Rotating machine fault diagnosis method based on Vision Transform and multi-strategy domain adaptation

Abstract

The invention provides a rotating machine fault diagnosis method based on Vision Transform and multi-strategy domain adaptation, and relates to the technical field of intelligent fault diagnosis. The method comprises the steps that vibration signals under different working conditions are collected through an acceleration sensor and preprocessed, and standardized time-frequency spectrogram data are obtained; a feature extraction network comprising a Vision Transform and a convolution attention module is utilized to obtain high-dimensional feature representation of a fusion attention mechanism, and a multi-strategy domain adaptation framework comprising a classification and domain alignment module is constructed. Through joint optimization of classification feature learning and domain feature alignment, feature distribution consistency between a source domain and a target domain is realized. And end-to-end joint training is carried out by taking the diagnosis feature set as input to obtain a fault diagnosis model with cross-working-condition generalization capability, so that the recognition accuracy and stability under complex working conditions are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent fault diagnosis, in particular to a rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation. BACKGROUND

[0002] Rotating machinery is the core equipment of modern industrial production, and its health status is directly related to the safety and efficiency of the entire production system. Internal faults of equipment will be reflected in specific vibration patterns, so vibration signal-based analysis is one of the most mainstream and effective methods for monitoring the health status of rotating machinery. Traditional fault diagnosis relies heavily on signal processing techniques and expert experience for manual feature extraction, which not only consumes time but also is difficult to scale, limiting its application in complex and variable industrial scenarios.

[0003] The advent of deep learning technology has brought about revolutionary changes. Models represented by convolutional neural networks and recurrent neural networks can automatically learn end-to-end features from signals, greatly improving the automation level and accuracy of diagnosis. However, although these models perform well in laboratory environments, they cannot adapt to varying operating conditions, making their practical application less feasible. In real industrial scenarios, the operating parameters of rotating machinery such as speed and load change frequently, causing significant shifts in the statistical distribution of vibration signals. This makes models trained under known operating conditions perform poorly under another operating condition, hindering the application of intelligent diagnosis technology in real-world scenarios. To solve this problem, unsupervised domain adaptation technology in transfer learning becomes a key. SUMMARY

[0004] To overcome the shortcomings of the prior art, the present application aims to provide a rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation, which solves the limitations of deep learning models in feature extraction and the problem of sharp decline in diagnosis performance under varying operating conditions.

[0005] To achieve the above-mentioned purpose, the present application provides the following solutions:

[0006] A rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation, comprising:

[0007] During the operation of the rotating machinery, vibration signals under different operating conditions are collected by an acceleration sensor, and the vibration signals are preprocessed to obtain a standardized time-frequency spectrum dataset;

[0008] The standardized time-frequency spectrogram dataset is input into a feature extraction network composed of a Vision Transformer model and a convolution attention module to obtain a high-dimensional feature representation fused with attention mechanism;

[0009] Based on the high-dimensional feature representation, a multi-strategy domain adaptation framework including a classification module and a domain alignment module is constructed.

[0010] The multi-strategy domain adaptation framework obtains a domain-aligned optimized diagnostic feature set through joint optimization of classification feature learning and domain feature alignment.

[0011] The diagnostic feature set is input to perform end-to-end training and parameter optimization using multiple loss functions to train and update the multi-strategy domain adaptation framework and the feature extraction network, and obtain a final fault diagnosis model.

[0012] The final fault diagnosis model is used to obtain fault recognition results.

[0013] Preferably, the working conditions include:

[0014] Source domain working conditions with labels and target domain working conditions without labels.

[0015] Preferably, during the operation of the rotating machinery, the vibration signals under different working conditions are collected by the acceleration sensor, and the vibration signals are preprocessed to obtain a standardized time-frequency spectrogram dataset, including:

[0016] An acceleration sensor is installed at a key part of the rotating machinery to collect vibration signals of the equipment in real time when it operates under different working conditions, and obtain an original vibration signal dataset.

[0017] The original vibration signal is subjected to amplitude standardization and normalization processing, and the normalized signal is segmented according to a preset step length using a fixed-length sliding window to obtain a normalized vibration signal segment set.

[0018] The signal energy distribution information changing with time in each normalized vibration signal segment is mapped to a two-dimensional time-frequency space to obtain an original time-frequency spectrogram set.

[0019] The original time-frequency spectrogram set is uniformly adjusted in size to obtain a standardized time-frequency spectrogram dataset.

[0020] Preferably, the expression of the normalized vibration signal segment set is:

[0021] ;

[0022] Wherein, and are the minimum and maximum values of the signal, respectively, represents a signal normalized to the interval [0, 1].

[0023] Preferably, the standardized time-frequency spectrogram dataset is input into a feature extraction network composed of a Vision Transformer model and a convolutional attention module to obtain a high-dimensional feature representation fused with attention mechanism, including:

[0024] The standardized time-frequency spectrogram dataset is segmented by using the Vision Transformer model to obtain a plurality of non-overlapping image blocks of a fixed size, and each non-overlapping image block is converted into a corresponding vector sequence.

[0025] A learnable classification identification embedding and a position encoding are added to the front end of the vector sequence to obtain an image embedding sequence with position information.

[0026] A convolutional attention module is introduced before the Transformer encoder to apply attention weighting to the channel dimension and the spatial dimension of the image embedding sequence respectively to obtain an enhanced feature sequence.

[0027] The enhanced feature sequence is input into the Transformer encoder to obtain a high-dimensional feature representation fused with attention mechanism.

[0028] The present application discloses the following technical effects:

[0029] This invention provides a rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation, comprising: during the operation of the rotating machinery, collecting vibration signals under different working conditions through an accelerometer, and preprocessing the vibration signals to obtain a standardized time-spectrum dataset; inputting the standardized time-spectrum dataset into a feature extraction network consisting of a Vision Transformer model and a convolutional attention module to obtain a high-dimensional feature representation fused with the attention mechanism; based on the high-dimensional feature representation, constructing a multi-strategy domain adaptation framework including a classification module and a domain alignment module; the multi-strategy domain adaptation framework obtaining a domain alignment-optimized diagnostic feature set through joint optimization of classification feature learning and domain feature alignment; using the diagnostic feature set as input, performing end-to-end training and parameter optimization using various loss functions to train and update the multi-strategy domain adaptation framework and the feature extraction network to obtain a final fault diagnosis model; and using the final fault diagnosis model to obtain a fault identification result. At the feature extraction level, this method uses the time-frequency representation of the vibration signal as input and employs the self-attention mechanism of the Vision Transformer (ViT) to capture the global long-range dependencies of the time-spectrum graph, overcoming the limitations of the local receptive field of traditional CNNs. Simultaneously, by fusing the Convolutional Block Attention Module (CBAM), the model can adaptively focus on key time-frequency features, significantly enhancing its sensitivity to weak fault signals. At the domain adaptation level, to address the data distribution shift caused by varying operating conditions, this method employs a dual-strategy mechanism. This mechanism combines Correlation Alignment (CORAL) for efficient coarse-grained second-order statistical alignment with Conditional Adversarial Domain Adaptation Network (CDAN) for refined category-aware alignment. The two complement each other, eliminating inter-domain differences caused by changes in operating conditions from coarse to fine, ensuring high generalization ability and diagnostic accuracy of the model under unknown operating conditions. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 A flowchart of a rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation is provided for an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the feature extraction and domain adaptation framework provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0034] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0035] As shown in Figure 1 The present application provides a rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation, comprising:

[0036] Step 100: During the operation of the rotating machinery, vibration signals under different working conditions are collected by an acceleration sensor, and the vibration signals are preprocessed to obtain a standardized time-frequency spectrum data set;

[0037] Step 200: The standardized time-frequency spectrum data set is input into a feature extraction network composed of a Vision Transformer model and a convolution attention module to obtain a high-dimensional feature representation fused with attention mechanism;

[0038] Step 300: Based on the high-dimensional feature representation, a multi-strategy domain adaptation framework containing a classification module and a domain alignment module is constructed;

[0039] Step 400: The multi-strategy domain adaptation framework obtains a domain alignment optimized diagnosis feature set through joint optimization of classification feature learning and domain feature alignment;

[0040] Step 500: The diagnosis feature set is input to perform end-to-end training and parameter optimization using multiple loss functions to train and update the multi-strategy domain adaptation framework and the feature extraction network, and obtain a final fault diagnosis model;

[0041] Step 600: The final fault diagnosis model is used to obtain a fault recognition result.

[0042] Further, the vibration signals under different working conditions are collected by an acceleration sensor during the operation of the rotating machinery, and the vibration signals are preprocessed to obtain a standardized time-frequency spectrum data set, comprising:

[0043] An acceleration sensor is installed at a key position of a rotating machine to collect vibration signals of the device in real time when the device is running under different working conditions, and an original vibration signal dataset is obtained;

[0044] The original vibration signal is subjected to amplitude standardization and normalization processing, and a sliding window with a fixed length is used to segment the normalized signal by a preset step length, and a normalized vibration signal segment set is obtained;

[0045] The signal energy distribution information changing with time in each normalized vibration signal segment is mapped to a two-dimensional time-frequency space, and an original time-frequency spectrum set is obtained;

[0046] The original time-frequency spectrum set is subjected to uniform size adjustment, and a standardized time-frequency spectrum dataset is obtained.

[0047] Specifically, first, data collection is performed. An acceleration sensor is installed at a key position of a rotating machine to collect vibration signals of the device in real time when the device is running under different working conditions, and these working conditions are divided into source domain working conditions and target domain working conditions, the former data has a fault label, and the latter data has no fault label.

[0048] Second, signal segmentation and normalization are performed. Since the original vibration signal is usually continuous long time series data, and the signal amplitude range under different working conditions differs greatly, direct processing will affect the model training effect. Therefore, the original one-dimensional vibration signal collected is subjected to normalization operation:

[0049]

[0050] Among them, and are the minimum and maximum values of the signal, denotes the signal normalized to the [0, 1] interval. Subsequently, a sliding window with a length of L is used to segment the normalized signal at a fixed step length, and a series of one-dimensional time series segment samples with a length of L are obtained.

[0051] Next, time-frequency transformation is performed. In order to fully exploit the non-stationary characteristics of the vibration signal, continuous wavelet transform (CWT) is applied to each one-dimensional vibration signal segment to convert it into a two-dimensional time-frequency image that can represent both time and frequency information. The calculation process of CWT is as follows:

[0052]

[0053] Among them, is the input one-dimensional signal segment, is the scale parameter, is the translation parameter,​ is the mother wavelet function. Through this transformation, each one-dimensional signal segment is mapped as a two-dimensional matrix, i.e., a time-frequency spectrogram .

[0054] Finally, image standardization is performed. To meet the input requirements of the subsequent ViT model, the generated time-frequency spectrograms are uniformly resized and pixel value normalized. All time-frequency spectrograms are uniformly adjusted to the same size HxW to ensure consistency in input dimensions. The adjusted image is represented as:

[0055]

[0056] wherein: is the original image, is the adjusted image, are the height and width of the target image, respectively.

[0057] To eliminate the difference in pixel value range, the pixel values of the infrared image are normalized to the range [0, 1]:

[0058]

[0059] To improve the robustness and generalization ability of the model, the obtained images are randomly rotated at angles :

[0060]

[0061] After preprocessing, each vibration signal segment is finally represented as a standardized two-dimensional image sample, serving as input for the subsequent network.

[0062] Further, the standardized time-frequency spectrogram dataset is input into a feature extraction network composed of a Vision Transformer model and a convolutional attention module, obtaining a high-dimensional feature representation that integrates attention mechanisms, including:

[0063] The Vision Transformer model is used to process the segmentation of the standardized time-frequency spectrogram dataset, obtaining multiple non-overlapping image blocks of fixed size, and converting each non-overlapping image block into a corresponding vector sequence.

[0064] A learnable classification label embedding and position encoding are added to the front of the vector sequence, obtaining an image embedding sequence with position information.

[0065] A convolutional attention module is introduced before the Transformer encoder to apply attention weighting to the channel dimension and spatial dimension of the image embedding sequence respectively, obtaining an enhanced feature sequence.

[0066] The enhanced feature sequence is input into the Transformer encoder to obtain a high-dimensional feature representation that fuses attention mechanisms.

[0067] Specifically, as shown in Figure 2 , the framework mainly includes the following modules: 1) input module: the input is a two-dimensional time-frequency spectrogram converted from the vibration signal in step one; 2) feature extractor: ViT is used as the only feature extractor to capture global patterns in the time-frequency spectrogram; 3) feature enhancement module: to make ViT pay more attention to key time-frequency information, a CBAM module is introduced after the linear embedding of ViT and before input into the Transformer encoder; 4) classifier and domain adaptation module: the high-level features extracted by ViT are input into the classification and domain adaptation module. Here, the domain refers to different operating conditions.

[0068] 1) Input module: let the source domain condition dataset be , and the target domain condition dataset be , where represents the time-frequency spectrogram, represents the fault label.

[0069] 2) Feature extractor: the input time-frequency spectrogram with a size of is divided into fixed-size non-overlapping small squares, each with a size of . Therefore, . Then each small square is flattened and linearly mapped into a dimensional vector. This process is implemented through a learnable linear projection matrix . At the front end of the image block embedding sequence, a learnable classification token embedding is spliced. To preserve spatial information, a learnable position encoding is added to the entire sequence. Finally, the sequence input into the Transformer encoder can be represented as:

[0070]

[0071] where, is the i-th image block, .

[0072] The embedded vector sequence is fed into a Transformer encoder stacked by L encoder layers. Each encoder layer contains a Multi-Head Self-Attention (MSA) module and a Multi-Layer Perceptron (MLP) module. The calculation process of MSA is as follows:

[0073]

[0074]

[0075] where h is the number of attention heads, is the query, key, value projection matrix of the i-th head. Through MSA, the model can capture long-distance dependencies between different time regions. Finally, the vector corresponding to the CLS token output by the Transformer encoder is regarded as the high-level feature representation of the entire signal segment.

[0076] 3) Feature enhancement module: CBAM sequentially performs attention weighting on the input block embedding sequence from the channel and spatial dimensions. The channel attention process is as follows: first, the input block embedding sequence is subjected to global average pooling and global maximum pooling in the spatial dimension to obtain two vectors. After passing through a shared MLP, the two vectors are added and then passed through a Sigmoid function to generate the channel attention weight . In addition, the spatial attention process is as follows: the channel attention weighted feature is subjected to global average pooling and global maximum pooling in the channel dimension to obtain two vectors. After concatenation, they are passed through a convolution layer and a Sigmoid function to generate the spatial attention weight . Finally, the feature enhanced by CBAM is:

[0077]

[0078] is taken as the actual input of the Transformer encoder, which enables the model to prioritize attention to the key frequency bands and time points with the most concentrated energy or the most abundant information before global analysis.

[0079] 4) Classifier and domain adaptation module: This module is the key to achieve high-precision cross-condition diagnosis, which handles two core tasks in parallel: one is to accurately classify the source domain condition data; the second is to bridge the data distribution gap between the source domain and the target domain, to ensure that the model has strong generalization ability. This module is composed of three parts: classification loss, CORAL loss and CDAN loss, which work together to optimize the entire network.

[0080] The basic task of the module is to perform fault classification under supervised learning. The classifier is usually composed of a multi-layer perceptron (MLP), which nonlinearly maps the high-dimensional feature space to the logical output space corresponding to a specific fault category. Then, through a Softmax function, these logical outputs are converted into an effective probability distribution vector , where and . Each element in the vector represents the predicted confidence that the input sample belongs to the corresponding fault category.

[0081] For source domain condition data with real fault labels , this method uses the standard cross-entropy loss function to quantify the gap between model predictions and reality. Cross-entropy is an effective indicator of the difference between two probability distributions, and in this task, it measures the dissimilarity between the model's predicted probability distribution and the one-hot encoded distribution representing the true class . Its mathematical expression is as follows:

[0082]

[0083] where is the number of samples in the source domain batch, is the total number of fault categories, represents the value of the true label of the th sample in the th category, is the predicted probability of the th sample belonging to the th category. By backpropagation and minimizing this loss function, the feature extractor ViT and the classifier can jointly learn a highly discriminative feature space for known fault patterns in the source domain, which is the fundamental guarantee for the model to obtain basic diagnostic capabilities.

[0084] To make the model preliminarily adapt to the data of the target domain working condition, a first domain adaptation strategy, i.e., CORAL loss, is introduced while the basic classification ability is established. CORAL is an efficient and stable difference measure method, which aims to realize the coarse-grained alignment between domains. The core statistical motivation is that the feature distribution difference under different working conditions can be largely characterized by the difference of its second-order statistics. The goal of CORAL loss is to directly align the covariance matrices of the source domain feature batch and the target domain feature batch.

[0085] The covariance matrix describes the linear relationship and correlation structure between different dimensions inside the eigenvector. By minimizing the difference between the covariance matrices of the two domains, CORAL can prompt the feature extractor ViT to learn a unified feature space, in which the correlation structure between new dimensions tends to remain consistent under any operating condition. This is equivalent to aligning the shape and direction of the feature space. Specifically, the covariance matrices of the source domain and the target domain are calculated as follows:

[0086]

[0087] wherein,

[0088] is the batch size, is the feature mean vector, is a column vector of all 1s. The CORAL loss is the square of the Frobenius norm of the difference between the two matrices:

[0089] wherein,

[0090] represents the square of the matrix Frobenius norm, i.e., the sum of the squares of all elements of the matrix, is a normalization factor. The advantage of this method is that it is simple to calculate and does not require complex adversarial training, which can provide a stable and fast global distribution preliminary alignment for the entire domain adaptation process, effectively preventing pattern collapse and other problems that may occur during subsequent adversarial training, and laying a good foundation for more refined alignment.

[0091] ​​To achieve more fine-grained and robust domain adaptation, the framework introduces Conditional Domain Adversarial Network (CDAN) as one of its core strategies. The core idea of CDAN is derived from the adversarial learning paradigm, which aims to build a minimax two-player game between the feature extractor and the domain discriminator In this game, the goal of the domain discriminator is to accurately identify the source domain of the input features, while the goal of the feature extractor is to learn a domain-invariant feature mapping that generates features that can confuse the domain discriminator, making it unable to effectively distinguish the source. When the adversarial process reaches a Nash equilibrium, the feature representation learned by the feature extractor possesses the desired domain invariance.

[0092] One major limitation of traditional adversarial domain adaptation methods is that their alignment process is usually class-agnostic, meaning that when minimizing the distribution difference between domains, the inherent class structure of the data may be ignored, leading to incorrect alignment of features from different classes. CDAN effectively overcomes this deficiency by introducing an innovative conditional strategy. Instead of directly inputting the feature vector to the discriminator, it performs a multilinear mapping of the feature with the predicted probability of the classifier to form a conditional feature representation , which is then input to the discriminator. The core value of this design is that it makes the decision of the domain discriminator dependent not only on the feature itself, but also on the class information contained in the feature. The task of the discriminator thus becomes a conditional problem: given that the current sample is predicted to be a certain class, is its feature representation more consistent with the source domain distribution or the target domain distribution? This conditional adversarial process tightly couples domain adaptation with the classification task, allowing for the alignment of features from the same class across domains while effectively maintaining the discriminative distance between features from different classes, thus preserving the discriminative information essential for the classification task.

[0093] This adversarial minimax optimization objective is efficiently implemented by introducing the Gradient Reversal Layer (GRL). GRL behaves as an identity transformation during the forward propagation of the network, without changing any numerical values; however, during backpropagation, it multiplies the gradient from the domain discriminator by a negative constant and passes it to the feature extractor. In this way, when the entire network is optimized to minimize a unified loss function, the domain discriminator the parameters of the feature extractor will be updated towards a direction that better distinguishes the domains, while the parameters of the feature extractor will be updated towards a direction that better confuses the domains. The loss function of CDAN is a standard binary cross-entropy loss:

[0094]

[0095] Through this refined conditional adversarial game, the model is driven to learn a high-level feature representation that is both domain-invariant and class-discriminative, thus providing a solid theoretical foundation for achieving accurate cross-condition fault diagnosis.

[0096] Furthermore, the method designs three types of loss functions to jointly guide the network parameters to converge to the optimal solution. The classification loss adopts a cross-entropy loss, which is used to supervise the fault classification performance of the model. It calculates the difference between the probability distribution of the fault class predicted by the model and the true label on the source domain condition data, to drive the feature extractor and the classifier to learn a feature representation that has high discriminability for fault patterns. This is the core to ensure the basic diagnostic accuracy of the model. The correlation alignment loss is used to achieve coarse-grained alignment between domains. It calculates and minimizes the difference between the second-order statistics of the feature distribution of the source domain condition and the target domain condition, to encourage the model to learn a representation space in which the feature correlation structure remains consistent under different conditions. The introduction of this loss function provides a stable and efficient global alignment basis for the entire domain adaptation process. The conditional adversarial loss is used to achieve fine-grained alignment between domains. Through the adversarial game between the feature extractor and the domain discriminator, and by introducing classification prediction as a condition, this loss function drives the model to learn a truly domain-invariant feature. Such a feature not only cannot be distinguished between different conditions, but also retains clear class boundaries, effectively preventing the problem of class confusion that may occur during domain alignment.

[0097] 2) Optimize model parameters

[0098] The goal of model training is to minimize a total loss composed of the above three types of loss functions weighted. This is a minimax optimization problem. For the feature extractor and the classifier, the optimization goal is to minimize the classification loss and the correlation alignment loss, while maximizing the conditional adversarial loss. For the domain discriminator, the optimization goal is to minimize the conditional adversarial loss. The total loss function can be represented as:

[0099]

[0100] where, and is a hyper-parameter for balancing the weights of different loss terms. During the training process, the model calculates the gradient of the loss function with respect to each part of the network parameters by the gradient backpropagation algorithm, and iteratively updates the parameters using optimizers such as Adam, so that the total loss function gradually converges.

[0101] 3) Training termination condition

[0102] The training process will continue until one of the following conditions is met: the training epoch reaches a set value; the loss function converges, that is, the loss changes very little after several consecutive training rounds, and the model is stable.

[0103] Through the above optimization process, the method can effectively learn fault features with domain invariance and class discriminability, thereby realizing high-precision cross-condition fault diagnosis.

[0104] The expression of the final fault diagnosis model is:

[0105] ;

[0106] wherein, is the Softmax probability vector output by the classifier for the feature , is the total number of fault categories. The function returns the class index that maximizes the probability .

[0107] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other.

[0108] The principles and implementation modes of the present application are described by applying specific examples in this paper. The above description of the embodiments is only used to help understand the method of the present application and its core idea; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A rotating machinery fault diagnosis method based on Vision Transformer and multi-strategy domain adaptation, characterized in that, The application relates to a method for realizing fault diagnosis of rotating machinery, and belongs to the technical field of fault diagnosis. During operation of the rotating machinery, vibration signals under different working conditions are collected by an acceleration sensor, and the vibration signals are preprocessed to obtain a standardized time-frequency spectrum data set; The standardized time-frequency spectrum data set is input into a feature extraction network composed of a Vision Transformer model and a convolution attention module to obtain high-dimensional feature representation fused with attention mechanisms; Based on the high-dimensional feature representation, a multi-strategy domain adaptation framework including a classification module and a domain alignment module is constructed; The multi-strategy domain adaptation framework obtains a diagnosis feature set optimized by domain alignment through joint optimization of classification feature learning and domain feature alignment; The diagnosis feature set is input, and end-to-end training and parameter optimization are performed by using multiple loss functions to train and update the multi-strategy domain adaptation framework and the feature extraction network, thereby obtaining a final fault diagnosis model; The final fault diagnosis model is used to obtain a fault recognition result.

2. The method according to claim 1, wherein, The working conditions include: Source domain working conditions with labels and target domain working conditions without labels.

3. The method according to claim 1, wherein, During operation of the rotating machinery, vibration signals under different working conditions are collected by an acceleration sensor, and the vibration signals are preprocessed to obtain a standardized time-frequency spectrum data set, including: An acceleration sensor is installed at a key position of the rotating machinery to collect vibration signals of the equipment in real time when the equipment is running under different working conditions, thereby obtaining an original vibration signal data set; The original vibration signal is subjected to amplitude standardization and normalization processing, and the normalized signal is segmented according to a preset step length by using a fixed-length sliding window, thereby obtaining a normalized vibration signal segment set; Signal energy distribution information changing with time in each normalized vibration signal segment is mapped to a two-dimensional time-frequency space to obtain an original time-frequency spectrum set; The original time-frequency spectrum set is subjected to unified size adjustment to obtain a standardized time-frequency spectrum data set.

4. The method according to claim 3, wherein, The expression of the normalized vibration signal segment set is: ； where, and are the minimum and maximum values of the signal, respectively, denotes the signal normalized to the interval [0, 1].

5. The method according to claim 1, wherein, The standardized time-frequency spectrum data set is input into a feature extraction network composed of a Vision Transformer model and a convolution attention module to obtain high-dimensional feature representation fused with attention mechanisms, including: The Vision Transformer model is used to segment the standardized time-frequency spectrum data set to obtain a plurality of non-overlapping image blocks of a fixed size, and each non-overlapping image block is converted into a corresponding vector sequence; A learnable classification identification embedding and position coding are added to the front end of the vector sequence to obtain an image embedding sequence with position information; A convolution attention module is introduced before the Transformer encoder to apply attention weighting to the channel dimension and the spatial dimension of the image embedding sequence respectively, thereby obtaining an enhanced feature sequence; The enhanced feature sequence is input into the Transformer encoder to obtain high-dimensional feature representation fused with attention mechanisms.

6. The method according to claim 1, wherein, The expression of the final fault diagnosis model is: ； where, is the classifier's output Softmax probability vector for the feature , is the total number of failure classes. The function returns the class index that maximizes the probability .

Images