Rolling bearing fault diagnosis method based on time-frequency fusion and double-branch deep network

Abstract

The application discloses a rolling bearing fault diagnosis method based on time-frequency fusion and double-branch deep network, and relates to the technical field of mechanical equipment state monitoring and fault diagnosis, and the method comprises the following steps: adopting time-frequency joint representation of VMD and Multi-taper PSD on a single-channel vibration signal, and rearranging the time sequence into a fixed-size gray matrix, so that the time sequence is retained and stable spectrum energy estimation is obtained; a parallel double branch is constructed on network modeling: the frequency domain branch performs layered window self-attention modeling, and the time domain branch strengthens long-term dependence and focuses on key moments by means of TCN-CBAM; an XPAF module is introduced, a pair of cross attention is taken as the core, a prior bias B and a directional mixed structure are combined, feature alignment and residual path are matched, and the time-frequency information consistency and information efficient flow are ensured. The method significantly improves the weak fault separability and the robustness under noise and variable working conditions, and is suitable for online monitoring and intelligent diagnosis of bearings and other rotating machines.

Description

Technical Field

[0001] This invention relates to the field of mechanical equipment condition monitoring and fault diagnosis technology, and in particular to a method for diagnosing rolling bearing faults based on time-frequency fusion and dual-branch deep networks. Background Technology

[0002] As industrial equipment develops towards high-speed, heavy-load, and long-cycle operation, bearing vibration signals exhibit characteristics such as strong non-stationarity, weak periodicity, and low signal-to-noise ratio. Traditional fault diagnosis methods either rely on manual experience to extract a small number of time-domain or frequency-domain indicators, or lack sufficient collaborative modeling and robust fusion of information from both domains under complex operating conditions. Consequently, it is difficult to balance diagnostic accuracy, stability, and online deployment costs.

[0003] For example, the scheme in publication number CN115758083A decomposes vibration signals using VMD, then constructs features using SVD / statistics and submits them to shallow classifiers such as SVM for identification. This process is simple to implement, but the features are manually designed and lack end-to-end optimization, limiting the feature representation ability when facing noise and variable loads; at the same time, the time domain and frequency domain lack organic coordination, making it difficult to ensure semantic consistency and complementary use between the two from a mechanism perspective, thus limiting inter-class separability.

[0004] For example, the scheme in publication CN118918424A uses the original signal and the reconstructed time-frequency image as input and fuses them through a multi-scale attention network. While such methods can improve the representation dimension, they rely on an additional image reconstruction pipeline and a relatively complex network structure, resulting in high training and deployment costs and limited online applicability. Furthermore, the fusion is mostly a weighted concatenation at the end, lacking constraints on the "time-frequency" correspondence, and the differentiation of early weak faults in strong noise scenarios remains unstable.

[0005] In summary, existing bearing fault diagnosis technologies still have significant shortcomings in terms of the comprehensiveness of time-frequency fusion feature extraction, model diagnostic accuracy, and adaptability to complex working conditions, making it difficult to meet the needs of efficient and accurate bearing fault diagnosis. How to effectively fuse the time-domain and frequency-domain features of vibration signals and automatically extract fault information using deep network models to improve the accuracy, robustness, and real-time performance of diagnosis remains a pressing technical challenge in the field of bearing diagnosis. Summary of the Invention

[0006] To address the aforementioned problems and technical requirements, the inventors propose a rolling bearing fault diagnosis method based on time-frequency fusion and a dual-branch deep network. Using only a single accelerometer, robust multi-scale time-frequency features are constructed by jointly employing variational mode decomposition (VMD) and multi-window power spectral density estimation (Multi-taper PSD). Dual-branch modeling is used, employing a Swing TransformerBlock module (frequency domain branch) and a TCN-CBAM Block module (time domain branch). Fine-grained interaction between the frequency and time domains is achieved through a priori weighted cross-attention fusion XPAF module, thereby improving the separability of weak faults, noise resistance, and online deployment efficiency. To achieve the above objectives, the technical solution of this invention is as follows:

[0007] A method for fault diagnosis of rolling bearings based on time-frequency fusion and dual-branch deep network, the method comprising the following steps:

[0008] Step 1: Acquire the vibration signal of the rolling bearing under test. Install at least one accelerometer at a radial measurement point on the bearing housing or casing (near the drive or non-drive end of the outer ring of the bearing under test). The measurement direction should be radial, and the sampling frequency should be 12–64 kHz (adjustable according to the model). The window length should be 1024 or an integer multiple thereof. Alternatively, one or a combination of triaxial sensors can be used. Then, the acquired vibration signal is detrended, bandpass / anti-aliasing filtered, normalized, and sliced ​​at fixed lengths.

[0009] Step 2: Combine variational mode decomposition (VMD) and multi-window power spectral density estimation (Multi-taper PSD) to process the vibration signal in order to construct multi-scale time-frequency features.

[0010] Step 3: Input the multi-scale time-frequency features into a pre-trained parallel dual-branch deep network for feature extraction and fusion to obtain fused features.

[0011] Parallel dual-branch deep networks include:

[0012] The frequency domain branch consists of N Swing Transformer Block modules and N-1 prior weighted cross-attention fusion XPAF modules, with the two modules arranged in a cross pattern.

[0013] The time-domain branch consists of N TCN-CBAM Block modules and one XPAF module, with the XPAF module arranged at the end;

[0014] Each XPAF module is used to fuse the features output by its connected Swin Transformer Block module and TCN-CBAMBlock module. A stage-by-stage fusion method is employed, with each stage's fused feature output only at the end through a single XPAF module and then fed to the next stage's Swin Transformer Block module. The final fused feature F is output by the XPAF module on the time-domain branch. out Before entering the XPAF module, the time length T of the two branches needs to be aligned and the channel mapping completed by 1×1 convolution.

[0015] Step 4: Fuse the features F out The bearing fault diagnosis results are output after passing through the Global Average Pooling (GAP) layer, the fully connected classifier, and the Softmax layer in sequence.

[0016] The further technical solution is that step 2 above specifically includes the following:

[0017] VMD decomposition: The modal decomposition number K and penalty factor α are preset (e.g., K = 4~8, α = 1000~4000). The original non-stationary vibration signal is decomposed into K intrinsic mode function (IMF) sequences with different center frequencies through the non-recursive iterative method of VMD, overcoming the modal aliasing and endpoint effects of EMD. The Top-k gating mechanism is applied to the K IMF sequences to generate the channel weight vector g and suppress the unselected IMF components.

[0018] Optionally, the number of selections for the Top-k gating mechanism is k′∈[2,6], and the amplitude attenuation of unselected IMF components is no greater than 0.1.

[0019] Robust spectral estimation: Perform Multi-taper PSD on K IMF sequences and subband normalize the frequency band (to suppress spectral leakage and operating condition drift) to obtain a robust frequency domain characterization (smooth, noise-resistant power spectrum).

[0020] Optionally, the Multi-taper PSD uses DPSS (Discrete Prolate Spheroidal Sequences) tapes to calculate the PSD, with a time-bandwidth product NW∈[2,4], a number of tapes Kt∈[3,7], and an overlap rate of 25%–50%; subband normalization is standardized according to the frequency band mean.

[0021] Feature Construction: The IMF components selected by the Top-k gating mechanism and the normalized frequency domain representation are combined to form multi-scale time-frequency features, which are then input into the frequency domain branch and the time domain branch, respectively. The constructed multi-scale time-frequency features are jointly represented by the multi-scale time domain and frequency domain to compensate for the insufficient capture of non-stationary features by single time domain analysis. To ensure the complete preservation of network input and temporal order, the multi-scale time-frequency features can be IMF1 to IMF stacked in time alignment. K It forms a (K+1)×T matrix with a PSD.

[0022] Optionally, before inputting multi-scale time-frequency features into the frequency domain branch, rearrange them in chronological order and fill them into a grayscale matrix of a predetermined size to maintain the temporal relationship and adapt to the two-dimensional input of the frequency domain branch. For example, perform linear interpolation / rearrangement on each row of the (K+1)×T matrix to form grayscale blocks of a fixed size, where the horizontal axis still corresponds to time (or frequency) and the vertical axis corresponds to the channel / sub-band. Before feeding them into the Swin Transformer Block module, perform channel alignment using 1×1 convolution. If the frequency domain is an H×W spectrum, use W as the time alignment dimension. If necessary, upsample / downsample W→T to ensure that T in the time domain branch is consistent with the "time direction" of the frequency domain branch.

[0023] The further technical solution is as follows: In step 3:

[0024] (1) The Swin Transformer Block module includes multi-head self-attention windowed multi-head self-attention (W-MSA) and shifted window multi-head self-attention window (SW-MSA) for hierarchical extraction, as well as feedforward sub-layers located between layers. The feedforward sub-layers use C-lite FFN to replace the traditional MLP. In C-lite FFN, directional mixing and dimensional convergence are achieved in the order of first convolution (1×k), first GLU gate, second convolution (k×1), and third convolution (1×1). The residual connections remain stable, significantly reducing parameters and computational cost, and better suiting frequency×time anisotropic data. Among them, the convolution kernel k is an odd number and takes k∈{3,5,7}; the first GLU gate divides the channels and performs gate, and the number of channels changes from 2C→C.

[0025] In the frequency domain branch, the two-dimensional gray matrix is ​​first divided into patches. W-MSA is performed at each stage in conjunction with SW-MSA to expand the receptive field. Layer normalization and feedforward sub-layers are used between layers to take into account both local details and global contextual relevance.

[0026] (2) The TCN-CBAM Block module contains several hierarchically connected dilated convolutional residual blocks (dilation rate d is {1,2,4,8...}) for long-term dependency modeling. Each residual block embeds an attention module CBAM. CBAM first applies channel attention (Global Average Pooling GAP / Global Max Pooling GMP to obtain the statistical features of each channel → MLP → Sigmoid to generate channel weights) to the features extracted by convolution to highlight key channels, and then applies spatial / temporal attention (performing k-axis multiplication on the feature map after channel enhancement in the spatial dimension). s A one-dimensional convolution operation of ×1 is used to generate spatial weights, emphasizing key temporal locations and suppressing irrelevant perturbations to enhance sensitivity to weak fault features and improve the separability of weak faults in complex noise backgrounds.

[0027] Optionally, the expansion rate d adopts an exponentially increasing scheme:

[0028] d l =2 l-1 ,l=1,…,L

[0029] For example, when L = 4, d = {1, 2, 4, 8}. Therefore, the effective receptive field of this branch is:

[0030]

[0031] (3) In the XPAF module, firstly, the features output by the connected Swin Transformer Block module and TCN-CBAMBlock module are received and input into the two cross-attention channels with different directions for calculation, and a prior bias B is added to the attention logits. Specifically, the features output by the Swin Transformer Block module are input into the cross-attention channel with the direction of time-to-frequency domain (Time→Freq), where Q is the time-domain feature output by the Swin Transformer Block module, and K and V are the frequency-domain features output by the Swin Transformer Block module. The features output by the TCN-CBAM Block module are input into the cross-attention channel with the direction of frequency-to-time domain (Freq→Time), where Q is the frequency-domain feature output by the TCN-CBAM Block module, and K and V are the time-domain features output by the TCN-CBAM Block module. The prior bias B is injected into the logits of the two cross-attention channels, and the calculation formula for each cross-attention channel satisfies:

[0032]

[0033] Where d represents the feature dimension, and the prior bias B originates from any one or a combination of the following:

[0034] (1) Estimated based on the known geometric characteristic frequency or rotational speed of the rolling bearing under test;

[0035] (2) Results of clustering the spectral peaks of Multi-taper PSD;

[0036] (3) The channel weight vector g generated by the Top-k gating mechanism is used for channel mapping.

[0037] Secondly, while keeping the time dimension T constant, the outputs of the two cross-attention paths are concatenated along the channel dimension and subjected to a fourth convolution I (1×1 projection). Then, they pass through a directional mixing stage consisting of Align, a fifth convolution (1×k), a second GLU gate, and a sixth convolution (k×1). The residuals are then added to the output of the fourth convolution I, and finally, the fused features are obtained through a fourth convolution II (1×1). In the directional mixing stage, the channel weight vector g generated by the Top-k gating mechanism is used for channel amplification.

[0038] Optionally, if the features output by the Swin Transformer Block and the TCN-CBAM Block are inconsistent in length along the temporal axis, linear interpolation / average pooling is used to align their lengths before concatenation. This module design clearly defines the weight generation location (channel attention within the fusion module) while ensuring cross-modal temporal consistency.

[0039] The further technical solution is as follows: In step 4:

[0040] For fusion feature F out A fixed-length vector is obtained by performing a gap analysis. This vector is then passed through a fully connected classifier with temperature scaling (which allows for calibration of logits) and a Softmax layer to output the probabilities of each fault category. When the probability of a certain category exceeds a set alarm threshold τ (e.g., 0.5–0.7), an alarm or maintenance command is triggered.

[0041] The further technical solution is that the attention weights and their prior bias weights B in the XPAF module, the channel weights of CBAM, and the channel weight vector g generated by the Top-k gating mechanism are all adaptively learned through end-to-end training, and are not fixed constants.

[0042] The further technical solution is that the training phase of the parallel dual-branch deep network uses cross-entropy as the loss function, adopts an adaptive first-order optimization algorithm, and combines learning rate annealing and regularization (including Dropout) to improve the network's convergence stability and generalization ability.

[0043] The beneficial technical effects of this invention are:

[0044] The method provided by this invention significantly improves the accuracy and robustness of rolling bearing fault identification through synergistic improvements in the feature construction layer, network modeling layer, and fusion layer. In the feature construction layer, VMD and Multi-taper power spectrum are introduced to jointly construct multi-scale time-frequency features, and Top-k differentiable gating of the IMF sequence is set to generate channel weights g, which preserves the time sequence and obtains stable spectral energy estimation, making it more comprehensive and robust than the manual process of "VMD + manual indicators / SVD". Based on the spectral peak and / or weights g, a priori bias B is generated for attention logits weighting in the subsequent fusion layer, so that weak fault bands can still be highlighted and spectral leakage and variance can be suppressed under strong noise and variable load conditions.

[0045] In the network modeling layer, the frequency domain branch uses Swing-Transformer and replaces the traditional MLP with C-lite FFN (1×k→GLU→k×1→1×1) to achieve anisotropic direction mixing and reduce parameter computation. The time domain branch uses TCN-CBAM to expand convolution to model long-term dependencies and enhance the response at key moments through channel-space attention. The two branches complement each other in terms of receptive field and inductive bias. Both branches can achieve end-to-end training and the number of parameters is controllable, which is convenient for online deployment.

[0046] In the fusion layer, XPAF (Prior Weighted Bidirectional Cross Attention) is proposed for implementation. The system guides and superimposes Align alignment and “1×k→GLU→k×1” directional mixing and residual convergence, which can achieve clearer inter-class separation and more robust cross-condition performance compared to simple splicing fusion. Finally, the system outputs the data through GAP and a Soft-max classification head with temperature scaling, and the threshold can be set to achieve immediate judgment and alarm.

[0047] In summary, the present invention outperforms existing technologies in terms of feature extraction comprehensiveness, diagnostic accuracy, anti-interference and edge reasoning efficiency. It can achieve timely, stable and accurate identification of multiple types of bearing faults under single sensor conditions, meeting the needs of engineering sites for intelligent operation and maintenance with high reliability, low cost and easy deployment. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the overall process of the rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network provided in this application;

[0049] Figure 2 This is a schematic diagram of converting multi-scale time-frequency feature data into grayscale image format, as provided in this application.

[0050] Figure 3 This is a schematic diagram of the structure of the Swin Transformer module used in the frequency domain branch provided in this application;

[0051] Figure 4 This is a schematic diagram of the TCN-CBAM module used in the time-domain branch provided in this application;

[0052] Figure 5 This is a schematic diagram of the prior weighted cross-attention XPAF module provided in this application;

[0053] Figure 6 These are the performance change curves during the network training process provided in this application, where (a) is a schematic diagram of the change in accuracy of the training set and validation set with the number of iterations; and (b) is a schematic diagram of the change in loss value of the training set and validation set with the number of iterations.

[0054] Figure 7 This is a schematic diagram of the confusion matrix of the method provided in this application on the test set, showing the correspondence between each true category and the predicted category and the distribution of classification accuracy.

[0055] Figure 8 The above are the feature visualization results based on clustering visualization t-SNE provided in this application, where (a) is the sample distribution at the input end (after time-frequency feature construction and before bi-branch extraction); (b) is the embedded feature distribution after time-frequency fusion and bi-branch deep network extraction by the method of this invention, and the comparison shows the improvement of inter-class separability. Detailed Implementation

[0056] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0057] One embodiment of this application provides a rolling bearing fault diagnosis method based on time-frequency fusion and a dual-branch deep network. This embodiment uses a rolling bearing in an industrial site as the test object, employing a single-channel accelerometer (preferably a piezoelectric accelerometer, radially attached or bolted to the bearing housing near the measuring point of the bearing under test) to collect vibration signals and input them into the fault diagnosis system. This method allows for online determination of the bearing's operating status. Firstly, as... Figure 1 As shown, the entire process consists of three stages: time-frequency feature construction (VMD+Multi-taper), dual-branch deep modeling (Swin-T+TCN-CBAM), and prior-weighted cross-attention fusion (XPAF) and fault classification.

[0058] Phase 1 (Construction of Time-Frequency Features): Combining Figure 1 and Figure 2As shown, the acquired raw vibration signal, after necessary detrending and normalization, is input into the feature extraction module: Variational Mode Decomposition (VMD) is used to decompose the signal into 5 Intrinsic Mode Functions (IMFs), resulting in a 5×1024 time-domain feature sequence; simultaneously, Multi-taper Power Spectrum Estimation (PSD) is performed on this time-domain feature sequence, and then sub-band normalization is performed on the frequency band to eliminate amplitude dimension differences, forming a "frequency domain feature map". Subsequently, the 5 IMFs are rearranged in chronological order and combined with the normalized frequency domain grayscale blocks to form a 5×32×32 grayscale tensor, which serves as the "frequency domain input", as shown below. Figure 2 As shown in the figure, "Inner_××", "Ball_××", and "Outer_××" represent inner race faults, rolling element faults, and outer race faults, respectively. The suffix numbers (e.g., 18 / 36 / 54) are the fault size / grade markers for this dataset (larger values ​​indicate more severe damage); "Normal_00" represents normal samples. The corresponding "Time Domain Input" is a 5×1024 frequency domain feature map. The output of this stage is... Figure 1 The process is consistent with "Step 1: Signal Acquisition and Data Preprocessing" on the left side of the document. For details, please refer to Steps 1 and 2 in the invention description. It will not be repeated here.

[0059] Second stage (feature extraction): such as Figure 3 and Figure 4 ,as well as Figure 1 As shown on the right, a parallel dual-branch deep network is constructed to extract the aforementioned multi-scale features in depth. Specifically, this includes:

[0060] (1) Frequency domain branching (using 4 Swin-Ts, including C-lite FFN): The Swin Transformer divides the grayscale tensor into patches and stacks them in two steps: “W-MSA→LN→C-lite FFN” and “SW-MSA→LN→C-lite FFN” (see Figure 3 ) Extract spectral information layer by layer. Among them, C-lite FFN replaces the traditional MLP with a factorized convolution-gated structure of "1×k→first GLU gate→k×1→1×1", and performs lightweight feedforward transformation with directional convolution (two axes along the frequency and time), which not only preserves the number of channels, but also better fits the time-frequency anisotropy.

[0061] (2) Time-domain branching (using 4 TCN-CBAMs): such as Figure 4As shown, the network consists of multi-level dilated convolutional residual blocks, each with weight normalization, ReLU activation, and Dropout. Each residual block then embeds a CBAM attention module. Specifically, the temporal branch first captures the long-term correlation features of the vibration signal through multiple layers of one-dimensional dilated convolutions. Each convolutional layer is followed by weight normalization and ReLU activation to ensure stable convergence, and Dropout regularization is introduced to prevent overfitting. Subsequently, the CBAM module in the residual block applies channel attention and spatial attention sequentially to the features extracted by the convolutions: channel attention reweights the features according to their importance, while spatial attention further highlights the feature responses at important temporal locations. This TCN-CBAM branch enhances the model's focus on key temporal fault information, accurately extracting subtle fault features under complex operating conditions.

[0062] The third stage (using four XPAF files for two-branch fusion, and finally classification): For example... Figure 5 and Figure 1 As shown in the shaded box below, the XPAF module (bidirectional cross-attention + prior bias B) is used to fuse the two branches and complete fault classification. Specifically, the "TCN-CBAM output" and "Swin-T output" are first fed into the corresponding cross-attention modules: the features output by the Swin Transformer Block module are input into the cross-attention module in the direction of time-to-frequency domain (Time→Freq), where Q is the time-domain feature output by the Swin Transformer Block module, and K and V are the frequency-domain features output by the Swin Transformer Block module. The features output by the TCN-CBAM Block module are input into the cross-attention module in the direction of frequency-to-time domain (Freq→Time), where Q is the frequency-domain feature output by the TCN-CBAM Block module, and K and V are the time-domain features output by the TCN-CBAM Block module. A prior bias B (weight g generated from the spectral peaks of the multi-taper and / or the IMF Top-k gating of the VMD) is added to the attention logits to a priori amplify the fault-related channels / bands. The results of the two cross-attention paths are fused and aligned via “Concat→1×1Conv”; subsequently, a directional hybrid chain of “Align→1×k→GLU→k×1→1×1” is applied to the output (see...). Figure 5 By generating residuals, intermediate channels are added in parallel with the initial input to ensure consistency between temporal and frequency semantics and convergence speed, resulting in the final fused feature F. out .

[0063] Fault classification: such as Figure 1 As shown, Global Average Pooling (GAP) is used to pool F.out The model is mapped to a fixed-length vector, fed into a feedforward layer (an FC classifier with temperature scaling) and a Softmax layer, outputting the probabilities of normal operation and various other fault categories. The network is trained by minimizing the cross-entropy loss to optimize model parameters and achieve high-accuracy classification output. In experimental tests, the method of this invention can accurately distinguish between various operating conditions, such as minor inner race faults and severe outer race faults, in sample data. For example, for a real bearing dataset, the model's diagnostic accuracy exceeds 95%, significantly outperforming traditional single-domain feature diagnostic methods, fully demonstrating the ability of this invention to extract and identify fault features in complex environments.

[0064] Performance validation: The method was trained and tested on the CWRU public bearing dataset. Figures 6-8 The training phase employs cross-entropy loss and a first-order adaptive optimization algorithm; evaluation metrics include accuracy, recall, and F1 score, and training / validation curves and test set confusion matrices are recorded. Figure 6 As shown in (a), the training and validation accuracy steadily increases with iteration and tends to converge after several iterations; Figure 6 As shown in (b), the training and validation losses decreased synchronously and then plateaued in the later stages, with no obvious signs of overfitting. This result demonstrates that the time-frequency fusion and dual-branch deep network architecture continuously improves generalization performance during optimization. Figure 7 As shown, the confusion matrix of the test set exhibits a high-intensity distribution along the diagonal, indicating a significant reduction in inter-class misclassifications. This demonstrates that the proposed method can effectively distinguish between different parts and degrees of bearing faults. Compared to control methods that only model in the time domain or frequency domain, this method achieves a higher correct classification rate in most categories. Figure 8 (a) and Figure 8 As shown in (b) above, after mapping the high-dimensional features to a two-dimensional space using t-SNE, it can be seen that the sample point cloud distribution at the input end (after time-frequency feature construction and before dual-branch extraction) is discrete. However, after extraction by the frequency domain branch Swin-T and the time domain branch TCN-CBAM of the present invention and weighted fusion by the XPAF module, the feature distribution forms a compact cluster, and the inter-class separability is significantly enhanced. This phenomenon confirms the effectiveness of the complementarity of the "frequency domain-time domain" and the prior weighting of the method of the present invention.

[0065] In summary, the method of this invention can accurately identify various faults in rolling bearings (such as inner ring, outer ring, and rollers); it maintains high diagnostic accuracy even under conditions of high noise and variable speed. This method is suitable for online condition monitoring of rotating machinery such as wind turbine generators and CNC machine tool spindles; due to the use of lightweight C-lite FFN and XPAF fusion, the number of model parameters is controlled, resulting in better engineering portability. By adjusting the sensor and preprocessing scheme, it can also be extended to fault identification in other equipment such as gearboxes and motors. Therefore, this invention improves the accuracy and reliability of bearing fault diagnosis while also meeting the urgent needs of modern industry for intelligent fault diagnosis technology.

[0066] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A method for fault diagnosis of rolling bearings based on time-frequency fusion and dual-branch deep network, characterized in that, The method includes: Collect vibration signals from the rolling bearing under test; The vibration signal is processed by combining variational mode decomposition (VMD) and multi-window power spectral density estimation (Multi-taper PSD) to construct multi-scale time-frequency features; The multi-scale time-frequency features are input into a pre-trained parallel dual-branch deep network for feature extraction and fusion to obtain fused features. The fused features are sequentially passed through a global average pooling layer, a fully connected classifier, and a Softmax layer to output bearing fault diagnosis results. The parallel dual-branch deep network includes: The frequency domain branch consists of N Swing Transformer Block modules and N-1 prior weighted cross-attention fusion XPAF modules, with the two modules arranged in a cross pattern. The time-domain branch consists of N TCN-CBAM Block modules and one XPAF module, with the XPAF module arranged at the end; Each XPAF module is used to fuse the features output by the connected Swin Transformer Block module and the TCN-CBAM Block module, and output them to the next Swin Transformer Block module. The XPAF module on the time-domain branch outputs the final fused feature. The process of combining variational mode decomposition (VMD) and multi-window power spectral density estimation (Multi-taper PSD) to process the vibration signal to construct multi-scale time-frequency features includes: Pre-set the modal decomposition number K and penalty factor α The vibration signal is decomposed into K intrinsic mode function (IMF) sequences with different center frequencies through the non-recursive iterative method of VMD. Multi-taperPSD is performed on the K IMF sequences and sub-band normalization is performed on the frequency band to obtain a robust frequency domain characterization. A Top-k gating mechanism is applied to the K IMF sequences to generate a channel weight vector g and suppress components of IMFs that were not selected. The selected IMF components and the normalized frequency domain representation together constitute the multi-scale time-frequency features; In the XPAF module: The features output by the Swin Transformer Block module are input into a cross-attention process directed from the time domain to the frequency domain, and Q The time-domain features output by the Swing Transformer Block module. K , V The frequency domain characteristics output by the SwinTransformer Block module; The features output by the TCN-CBAM Block module are input into a cross-attention process directed from the frequency domain to the time domain, and Q The frequency domain characteristics output by the TCN-CBAM Block module, K , V The time-domain characteristics output by the TCN-CBAM Block module; Injecting a prior bias B at the logits of the two cross-attention paths, the calculation formula for each cross-attention path satisfies: in, d Representing the feature dimension, the prior bias B originates from any one or a combination of the following: (1) Estimated based on the known geometric characteristic frequency or rotational speed of the rolling bearing under test; (2) The results of clustering the spectral peaks of the Multi-taper PSD; (3) The channel weight vector g generated by the Top-k gating mechanism is used for channel mapping; After concatenating the outputs of the two cross-attention paths and performing a fourth convolution I, the system sequentially passes through a directional mixing stage consisting of alignment, a fifth convolution, a second GLU gate, and a sixth convolution. The residual is then added to the output of the fourth convolution I, and finally, the system passes through a fourth convolution II to obtain the fused features. In the directional mixing stage, the channel weight vector g generated by the Top-k gating mechanism is used for channel amplification.

2. The rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network according to claim 1, characterized in that, The Multi-taper PSD is calculated using DPSS taper, with time-bandwidth product NW∈[2,4], number of tapers Kt∈[3,7], and overlap rate of 25%–50%; subband normalization is standardized according to the frequency band mean.

3. The rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network according to claim 1, characterized in that, The number of selections for the Top-k gating mechanism is k′∈[2,6], and the amplitude attenuation of the unselected IMF component is no greater than 0.

1.

4. The rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network according to claim 1, characterized in that, The method further includes: The multi-scale time-frequency features are input into the frequency domain branch, rearranged in chronological order, and filled into a grayscale matrix of a predetermined size to maintain the temporal relationship and adapt to the two-dimensional input of the frequency domain branch.

5. The rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network according to claim 1, characterized in that, The Swin Transformer Block module includes layered extraction of window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SW-MSA), as well as feedforward sub-layers located between layers; The feedforward sublayer adopts C-lite FFN, in which directional mixing and dimensional convergence are achieved in the order of first convolution, first GLU gate, second convolution, and third convolution.

6. The rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network according to claim 1, characterized in that, The TCN-CBAM Block module contains several hierarchically connected dilated convolutional residual blocks, with an attention module CBAM embedded in each residual block. The CBAM applies channel attention and spatial attention to the features extracted by convolution in sequence to enhance the sensitivity to weak fault features.

7. The rolling bearing fault diagnosis method based on time-frequency fusion and dual-branch deep network according to claim 1, characterized in that, The method further includes: The fully connected classifier uses temperature scaling to calibrate logits; When the probability of any one of the fault categories output by the Softmax layer exceeds the set alarm threshold... τ It can trigger alarms or maintenance commands at any time.

Images