning-based pseudo-label generation mechanism generates pseudo-labels on the target domain, including: Using the feature extractor High-level feature representations are extracted from the target domain samples, and then the model parameters are frozen so that the model does not perform backpropagation and gradient updates. The label classifier is then used. Output the predicted class probabilities for high-level features and convert them into one-hot vectors as pseudo-labels for the target domain, represented as: in, This represents a pseudo-label refinement function that incorporates prediction confidence information. This refers to the label classifier. Confidence estimator for prediction The feature extractor Extracted target domain features W This indicates the number of fault categories.

3. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to claim 1, characterized in that, The intermediate mixing domain Constructed based on the interpolation method of optimal transmission: in, Represents parameters Hybrid transformation operator, Used to control noise Strength to improve robustness The covariance matrix representing the feature estimation of the target domain, and the pseudo-label. The predicted class assignment representing the target domain is encoded as a one-hot vector; the mixing ratio parameter To control the contribution weight of source domain information during the migration process, it is represented as: in, This indicates the current training iteration number. This represents the total number of iterations. To control the weight of the exponential decay term, Used to control the exponential decay rate. Used to adjust the order of the exponential decay curve. Used to adjust the order of the polynomial decay curve. To control the weight of the polynomial decay term, and It is fixed at 1 to ensure that the hybrid domain is aligned with the source domain in the early stages of training, so as to achieve full learning of the source domain information.

4. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to claim 1, characterized in that, In step five, the classification loss in the total loss function includes: The classification loss for labeled samples in the source domain, expressed using the standard cross-entropy loss, is as follows: in, N s The number of labeled samples in the source domain. W It is the number of fault categories; It is the feature extractor Extracted source domain features, The label classifier Output category probability prediction, for Network parameters; It is an indicator function; its value is 1 when the subscript and superscript are equal, and 0 otherwise. The classification loss for mixed-domain samples, expressed using the standard cross-entropy loss, is as follows: in, N t It is the number of unlabeled samples in the target domain. The feature extractor Extracted target domain features The mixing ratio parameter changes dynamically with each training round. It is also an indicator function; The total classification loss is the sum of the classification losses from the source domain and the mixed domain, that is: 。 5. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to claim 1, characterized in that, In step five, the domain adversarial loss in the total loss function includes: The adversarial loss between the source and target domains, expressed as a binary classification cross-entropy loss, is as follows: in, N s It refers to the number of labeled samples in the source domain. It is the feature extractor Extracted source domain features, N t It is the number of unlabeled samples in the target domain. yes Extracted target domain features This refers to the neighborhood discriminator. The predicted output, for Network parameters, This represents the gradient reversal backpropagation function; The domain adversarial loss for the mixed domain is expressed using the binary classification cross-entropy loss as follows: in N m It is the number of samples in the mixed domain; yes Extracted hybrid domain features; The total domain adversarial loss is the sum of the two losses mentioned above, namely: 。 6. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to claim 1, characterized in that, In step five, the distribution difference loss in the total loss function is the multi-kernel maximum mean difference loss (MK-MMD) between the feature vectors of the source domain samples and the feature vectors of the target domain samples, expressed as: in, Representing Hilbert space, This represents the kernel function that maps the probability distribution to the reproducing kernel Hilbert space, and employs a multi-kernel approach. express; Indicates the expected value. These represent samples from the source and target domains, respectively. For vector dimensions; These are weighting parameters for different cores. It refers to the number of cores. Represents the kernel function.

7. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to claim 1, characterized in that, In step five, the total loss function includes the target domain pseudo-label conditional consistency loss, and the label classifier is used when the model parameters are not frozen. When predicting outputs in the target domain and freezing model parameters The positive KL divergence representation between pseudo-labels generated in the target domain serves to improve the quality of pseudo-label predictions in the target domain by stabilizing the model's output in the target domain; the target domain pseudo-label conditional consistency loss is expressed as: in, and These represent the parameters before and after freezing, respectively. j The predicted probability distribution of a sample in the target domain. This represents the expected value of the variance of the probability distribution. This represents the positive KL divergence function.

8. The method for cross-domain fault diagnosis of rotating machinery based on dynamic evolution and wavelet dual-path structure according to any one of claims 1-7, characterized in that, In step five: Minimize the cross-entropy classification loss of source and mixed domain samples during model training. Loss due to distribution differences between source and target domain samples Conditional consistency loss with target domain pseudo-labels And the domain adversarial loss that maximizes the number of samples from the source domain, target domain, and mixed domain. The total loss function is then expressed as: in To balance hyperparameters, reduce them in the early stages of training. The weights are gradually increased with each training iteration of the model. value, Represented as: in This represents the current training iteration number. This represents the total number of iterations.

Images