Rotary machinery fault diagnosis method based on transfer learning

Abstract

The invention provides a rotary machine fault diagnosis method based on transfer learning, and the method comprises the steps: carrying out the multi-mode noise suppression processing of an obtained original vibration signal, and generating a noise reduction data set; inputting the noise reduction data set into a multi-layer deep residual network for feature extraction to obtain a time-space synchronization feature set; wherein the space-time synchronization feature set comprises source domain data and target domain data; inputting the time-space synchronization feature set into a multi-core domain alignment model based on physical constraints to obtain domain invariant features; and performing fault diagnosis on the domain invariant features through an adversarial migration fault classifier to obtain a target domain fault type. By adopting the method, high-robustness diagnosis of cross-working-condition, cross-sensor and other multi-domain difference scenes in a strong noise environment can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of fault diagnosis, and in particular relates to a rotating machinery fault diagnosis method based on transfer learning. Background Art

[0002] In recent years, rotating machinery fault diagnosis technology based on transfer learning has become a research hotspot in industrial intelligent operation and maintenance. Its core approach is to narrow the feature distribution gap between the source domain (e.g., laboratory standard data) and the target domain (field operating data) through domain adaptation methods. Typical technical solutions include deep networks based on adversarial training (e.g., DANN), feature distribution alignment (e.g., MMD metric), and subspace mapping. These methods extract domain-invariant features through shared network layers, enabling 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 lack adaptability to nonlinear domain offsets caused by dynamic changes in operating parameters (such as variable speed and variable load), resulting in feature alignment failure; second, strong noise interference in actual industrial environments will destroy the robustness of migrated 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 makes it difficult to meet the engineering needs of completely unlabeled scenarios. Summary of the Invention

[0004] Based on this, it is necessary to provide a rotating machinery fault diagnosis method based on transfer learning to address the above technical problems, which can achieve highly robust diagnosis in multiple domain difference scenarios such as across working conditions and sensors in a strong noise environment.

[0005] This application provides a rotating machinery fault diagnosis method based on transfer learning, including: Perform multimodal noise suppression on the acquired original vibration signal to generate a noise reduction dataset; The denoised dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronized feature set; the spatiotemporal synchronized feature set includes source domain data and target domain data; The spatiotemporal synchronized feature set is input into a multi-core domain alignment model based on physical constraints to obtain domain-invariant features. Fault diagnosis is performed on domain-invariant features using an adversarial transfer fault classifier to obtain the target domain fault type.

[0006] In one embodiment, performing multimodal noise suppression processing on the acquired original vibration signal to generate a noise reduction dataset includes: Based on the industrial noise dataset, the original vibration signal is enhanced to generate a noise mixed signal; Normalizing the noise mixed signal to obtain a standardized noise signal; Dynamically adjust the sliding window length through signal spectrum analysis, segment the standardized noise signal, and generate an anti-interference noise signal sample set; Based on the anti-interference noise signal sample set, the original vibration signal is subjected to wavelet threshold denoising to generate a denoised data set.

[0007] In one embodiment, the denoised dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronized feature set, including: The spatial dimension of the denoising dataset is extracted using the ResNet-CBAM composite structure to obtain spatial features, where the channel attention weight is calculated using the following formula: ; in, is the channel attention weight, is the channel dimension average pooling result, is the maximum pooling result in the channel dimension, is the weight matrix of the fully connected layer, represents the activation function; The spatial features are input into the LSTM time series modeling layer, and the time series dependency is established through the gating mechanism to obtain the temporal features. The forget gate output is calculated by the following formula: ; in, is the output of the forget gate, is the forget gate weight matrix, is the hidden state vector at the previous moment, Input features for the current time step, is the bias term; The spatial and temporal features are coupled using the following formula to obtain the spatiotemporal synchronization feature set: ; in, is the spatiotemporal synchronization feature set, is the spatial feature, is the time feature, is the coupling coefficient obtained by back-propagation optimization.

[0008] In one embodiment, the spatiotemporal synchronization feature set is input into a multi-core domain alignment model based on physical constraints to obtain domain-invariant features, including: Perform multi-core maximum mean difference calculation on the spatiotemporal synchronization feature set to obtain the distribution difference; Dynamically generate domain alignment loss weight coefficients based on distribution difference metrics and classification task gradient standard deviation; According to the rotating machinery dynamics equation, a physical constraint regularization term is constructed for the spatiotemporal synchronization feature set; Through the adversarial training mechanism, the parameters of the multi-layer deep residual network and the domain alignment loss weight coefficient are alternately optimized, combined with the physical constraint regularization term to generate domain-invariant features.

[0009] In one embodiment, noise enhancement is performed on an original vibration signal based on an industrial noise dataset to generate a noise mixed signal, including: The industrial noise dataset is modally expanded using a generative adversarial network to generate multimodal industrial noise including Gaussian white noise, impact noise, and gear meshing harmonic noise. Bandpass filtering is performed on multimodal industrial noise based on the spectral characteristics of the target domain equipment to generate spectrum-matched noise. The spectrum matching noise is superimposed on the original vibration signal to generate a noise mixed signal.

[0010] In one embodiment, spatial dimension feature extraction is performed on the denoised dataset using a ResNet-CBAM composite structure to obtain spatial features, including: Perform convolution feature extraction on the denoised dataset through the residual network to generate initial spatial features; The initial spatial features are weighted in the channel dimension by the channel attention module to generate channel weighted features; The spatial attention module is used to focus on the local area of ​​the channel weighted features to generate spatial features.

[0011] In one embodiment, the spatial features are input into the LSTM time series modeling layer, and the time series dependency is established through a gating mechanism to obtain the temporal features, including: Input the spatial features into the bidirectional LSTM network, extract the forward temporal dependency, and generate the forward temporal features; Extract backward temporal dependencies through the reverse LSTM network and generate backward temporal features; The forward time series features and the backward time series features are concatenated to generate time features.

[0012] In one embodiment, a multi-core maximum mean difference calculation is performed on the spatiotemporal synchronization feature set to obtain a distribution difference, including: Calculate the distribution distance of the Gaussian kernel function between the source domain data and the target domain data respectively to 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 to generate the Laplace kernel difference; The Gaussian kernel difference and the Laplace kernel difference are weightedly fused to generate the distribution difference.

[0013] In one embodiment, a physical constraint regularization term is constructed for the spatiotemporal synchronization feature set according to the rotating machinery dynamics equation, including: Extract the vibration signal spectrum from the spatiotemporal synchronization feature set to generate characteristic frequency components; Calculate the matching degree between the characteristic frequency component and the theoretical fault characteristic frequency and generate frequency constraint conditions; An energy-preserving regularization term is constructed based on the frequency constraint condition to generate a physical constraint regularization term.

[0014] In one embodiment, the parameters of a multi-layer deep residual network and the domain alignment loss weight coefficients are alternately optimized through an adversarial training mechanism, combined with a physical constraint regularization term, to generate domain-invariant features, including: Calculate the cross entropy loss of the source domain data through the classifier and generate the classification loss gradient; Calculate domain alignment loss based on distribution difference and generate domain adversarial gradient; The classification loss gradient and domain adversarial gradient are projected and eliminated to generate optimized domain invariant features.

[0015] The above-mentioned rotating machinery fault diagnosis method based on transfer learning generates a denoised data set by performing multimodal noise suppression processing on the original vibration signal to eliminate strong noise interference; a multi-layer deep residual network is used to extract a spatiotemporal synchronization feature set from the denoised data. This feature set integrates the spatiotemporal correlation of the source domain and target domain data, effectively capturing common features across working conditions and sensors; the spatiotemporal synchronization feature set is further input into a multi-core domain alignment model based on physical constraints, and domain-invariant features are generated through joint optimization of multi-core function metrics and dynamic constraints to eliminate multiple domain differences; the domain-invariant features are diagnosed by an adversarial transfer fault classifier to achieve accurate classification of the target domain fault type. The above-mentioned method suppresses the coupling interference of noise and domain differences while retaining the essential characteristics of the fault through the synergistic effect of noise suppression, cross-domain feature alignment and physical constraints, thereby achieving highly robust diagnosis in complex industrial scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0017] Figure 1 A schematic flow chart of a rotating machinery fault diagnosis method based on transfer learning provided by the present invention; Figure 2A schematic flow chart of a multi-layer deep residual network feature extraction method provided by the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0019] First, a brief introduction is given to the terms involved in the embodiments of this application.

[0020] Transfer learning is a machine learning method whose core is to transfer learned knowledge (such as feature representations, model parameters, etc.) from the source domain to the target domain to address the problem of model performance degradation caused by data scarcity or distribution differences in the target domain. In industrial fault diagnosis scenarios, this method uses the rich labeled data in the source domain to drive the training of the target domain model by exploring the potential correlations between different working conditions, sensors, or equipment, thereby breaking through the traditional method's dependence on the amount of target domain data and the quality of annotations, and significantly improving the diagnostic generalization ability in cross-domain scenarios. The core mechanism of transfer learning is to suppress the interference of domain differences while retaining the essential characteristics of the fault through technical means such as domain adaptation and feature alignment, providing a theoretical framework for robust diagnosis under multiple domain differences.

[0021] The Deep Residual Network (ResNet) is an architecture that optimizes the deep neural network training process by introducing a residual learning mechanism. Its core idea is to construct a residual mapping relationship between layers, allowing the network to effectively alleviate the problems of gradient vanishing and network degradation while increasing the number of layers. In the field of rotating machinery fault diagnosis, the network establishes a direct mapping path between input signals and high-level features through a cross-layer connection structure. It can stably extract discriminative spatiotemporal features from vibration signals with strong noise interference, and gradually enhance the feature expression capability through hierarchical stacked residual modules, thereby improving the diagnostic robustness in complex scenarios such as cross-working conditions and cross-sensors. Compared with traditional convolutional neural networks, deep residual networks achieve more efficient gradient propagation and deeper network structure design through a residual learning mechanism, providing a reliable technical foundation for multi-domain fault feature extraction of industrial equipment.

[0022] Multi-kernal Maximum Mean Discrepency (MK-MMD) is a distribution difference measurement method based on statistical learning theory. It jointly measures the probability distribution difference between source and target domain data in the reproducing kernel Hilbert space by fusing multiple kernel functions (such as Gaussian kernels and Laplace kernels). Its core principle is to leverage the complementary properties of multi-kernel combinations to dynamically capture distribution shifts of different scales and types, overcoming the shortcomings of single kernel functions in modeling complex domain differences. In transfer learning scenarios, this method maximizes the mean difference measurement under multi-kernel mapping, driving the model to learn domain-invariant features. This effectively addresses the problem of diagnostic performance degradation caused by data distribution shifts in scenarios such as cross-working conditions and cross-sensors, and provides theoretical support for robust alignment in scenarios with multiple domain differences.

[0023] Based on the above explanations, the implementation environment of the transfer learning-based rotating machinery fault diagnosis method provided in the embodiments of this application is described. Schematically, the implementation environment includes: a multimodal sensor, a terminal, and a processor. The processor, multimodal sensor, and terminal are connected via network signals. 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, among others, without limitation here.

[0024] In combination with the above explanations of terms and implementation environments, the application scenarios of the embodiments of this application are explained. The rotating machinery fault diagnosis method based on transfer learning provided in the embodiments of this application can be applied to, including but not limited to, the following scenarios: In the fault diagnosis of reciprocating pumps under variable operating conditions, multimodal noise suppression is used to effectively separate the piston rod wear characteristics and fluid pulsation interference due to load fluctuations and impact noise caused by changes in the viscosity of the conveying medium (such as switching between petroleum and chemical fluids). Combined with the alignment of cross-operating-condition spatiotemporal features, stable detection of plunger seal failures at different pressure levels is achieved, solving the problem of increased missed reporting rate caused by sudden changes in operating conditions in traditional threshold methods.

[0025] In the cross-sensor diagnosis of marine diesel engines, in view of the data heterogeneity of the marine diesel engine cylinder head vibration sensor and the crankcase acoustic emission sensor, the high-frequency impact and low-frequency vibration signals are fused based on the spatiotemporal synchronous feature extraction network, and the fault knowledge transfer between sensors is realized through the adversarial transfer classifier, so as to accurately identify hidden faults such as abnormal connecting rod bearing clearance.

[0026] Illustratively, the rotating machinery fault diagnosis method based on transfer learning provided in the embodiments of the present application can also be applied to other application scenarios. This is only used as an example and is not limited to the specific application scenario.

[0027] In an exemplary embodiment, Figure 1 As shown, a method for diagnosing rotating machinery faults based on transfer learning is provided. This embodiment uses the method as an example of a terminal in the aforementioned implementation environment. It is understood that the method can also be applied to a server, or to a system including a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 104: Step 101: Perform multimodal noise suppression processing on the acquired original vibration signal to generate a noise reduction data set.

[0028] For example, the original vibration signal can be collected by a vibration sensor installed on the rotating machinery, such as an accelerometer, a velocity sensor, or a displacement sensor. The signals collected by these sensors are usually mixed with various noise components such as environmental noise, electromagnetic interference noise, and background noise generated by mechanical operation. Multimodal noise suppression processing can be achieved through a variety of methods. For example, the threshold denoising method based on wavelet transform removes noise by setting appropriate thresholds for different frequency bands after wavelet decomposition of the signal; or an adaptive filtering method is used to offset noise components by constructing a reference signal and an adaptive filter. The denoised data set generated after denoising can more clearly reflect the actual vibration state of the rotating machinery and provide a more reliable data foundation for subsequent feature extraction. Through the above technical solution, 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.

[0029] In step 102, the denoised 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.

[0030] Specifically, a multi-layer deep residual network (MDRN) is a deep neural network structure composed of multiple residual blocks. Its skip connections effectively alleviate the vanishing gradient problem during deep network training, enabling the network to more deeply learn data features. During feature extraction, the network's convolutional and pooling layers perform layer-by-layer abstraction and feature mapping on the denoised vibration signal, thereby extracting spatiotemporal synchronized features that characterize the vibration characteristics of rotating machinery. For example, this spatiotemporal synchronized feature set integrates the dynamic evolution of the vibration signal in the temporal dimension and the 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 operating condition or sensor configuration, while target domain data refers to data to be diagnosed collected under different operating conditions or sensor conditions. Feature extraction using a MDRN can transform the original vibration signal into a more discriminative feature representation, providing strong support for subsequent domain alignment and fault diagnosis. This technical solution enables deep mining of feature information from the denoised dataset, extracting a feature set with spatiotemporal synchronized characteristics, and providing a high-quality feature foundation for subsequent domain alignment and fault classification.

[0031] In step 103 , the spatiotemporal synchronization feature set is input into a multi-core domain alignment model based on physical constraints to obtain domain-invariant features.

[0032] For example, due to the difference in data distribution between the source domain and the target domain, for example, different working conditions will cause changes in the frequency, amplitude and other characteristics of the vibration signal, and different sensor characteristics will also introduce differences, which requires domain alignment processing. The multi-core domain alignment model based on physical constraints guides the domain alignment process by introducing knowledge of the physical characteristics of rotating machinery, such as considering the resonant frequency of the mechanical structure, the relationship between the fault characteristic frequency and the rotational speed, and other physical constraints, so that the model can more accurately identify and align the common features between different domains. The multi-core learning method can better adapt to complex data distribution and feature structure by combining a variety of different kernel functions, thereby effectively narrowing the gap between the feature distribution of the source domain and the target domain, and obtaining a feature representation with domain invariance, that is, domain invariant features. Through the above technical solution, the feature distribution of the source domain and the target domain can be effectively aligned, the influence of different working conditions and sensor conditions can be eliminated, the domain adaptability and consistency of the features can be improved, and a robust feature representation can be provided for subsequent fault diagnosis.

[0033] Step 104 , performing fault diagnosis on the domain-invariant features by using an adversarial transfer fault classifier to obtain the target domain fault type.

[0034] Specifically, the adversarial transfer fault classifier is built based on the principles of generative adversarial networks (GANs) and consists of a feature extractor and a classifier. During training, the classifier attempts to accurately distinguish domain-invariant features of different fault types. While generating domain-invariant features, the feature extractor 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, during actual diagnosis, the domain-invariant features of the target domain are input into the adversarial transfer fault classifier. Based on the learned fault feature patterns, the classifier identifies and classifies the target domain fault type and outputs corresponding fault diagnosis results, such as normal status, specific fault type (such as bearing wear, gear fracture), and its severity. This technical solution enables highly robust fault diagnosis in scenarios with multiple domain differences, including cross-operating conditions and cross-sensor, in strong noise environments. This effectively improves the accuracy and reliability of rotating machinery fault diagnosis and reduces the impact of factors such as operating condition variations and sensor differences on diagnostic results, providing strong technical support for rotating machinery health management and fault early warning.

[0035] The above-mentioned rotating machinery fault diagnosis method based on transfer learning generates a denoised data set by performing multimodal noise suppression processing on the original vibration signal to eliminate strong noise interference; a multi-layer deep residual network is used to extract a spatiotemporal synchronization feature set from the denoised data. This feature set integrates the spatiotemporal correlation of the source domain and target domain data, effectively capturing common features across working conditions and sensors; the spatiotemporal synchronization feature set is further input into a multi-core domain alignment model based on physical constraints, and domain-invariant features are generated through joint optimization of multi-core function metrics and dynamic constraints to eliminate multiple domain differences; the domain-invariant features are diagnosed by an adversarial transfer fault classifier to achieve accurate classification of the target domain fault type. The above-mentioned method suppresses the coupling interference of noise and domain differences while retaining the essential characteristics of the fault through the synergistic effect of noise suppression, cross-domain feature alignment and physical constraints, thereby achieving highly robust diagnosis in complex industrial scenarios.

[0036] In one embodiment, performing multimodal noise suppression processing on the acquired original vibration signal to generate a noise reduction dataset includes: Based on the industrial noise dataset, the original vibration signal is enhanced to generate a noise mixed signal.

[0037] Specifically, for example, the original vibration signal can be collected by a vibration sensor installed on rotating machinery, while the industrial noise dataset can be derived from background noise records in real industrial environments or public databases. By mixing the noise signal in the industrial noise dataset with the original vibration signal in a certain proportion, it is possible to simulate working scenarios under different noise intensities, thereby generating a noise mixed signal. This process enhances the model's adaptability to different noise environments and provides a more challenging data foundation for subsequent noise reduction processing.

[0038] The noise mixed signal is normalized to obtain a standardized noise signal.

[0039] Specifically, the generated noise mixed signal is normalized to obtain a standardized noise signal. The purpose of normalization is to unify signals of different amplitude ranges into a standard scale for subsequent processing and analysis. For example, the z-score normalization method can be used to normalize the signal to zero mean and unit variance so that the signal amplitude is distributed within a relatively fixed range. In addition, the Min-Max normalization method can also 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 and improve the stability and accuracy of subsequent processing.

[0040] The sliding window length is dynamically adjusted through signal spectrum analysis, and the standardized noise signal is segmented to generate an anti-interference noise signal sample set.

[0041] Specifically, signal spectrum analysis can reveal the energy distribution of the signal in different frequency bands. Based on the spectrum analysis results, the length of the sliding window is dynamically adjusted so that the window can cover the main characteristic frequency band of the signal while avoiding the inclusion of excessive noise frequency components. For example, when a sudden energy change of the signal in a certain frequency band is detected, the sliding window length can be appropriately shortened to improve the 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 of the signal as a sample, and ultimately generating an anti-interference noise signal sample set. This segmented processing method can adaptively adjust the window parameters according to the signal characteristics, effectively improving the signal's anti-interference ability and providing more meaningful sample data for subsequent noise reduction processing.

[0042] Based on the anti-interference noise signal sample set, the original vibration signal is subjected to wavelet threshold denoising to generate a denoised data set.

[0043] Specifically, wavelet threshold denoising is performed on the original vibration signal based on a sample set of anti-interference noise signals to generate a denoised dataset. Wavelet threshold denoising is a widely used signal denoising technique. Its basic concept is to decompose the signal into wavelet domains of different scales and then remove noise components by thresholding the wavelet coefficients. For example, the wavelet coefficients can be processed using either a hard thresholding method or a soft thresholding method. In the hard thresholding method, 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. The soft thresholding method applies a certain degree of shrinkage to the retained coefficients based on the hard thresholding. By selecting appropriate wavelet basis functions and threshold parameters, noise can be effectively suppressed while retaining important characteristic 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, 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 rotating machinery fault diagnosis.

[0044] 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 synchronized feature set, including: Step 201: extract spatial features from the denoised dataset using the ResNet-CBAM composite structure to obtain spatial features, where the channel attention weight is calculated using the following formula: ; in, is the channel attention weight, is the channel dimension average pooling result, is the maximum pooling result in the channel dimension, is the weight matrix of the fully connected layer, Represents the activation function.

[0045] 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 convolutional attention module (CBAM) within the residual block and enhances fault-sensitive frequency band features through a channel-wise attention mechanism. The input feature map undergoes channel-wise average pooling and maximum pooling, respectively. The two pooling results are fed into a shared fully connected layer, where an activation function generates a channel-wise attention weight matrix. This enhances characteristic channel responses related to the fault mechanism (such as the characteristic frequency of bearing outer race faults). Furthermore, a spatial attention module focuses on localized sudden changes in the vibration signal (such as the impact waveform of a broken gear tooth), using a deformable convolution kernel to adaptively adjust the receptive field and generate spatial features. Through the synergistic mechanism of channel and spatial attention, the ResNet-CBAM composite structure enhances fault-sensitive frequency bands and local defect features, improving the separability of spatial features.

[0046] In step 202, the spatial features are input into the LSTM time series modeling layer, and the time series dependency is established through the gating mechanism to obtain the time features, where the forget gate output is calculated by the following formula: ; in, is the output of the forget gate, is the forget gate weight matrix, is the hidden state vector at the previous moment, Input features for the current time step, is the bias term.

[0047] Specifically, spatial features are input into a long-short-term memory (LSTM) network to model temporal dependencies. In the forget gate, the retention ratio of historical hidden states is calculated using an activation function and a weight matrix, and the memory cell state is updated based on the input features of 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 nature of the fault impact. Furthermore, a bidirectional LSTM structure is used to synchronously extract forward and backward temporal features, enhancing the modeling capabilities for complex operating conditions (such as variable speed and sudden load changes). This step effectively characterizes the temporal evolution of fault characteristics (such as the periodic impact interval of a rolling bearing spalling defect), improving the model's adaptability to dynamic operating conditions.

[0048] Step 203: The spatial features and the temporal features are coupled to obtain a spatiotemporal synchronization feature set using the following formula: ; in, is the spatiotemporal synchronization feature set, is the spatial feature, is the time feature, is the coupling coefficient obtained by back-propagation optimization.

[0049] For example, dynamic weight coupling is performed on spatial and temporal features. The coupling coefficient is optimized through a back-propagation algorithm, and the fusion ratio of spatial and temporal features is adaptively adjusted based on the noise level of the input signal and the operating conditions. For example, the weight of temporal features is increased in strong noise scenarios to suppress spatial noise interference, while the weight of spatial features is increased in stable operating conditions to enhance the discriminability of local defects. Ultimately, a spatiotemporal synchronized feature set is generated that simultaneously preserves the fault physical pattern and the temporal evolution law. This process achieves complementary enhancement of spatiotemporal features across operating conditions through an adaptive feature fusion mechanism, providing highly robust input for subsequent domain alignment.

[0050] In one embodiment, the spatiotemporal synchronization feature set is input into a multi-core domain alignment model based on physical constraints to obtain domain-invariant features, including: The multi-core maximum mean difference calculation is performed on the spatiotemporal synchronization feature set to obtain the distribution difference.

[0051] Specifically, the multi-kernel maximum mean difference (MK-MMD) is calculated on the spatiotemporal synchronized feature set to obtain the distribution difference. For example, the spatiotemporal synchronized feature set contains feature data from the source and target domains, which may have distribution differences due to factors such as operating conditions and sensor differences. By combining multiple kernel functions (such as Gaussian kernels and polynomial kernels), the multi-kernel maximum mean difference can effectively measure the difference in the feature distributions of the source and target domains in the regenerated Hilbert space. For example, using multiple Gaussian kernels can capture distribution differences at different scales, thereby more comprehensively reflecting the feature distribution mismatch between the source and target domains and providing a quantitative basis for subsequent domain alignment.

[0052] The domain alignment loss weight coefficient is dynamically generated based on the distribution difference metric and the classification task gradient standard deviation.

[0053] Specifically, the standard deviation of the classifier gradient is calculated in real time during training. When the classification task has difficulty converging (e.g., gradient fluctuations increase significantly), the domain alignment loss weight is reduced to avoid overfitting. Conversely, when domain differences become the main contradiction, the domain alignment weight is increased to strengthen distribution alignment. For example, the historical mean of the gradient standard deviation is calculated through a sliding window, and the adaptive weight coefficient is generated in combination with the current distribution difference. This process can balance the optimization goals of classification accuracy and domain alignment, improving the generalization stability of the model in cross-domain scenarios.

[0054] A physical constraint regularization term is constructed for the spatiotemporal synchronization feature set based on the dynamic equations of rotating machinery.

[0055] Specifically, physical constraint regularization terms can be constructed for the spatiotemporal synchronized 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-damping 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 that the extracted domain-invariant features conform to the physical behavior of the mechanical system. For example, this can be achieved by constraining the energy distribution of the features in a specific frequency band to match the resonant frequency of the mechanical system, or by constraining the time evolution characteristics of the features to 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.

[0056] Through the adversarial training mechanism, the parameters of the multi-layer deep residual network and the domain alignment loss weight coefficient are alternately optimized, combined with the physical constraint regularization term to generate domain-invariant features.

[0057] Specifically, the domain discriminator parameters are fixed, and the classification loss and physical constraint regularization term of the feature extraction network are optimized to improve the fault discrimination ability; the feature extraction network parameters are fixed, and 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 scenarios (such as variable speed and variable load), this training mechanism can effectively eliminate the distribution offset caused by operating condition differences and generate domain-invariant features that meet both classification accuracy and physical interpretability. The above technical solution achieves stable expression and cross-domain migration of fault features under multiple challenges such as noise interference, sensor heterogeneity, and sudden changes in operating conditions through the deep integration of multi-core metrics, dynamic optimization, physical constraints, and adversarial training, providing a highly robust solution for intelligent diagnosis of industrial equipment, especially suitable for predictive maintenance scenarios of complex rotating machinery such as wind power gearboxes and petrochemical centrifugal compressors.

[0058] In one embodiment, noise enhancement is performed on an original vibration signal based on an industrial noise dataset to generate a noise mixed signal, including: The industrial noise dataset is modally expanded through a generative adversarial network to generate multimodal industrial noise including Gaussian white noise, impact noise and gear meshing harmonic noise.

[0059] Specifically, industrial noise datasets can be obtained by collecting background noise from real industrial environments. This noise data covers noise characteristics under different working conditions and equipment operating states. A generative adversarial network consists of a generator and a discriminator. The generator is responsible for generating noise samples of multiple modalities, while the discriminator is responsible for distinguishing the generated noise samples from real noise samples. For example, the generator can generate Gaussian white noise with zero mean and constant power spectral density; shock noise can be generated by simulating sudden shock events during equipment operation; and gear mesh harmonic noise can be 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, generate more comprehensive and realistic noise samples, and provide a diverse noise source for subsequent noise enhancement processing. In this way, the diversity and authenticity of the noise data can be significantly improved, ensuring that the generated noise samples better simulate the complex noise conditions found in real industrial environments.

[0060] Based on the spectral characteristics of the target domain equipment, multimodal industrial noise is bandpass filtered to generate spectrum-matched noise.

[0061] Specifically, the spectral characteristics of the target domain equipment can be obtained by performing spectral analysis on the vibration signal under normal operating conditions to determine its main operating frequency range and noise-sensitive frequency band. The design of the bandpass filter should be adjusted according to these spectral characteristics so that the filtered noise is mainly concentrated in the noise-sensitive frequency band of the target domain equipment. For example, if the main operating frequency range of the target domain equipment is 500Hz to 2000Hz, the bandpass filter can be set to only allow the noise within this frequency band to pass. In this way, the generated spectrum-matched noise can be closer to the actual noise environment of the target domain equipment, improving the pertinence and effectiveness of the noise enhancement processing. This method can ensure that the generated noise matches the spectral characteristics of the target domain equipment, making the subsequent noise enhancement processing more in line with the actual application scenario and enhancing the model's adaptability to the noise of the target domain equipment. In this way, the accuracy and effectiveness of the noise enhancement processing can be significantly improved, ensuring that the generated noise mixed signal is more in line with the operating environment of actual industrial equipment.

[0062] The spectrum matching noise is superimposed on the original vibration signal to generate a noise mixed signal.

[0063] Specifically, the original vibration signal can be collected by a vibration sensor installed on the rotating machinery, and the spectrum 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 spectrum matching noise, it is made to reach a certain proportional relationship with the amplitude of the original vibration signal. The generated noise mixed signal can more realistically reflect the situation of the rotating machinery being disturbed by noise in an actual industrial environment. In this way, a more challenging and practical data foundation is provided for the subsequent training of the fault diagnosis algorithm. The denoised dataset after denoising can more clearly reflect the true vibration characteristics of the rotating machinery, providing high-quality data support for subsequent fault diagnosis. Through the above-mentioned 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 the fault characteristics can be enhanced, thereby providing a more reliable data foundation for the fault diagnosis of rotating machinery.

[0064] In one embodiment, spatial dimension feature extraction is performed on the denoised dataset using a ResNet-CBAM composite structure to obtain spatial features, including: Convolutional feature extraction is performed on the denoised dataset through the residual network to generate initial spatial features.

[0065] For example, the denoised dataset is derived from vibration signals that have undergone multimodal noise suppression processing. These signals are collected by vibration sensors mounted on rotating machinery. The residual network consists of multiple residual blocks, each of which contains two convolutional layers and a skip connection. During convolutional feature extraction, the denoised dataset is first convolved with a convolutional layer to extract local features from the signal. For example, convolution kernels of different sizes (such as 3×3 or 5×5) can be used to capture features at different scales. Skip connections pass the input directly to subsequent layers, addressing the vanishing gradient 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 denoised dataset, providing a foundation for subsequent feature processing. This method effectively extracts local features from the denoised dataset while preserving the global information of the original data through skip connections, improving the efficiency and accuracy of feature extraction.

[0066] The initial spatial features are weighted by the channel dimension through the channel attention module to generate channel weighted features.

[0067] Specifically, the channel-dimensional weights of the initial spatial features are assigned through the channel attention module to generate channel-weighted features. The input of 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 maximum pooling are performed on the initial spatial features to obtain the average response and maximum response of each channel, respectively. The pooling results are processed through two fully connected layers, and activation functions (such as ReLU and Sigmoid) are used to generate channel attention weights. For example, the fully connected layer can map the pooled features to a lower-dimensional space and then map them back to the original dimension, thereby learning the importance weight of each channel. The channel attention weights are multiplied by the initial spatial features channel by channel to generate channel-weighted features. This method can enhance the expressiveness of key feature channels and improve the model's sensitivity to important features by adaptively adjusting channel weights.

[0068] The spatial attention module is used to focus on the local area of ​​the channel weighted features to generate spatial features.

[0069] Specifically, the input of the spatial attention module is the channel-weighted feature, and its goal is to further enhance the key areas related to the fault in the feature map. For example, the channel-weighted features are average pooled and max pooled in the spatial dimension to obtain the average response and maximum response of each position; the two pooling results are spliced ​​according to the channel dimension, and a spatial attention map is generated through a convolution layer. For example, a 1×1 convolution kernel can be used to map the spliced ​​features to a two-dimensional attention map; the spatial attention map is element-wise multiplied with the channel-weighted features to generate spatial features. This method can dynamically adjust the weights of different areas according to the spatial distribution of the feature map, highlight the key areas and suppress background noise. In this way, the key areas related to the fault in the feature map can be further enhanced, and the locality and discriminability of the features can be improved, thereby providing higher-quality feature representation for subsequent fault diagnosis.

[0070] In one embodiment, the spatial features are input into the LSTM time series modeling layer, and the time series dependency is established through a gating mechanism to obtain the temporal features, including: The spatial features are input into the bidirectional LSTM network to extract the forward temporal dependencies and generate forward temporal features.

[0071] Specifically, spatial features are derived from feature data extracted by the ResNet-CBAM composite structure. These feature data capture the vibration characteristics of rotating machinery in the spatial dimension. The bidirectional LSTM network comprises two LSTM layers in two directions. The forward LSTM layer processes 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 vibration signals from rotating machinery, the forward LSTM layer can capture the evolution of features as the machinery gradually enters a stable operating state after startup. The forward time series features generated in this way can reflect the forward dependencies of the time series, providing important information for subsequent temporal feature fusion. This approach effectively captures long-term dependencies in time series and improves the model's efficient use of temporal information.

[0072] The backward temporal dependencies are extracted through the reverse LSTM network to generate backward temporal features.

[0073] Specifically, the reverse LSTM layer processes input features in reverse order, from the end to the beginning of the time series, capturing future information beyond the current time step. For example, when processing vibration signals from rotating machinery, the reverse LSTM layer can capture the changing trends in mechanical features before a failure occurs. The backward time series features generated in this way reflect the backward dependencies of the time series, providing supplementary information for subsequent time feature fusion. This approach effectively supplements future information that the forward LSTM cannot capture, enhancing the model's overall understanding of the time series.

[0074] The forward time series features and the backward time series features are concatenated to generate time features.

[0075] Specifically, the forward and reverse time series features are concatenated along the channel dimension and compressed to a unified dimension through a fully connected layer to generate time features that incorporate bidirectional time series dependencies. This processing can integrate forward causal relationships and reverse propagation characteristics (such as the generation and reflection process of the shock waveform), improving the global expressiveness of time series features. This embodiment uses a bidirectional long short-term memory network (Bi-LSTM) to perform bidirectional time series modeling on spatial features, fully capturing the forward causal relationships and reverse propagation characteristics of vibration signals, such as the periodic interval pattern of bearing spalling impact and the reflection attenuation effect of the shock waveform. By splicing and fusing the features of the forward LSTM and reverse LSTM, the global expressiveness of time series features is enhanced to adapt to complex time series evolution patterns under non-stationary operating conditions such as sudden load changes and speed fluctuations. This solution combines the dynamic adjustment of the gating mechanism with the equipment speed parameters to make the timing modeling process conform to the physical motion laws of rotating machinery, improve the feature interpretability, and maintain high diagnostic stability across operating scenarios. In particular, the detection sensitivity for early weak periodic shocks (such as early cracks in gears) is significantly improved, providing a highly robust timing feature expression basis for rotating machinery fault diagnosis.

[0076] Furthermore, the multi-core maximum mean difference calculation is performed on the spatiotemporal synchronization feature set to obtain the distribution difference, including: The distribution distance of the Gaussian kernel function between the source domain data and the target domain data is calculated respectively to generate the Gaussian kernel difference.

[0077] Specifically, the spatiotemporal synchronization feature set contains feature data from the source and target domains, which have distribution differences due to factors such as working conditions and sensor differences. The Gaussian kernel function can effectively capture the nonlinear distribution characteristics of the data by mapping the data into a high-dimensional space. For example, a radial basis function (RBF) can be used as a Gaussian kernel function, which is in the form of converting the Euclidean distance between data points into similarity through an exponential function. The Gaussian kernel difference is obtained by calculating the maximum mean difference (MMD) of the source and target domain data under the Gaussian kernel. The Gaussian kernel difference generated in this way can effectively reflect the nonlinear distribution differences between the source and target domains, and improve the modeling ability of complex data distributions.

[0078] The distribution distance of the Laplace kernel function between the source domain data and the target domain data is calculated respectively to generate the Laplace kernel difference.

[0079] 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 into similarity through an exponential function. The Laplace kernel difference is obtained by calculating the maximum mean difference between the source domain and the target domain data under the Laplace kernel. This step can further quantify the distribution difference between the source domain and the target domain in the Laplace kernel space. For example, when processing vibration signals containing sudden fault characteristics, the Laplace kernel can effectively capture the sparse feature changes in the signal and provide supplementary information for subsequent domain alignment. The Laplace kernel difference generated in this way can enhance the modeling ability of data sparse areas and improve the robustness of the model.

[0080] The Gaussian kernel difference and the Laplace kernel difference are weightedly fused to generate the distribution difference.

[0081] Specifically, the Gaussian kernel difference and the Laplace kernel difference can be linearly combined by introducing a weight coefficient. For example, the distribution difference can be expressed as the weighted sum of the Gaussian kernel difference and the Laplace kernel difference, and the weight coefficient is adjusted according to actual needs. For example, 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. These two characteristics can be balanced through weighted fusion. Through the above technical solution, the quantification accuracy of the distribution difference between the source domain and the target domain can be significantly improved, providing more accurate guidance for subsequent domain alignment. This method can effectively integrate the advantages of multi-kernel functions, improve the modeling ability of complex data distributions, enhance the robustness and adaptability of the model, and thus provide more reliable support for fault diagnosis of rotating machinery.

[0082] Furthermore, a physical constraint regularization term is constructed for the spatiotemporal synchronization feature set based on the rotating machinery dynamics equations, including: The vibration signal spectrum in the spatiotemporal synchronization feature set is extracted to generate characteristic frequency components.

[0083] Specifically, a fast Fourier transform can be used to convert the time-domain vibration signal into a frequency-domain representation, extracting characteristic frequency components. These components include the characteristic frequencies of rolling bearing faults (e.g., outer race fault frequency BPFO and inner race fault frequency BPFI), gear meshing frequency, and its sideband components. This step, through frequency-domain energy distribution analysis, reveals physical characteristics directly related to the rotating machinery's fault mechanism, providing a quantitative basis for subsequent constraint development.

[0084] Calculate the matching degree between the characteristic frequency component and the theoretical fault characteristic frequency and generate the frequency constraint condition.

[0085] Specifically, a theoretical fault frequency is generated based on rotating machinery dynamics equations (such as the formula for calculating bearing fault characteristic frequencies). The degree of match between the actual characteristic frequency and the theoretical value is measured using cosine similarity or energy fraction, generating frequency constraints. For example, when a rolling bearing outer race fault is detected, the energy fraction of the domain-invariant feature in the BPFO frequency band must be no less than a theoretical threshold. This step constrains the model's learning direction by applying physical laws to prevent interference from noise in irrelevant frequency bands.

[0086] An energy-preserving regularization term is constructed based on the frequency constraint condition to generate a physical constraint regularization term.

[0087] Specifically, the characteristic frequency matching degree is used as the penalty coefficient of the regularization term, and the deviation between the actual feature and the theoretical value is quantified by the mean square error or KL divergence, and added to the total loss function of the model. For example, in bearing fault diagnosis, the regularization term forces the energy distribution of the domain-invariant feature in the BPFO frequency band to be consistent with the theoretical fault feature frequency. This process can suppress the migration of invalid features and ensure that the model retains physical features that are strongly related to the mechanical failure mechanism. This embodiment can significantly improve the physical interpretability and cross-domain generalization ability of the fault diagnosis model, force the domain-invariant features to retain the key frequency band energy that is strongly related to the mechanical failure mechanism, and suppress noise and irrelevant harmonic interference; at the same time, the energy-retention regularization term is used to guide the model to focus on common physical features, reduce the sensitivity of distribution offset under cross-working conditions and cross-sensor scenarios, avoid invalid feature migration, and ultimately achieve high robustness and high reliability of rotating machinery fault diagnosis.

[0088] In one embodiment, the parameters of a multi-layer deep residual network and the domain alignment loss weight coefficients are alternately optimized through an adversarial training mechanism, combined with a physical constraint regularization term, to generate domain-invariant features, including: The cross entropy loss of the source domain data is calculated through the classifier to generate the classification loss gradient.

[0089] The domain alignment loss is calculated based on the distribution difference to generate domain adversarial gradients.

[0090] The classification loss gradient and domain adversarial gradient are projected and eliminated to generate optimized domain invariant features.

[0091] Specifically, this embodiment uses 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 through the classifier, and the classification loss gradient is generated by backpropagation to optimize the network parameters and improve the accuracy of fault category discrimination; based on the domain alignment loss, the domain adversarial gradient is generated, and the gradient of the domain discriminator is inverted through the gradient reversal layer, driving the feature extraction network to generate feature representations of the confused domain discriminator, eliminating distribution differences across sensors and working conditions; further, the gradient projection algorithm is used to eliminate the directional conflict between the classification gradient and the domain adversarial gradient, and the parameters are jointly updated in combination with the physical constraint regularization term, ensuring classification accuracy while forcing the features to conform to physical laws. This solution achieves high discriminability and cross-domain generalization capabilities of domain-invariant features through dynamic gradient projection and multi-objective collaborative optimization, and simultaneously improves diagnostic accuracy and model robustness in complex scenarios.

[0092] In summary, the present application provides a method for rotating machinery fault diagnosis based on transfer learning. This method generates a high-fidelity mixed signal by performing multimodal noise enhancement and spectrum matching filtering on the original vibration signal based on a generative adversarial network. This method then combines wavelet threshold denoising to generate an interference-resistant 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. The channel attention mechanism is used to enhance fault-sensitive frequency bands. The spatial attention module is used to focus on local defect areas. A bidirectional LSTM network is combined to model temporal dependencies, achieving a spatiotemporal joint expression of fault features. The 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 weight. A physical constraint regularization term (such as a bearing fault feature frequency energy preservation constraint) is introduced into the rotating machinery dynamics equation to suppress invalid 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. This generates highly discriminative and physically invariant domain features, driving the adversarial transfer classifier to achieve cross-domain fault diagnosis. This technical solution achieves highly robust diagnosis in multiple domain difference scenarios such as across working conditions and sensors in strong noise environments through the synergistic effect of four core technologies: multimodal noise suppression, spatiotemporal synchronous feature extraction, multi-core domain alignment, and physical constraint regularization.

[0093] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0094] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A rotating machinery fault diagnosis method based on transfer learning, characterized in that: The method comprises: Perform multimodal noise suppression on the acquired original vibration signal to generate a noise reduction dataset; Inputting the denoised dataset 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; Inputting the spatiotemporal synchronization feature set into a multi-core domain alignment model based on physical constraints to obtain domain-invariant features; Fault diagnosis is performed on the domain-invariant features by adversarially migrating the fault classifier to obtain the target domain fault type.

2. The method according to claim 1, characterized in that The multimodal noise suppression processing is performed on the acquired original vibration signal to generate a noise reduction data set, including: Performing noise enhancement on the original vibration signal based on an industrial noise dataset to generate a noise mixed signal; Normalizing the noise mixed signal to obtain a standardized noise signal; Dynamically adjust the sliding window length through signal spectrum analysis, perform segmentation processing on the standardized noise signal, and generate an anti-interference noise signal sample set; The original vibration signal is subjected to wavelet threshold denoising based on the anti-interference noise signal sample set to generate the denoised data set.

3. The method according to claim 1, characterized in that The denoised dataset is input into a multi-layer deep residual network for feature extraction to obtain a spatiotemporal synchronization feature set, including: The spatial dimension of the denoised dataset is extracted using the ResNet-CBAM composite structure to obtain spatial features, where the channel attention weight is calculated using the following formula: ； in, is the channel attention weight, is the channel dimension average pooling result, is the maximum pooling result in the channel dimension, is the weight matrix of the fully connected layer, represents the activation function; The spatial features are input into the LSTM time series modeling layer, and the time series dependency is established through the gating mechanism to obtain the time features, where the forget gate output is calculated by the following formula: ； in, is the output of the forget gate, is the forget gate weight matrix, is the hidden state vector at the previous moment, Input features for the current time step, is the bias term; The spatial feature and the temporal feature are coupled by the following formula to obtain the spatiotemporal synchronization feature set: ； in, is the spatiotemporal synchronization feature set, is the spatial feature, is the time feature, is the coupling coefficient obtained by back-propagation optimization.

4. The method according to claim 1, wherein Inputting the spatiotemporal synchronization feature set into a multi-core domain alignment model based on physical constraints to obtain domain-invariant features includes: Performing multi-core maximum mean difference calculation on the spatiotemporal synchronization feature set to obtain distribution difference; Dynamically generate domain alignment loss weight coefficients based on the distribution difference metric and the classification task gradient standard deviation; constructing a physical constraint regularization term for the spatiotemporal synchronization feature set according to the rotating machinery dynamics equation; The domain-invariant features are generated by alternately optimizing the parameters of the multi-layer deep residual network and the domain alignment loss weight coefficient through an adversarial training mechanism and combining the physical constraint regularization term.

5. The method according to claim 2, characterized in that The performing noise enhancement on the original vibration signal based on the industrial noise dataset to generate a noise mixed signal includes: Performing modal expansion on the industrial noise dataset through a generative adversarial network to generate multimodal industrial noise including Gaussian white noise, impact noise, and gear meshing harmonic noise; performing bandpass filtering on the multimodal industrial noise based on the spectrum characteristics of the target domain device to generate spectrum-matched noise; The spectrum matching noise is superimposed on the original vibration signal to generate the noise mixed signal.

6. The method according to claim 3, characterized in that The spatial dimension feature extraction of the denoised dataset is performed using the ResNet-CBAM composite structure to obtain spatial features, including: Performing convolution feature extraction on the denoised dataset through a residual network to generate initial spatial features; Performing channel dimension weight assignment on the initial spatial features through a channel attention module to generate channel weighted features; The channel weighted features are focused on a local area through a spatial attention module to generate the spatial features.

7. The method according to claim 3, characterized in that The spatial features are input into the LSTM time series modeling layer, and a time series dependency relationship is established through a gating mechanism to obtain time features, including: Input the spatial features into a bidirectional LSTM network, extract the forward temporal dependency, and generate forward temporal features; Extract backward temporal dependencies through the reverse LSTM network and generate backward temporal features; The forward time series feature and the backward time series feature are concatenated to generate the time feature.

8. The method according to claim 4, characterized in that The performing multi-core maximum mean difference calculation on the spatiotemporal synchronization feature set to obtain the distribution difference includes: respectively calculating the distribution distance of the Gaussian kernel function between the source domain data and the target domain data to generate a Gaussian kernel difference; respectively calculating the distribution distance of the Laplace kernel function between the source domain data and the target domain data to generate a Laplace kernel difference; The Gaussian kernel difference and the Laplace kernel difference are weightedly fused to generate the distribution difference.

9. The method according to claim 4, characterized in that The constructing of a physical constraint regularization term for the spatiotemporal synchronization feature set according to the rotating machinery dynamics equation includes: Extracting the vibration signal spectrum in the spatiotemporal synchronization feature set to generate characteristic frequency components; Calculating the matching degree between the characteristic frequency component and the theoretical fault characteristic frequency to generate a frequency constraint condition; An energy preservation regularization term is constructed based on the frequency constraint condition to generate the physical constraint regularization term.

10. The method according to claim 4, characterized in that The alternately optimizing the parameters of the multi-layer deep residual network and the domain alignment loss weight coefficient through the adversarial training mechanism, combined with the physical constraint regularization term, to generate the domain invariant features, includes: Calculate the cross entropy loss of the source domain data through the classifier and generate the classification loss gradient; Calculating domain alignment loss based on the distribution difference and generating domain adversarial gradient; Projection and elimination are performed on the classification loss gradient and the domain adversarial gradient to generate the optimized domain invariant feature.

Images