A Fault Diagnosis Method for Rotating Machinery Based on Transfer Learning

Abstract

This application provides a fault diagnosis method for rotating machinery based on transfer learning, comprising: performing multimodal noise suppression processing on the acquired raw vibration signal to generate a denoised dataset; inputting the denoised dataset into a multilayer deep residual network for feature extraction to obtain a spatiotemporal synchronized feature set; wherein the spatiotemporal synchronized feature set includes source domain data and target domain data; inputting the spatiotemporal synchronized feature set into a physically constrained multi-kernel domain alignment model to obtain domain-invariant features; and performing fault diagnosis on the domain-invariant features through an adversarial transfer fault classifier to obtain the target domain fault type. This method can achieve highly robust diagnosis in scenarios with multiple domain differences, such as across operating conditions and across sensors, under strong noise environments.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, and in particular relates to a fault diagnosis method for rotating machinery based on transfer learning. Background Technology

[0002] In recent years, fault diagnosis technology for rotating machinery based on transfer learning has become a research hotspot in industrial intelligent operation and maintenance. Its core idea is to reduce the feature distribution differences between the source domain (such as laboratory standard data) and the target domain (field operating condition data) through domain adaptation methods. Typical technical solutions include deep networks based on adversarial training (such as DANN), feature distribution alignment (such as MMD metric), and subspace mapping. These methods extract domain-invariant features by sharing network layers, achieving limited knowledge transfer across equipment or operating conditions in scenarios such as bearings and gearboxes.

[0003] However, existing methods still face three challenges: First, they are not adaptable enough to nonlinear domain shifts caused by dynamic changes in operating parameters (such as variable speed and variable load), leading to feature alignment failure. Second, strong noise interference in real industrial environments can damage the robustness of the transferred features and reduce diagnostic stability. Third, most methods rely on a small amount of labeled data in the target domain to fine-tune the model, which is difficult to meet the engineering needs of completely unlabeled scenarios. Summary of the Invention

[0004] Therefore, it is necessary to provide a fault diagnosis method for rotating machinery based on transfer learning to address the above-mentioned technical problems. This method can achieve highly robust diagnosis in scenarios with multiple domain differences, such as cross-operating conditions and cross-sensor differences, under strong noise conditions.

[0005] This application provides a method for fault diagnosis of rotating machinery based on transfer learning, including:

[0006] The acquired raw vibration signal is subjected to multimodal noise suppression processing to generate a denoised dataset;

[0007] The denoised dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set, which includes source domain data and target domain data.

[0008] By inputting the spatiotemporal synchronization feature set into a physically constrained multi-kernel domain alignment model, domain-invariant features are obtained.

[0009] By using an adversarial migration fault classifier to diagnose faults invariant features in the target domain, the fault type in the target domain can be obtained.

[0010] In one embodiment, the acquired raw vibration signal is subjected to multimodal noise suppression processing to generate a denoised dataset, including:

[0011] Noise enhancement is performed on the original vibration signal based on an industrial noise dataset to generate a mixed noise signal.

[0012] The noise mixture signal is normalized to obtain a standardized noise signal;

[0013] By dynamically adjusting the sliding window length through signal spectrum analysis, the standardized noise signal is segmented and processed to generate an anti-interference noise signal sample set.

[0014] Wavelet threshold denoising is performed on the original vibration signal based on the anti-interference noise signal sample set to generate a denoised dataset.

[0015] In one embodiment, the denoised dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set, including:

[0016] Spatial features are extracted from the denoising dataset using a ResNet-CBAM composite structure. The channel attention weights are calculated using the following formula:

[0017] ;

[0018] in, For channel attention weights, This is the average pooling result along the channel dimension. This is the result of max pooling along the channel dimension. This is the weight matrix of the fully connected layer. Indicates the activation function;

[0019] Spatial features are input into the LSTM temporal modeling layer, and time series dependencies are established through a gating mechanism to obtain temporal features. The output of the forget gate is calculated using the following formula:

[0020] ;

[0021] in, Output for the forget gate. Here is the forget gate weight matrix. This is the hidden state vector from the previous time step. Input features for the current time step. For bias terms;

[0022] The spatiotemporal synchronization feature set is obtained by coupling spatial and temporal features using the following formula:

[0023] ;

[0024] in, For spatiotemporal synchronization feature set, For spatial features, As a time feature, The coupling coefficient is obtained by optimization based on backpropagation.

[0025] In one embodiment, the spatiotemporal synchronization feature set is input into a physically constrained multi-kernel domain alignment model to obtain domain-invariant features, including:

[0026] The distribution difference degree is obtained by calculating the multi-kernel maximum mean difference of the spatiotemporal synchronization feature set.

[0027] Based on the distribution difference metric and the standard deviation of the classification task gradient, dynamically generate the domain alignment loss weight coefficients;

[0028] Physical constraint regularization terms are constructed for the spatiotemporal synchronization feature set based on the equations of rotational machinery dynamics.

[0029] By alternately optimizing the parameters and domain alignment loss weight coefficients of the multilayer deep residual network through an adversarial training mechanism, and combining the physical constraint regularization term, domain-invariant features are generated.

[0030] In one embodiment, noise enhancement is performed on the original vibration signal based on an industrial noise dataset to generate a mixed noise signal, including:

[0031] By using an adversarial generative network to modally augment an industrial noise dataset, multimodal industrial noise including Gaussian white noise, impulse noise, and gear meshing harmonic noise is generated.

[0032] Bandpass filtering is performed on multimodal industrial noise based on the spectral characteristics of the target domain device to generate spectral matched noise;

[0033] The spectral matching noise is superimposed on the original vibration signal to generate a noise-mixed signal.

[0034] In one embodiment, spatial features are extracted from the denoising dataset using a ResNet-CBAM composite structure to obtain spatial features, including:

[0035] Initial spatial features are generated by extracting convolutional features from the denoised dataset using a residual network.

[0036] The initial spatial features are weighted by channel dimension allocation using the channel attention module to generate channel-weighted features.

[0037] Spatial features are generated by focusing local areas on channel-weighted features using a spatial attention module.

[0038] In one embodiment, spatial features are input into the LSTM temporal modeling layer, and time series dependencies are established through a gating mechanism to obtain temporal features, including:

[0039] Spatial features are input into a bidirectional LSTM network to extract forward temporal dependencies and generate forward temporal features;

[0040] Backward temporal dependencies are extracted using a reverse LSTM network to generate backward temporal features;

[0041] The forward and backward time series features are concatenated to generate time features.

[0042] In one embodiment, a multi-kernel maximum mean difference calculation is performed on the spatiotemporal synchronization feature set to obtain the distribution difference degree, including:

[0043] Calculate the distribution distance of the Gaussian kernel function between the source domain data and the target domain data respectively, and generate the Gaussian kernel difference.

[0044] Calculate the distribution distance of the Laplace kernel function between the source domain data and the target domain data respectively, and generate the Laplace kernel dissimilarity.

[0045] The Gaussian kernel dissimilarity and the Laplace kernel dissimilarity are weighted and fused to generate the distribution dissimilarity.

[0046] In one embodiment, a physical constraint regularization term is constructed on the spatiotemporal synchronization feature set based on the equations of rotational machinery dynamics, including:

[0047] Extract the vibration signal spectrum from the spatiotemporal synchronization feature set and generate characteristic frequency components;

[0048] Calculate the matching degree between the characteristic frequency components and the theoretical fault characteristic frequencies, and generate frequency constraints.

[0049] An energy conservation regularization term is constructed based on frequency constraints, and a physical constraint regularization term is generated.

[0050] In one embodiment, the parameters and domain alignment loss weight coefficients of a multi-layer deep residual network are alternately optimized through an adversarial training mechanism, and combined with a physical constraint regularization term, domain-invariant features are generated, including:

[0051] The cross-entropy loss of the source domain data is calculated using a classifier to generate the classification loss gradient.

[0052] Calculate the domain alignment loss based on the distribution dissimilarity and generate domain adversarial gradients;

[0053] Projection resolution is performed on the classification loss gradient and the domain adversarial gradient to generate optimized domain-invariant features.

[0054] The aforementioned transfer learning-based rotating machinery fault diagnosis method generates a denoised dataset by performing multimodal noise suppression processing on the original vibration signal to eliminate strong noise interference. A multi-layer deep residual network is then used to extract a spatiotemporal synchronization feature set from the denoised data. This feature set integrates the spatiotemporal correlation between source and target domain data, effectively capturing common features across operating conditions and sensors. Furthermore, the spatiotemporal synchronization feature set is input into a physically constrained multi-kernel domain alignment model. Through joint optimization using multi-kernel function metrics and dynamic constraints, domain-invariant features are generated to eliminate multi-domain differences. Finally, an adversarial transfer fault classifier is used to diagnose the domain-invariant features, achieving accurate classification of the target domain fault type. This method, through the synergistic effect of noise suppression, cross-domain feature alignment, and physical constraints, preserves the essential characteristics of the fault while suppressing the coupling interference of noise and domain differences, thus achieving highly robust diagnosis in complex industrial scenarios. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 A flowchart illustrating a method for diagnosing rotating machinery faults based on transfer learning, provided by the present invention.

[0057] Figure 2 This is a flowchart illustrating a multilayer deep residual network feature extraction method provided by the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] First, a brief introduction to the terms used in the embodiments of this application will be given.

[0060] Transfer learning is a machine learning method that transfers learned knowledge (such as feature representations and model parameters) from the source domain to the target domain to address the performance degradation caused by data scarcity or distribution differences in the target domain. In industrial fault diagnosis scenarios, this method leverages the rich labeled data in the source domain to drive the training of the target domain model by mining potential correlations between different operating conditions, sensors, or equipment. This overcomes the dependence of traditional methods on the amount and quality of target domain data, significantly improving diagnostic generalization capabilities across domains. The core mechanism of transfer learning uses techniques such as domain adaptation and feature alignment to suppress domain-specific interference while preserving the essential characteristics of the fault, providing a theoretical framework for robust diagnosis under multi-domain differences.

[0061] Deep Residual Networks (ResNets) are architectures that optimize the training process of deep neural networks by introducing residual learning mechanisms. Their core idea is to construct residual mapping relationships between layers, allowing the network to effectively mitigate gradient vanishing and network degradation problems while increasing the number of layers. In the field of rotating machinery fault diagnosis, this network establishes a direct mapping path between input signals and higher-level features through cross-layer connection structures. It can stably extract discriminative spatiotemporal features from vibration signals with strong noise interference, and enhances feature representation capabilities level by level through stacked residual modules, thereby improving diagnostic robustness in complex scenarios such as cross-operating conditions and cross-sensor environments. Compared to traditional convolutional neural networks, deep residual networks achieve more efficient gradient propagation and deeper network structure design through residual learning mechanisms, providing a reliable technical foundation for multi-domain fault feature extraction in industrial equipment.

[0062] Multi-kernel Maximum Mean Discrepancy (MK-MMD) is a distribution discrepancy measurement method based on statistical learning theory. It jointly measures the probability distribution differences between source and target domain data in a regenerating kernel Hilbert space by fusing multiple kernel functions (such as Gaussian and Laplace kernels). Its core principle lies in leveraging the complementary properties of multiple kernel combinations to dynamically capture distribution shifts of different scales and types, overcoming the limitations of single kernel functions in modeling complex domain discrepancies. In transfer learning scenarios, this method drives the model to learn domain-invariant features by maximizing the mean discrepancy measure under multi-kernel mapping, effectively addressing the diagnostic performance degradation caused by data distribution shifts in scenarios involving cross-conditions and cross-sensor environments. This provides theoretical support for robust alignment in scenarios with multiple domain discrepancies.

[0063] Based on the above definitions, the implementation environment of the rotating machinery fault diagnosis method based on transfer learning provided in this application embodiment will be described. Indicatively, the implementation environment includes: multimodal sensors, a terminal, and a processor. The processor, multimodal sensors, and terminal are connected via network signals. The multimodal sensors include, but are not limited to, vibration sensors, temperature sensors, acoustic sensors, speed sensors, and current sensors. The processor can be a central processing unit, a multi-core parallel processor, or an artificial intelligence chip, etc., and is not limited here.

[0064] Based on the above explanations of terms and implementation environments, the application scenarios of the embodiments of this application are described. The rotating machinery fault diagnosis method based on transfer learning provided in the embodiments of this application can be applied to scenarios including but not limited to the following:

[0065] In the fault diagnosis of reciprocating pumps under varying operating conditions, for load fluctuations and impact noise caused by changes in the viscosity of the conveyed medium (such as switching between petroleum and chemical fluids), multimodal noise suppression is used to effectively separate piston rod wear characteristics from fluid pulsation interference. Combined with cross-operating condition spatiotemporal feature alignment, stable detection of plunger seal failure under different pressure levels is achieved, solving the problem of increased false negative rate caused by sudden changes in operating conditions in the traditional threshold method.

[0066] In cross-sensor diagnostics of marine diesel engines, considering the data heterogeneity between the cylinder head vibration sensor and the crankcase acoustic emission sensor, a spatiotemporal synchronous feature extraction network is used to fuse high-frequency impact and low-frequency vibration signals. An adversarial transfer classifier is used to achieve fault knowledge transfer between sensors, accurately identifying hidden faults such as abnormal connecting rod bearing clearance.

[0067] As an illustration, the rotating machinery fault diagnosis method based on transfer learning provided in this application embodiment can also be applied to other application scenarios. This is only an example and does not limit the specific application scenarios.

[0068] In one exemplary embodiment, such as Figure 1 As shown, a method for diagnosing rotating machinery faults based on transfer learning is provided. This embodiment illustrates the application of this method to a terminal in the aforementioned implementation environment. It is understood that this method can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 104:

[0069] Step 101: Perform multimodal noise suppression processing on the acquired original vibration signal to generate a denoised dataset.

[0070] For example, the original vibration signal can be acquired by vibration sensors installed on rotating machinery, such as accelerometers, velocity sensors, or displacement sensors. The signals acquired by these sensors typically contain various noise components, including environmental noise, electromagnetic interference noise, and background noise generated by the machinery's operation. Multimodal noise suppression can be achieved through various methods. For instance, wavelet transform-based threshold denoising removes noise by setting appropriate thresholds for different frequency bands after wavelet decomposition of the signal; or adaptive filtering methods can be used to cancel out noise components by constructing a reference signal and an adaptive filter. The denoised dataset generated after denoising processing can more clearly reflect the true vibration state of the rotating machinery, providing a more reliable data foundation for subsequent feature extraction. Through the above technical solutions, noise interference can be effectively reduced, the quality of the original vibration signal can be improved, and a good data foundation can be laid for subsequent feature extraction and fault diagnosis.

[0071] Step 102: Input the denoised dataset into a multilayer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set; wherein the spatiotemporal synchronization feature set includes source domain data and target domain data.

[0072] Specifically, a multi-layer deep residual network is a deep neural network structure composed of multiple residual blocks. Through skip connections, it effectively alleviates the gradient vanishing problem during deep network training, enabling the network to learn data features more deeply. During feature extraction, the network's convolutional and pooling layers abstract and map features layer by layer from the denoised vibration signal, thereby extracting spatiotemporal synchronous features that characterize the vibration properties of rotating machinery. For example, the spatiotemporal synchronous feature set integrates the dynamic evolution features of the vibration signal in the time dimension and relevant features collected by sensors at different locations in the spatial dimension. Source domain data typically refers to known fault feature data collected under a specific working condition or sensor configuration, while target domain data is diagnostic data collected under different working conditions or sensor conditions. Feature extraction through a multi-layer deep residual network can transform the original vibration signal into a more discriminative feature representation, providing strong support for subsequent domain alignment and fault diagnosis. Through the above technical solution, the feature information in the denoised dataset can be deeply mined, and a feature set with spatiotemporal synchronous characteristics can be extracted, providing a high-quality feature foundation for subsequent domain alignment and fault classification.

[0073] Step 103: Input the spatiotemporal synchronization feature set into the physical constraint-based multi-kernel domain alignment model to obtain domain-invariant features.

[0074] For example, due to differences in data distribution between the source and target domains—such as variations in vibration signal frequency and amplitude caused by different operating conditions and differences introduced by different sensor characteristics—domain alignment is necessary. A multi-kernel domain alignment model based on physical constraints introduces knowledge of the physical characteristics of rotating machinery, such as considering the resonant frequency of the mechanical structure and the relationship between fault characteristic frequencies and rotational speed, to guide the domain alignment process. This allows the model to more accurately identify and align common features between different domains. Multi-kernel learning methods, by combining various kernel functions, can better adapt to complex data distributions and feature structures, effectively narrowing the gap between the feature distributions of the source and target domains, resulting in domain-invariant feature representations, i.e., domain-invariant features. Through the above technical solutions, the feature distributions of the source and target domains can be effectively aligned, eliminating the influence of different operating conditions and sensor conditions, improving the domain adaptability and consistency of features, and providing robust feature representations for subsequent fault diagnosis.

[0075] Step 104: Perform fault diagnosis on domain-invariant features using an adversarial migration fault classifier to obtain the fault type of the target domain.

[0076] Specifically, the adversarial transfer fault classifier is built based on the principles of Generative Adversarial Networks (GANs), comprising a feature extractor and a classifier. During training, the classifier attempts to accurately distinguish domain-invariant features of different fault types, while the feature extractor, while generating domain-invariant features, uses adversarial training to make these features as difficult for the classifier to distinguish between the source and target domains, thereby further enhancing the domain adaptability of the features and the robustness of the classification. For example, in actual diagnosis, the domain-invariant features of the target domain are input into the adversarial transfer fault classifier. The classifier identifies and classifies the fault type of the target domain based on the learned fault feature patterns and outputs corresponding fault diagnosis results, such as normal state, specific fault type (e.g., bearing wear, gear breakage), and its severity. Through this technical solution, highly robust fault diagnosis can be achieved in high-noise environments under multiple domain differences, including cross-operating conditions and cross-sensor scenarios. This effectively improves the accuracy and reliability of rotating machinery fault diagnosis, reduces the impact of factors such as changes in operating conditions and sensor differences on the diagnostic results, and provides strong technical support for the health management and fault early warning of rotating machinery.

[0077] The aforementioned transfer learning-based rotating machinery fault diagnosis method generates a denoised dataset by performing multimodal noise suppression processing on the original vibration signal to eliminate strong noise interference. A multi-layer deep residual network is then used to extract a spatiotemporal synchronization feature set from the denoised data. This feature set integrates the spatiotemporal correlation between source and target domain data, effectively capturing common features across operating conditions and sensors. Furthermore, the spatiotemporal synchronization feature set is input into a physically constrained multi-kernel domain alignment model. Through joint optimization using multi-kernel function metrics and dynamic constraints, domain-invariant features are generated to eliminate multi-domain differences. Finally, an adversarial transfer fault classifier is used to diagnose the domain-invariant features, achieving accurate classification of the target domain fault type. This method, through the synergistic effect of noise suppression, cross-domain feature alignment, and physical constraints, preserves the essential characteristics of the fault while suppressing the coupling interference of noise and domain differences, thus achieving highly robust diagnosis in complex industrial scenarios.

[0078] In one embodiment, the acquired raw vibration signal is subjected to multimodal noise suppression processing to generate a denoised dataset, including:

[0079] The original vibration signal is enhanced with noise based on an industrial noise dataset to generate a mixed noise signal.

[0080] Specifically, and exemplarily, the original vibration signal can be acquired by a vibration sensor installed on rotating machinery, while the industrial noise dataset can be derived from background noise records in actual industrial environments or from public databases. By mixing the noise signal from the industrial noise dataset with the original vibration signal in a certain proportion, working scenarios under different noise intensities can be simulated, thereby generating a mixed noise signal. This process enhances the model's adaptability to different noise environments, providing a more challenging data foundation for subsequent noise reduction processing.

[0081] The noise mixture signal is normalized to obtain a standardized noise signal.

[0082] Specifically, the generated mixed noise signal is normalized to obtain a standardized noise signal. The purpose of normalization is to unify signals with different amplitude ranges into a standard scale to facilitate subsequent processing and analysis. For example, the z-score normalization method can be used to standardize the signal to zero mean and unit variance, making the signal amplitude distribution within a relatively fixed range. Alternatively, the Min-Max normalization method can be used to linearly map the signal amplitude to a predefined interval, such as [0,1] or [-1,1]. The standardized noise signal after normalization can eliminate the dimensional differences between different signals, improving the stability and accuracy of subsequent processing.

[0083] By dynamically adjusting the sliding window length through signal spectrum analysis, the standardized noise signal is segmented and processed to generate an anti-interference noise signal sample set.

[0084] Specifically, signal spectrum analysis reveals the energy distribution of a signal across different frequency bands. Based on the spectrum analysis results, the length of the sliding window is dynamically adjusted to cover the main characteristic frequency bands of the signal while avoiding the inclusion of excessive noise frequency components. For example, when a sudden energy change in a certain frequency band is detected, the sliding window length can be appropriately shortened to improve time resolution, thereby more accurately capturing the transient characteristics of the signal. The sliding window moves along the signal time axis, segmenting the standardized noise signal, with each segment serving as a sample, ultimately generating an anti-interference noise signal sample set. This segmentation method can adaptively adjust window parameters according to signal characteristics, effectively improving the signal's anti-interference capability and providing more meaningful sample data for subsequent noise reduction processing.

[0085] Wavelet threshold denoising is performed on the original vibration signal based on the anti-interference noise signal sample set to generate a denoised dataset.

[0086] Specifically, wavelet thresholding denoising is performed on the original vibration signal based on an anti-interference noise signal sample set to generate a denoised dataset. Wavelet thresholding denoising is a widely used signal denoising technique. Its basic idea is to decompose the signal into wavelet domains of different scales and then remove noise components by thresholding the wavelet coefficients. For example, hard thresholding or soft thresholding can be used to process the wavelet coefficients. In hard thresholding, wavelet coefficients with absolute values ​​less than the threshold are directly set to zero, while coefficients greater than or equal to the threshold remain unchanged. Soft thresholding, on the other hand, performs a certain degree of shrinkage on the retained coefficients based on hard thresholding. By selecting appropriate wavelet basis functions and threshold parameters, noise can be effectively suppressed while retaining important feature information in the signal. The denoised dataset, after denoising, can more clearly reflect the true vibration characteristics of rotating machinery, providing high-quality data support for subsequent fault diagnosis. Through the above multimodal noise suppression processing steps, noise interference in the original vibration signal can be significantly reduced, the signal-to-noise ratio can be improved, and the identifiability of fault features can be enhanced, thus providing a more reliable data foundation for fault diagnosis of rotating machinery.

[0087] like Figure 2 As shown, in one embodiment, the denoised dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set, including:

[0088] Step 201: Extract spatial features from the denoising dataset using the ResNet-CBAM composite structure to obtain spatial features. The channel attention weights are calculated using the following formula:

[0089] ;

[0090] in, For channel attention weights, This is the average pooling result along the channel dimension. This is the result of max pooling along the channel dimension. This is the weight matrix of the fully connected layer. This represents the activation function.

[0091] Specifically, the ResNet-CBAM composite structure is a feature extraction architecture that combines a deep residual network (ResNet) with a convolutional block attention module (CBAM). This network embeds a CBAM within the residual block, enhancing fault-sensitive frequency band features through a channel attention mechanism: average pooling and max pooling are performed on the input feature map along the channel dimension, and the results of both pooling are input into a shared fully connected layer. An activation function generates a channel attention weight matrix, strengthening the feature channel responses related to the fault mechanism (such as the characteristic frequency of bearing outer ring faults). Furthermore, a spatial attention module focuses on local abrupt changes in the vibration signal (such as the impact waveform of a broken gear tooth), using deformable convolutional kernels to adaptively adjust the receptive field and generate spatial features. The ResNet-CBAM composite structure, through the collaborative mechanism of channel and spatial attention, enhances fault-sensitive frequency bands and local defect features, improving the separability of spatial dimension features.

[0092] Step 202: Input the spatial features into the LSTM temporal modeling layer, establish the time series dependencies through a gating mechanism, and obtain the temporal features. The output of the forget gate is calculated using the following formula:

[0093] ;

[0094] in, Output for the forget gate. Here is the forget gate weight matrix. This is the hidden state vector from the previous time step. Input features for the current time step. This is a bias term.

[0095] Specifically, spatial features are input into a Long Short-Term Memory (LSTM) network for temporal dependency modeling. In the forget gate, the retention ratio of historical hidden states is calculated using an activation function and weight matrix, and the memory unit state is updated based on the input features at the current time step. For example, when a periodic impact signal is detected, the forget gate dynamically adjusts the decay rate of historical information to capture the periodic characteristics of the fault impact. Furthermore, a bidirectional LSTM structure is used to simultaneously extract forward and backward temporal features, enhancing the modeling capability for complex operating conditions (such as variable speed and sudden load changes). This step effectively characterizes the temporal evolution of fault features (such as the periodic impact interval of a rolling bearing spalling defect), improving the model's adaptability to dynamic operating conditions.

[0096] Step 203: Couple spatial and temporal features using the following formula to obtain the spatiotemporal synchronization feature set:

[0097] ;

[0098] in, For spatiotemporal synchronization feature set, For spatial features, As a time feature, The coupling coefficient is obtained by optimization based on backpropagation.

[0099] For example, spatial and temporal features are dynamically weighted and coupled. The coupling coefficient is optimized using a backpropagation algorithm, and the fusion ratio of spatial and temporal features is adaptively adjusted according to the noise level and operating conditions of the input signal: for example, the weight of temporal features is increased in noisy scenarios to suppress spatial noise interference, while the weight of spatial features is increased in stable operating conditions to enhance the discriminative power of local defects; ultimately, a spatiotemporal synchronous feature set that simultaneously preserves the physical mode of the fault and the temporal evolution law is generated. This processing achieves complementary enhancement of spatiotemporal features across operating conditions through an adaptive feature fusion mechanism, providing a highly robust input for subsequent domain alignment.

[0100] In one embodiment, the spatiotemporal synchronization feature set is input into a physically constrained multi-kernel domain alignment model to obtain domain-invariant features, including:

[0101] The distribution difference degree is obtained by calculating the maximum mean difference of the spatiotemporal synchronization feature set using multi-kernel methods.

[0102] Specifically, the spatiotemporal synchronization feature set is calculated using multi-kernel maximum mean difference (MK-MMD) to obtain the distribution difference degree. For example, the spatiotemporal synchronization feature set contains feature data from the source and target domains, which exhibit distribution differences due to factors such as operating conditions and sensor variations. MK-MMD, by combining various kernel functions (such as Gaussian kernels and polynomial kernels), can effectively measure the difference in feature distribution between the source and target domains in the regenerated Hilbert space. For instance, using multiple Gaussian kernels can capture distribution differences at different scales, thus more comprehensively reflecting the feature distribution mismatch between the source and target domains and providing a quantitative basis for subsequent domain alignment.

[0103] Based on the distribution difference metric and the standard deviation of the classification task gradient, the domain alignment loss weight coefficients are dynamically generated.

[0104] Specifically, the classifier gradient standard deviation is calculated in real time during training. When the classification task becomes difficult to converge (e.g., gradient fluctuations increase significantly), the domain alignment loss weights are reduced to avoid overfitting; conversely, when domain discrepancies become the primary concern, the domain alignment weights are increased to strengthen distribution alignment. For example, the historical mean of the gradient standard deviation is statistically analyzed using a sliding window, and adaptive weight coefficients are generated by combining this with the current distribution discrepancy. This process balances the optimization objectives of classification accuracy and domain alignment, improving the model's generalization stability in cross-domain scenarios.

[0105] Physical constraint regularization terms are constructed for the spatiotemporal synchronization feature set based on the equations of rotational machinery dynamics.

[0106] Specifically, physical constraint regularization terms can be constructed for the spatiotemporal synchronization feature set based on the dynamic equations of rotating machinery. For example, the dynamic equations of rotating machinery can be described as a mass-stiffness-damped system, whose vibration response is closely related to the system's physical parameters (such as mass, stiffness, and damping coefficient) and excitation conditions (such as rotational speed and load). Based on these dynamic equations, physical constraint regularization terms can be constructed, requiring the extracted domain-invariant features to conform to the physical behavior of the mechanical system. For example, the energy distribution of the constraint features in a specific frequency range can be matched with the resonant frequency of the mechanical system, or the time evolution characteristics of the constraint features can be consistent with the dynamic response characteristics of the mechanical system. This physical constraint regularization can guide the model to extract more physically meaningful features, reduce the impact of data distribution differences, and improve the interpretability and robustness of the features.

[0107] By alternately optimizing the parameters and domain alignment loss weight coefficients of the multilayer deep residual network through an adversarial training mechanism, and combining the physical constraint regularization term, domain-invariant features are generated.

[0108] Specifically, by fixing the parameters of the domain discriminator, the classification loss and physical constraint regularization term of the feature extraction network are optimized to improve fault detection capability; by fixing the parameters of the feature extraction network, the adversarial loss of the domain discriminator is optimized to drive the generation of domain-invariant features; the above process is iteratively executed until the model converges. For example, in cross-operating condition scenarios (such as variable speed and variable load), this training mechanism can effectively eliminate distribution shifts caused by differences in operating conditions, generating domain-invariant features that simultaneously satisfy classification accuracy and physical interpretability. The above technical solution, through the deep integration of multi-kernel metrics, dynamic optimization, physical constraints, and adversarial training, achieves stable expression and cross-domain transfer of fault features under multiple challenges such as noise interference, sensor heterogeneity, and sudden changes in operating conditions. It provides a highly robust solution for intelligent diagnosis of industrial equipment, and is particularly suitable for predictive maintenance scenarios of complex rotating machinery such as wind turbine gearboxes and petrochemical centrifugal compressors.

[0109] In one embodiment, noise enhancement is performed on the original vibration signal based on an industrial noise dataset to generate a mixed noise signal, including:

[0110] By using an adversarial generative network to modally augment an industrial noise dataset, multimodal industrial noise is generated, including Gaussian white noise, impulse noise, and gear meshing harmonic noise.

[0111] Specifically, industrial noise datasets can be obtained by collecting background noise data from real industrial environments, covering noise characteristics under different operating conditions and equipment states. The Generative Adversarial Network (GAN) consists of a generator and a discriminator. The generator is responsible for generating noise samples with multiple modalities, while the discriminator distinguishes between the generated noise samples and real noise samples. For example, the generator can generate Gaussian white noise with zero mean and constant power spectral density; impact noise is generated by simulating sudden impact events during equipment operation; and gear meshing harmonic noise is generated based on parameters such as gear speed and number of teeth. Through adversarial training, the generator continuously learns and generates more realistic multimodal noise. This method can expand the modalities of the industrial noise dataset, generating more comprehensive and realistic noise samples, providing diverse noise sources for subsequent noise enhancement processing. In this way, the diversity and realism of the noise data can be significantly improved, ensuring that the generated noise samples can better simulate the complex noise conditions in real industrial environments.

[0112] Bandpass filtering is performed on multimodal industrial noise based on the spectral characteristics of the target domain device to generate spectral matched noise.

[0113] Specifically, the spectral characteristics of the target domain device can be obtained through spectral analysis of its vibration signals under normal operating conditions, determining its main operating frequency range and noise-sensitive frequency band. The design of the bandpass filter should be adjusted according to these spectral characteristics, ensuring that the filtered noise is mainly concentrated within the noise-sensitive frequency band of the target domain device. For example, if the main operating frequency range of the target domain device is 500Hz to 2000Hz, the bandpass filter can be set to only allow noise within this frequency band to pass. In this way, the generated spectral-matched noise can more closely resemble the actual noise environment of the target domain device, improving the targeting and effectiveness of noise enhancement processing. This method ensures that the generated noise matches the spectral characteristics of the target domain device, making subsequent noise enhancement processing more consistent with actual application scenarios and enhancing the model's adaptability to the noise of the target domain device. This approach significantly improves the accuracy and effectiveness of noise enhancement processing, ensuring that the generated noise mixture signal better matches the actual operating environment of industrial equipment.

[0114] The spectral matching noise is superimposed on the original vibration signal to generate a noise-mixed signal.

[0115] Specifically, the original vibration signal can be acquired by a vibration sensor installed on the rotating machinery, while the spectral matching noise is the noise signal after the above processing. During the superposition process, the intensity of the noise can be controlled to simulate working scenarios under different signal-to-noise ratio conditions. For example, by adjusting the amplitude of the spectral matching noise to achieve a certain proportional relationship with the amplitude of the original vibration signal, the generated noise-mixed signal can more realistically reflect the noise interference situation of the rotating machinery in the actual industrial environment. In this way, a more challenging and practical data foundation is provided for the subsequent training of fault diagnosis algorithms. The denoised dataset after noise reduction processing can more clearly reflect the real vibration characteristics of the rotating machinery, providing high-quality data support for subsequent fault diagnosis. Through the above multimodal noise suppression processing steps, the noise interference in the original vibration signal can be significantly reduced, the signal-to-noise ratio of the signal can be improved, and the identifiability of fault characteristics can be enhanced, thereby providing a more reliable data foundation for the fault diagnosis of rotating machinery.

[0116] In one embodiment, spatial features are extracted from the denoising dataset using a ResNet-CBAM composite structure to obtain spatial features, including:

[0117] The initial spatial features are generated by extracting convolutional features from the denoised dataset using a residual network.

[0118] For example, the denoising dataset originates from vibration signals that have undergone multimodal noise suppression processing, acquired by vibration sensors mounted on rotating machinery. The residual network consists of multiple residual blocks, each containing two convolutional layers and one skip connection. During convolutional feature extraction, the denoising dataset is first convolved by the convolutional layers to extract local features from the signals. For example, different sized convolutional kernels (such as 3×3 or 5×5) can be used to capture features at different scales. The skip connection directly passes the input to subsequent layers, solving the gradient vanishing problem during deep network training and enabling the network to learn data features more deeply. The initial spatial features generated in this way effectively preserve the spatial information in the denoising dataset, providing a foundation for subsequent feature processing. This method effectively extracts local features from the denoising dataset while preserving global information of the original data through skip connections, improving the efficiency and accuracy of feature extraction.

[0119] The initial spatial features are weighted by channel dimension through the channel attention module to generate channel-weighted features.

[0120] Specifically, a channel attention module assigns channel-dimensional weights to the initial spatial features, generating channel-weighted features. The input to the channel attention module is the initial spatial features, and its goal is to highlight key feature channels and suppress irrelevant information. For example, global average pooling and global max pooling are performed on the initial spatial features to obtain the average and maximum responses for each channel, respectively. The pooling results are then processed through two fully connected layers, and activation functions (such as ReLU and Sigmoid) are used to generate channel attention weights. For instance, the fully connected layers can map the pooled features to a lower-dimensional space and then back to the original dimension, thereby learning the importance weights of each channel. The channel attention weights are then multiplied channel-by-channel with the initial spatial features to generate channel-weighted features. This method, by adaptively adjusting the channel weights, enhances the expressive power of key feature channels and improves the model's sensitivity to important features.

[0121] Spatial features are generated by focusing local areas on channel-weighted features using a spatial attention module.

[0122] Specifically, the spatial attention module takes channel-weighted features as input and aims to further enhance key fault-related regions in the feature map. For example, spatial-dimensional average pooling and max pooling are performed on the channel-weighted features to obtain the average and maximum responses at each location. These pooling results are then concatenated along the channel dimension and passed through a convolutional layer to generate a spatial attention map. For instance, a 1×1 convolutional kernel can be used to map the concatenated features onto a two-dimensional attention map. The spatial attention map is then multiplied element-wise with the channel-weighted features to generate spatial features. This method dynamically adjusts the weights of different regions based on the spatial distribution of the feature map, highlighting key regions and suppressing background noise. In this way, key fault-related regions in the feature map can be further enhanced, improving the locality and discriminative power of the features, thereby providing higher-quality feature representations for subsequent fault diagnosis.

[0123] In one embodiment, spatial features are input into the LSTM temporal modeling layer, and time series dependencies are established through a gating mechanism to obtain temporal features, including:

[0124] Spatial features are input into a bidirectional LSTM network to extract forward temporal dependencies and generate forward temporal features.

[0125] Specifically, the spatial features originate from feature data extracted by the ResNet-CBAM composite structure, capturing the vibration characteristics of rotating machinery in the spatial dimension. The bidirectional LSTM network comprises LSTM layers in two directions. The forward LSTM layer processes the input features sequentially from the beginning to the end of the time series, capturing historical information prior to the current time step. For example, when processing the vibration signal of rotating machinery, the forward LSTM layer can capture the feature evolution process of the machinery gradually entering a stable operating state after startup. The forward temporal features generated in this way reflect the forward dependencies of the time series, providing important information for subsequent temporal feature fusion. This method effectively captures long-term dependencies in the time series, improving the model's efficiency in utilizing temporal information.

[0126] Backward temporal dependencies are extracted using a reverse LSTM network to generate backward temporal features.

[0127] Specifically, the inverse LSTM layer processes input features in reverse order from the end to the beginning of the time series, capturing future information after the current time step. For example, when processing vibration signals from rotating machinery, the inverse LSTM layer can capture the characteristic change trends of the machinery before the failure. The backward time series features generated in this way can reflect the backward dependencies of the time series, providing supplementary information for subsequent time feature fusion. This method effectively supplements the future information that the forward LSTM cannot capture, enhancing the model's overall understanding of the time series.

[0128] The forward and backward time series features are concatenated to generate time features.

[0129] Specifically, forward and backward temporal features are connected along the channel dimension and compressed to a unified dimension through a fully connected layer to generate temporal features that fuse bidirectional temporal dependencies. This processing can integrate forward causal relationships and backward propagation characteristics (such as the generation and reflection process of impact waveforms), improving the global expressive power of temporal features. In this embodiment, a bidirectional long short-term memory network (Bi-LSTM) is used to perform bidirectional temporal modeling of spatial features, which can comprehensively capture the forward causal relationships and backward propagation characteristics of vibration signals, such as the periodic interval pattern of bearing spalling impact and the reflection attenuation effect of impact waveforms; by splicing and fusing features from forward LSTM and backward LSTM, the global expressive power of temporal features is enhanced, adapting to complex temporal evolution patterns under non-stationary conditions such as sudden load changes and speed fluctuations. This scheme combines a dynamic adjustment gating mechanism based on equipment rotation speed parameters to make the time-series modeling process conform to the physical motion laws of rotating machinery, improve feature interpretability, and maintain high diagnostic stability across different operating conditions. In particular, it significantly improves the detection sensitivity for early weak periodic impacts (such as early gear cracks), providing a highly robust time-series feature representation basis for rotating machinery fault diagnosis.

[0130] Furthermore, multi-kernel maximum mean difference calculation is performed on the spatiotemporal synchronization feature set to obtain the distribution difference degree, including:

[0131] Calculate the distribution distance of the Gaussian kernel function between the source domain data and the target domain data respectively, and generate the Gaussian kernel difference.

[0132] Specifically, the spatiotemporal synchronization feature set contains feature data from the source and target domains, which exhibit distributional differences due to factors such as operating conditions and sensor variations. The Gaussian kernel function, by mapping the data to a high-dimensional space, effectively captures the nonlinear distribution characteristics of the data. For example, the radial basis function (RBF) can be used as the Gaussian kernel function, which transforms the Euclidean distance between data points into similarity using an exponential function. The Gaussian kernel dissimilarity is obtained by calculating the maximum mean difference (MMD) between the source and target domain data under the Gaussian kernel. The Gaussian kernel dissimilarity generated in this way effectively reflects the nonlinear distributional differences between the source and target domains, improving the ability to model complex data distributions.

[0133] Calculate the distribution distance of the Laplace kernel function between the source domain data and the target domain data respectively, and generate the Laplace kernel dissimilarity.

[0134] Specifically, the Laplace kernel function can also map data to a high-dimensional space, but it is more robust to data sparsity and outliers. For example, the Laplace kernel function calculates the first-order distance between data points and converts it to similarity using an exponential function. The Laplace kernel dissimilarity is obtained by calculating the maximum mean difference between the source and target domain data under the Laplace kernel. This step further quantifies the distributional differences between the source and target domains in the Laplace kernel space. For instance, when processing vibration signals containing characteristics of sudden faults, the Laplace kernel can effectively capture sparse feature changes in the signal, providing supplementary information for subsequent domain alignment. The Laplace kernel dissimilarity generated in this way enhances the modeling ability for sparse regions of data and improves the robustness of the model.

[0135] The Gaussian kernel dissimilarity and the Laplace kernel dissimilarity are weighted and fused to generate the distribution dissimilarity.

[0136] Specifically, Gaussian kernel dissimilarity and Laplace kernel dissimilarity can be linearly combined by introducing weighting coefficients. For example, distribution dissimilarity can be expressed as a weighted sum of Gaussian kernel dissimilarity and Laplace kernel dissimilarity, with the weighting coefficients adjusted according to actual needs. For instance, in some cases, the Gaussian kernel may be more sensitive to the local structure of the data, while the Laplace kernel is more sensitive to the global structure; weighted fusion can balance these two characteristics. This technical approach significantly improves the quantification accuracy of the distribution differences between the source and target domains, providing more accurate guidance for subsequent domain alignment. This method effectively integrates the advantages of multi-kernel functions, improves the modeling ability for complex data distributions, enhances the robustness and adaptability of the model, and thus provides more reliable support for fault diagnosis of rotating machinery.

[0137] Furthermore, physical constraint regularization terms are constructed for the spatiotemporal synchronization feature set based on the equations of rotational machinery dynamics, including:

[0138] Extract the vibration signal spectrum from the spatiotemporal synchronization feature set to generate characteristic frequency components.

[0139] Specifically, a fast Fourier transform can be used to convert the time-domain vibration signal into a frequency-domain representation, extracting characteristic frequency components, such as the fault characteristic frequencies of rolling bearings (e.g., outer ring fault frequency BPFO, inner ring fault frequency BPFI), gear meshing frequencies, and their sideband components. This step, through frequency-domain energy distribution analysis, reveals physical characteristics directly related to the fault mechanism of rotating machinery, providing a quantitative basis for subsequent constraint construction.

[0140] Calculate the matching degree between the characteristic frequency components and the theoretical fault characteristic frequencies, and generate frequency constraints.

[0141] Specifically, theoretical fault frequencies are generated based on the dynamic equations of rotating machinery (such as the formula for calculating the characteristic frequencies of bearing faults). The degree of matching between the actual characteristic frequencies and theoretical values ​​is measured using cosine similarity or energy proportion, thus generating frequency constraints. For example, when a fault is detected in the outer ring of a rolling bearing, the energy proportion of the domain-invariant feature in the BPFO band is forced to be no less than the theoretical threshold. This step constrains the model's learning direction through physical laws, avoiding interference from irrelevant frequency band noise.

[0142] An energy conservation regularization term is constructed based on frequency constraints, and a physical constraint regularization term is generated.

[0143] Specifically, the feature frequency matching degree is used as the penalty coefficient of the regularization term. The deviation between the actual features and the theoretical values ​​is quantified by the mean square error or KL divergence and added to the model's total loss function. For example, in bearing fault diagnosis, the regularization term forces the energy distribution of the domain-invariant features in the BPFO band to be consistent with the theoretical fault feature frequencies. This processing can suppress invalid feature migration and ensure that the model retains physical features strongly correlated with the mechanical failure mechanism. This embodiment can significantly improve the physical interpretability and cross-domain generalization ability of the fault diagnosis model. The forced domain-invariant features retain the key frequency band energy strongly correlated with the mechanical failure mechanism and suppress noise and irrelevant harmonic interference. At the same time, the energy retention regularization term guides the model to focus on common physical features, reduces the sensitivity of distribution shift under cross-operating conditions and cross-sensor scenarios, avoids invalid feature migration, and ultimately achieves highly robust and reliable rotating machinery fault diagnosis.

[0144] In one embodiment, the parameters and domain alignment loss weight coefficients of a multi-layer deep residual network are alternately optimized through an adversarial training mechanism, and combined with a physical constraint regularization term, domain-invariant features are generated, including:

[0145] The cross-entropy loss of the source domain data is calculated using a classifier to generate the classification loss gradient.

[0146] The domain alignment loss is calculated based on the distribution difference, and the domain adversarial gradient is generated.

[0147] Projection resolution is performed on the classification loss gradient and the domain adversarial gradient to generate optimized domain-invariant features.

[0148] Specifically, this embodiment utilizes an adversarial training mechanism to collaboratively optimize classification performance and domain-invariant feature generation: First, the cross-entropy loss of the source domain data is calculated using a classifier, and backpropagation generates a classification loss gradient to optimize network parameters and improve the accuracy of fault category discrimination. Then, a domain adversarial gradient is generated based on the domain alignment loss. A gradient inversion layer negates the gradient of the domain discriminator, driving the feature extraction network to generate a feature representation for the confused domain discriminator, eliminating distribution differences across sensors and operating conditions. Furthermore, a gradient projection algorithm is employed to eliminate directional conflicts between the classification gradient and the domain adversarial gradient. Combined with a physical constraint regularization term, parameters are jointly updated, ensuring classification accuracy while forcing features to conform to physical laws. This scheme achieves high discriminative power and cross-domain generalization ability of domain-invariant features through dynamic gradient projection and multi-objective collaborative optimization, simultaneously improving diagnostic accuracy and model robustness in complex scenarios.

[0149] In summary, the rotating machinery fault diagnosis method based on transfer learning provided in this application generates a high-fidelity mixed signal by performing multimodal noise enhancement and spectral matching filtering on the original vibration signal using an adversarial generative network. This is combined with wavelet threshold denoising to generate an anti-interference denoised dataset, effectively suppressing the interference of strong industrial noise on signal quality. An improved ResNet-CBAM composite network is used to extract spatiotemporal synchronization features. A channel attention mechanism is used to strengthen the fault-sensitive frequency band, and a spatial attention module is used to focus on local defect regions. A bidirectional LSTM network is combined to model temporal dependencies, achieving a spatiotemporal joint representation of fault features. Multi-kernel maximum mean difference (MK-MMD) is used to quantify the global-local distribution difference between the source and target domains, dynamically adjusting the domain alignment loss weights. Physical constraint regularization terms (such as bearing fault feature frequency energy retention constraints) are introduced from the rotating machinery dynamics equations to suppress ineffective feature transfer. An adversarial training mechanism is used to alternately optimize the classification loss and domain adversarial gradient, combined with gradient projection to resolve optimization direction conflicts, generating highly discriminative and physically consistent domain-invariant features that drive the adversarial transfer classifier to achieve cross-domain fault diagnosis. This technical solution achieves highly robust diagnosis across multiple domain differences, including cross-operating conditions and cross-sensor scenarios, through the synergistic effect of four core technologies: multimodal noise suppression, spatiotemporal synchronization feature extraction, multi-kernel domain alignment, and physical constraint regularization.

[0150] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0151] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A method for fault diagnosis of rotating machinery based on transfer learning, characterized in that, The method includes: The acquired raw vibration signal is subjected to multimodal noise suppression processing to generate a denoised dataset; The noise reduction dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set; wherein the spatiotemporal synchronization feature set includes source domain data and target domain data. The spatiotemporal synchronization feature set is input into a physically constrained multi-kernel domain alignment model to obtain domain-invariant features; Fault diagnosis is performed on the domain-invariant features using an adversarial migration fault classifier to obtain the target domain fault type; The step of inputting the denoised dataset into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set includes: Spatial features are extracted from the denoised dataset using a ResNet-CBAM composite structure, where the channel attention weights are calculated using the following formula: ； in, The channel attention weights, This is the average pooling result along the channel dimension. This is the result of max pooling along the channel dimension. This is the weight matrix of the fully connected layer. Indicates the activation function; The spatial features are input into the LSTM temporal modeling layer, and time series dependencies are established through a gating mechanism to obtain temporal features. The forget gate output is calculated using the following formula: ； in, The output of the forget gate, Here is the forget gate weight matrix. This is the hidden state vector from the previous time step. Input features for the current time step. For bias terms; The spatiotemporal synchronization feature set is obtained by coupling the spatial features and the temporal features using the following formula: ； in, For the spatiotemporal synchronization feature set, For the aforementioned spatial features, For the time feature, The coupling coefficient is obtained by optimization based on backpropagation; The step of inputting the spatiotemporal synchronization feature set into a physically constrained multi-kernel domain alignment model to obtain domain-invariant features includes: The spatiotemporal synchronization feature set is subjected to multi-kernel maximum mean difference calculation to obtain the distribution difference degree; Based on the distribution difference metric and the classification task gradient standard deviation, dynamically generate domain alignment loss weight coefficients; A physical constraint regularization term is constructed for the spatiotemporal synchronization feature set based on the equations of rotational machinery dynamics. The parameters of the multilayer deep residual network and the domain alignment loss weight coefficients are alternately optimized through an adversarial training mechanism, and the domain invariant features are generated by combining the physical constraint regularization term.

2. The method according to claim 1, characterized in that, The step of performing multimodal noise suppression processing on the acquired original vibration signal to generate a denoised dataset includes: The original vibration signal is enhanced with noise based on an industrial noise dataset to generate a mixed noise signal. The noise mixture signal is normalized to obtain a standardized noise signal; By dynamically adjusting the sliding window length through signal spectrum analysis, the standardized noise signal is segmented to generate an anti-interference noise signal sample set. Based on the anti-interference noise signal sample set, wavelet threshold denoising is performed on the original vibration signal to generate the denoised dataset.

3. The method according to claim 2, characterized in that, The noise enhancement of the original vibration signal based on the industrial noise dataset to generate a mixed noise signal includes: The industrial noise dataset is modally augmented using an adversarial generative network to generate multimodal industrial noise that includes Gaussian white noise, impulse noise, and gear meshing harmonic noise. The multimodal industrial noise is bandpass filtered based on the spectral characteristics of the target domain device to generate spectral matched noise; The spectral matching noise is superimposed on the original vibration signal to generate the noise mixture signal.

4. The method according to claim 1, characterized in that, The spatial features are extracted from the denoised dataset using the ResNet-CBAM composite structure, including: The noise reduction dataset is subjected to convolutional feature extraction using a residual network to generate initial spatial features; The initial spatial features are weighted by channel dimension allocation using the channel attention module to generate channel-weighted features. The spatial features are generated by focusing local regions on the channel-weighted features using a spatial attention module.

5. The method according to claim 1, characterized in that, The step of inputting the spatial features into the LSTM temporal modeling layer and establishing time series dependencies through a gating mechanism to obtain temporal features includes: The spatial features are input into a bidirectional LSTM network to extract forward temporal dependencies and generate forward temporal features. Backward temporal dependencies are extracted using a reverse LSTM network to generate backward temporal features; The forward temporal features and the backward temporal features are concatenated to generate the temporal features.

6. The method according to claim 1, characterized in that, The step of performing multi-kernel maximum mean difference calculation on the spatiotemporal synchronization feature set to obtain the distribution difference degree includes: Calculate the distribution distance of the Gaussian kernel function between the source domain data and the target domain data respectively, and generate the Gaussian kernel difference. Calculate the distribution distance of the Laplace kernel function between the source domain data and the target domain data respectively, and generate the Laplace kernel difference. The Gaussian kernel difference and the Laplace kernel difference are weighted and fused to generate the distribution difference.

7. The method according to claim 1, characterized in that, The construction of physical constraint regularization terms for the spatiotemporal synchronization feature set based on the equations of rotational machinery dynamics includes: Extract the vibration signal spectrum from the spatiotemporal synchronization feature set to generate characteristic frequency components; Calculate the matching degree between the characteristic frequency components and the theoretical fault characteristic frequencies to generate frequency constraints; An energy conservation regularization term is constructed based on the frequency constraint, and the physical constraint regularization term is generated.

8. The method according to claim 1, characterized in that, The process of alternately optimizing the parameters of the multi-layer deep residual network and the domain alignment loss weight coefficients through an adversarial training mechanism, combined with the physical constraint regularization term, to generate the domain-invariant features includes: The cross-entropy loss of the source domain data is calculated using a classifier to generate the classification loss gradient. Calculate the domain alignment loss based on the distribution difference degree, and generate the domain adversarial gradient; The classification loss gradient and the domain adversarial gradient are projected and resolved to generate the optimized domain-invariant features.

Images