Dual-stream fault diagnosis method based on state space modeling and dynamic edge context awareness

Abstract

The application discloses a double-flow fault diagnosis method based on state space modeling and dynamic edge context perception, which constructs a time domain state flow and a time-frequency edge flow, and the specific implementation process is as follows: the time domain state flow is modeled in a Transform framework to form a MambaT encoder for efficiently capturing long and short term dependencies with linear complexity and accurately representing time domain sensitive features; the time-frequency edge flow transforms vibration data into a two-dimensional time-frequency image through wavelet transform, designs a differentiable edge sensitive mask to adaptively separate impact features and background noise of the time-frequency image, simultaneously designs a boundary content perception adaptive padding to maintain the topological integrity of the image boundary, and cooperatively constitutes a dynamic edge context perception convolution to effectively extract time-frequency edge features; a channel enhancement and channel gating module is designed to establish a dynamic channel interaction gating mechanism, realize time domain-time frequency domain double-flow cascaded representation, and significantly improve fault discrimination ability. Experiments prove that the method has higher fault diagnosis performance.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment fault diagnosis technology, specifically to a dual-stream fault diagnosis method (DECAMambaT) based on state-space modeling and dynamic edge context awareness, which is particularly suitable for bearing fault diagnosis scenarios. This method addresses problems such as low efficiency in long-time-series signal modeling, insufficient extraction of time-frequency feature edges, and inadequate multimodal information fusion by fusing temporal state-space modeling with time-frequency image edge awareness technology. Background Technology

[0002] As industrial equipment becomes increasingly intelligent and complex, fault diagnosis technology for rotating machinery such as bearings has become a core means of ensuring safe equipment operation. However, traditional methods have significant limitations: time-frequency analysis methods based on signal processing are inefficient at modeling long-term vibration signals; recurrent neural networks struggle to capture long-range dependencies due to serial computation; and while Transformers possess global perception capabilities, their quadratic time complexity cannot adapt to high-sampling-rate industrial signals. Furthermore, traditional convolutional neural networks use static convolution kernels when processing time-frequency images, lacking an adaptive separation mechanism for impact features and background noise, and zero-padding operations disrupt boundary topology, resulting in insufficient boundary feature extraction. In addition, existing methods do not fully integrate long-range temporal dependencies with time-frequency spatial context information, limiting the collaborative discrimination capability of multimodal features. To address the aforementioned issues, this invention proposes a dual-stream fault diagnosis method (DECAMambaT) based on state-space modeling and dynamic edge context awareness. This method achieves linear complexity temporal modeling by constructing a MambaT encoder, while simultaneously designing dynamic edge context-aware convolution (DECAC) to adaptively separate impact features from background noise. Finally, a dynamic channel interaction gating (DCIG) mechanism is proposed to realize time-domain and time-frequency domain dual-stream cascaded representation, significantly improving fault diagnosis capabilities. Summary of the Invention

[0003] This invention addresses the problems of low efficiency in long-term time-domain information modeling, insufficient extraction of time-frequency image edge features, and inadequate time-domain-time-frequency domain information fusion in existing fault diagnosis methods. It proposes a dual-stream fault diagnosis method based on state-space modeling and dynamic edge context awareness. The specific implementation steps of this invention are as follows:

[0004] 1. Construct a time-domain state flow, perform state-space modeling within the Transformer framework to form a MambaT encoder, and perform long-term time-domain modeling with linear time complexity to accurately represent time-sensitive features. The steps are as follows:

[0005] (1a) Construction of dynamically parameterized state space

[0006] For upper-layer output or embedded vectors Where B is the batch size, L is the sequence length, and D is the sequence length. model As the feature dimension, it is first obtained by applying LayerNorm normalization. Then, using linear mapping... Projection into two equal-dimensional vectors

[0007]

[0008] Where D inner For internal channel dimensions, W in ,b in The weights and biases of the linear mapping are then used to perform a Conv transformation on x in the time dimension. dw One-dimensional convolution, kernel width d conv Each channel is independent and is obtained through SiLU activation.

[0009]

[0010] After that Dividing the channel into three groups through another linear layer, we obtain the time step δ∈R. B×L×r Input gain Output gain

[0011]

[0012] Where r is the low-rank dimension used to generate the step size, and W x ,b x For the linear layer weights and biases, d state The hidden state dimension is modeled in the state space, and then δ is mapped to the actual step size Δt = Softplus(W) per channel using a biased linear mapping and Softplus activation. δ δ+b δ The state transition matrix is ​​parameterized in its logarithmic form as follows: The actual matrix is ​​A = -exp(A log );

[0013] (1b) Selective state recursive update

[0014] For each time step t, update the hidden state for all channels.

[0015]

[0016] Where ⊙ represents the dot product operation, the SSM output is:

[0017]

[0018] in For jump connection gain, then the output of the entire sequence. After performing channel-wise SiLU gating with z, and then passing it through a projection layer to map the output back to the original dimension, we obtain the output modeled in the state space.

[0019]

[0020] in W represents the multiplication operation. out ,b out For projection layer weights and biases;

[0021] (1c) Residual connection and feedforward calculation

[0022] The result o from the state-space modeling is added to the original input h via Dropout and residual concatenation:

[0023] o′=h+Dropout1(o)

[0024] Dropout1 represents the first random drop, followed by LayerNorm, and then two layers of feedforward network consisting of linear transformation and ReLU activation:

[0025] F = W2ReLU(W1LN(o′)+b1)+b2

[0026] Where W1 and W2 are weight matrices, b1 and b2 are bias vectors, and finally the output Y of the MambaT encoder is obtained through the second residual and Dropout:

[0027] Y = o′ + Dropout2(F)

[0028] Dropout2 represents the second random drop.

[0029] 2. Construct a time-frequency edge stream, design a differentiable edge-sensitive mask to adaptively separate time-frequency image impact features from background noise, and design boundary content-aware adaptive padding to maintain the topological integrity of image boundaries. The two work together to form a dynamic edge context-aware convolution, effectively extracting time-frequency edge features. The steps are as follows:

[0030] (2a) Design a differentiable edge-sensitive mask

[0031] For the input feature map X∈R B×C×H×W The grayscale image G∈R is obtained by averaging along the channel dimension. B×1×H×W :

[0032]

[0033] Where C is the number of channels, and H and W are the spatial dimensions; then, a Sobel convolution kernel K with fixed weights is used. x ,K y ∈R 3×3 Calculate the horizontal and vertical gradients G x G y :

[0034] G x =K x *G

[0035] G y =K y *G

[0036] Where * denotes a convolution operation, and then the gradient magnitude A of each pixel is calculated. pixel ∈R B×1×H×W :

[0037]

[0038] Then, using the learnable gain scalar γ and offset scalar β, and utilizing the element-order sigmoid function σ(z) = 1 / (1+e^(-1 / 2))... -z Map it to a smoothness mask M∈(0,1) B×1×H×W :

[0039] M=σ(γA pixel +β)

[0040] To further simplify branch selection, this method binarizes the smoothness mask M into an indicator function form:

[0041] S = II (M > 0.5)

[0042] Where S∈{0,1} B×1×H×W II(·) is an indicator function, which is 1 when M > 0.5 and 0 otherwise; finally, the binary mask S is interpolated to the nearest neighbor to the size (H′, W′) and copied along the channel to obtain the result.

[0043] (2b) Design Boundary Content-Aware Adaptive Fill

[0044] First, extract the image boundary slices. Let the fill scale be p = (k-1) / 2, and take the leftmost and rightmost p column boundaries of the image:

[0045] L = X :,:,:,1:p R = X :,:,:,W-p+1:W

[0046] Where L and R are the extracted leftmost and rightmost boundary slice contents, respectively, and then the boundary slices are input into the boundary generator G:

[0047]

[0048] in Fill the generated boundary with content; then horizontally stitch them together to obtain the intermediate extended image X. hor :

[0049]

[0050] Similarly, generate top and bottom boundary fill content in the vertical direction. Vertically splice it to the middle extended image X hor Finally, a boundary content-aware adaptive padding feature map X with a size of (H+2p)×(W+2p) is obtained. pad :

[0051]

[0052] (2c) Design a dynamic branch selection mechanism

[0053] Dynamic edge-context-aware convolution dynamically switches between a main branch k×k convolution in the full receptive field and a side branch 1×1 lightweight projective convolution based on a differentiable edge-sensitive mask. The main branch performs boundary-content-aware adaptive padding on the input X. pad Performing a k×k convolution on the top yields the output Y. k ∈R B×D×H′×W′ :

[0054] Y k =Conv k×k (X pad )

[0055] Where D is the number of output channels, and (H′, W′) is the output space size; the bypass branch performs a 1×1 lightweight projective convolution on the original input X to obtain the output Y1∈R. B×D×H×W :

[0056] Y1 = Conv 1×1 (X)

[0057] If stride > 1, the bypass branch is downsampled to (H′, W′) using a max pooling operation to match the spatial dimensions; then, dynamic branch selection is performed based on a differentiable edge-sensitive mask to obtain the final output Y.

[0058]

[0059] In this way, dynamic edge context-aware convolution focuses on preserving the details of the full receptive field convolution in high-gradient edge regions, while relying on lightweight projection branches in low-gradient smooth regions to reduce redundant computation and noise impact. Ultimately, it achieves accurate separation of transient impact features from background noise in fault time-frequency images, and completes the extraction of transient impact features from time-frequency edge flows.

[0060] 3. Construct a channel enhancement and channel gating module to establish a dynamic channel interaction gating mechanism, generate feature enhancement vectors and dynamic gating weights, adaptively cascade dual-stream information, and realize joint representation of time-domain and time-frequency domain bidirectional enhancement, significantly improving fault detection capability. The steps are as follows:

[0061] (3a) Channel disturbance and initial splicing

[0062] The long-range time-domain features output from the time-domain state stream are concatenated with the time-frequency image features output from the time-frequency edge stream in the channel dimension to obtain the feature vector x∈R. B×C The channel dimensions of each sample are randomly shuffled:

[0063] x ch =x[:,randperm(C)]

[0064] To construct a "channel perspective" different from the original features, the original feature x and the shuffled feature x are then compared. ch Concatenating along the channel dimension yields u = [x; x ch ]∈R B×2C ;

[0065] (3b) Channel enhancement and channel gating

[0066] The channel enhancement vector f∈R is generated by parallel channel enhancement and channel gating modules respectively. B×C and dynamic gating weights g∈(0,1) B×C :

[0067] f = MLP enhance (u),g=σ(MLP gate (u))

[0068] MLP enhance Used to extract enhanced information, MLP gate Used to generate the gate weights for each channel, σ(·) represents the element-wise sigmoid mapping;

[0069] (3c) Gated fusion yields cascaded outputs

[0070] The enhanced vector and the original features are adaptively concatenated by channel through gated fusion to obtain the fused output y∈R. B×C :

[0071] y = g⊙f + (1-g)x

[0072] It can dynamically decide whether to extract supplementary information or retain the original features on each channel, realize the cascade enhancement and complementary fusion of time-domain and time-frequency domain features, thereby improving the fault discrimination ability and model robustness of the final joint representation. Finally, the dual-stream cascaded features are input into the linear classification head to project onto the category dimension and thus realize fault diagnosis.

[0073] The method of the present invention has the following advantages:

[0074] (1) High-efficiency long sequence modeling capability: Through dynamic parameterized state space construction and selective recursive update mechanism, long time domain dependency modeling with linear complexity (O(L)) is realized in the Transformer framework, which significantly improves the processing efficiency of high sampling rate industrial vibration signals and solves the bottleneck problem of the secondary computation complexity of traditional Transformer.

[0075] (2) Adaptive edge feature extraction capability: Differentiable edge sensitive mask and boundary content-aware filling technology are used to dynamically separate the impact features and background noise in the time-frequency image, effectively maintain the topological integrity of the feature boundary, and overcome the defects of traditional convolution in the loss of features or excessive smoothing in the edge region.

[0076] (3) Multimodal deep collaborative fusion capability: Based on the channel disturbance and gating weighting mechanism, the complementary features of the cascaded time-domain state flow and time-frequency edge flow are adaptively combined to generate a more discriminative joint representation, which significantly enhances the robustness and accuracy of fault classification under complex working conditions. Attached Figure Description

[0077] Figure 1 This is the overall framework diagram of the present invention.

[0078] Figure 2 This is a classification result diagram based on the confusion matrix of a random experiment on the CWRU dataset.

[0079] Figure 3 This is a classification result using a confusion matrix from a random experiment on the AUST dataset. Detailed Implementation

[0080] 1. Data Preprocessing

[0081] Data acquisition and standardization: Collect raw vibration signals of industrial bearings, calculate the mean and standard deviation of training set signals, perform min-max normalization on raw signals to eliminate dimensional differences and improve model convergence efficiency;

[0082] One-dimensional time domain signal acquisition: A fixed-length sliding window is used to segment the time series data into samples. The window length is set to 1024 sampling points and the step size is 512 points to obtain a one-dimensional time domain signal, ensuring that the sample covers the complete vibration cycle.

[0083] Two-dimensional time-frequency image acquisition: The standardized vibration signal is converted into a two-dimensional time-frequency image through continuous wavelet transform. The cmor100-1 wavelet basis function is selected, and a fixed scale of 128 is set to generate a time-frequency representation at a sampling frequency of 12kHz. The absolute values ​​of the wavelet coefficients are extracted to represent the signal energy. A 224×224 pixel time-frequency image is generated using contour plots, where the horizontal axis represents the time series, the vertical axis represents the frequency distribution, and the image color depth corresponds to the signal energy intensity, thus forming a two-dimensional time-frequency image that combines transient characteristics in the time domain with frequency resonance characteristics.

[0084] 2. Construct the time-domain state flow

[0085] The preprocessed one-dimensional time-domain signal is input into the MambaT encoder, and the following operations are performed sequentially:

[0086] The input sequence is normalized layer by layer to stabilize the training process. Two independent vectors are generated through dual-linear projection, used for state space computation and gating operations respectively. A depthwise separable one-dimensional convolution is performed on the first set of vectors to capture local waveform patterns, and the nonlinearity is enhanced by SiLU activation. The convolution output is then divided into three sets of parameters: time step, input gain, and output gain. The time step is activated by Softplus to ensure a positive value. The hidden state is updated in parallel and recursively according to the time step: the state transition matrix is ​​calculated in real time based on dynamic parameters to achieve long-range dependency modeling, while retaining the original input features through skip connections. Finally, the state space output and the second set of projection vectors are fused through channel-wise SiLU gating, and the feature dimension is restored through linear layers.

[0087] 3. Construct time-frequency edge streams

[0088] The generated two-dimensional time-frequency image is input into a dynamic edge context-aware convolutional layer for time-frequency feature extraction. The following operations are performed sequentially:

[0089] Dynamically generate edge-sensitive masks: Calculate the mean of the input image in the channel dimension to obtain a single-channel grayscale image; use the Sobel operator with fixed weights to calculate the gradients in the horizontal and vertical directions respectively, and fuse them into a gradient magnitude map; through the adjustment of learnable gain scalar and offset scalar, generate a smoothness probability map through the Sigmoid function, and binarize it with a threshold of 0.5 as a mask, and then interpolate it to the output size.

[0090] Content-aware adaptive fill: The left and right boundary pixel columns of the image are cropped and input into a lightweight convolutional generator to predict the fill content; the generated fill content is horizontally stitched with the original image to form an intermediate image with expanded width; the upper and lower boundaries are processed in the same way, and the final output is the filled image.

[0091] Dynamic convolution branch selection based on mask: For high gradient regions (mask value of 1), a 3×3 convolution is performed on the image after boundary content-aware adaptive padding to preserve "sharp" details; for low gradient regions (mask value of 0), a 1×1 convolution projection is performed on the original image to reduce redundant interference. When the convolution stride is greater than 1, max pooling is added to the output to align the spatial dimensions; then, dynamic branch selection is performed based on a differentiable edge-sensitive mask to obtain the dual-branch fusion result; finally, global average pooling is used to aggregate the spatial features into a time-frequency edge feature vector.

[0092] 4. Fault diagnosis using a dynamic channel interaction gating mechanism with cascaded dual-stream characteristics.

[0093] The feature vectors of the temporal state flow and the time-frequency edge flow are concatenated along the channel dimension. Then, for each sample, the channel dimension is randomly shuffled independently to generate a perturbed feature vector, which is then concatenated with the original features to form a high-dimensional extended feature. Next, channel enhancement vectors and dynamic gating weights are generated through parallel channel enhancement and channel gating modules, respectively. Then, the enhancement vectors and the original features are adaptively concatenated along the channel dimension through gating fusion to obtain a concatenated fusion output. Finally, the output is processed by a classification head for fault diagnosis.

[0094] The effectiveness of this invention was further verified through the following experiments:

[0095] On the CWRU and self-built AUST bearing datasets, the average accuracy of the proposed method reached 99.87% and 99.33%, respectively, significantly outperforming the comparison models (VGG1D, CNN-LSTM, and Transformer, etc.). In noise resistance experiments, under a strong noise environment of -10dB, the proposed method still maintained an accuracy of 84% on the CWRU dataset, and 100% accuracy could be restored with a signal-to-noise ratio of 0dB. Ablation experiments confirmed the necessity of the collaboration of each module. The accuracy of the time-domain stream alone on the CWRU and AUST datasets was 98.17% and 98.67%, respectively. Adding the time-frequency edge stream improved the accuracy to 99.83% and 99.17%, respectively. After further integrating the DCIG mechanism, the accuracy reached 100% and 99.67%, respectively, verifying the absolute advantage of dynamic edge perception and dual-stream cascading for complex fault feature extraction. Through the above experiments, the fault diagnosis effect of the proposed method has been further verified.

Claims

1. A dual-stream fault diagnosis method based on state-space modeling and dynamic edge context awareness, characterized in that, The method includes the following steps: (1) Construct the temporal state flow, perform state space modeling under the Transformer framework to form the MambaT encoder, and perform long temporal modeling with linear time complexity. Specifically, the input sequence is normalized by LayerNorm, linearly mapped, one-dimensional convolution, and then the time step, input gain and output gain are generated, selective state is recursively updated, and residual connection and feedforward calculation are performed to obtain the temporal state flow output. (2) Construct a time-frequency edge stream, design a differentiable edge-sensitive mask to adaptively separate the impact features of the time-frequency image from the background noise, and design boundary content-aware adaptive filling to maintain the topological integrity of the image boundary. The two work together to form a dynamic edge context-aware convolution. Specifically, the input feature map is processed by channel mean processing, Sobel convolution kernel gradient calculation, gradient magnitude calculation, smoothness mask generation and binarization, then boundary slice extraction and boundary generation are performed, and dynamic branch selection is performed between the main branch full receptive field convolution and the side branch lightweight projection convolution according to the mask to obtain the time-frequency edge stream output. (3) Construct a channel enhancement and channel gating module to establish a dynamic channel interaction gating mechanism, generate feature enhancement vectors and dynamic gating weights, adaptively cascade dual-stream information, and realize the joint representation of time-domain and time-frequency domain bidirectional enhancement. Specifically, the output features of the time-domain state stream and the output features of the time-frequency edge stream are concatenated by channel splicing. Then, the spliced ​​features are randomly shuffled by channel and spliced ​​again. Then, the channel enhancement vectors and dynamic gating weights are generated by parallel channel enhancement and channel gating modules respectively. Finally, the cascaded output is obtained by gating fusion and fault discrimination is performed.

2. The dual-stream fault diagnosis method based on state-space modeling and dynamic edge context awareness according to claim 1, characterized in that... The construction of the temporal state flow described in step (1), the state space modeling under the Transformer framework to form the MambaT encoder, and the long temporal modeling with linear time complexity to accurately represent temporal sensitive features are carried out as follows: (2a) Construction of dynamically parameterized state space For upper-layer output or embedded vectors ,in For batch size, For sequence length, As the feature dimension, it is first obtained by applying LayerNorm normalization. Then use linear mapping to Projection into two equal-dimensional vectors : in For internal channel dimensions, For the weights and biases of the linear mapping, then... Do it in the time dimension One-dimensional convolution, kernel width Each channel is independent and is obtained through SiLU activation. : After that Dividing the data into three channels through another linear layer, we obtain the time step. Input gain Output gain : in It is the low-rank dimension used to generate the step size. For linear layer weights and biases, It involves modeling the hidden state dimension in the state space, and then using a biased linear mapping and Softplus activation to... Actual step size mapped to each channel The state transition matrix is ​​parameterized in its logarithmic form as The actual matrix is ; (2b) Selective state recursive update For each time step And update the hidden state of all channels. : in If we represent the dot product operation, then the SSM output is: in For jump connection gain, then the output of the entire sequence. and By performing channel-wise SiLU gating and then passing the output through a projection layer to map it back to the original dimension, we obtain the output modeled in state space. : in This indicates a multiplication operation. For projection layer weights and biases; (2c) Residual connection and feedforward calculation The result of state-space modeling After Dropout and residual connection to the original input Add: in This indicates that the first iteration involves random discarding, followed by LayerNorm, and then two layers of feedforward networks consisting of linear transformations and ReLU activations. in and This is the weight matrix. The bias vector is used as the final output of the MambaT encoder, obtained by combining the second residual with Dropout. : in This indicates a second random discard.

3. The dual-stream fault diagnosis method based on state-space modeling and dynamic edge context awareness according to claim 1, characterized in that... Step (2) involves constructing a time-frequency edge stream, designing a differentiable edge-sensitive mask to adaptively separate time-frequency image impact features from background noise, and simultaneously designing boundary content-aware adaptive padding to maintain the topological integrity of image boundaries. The two work together to form a dynamic edge context-aware convolution, effectively extracting time-frequency edge features. The steps are as follows: (3a) Design a differentiable edge-sensitive mask For the input feature map The grayscale image is obtained by averaging along the channel dimension. : in For the number of channels, The spatial dimensions are defined; then, a Sobel convolution kernel with fixed weights is used. Calculate horizontal and vertical gradients : in This represents a convolution operation, and then the gradient magnitude of each pixel is calculated. : Then through a learnable gain scalar With offset scalar And using the element-wise sigmoid function Map it to a smoothness mask : To further simplify branch selection, this method uses a smoothness mask. Binarization to indicator function form: in , For indicator functions, when The value is 1 if it is true, and 0 otherwise. Finally, the binary mask Nearest neighbor interpolation to size And copied along the channel ; (3b) Design boundary content-aware adaptive filling First, extract the image boundary slices, and set the fill scale to... Take the leftmost and rightmost parts of the image. Column boundaries: in The leftmost and rightmost boundary slices are extracted respectively, and then the boundary slices are input into the boundary generator. middle: in Fill the generated boundaries with content; then horizontally stitch them together to obtain the intermediate extended image. : Similarly, generate top and bottom boundary fill content in the vertical direction. Vertically splice it to the middle extended image The final size is obtained as follows: Boundary content-aware adaptive filling feature map : (3c) Design a dynamic branch selection mechanism Dynamic edge context-aware convolutions rely on differentiable edge-sensitive masks in the full receptive field of the main branch. Convolution and Bypass Branch Dynamic switching between lightweight projective convolutions; the main branch uses the input after boundary-aware adaptive padding. Execution Convolution yields the output : in Number of output channels Output space size; bypass branch at original input Execution Lightweight projective convolution yields the output : If step size The bypass branch then goes through max pooling. Operation downsampling to To match the spatial dimensions; then, dynamic branch selection is performed based on a differentiable edge-sensitive mask to obtain the final output. : In this way, dynamic edge context-aware convolution focuses on preserving the details of the full receptive field convolution in high-gradient edge regions, while relying on lightweight projection branches in low-gradient smooth regions to reduce redundant computation and noise impact. Ultimately, it achieves accurate separation of transient impact features from background noise in fault time-frequency images, and completes the extraction of transient impact features from time-frequency edge flows.

4. The dual-stream fault diagnosis method based on state-space modeling and dynamic edge context awareness according to claim 1, characterized in that... Step (3) involves constructing a channel enhancement and channel gating module to establish a dynamic channel interaction gating mechanism, generating feature enhancement vectors and dynamic gating weights, adaptively cascading dual-stream information, and realizing joint representation of time-domain and time-frequency domain bidirectional enhancement, significantly improving fault discrimination capability. The steps are as follows: (4a) Channel disturbance and initial splicing The long-range time-domain features output from the time-domain state stream and the time-frequency image features output from the time-frequency edge stream are concatenated in the channel dimension to obtain the feature vector. The channel dimensions of each sample are randomly shuffled: To construct a "channel perspective" different from the original features, and then use the original features... With scrambled features By splicing along the channel dimension, we obtain... ; (4b) Channel enhancement and channel gating Channel enhancement vectors are generated through parallel channel enhancement and channel gating modules. and dynamic gating weights : in Used to extract enhanced information, Used to generate the gating weights for each channel. Represents an element-wise sigmoid mapping; (4c) Gated fusion yields cascaded output The enhanced vector and the original features are adaptively concatenated by channel using gated fusion to obtain the fused output. : It can dynamically decide whether to extract supplementary information or retain the original features on each channel, realizing the cascaded enhancement and complementary fusion of time-domain and time-frequency domain features, thereby significantly improving the fault discrimination capability and model robustness of the final joint representation.

Images