Cross-machine fault diagnosis method based on laplace renormalization and manifold causal mamba
By employing Laplace renormalization and manifold causality Mamba, the problems of sensor ill-conditioned dependence and domain offset in cross-machine fault diagnosis are solved, achieving high-precision cross-machine fault diagnosis, especially with significant diagnostic performance under conditions of few samples.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- YANSHAN UNIV
- Filing Date
- 2026-04-29
- Publication Date
- 2026-07-24
AI Technical Summary
In cross-machine fault diagnosis tasks, existing methods suffer from problems such as sensor ill-conditioning, difficulty in multi-sensor fusion, and low fault migration accuracy due to domain offset. In particular, when different machines and equipment face complex physical environment changes, it is difficult to achieve high-precision fault diagnosis.
A cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba is adopted. Multi-sensor signal fusion is performed through fast Fourier transform, random masking and high-order Laplace renormalization algorithm. Combined with saliency mixing extraction network and manifold causal routing mechanism, a combined loss function is constructed for joint optimization to realize cross-machine fault diagnosis.
It reduces the ill-conditioned dependence on a single sensor, improves cross-domain diagnostic accuracy under few-sample conditions, enhances the generalization and transfer accuracy of the model across machine fault types, and can effectively capture high-energy impact signals and perform accurate fault diagnosis.
Smart Images
Figure CN122451633A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the interdisciplinary field of industrial equipment health management and artificial intelligence, specifically to a cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba. Background Technology
[0002] Rotating machinery is a core piece of equipment in modern industrial production. While data-driven fault diagnosis technology based on deep learning has made significant progress, in real-world industrial scenarios, different machines face extremely complex changes in their physical environment. Due to the drastic fluctuations in rotational speed and load across different machines, as well as the differences in the placement and sensitivity of sensors on different devices, the collected multidimensional monitoring data suffers from severe domain shift.
[0003] Currently, in cross-machine fault diagnosis tasks, the source and target machines often face the dilemma of incomplete alignment of fault types and insufficient or even zero-sample labeling. Existing methods have the following significant drawbacks: First, traditional multi-sensor fusion often only uses channel splicing, and once a sensor experiences local failure or is mixed with strong background noise, the entire network is prone to falling into ill-conditioned local optima. Second, standard visual transformers (ViT) or convolutional neural networks (CNNs) struggle to dynamically focus on key high-frequency impact signals in extremely long sequences. Third, most existing prototypical networks rely solely on a single Euclidean distance for classification, ignoring the manifold topology and causal relationships in the feature space, resulting in poor model generalization performance and persistent transfer difficulties.
[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention
[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a cross-machine fault diagnosis method based on Laplace renormalization and manifold causality Mamba. This method has the advantages of low ill-conditioning dependence on a single sensor and high cross-domain diagnostic accuracy under conditions of few samples. It thus solves the problems of difficult cross-machine multi-sensor data fusion and low fault migration accuracy due to domain offset from the source.
[0006] (II) Technical Solution To achieve the advantages of low pathological dependence on a single sensor and high cross-domain diagnostic accuracy under conditions of few samples, the specific technical solution adopted in this invention is as follows: According to a first aspect of the present invention, a cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba is provided, the method comprising: S1. Perform Fast Fourier Transform and data augmentation on the acquired one-dimensional time series signal to obtain the enhanced frequency domain signal; perform masking on the enhanced frequency domain signal, and use a high-order Laplace renormalization algorithm to perform multi-sensor dynamic fusion on the masked frequency domain signal to obtain the fused frequency domain signal. S2. The fused frequency domain signal is input into a saliency hybrid extraction network containing a state-space model to extract local and global deep features of the sequence, resulting in global dense features and unpooled features. Based on the global dense features and unpooled features, a class prototype for few-shot tasks is constructed. The manifold causal routing mechanism is used to calculate the manifold probability distribution of the query samples for each fault category. The unpooled features are then amplified and mapped using a lightweight decoder to decode and restore the reconstructed frequency domain signal. S3. Using a combined loss function consisting of manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss constructed based on the reconstructed frequency domain signal, the manifold probability distribution of the saliency hybrid extraction network is jointly optimized; the final manifold probability distribution is calculated using the jointly optimized saliency hybrid extraction network, and the fault diagnosis results of the target machine are output.
[0007] Preferably, the step of performing Fast Fourier Transform and data augmentation processing on the acquired one-dimensional time series signal to obtain an enhanced frequency domain signal; performing masking processing on the enhanced frequency domain signal, and using a high-order Laplace renormalization algorithm to perform multi-sensor dynamic fusion on the masked frequency domain signal to obtain a fused frequency domain signal includes the following steps: S11. Perform a one-dimensional fast Fourier transform on the one-dimensional time series signal collected by each sensor; take the modulus of the one-dimensional fast Fourier transform result to obtain the amplitude spectrum, and perform normalization and truncation processing to obtain the initial frequency domain signal. S12. Use a sliding window to traverse the initial frequency domain signal, perform data augmentation processing including frequency domain elastic deformation and amplitude jitter, and generate an enhanced frequency domain signal. S13. Use a random masking mechanism to mask the enhanced frequency domain signal to obtain the masked frequency domain signal; S14. Construct a causal diffusion matrix based on the high-order Laplace matrix of the multi-sensor topology and the learnable diffusion weights. Use the causal diffusion matrix to perform causal diffusion modulation on the mask frequency domain signal and output the fused frequency domain signal.
[0008] Preferably, the step of masking the enhanced frequency domain signal using a random masking mechanism to obtain the masked frequency domain signal includes the following steps: S131. For the input enhanced frequency domain signal, generate a standard pure random noise matrix corresponding to the length of the frequency domain signal sequence; S132. Sort the standard pure random noise matrix in ascending order to obtain the corresponding position index, and truncate the position index of the first proportion as the occlusion area, and the remaining position index as the reserved area. S133. Using a Boolean substitution algorithm, the occluded region is assigned a value of zero, and the retained region is assigned a value of one to generate a binary mask matrix. The binary mask matrix is then subjected to a Hadamard product operation with the enhanced frequency domain signal to obtain the mask frequency domain signal.
[0009] Preferably, the construction of a causal diffusion matrix based on a high-order Laplacian matrix and learnable diffusion weights using the multi-sensor topology, and the subsequent causal diffusion modulation of the masked frequency domain signal using the causal diffusion matrix to output the fused frequency domain signal, includes the following steps: S141. To address the coupling failure mechanism of multiple sensors, a high-order Laplace matrix is constructed. S142. Introduce the diffusion coefficient that the network can learn dynamically, and combine it with the identity matrix and the higher-order Laplacian matrix to calculate the causal diffusion matrix. S143. Multiply the transpose of the mask frequency domain signal with the causal diffusion matrix to obtain the smoothed intermediate features; S144. Perform one-dimensional convolution embedding processing on the smoothed intermediate features, expand the number of channels of the smoothed intermediate features to the preset feature dimension, and sequentially perform instance normalization and corrected linear unit nonlinear activation processing to output the fused frequency domain signal.
[0010] Preferably, the step of inputting the fused frequency domain signal into a saliency fusion extraction network containing a state-space model to extract local and global deep features of the sequence, and obtaining global dense features and unpooled features, includes the following steps: Local texture features of the fused frequency domain signal are extracted by local convolution and activation functions; adaptive average pooling is applied to the local texture features, and the pooling result is input into a two-layer fully connected network to generate channel attention weight vectors; the channel attention weight vectors are multiplied element-wise with the local texture features to obtain weighted local features. Calculate the local energy scalar of the fused frequency domain signal along the channel dimension at each time step, and obtain the reordering index matrix by sorting them in descending order of energy; use the reordering index matrix to reorder the fused frequency domain signal in time steps to obtain the reordered signal with high energy feature priority. The rearranged signal is sequentially input into the layer normalization and Mamba state space model to extract global long-range dependency features; the global long-range dependency features are restored in time step by time using the inverse mapping of the rearranged index matrix; the global features restored in time step are added with the fused frequency domain signal and the weighted local features by residual addition to obtain the enhanced output features. The enhanced output features are adaptively pooled and flattened to obtain a globally dense feature vector; and the unpooled 3D sequence features in the last layer are taken as unpooled features.
[0011] Preferably, the step of constructing a class prototype for a few-shot task based on globally dense features and unpooled features, and calculating the manifold probability distribution of query samples for each fault category using a manifold causal routing mechanism includes the following steps: For each fault category, aggregate the global dense features of the corresponding support set and calculate the class prototype of the current fault category; use exponential moving average to fuse the class prototype with the corresponding historical round information to obtain the updated global class prototype. Calculate the Euclidean distance matrix between the global dense features of the query sample and the updated global class prototype, and transform the Euclidean distance matrix into the initial similarity distribution through negative exponential Softmax mapping; Calculate the Euclidean distance matrix between each global class prototype and introduce an infinite diagonal mask matrix to construct the prototype manifold adjacency matrix. The initial similarity distribution is multiplied by the adjacency matrix of the prototype manifold to obtain the modulated similarity distribution. The modulated similarity distribution and the initial similarity distribution are then weighted and summed using hyperparameters to obtain the manifold probability distribution.
[0012] Preferably, the construction of the combined loss function includes the following steps: The manifold routing loss is obtained by calculating the negative log-likelihood between the manifold probability distribution and the true label. The initial similarity distribution of the query samples to each type of prototype is multiplied by the prototype manifold adjacency matrix between the prototypes to generate the causal intervention probability distribution. The classification cross-entropy is calculated based on the causal intervention probability distribution and the true label to obtain the causal intervention loss. L2 normalization is performed on the global dense features of the query sample and the updated global class prototype, respectively. The cosine similarity matrix with temperature coefficient between the global dense features of the query sample and the updated global class prototype is calculated, and the cosine similarity matrix is transformed into feature discrimination loss. Based on the manifold probability distribution, the prediction entropy of each sample prediction result and the global marginal distribution entropy of all categories are calculated, and the sum of the prediction entropy and the global marginal distribution entropy is used as the confidence entropy loss. For the positions marked as occlusion areas in the mask matrix, the corresponding enhanced frequency domain signals and the reconstructed frequency domain signals output by the lightweight decoder are extracted, normalized, and then the mean square error is calculated to obtain the MAE reconstruction penalty loss. The combined loss function is obtained by weighting and summing the manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss according to preset weights.
[0013] According to a second aspect of the present invention, a cross-machine fault diagnosis system based on Laplace renormalization and manifold causal Mamba is provided, the system comprising: The frequency domain masking and Laplace topology fusion module is used to perform fast Fourier transform and data enhancement processing on the acquired one-dimensional time series signal to obtain the enhanced frequency domain signal; the enhanced frequency domain signal is masked, and a high-order Laplace renormalization algorithm is used to perform multi-sensor dynamic fusion on the masked frequency domain signal to obtain the fused frequency domain signal; The energy rearrangement Mamba and manifold causal inference modules are used to input the fused frequency domain signal into a saliency hybrid extraction network containing a state-space model to extract local and global deep features of the sequence, obtaining global dense features and unpooled features. Based on the global dense features and unpooled features, a class prototype for few-shot tasks is constructed, and a manifold causal routing mechanism is used to calculate the manifold probability distribution of query samples for each fault category. Finally, a lightweight decoder is used to amplify and map the unpooled features, and the reconstructed frequency domain signal is obtained by decoding and restoring them. The five-element manifold regularization loss optimization module is used to jointly optimize the manifold probability distribution of the saliency hybrid extraction network using a combined loss function consisting of manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss constructed based on the reconstructed frequency domain signal. The module then uses the jointly optimized saliency hybrid extraction network to calculate the final manifold probability distribution and outputs the fault diagnosis results of the target machine.
[0014] According to a third aspect of the present invention, a computer device is provided.
[0015] In some embodiments, the computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to implement the steps of the method described above.
[0016] According to a fourth aspect of the present invention, a computer-readable storage medium is provided.
[0017] In one embodiment, a computer program is stored on the computer-readable storage medium, which, when executed by a processor, implements the steps of the above method.
[0018] (III) Beneficial Effects Compared with existing technologies, this invention provides a cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba, which has the following beneficial effects: (1) This invention constructs a signal fusion architecture based on FFT, random masking mechanism and higher-order Laplace renormalization. Random masking abandons the traditional energy guidance and forces the model to learn redundant physical features with 50% feature loss, which greatly reduces the ill-conditioned dependence on a single sensor; the higher-order Laplace diffusion matrix completes the physical smoothing and dynamic weight allocation of multi-dimensional signals through the principle of manifold graph theory, which effectively resists the attenuation of sensor sensitivity.
[0019] (2) This invention proposes a saliency hybrid block that deeply integrates local CNN and global Mamba models. In particular, the introduced sequence energy reordering mechanism enables the Mamba state space equation to prioritize the calculation and propagation of effective features carrying high-energy impact signals, achieving extremely fast causal feature capture across thousands of time steps while ensuring linear computational complexity.
[0020] (3) A manifold causal network and a five-element combination loss function were designed. By calculating the manifold adjacency matrix and causal multiplication distribution of the class prototype, the limitations of pure Euclidean distance classification were overcome. At the same time, MAE signal reconstruction was used as a regularization penalty term, and in conjunction with multi-objective optimization such as manifold, causality, contrast and entropy, the generalization transfer accuracy of the model under different fault types (few samples) across machines was greatly improved. Attached Figure Description
[0021] 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.
[0022] Figure 1 This is a flowchart illustrating the overall framework of a cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the high-order Laplace renormalization and random mask multi-sensor fusion mechanism according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the significant mixed block and state space sequence energy reordering mechanism according to an embodiment of the present invention; Figure 4 These are the confusion matrix and T-SNE visualization results of the cross-machine migration task according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation
[0023] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.
[0024] According to embodiments of the present invention, a cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba is provided.
[0025] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, according to a first embodiment of the present invention, a cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba is provided, the method comprising: S1. Based on multiple sensors pre-deployed on different rotating machines, acquire one-dimensional time-series signals reflecting the operating status of the machinery; perform fast Fourier transform and data enhancement processing on the one-dimensional time-series signals to obtain enhanced frequency domain signals; use a random masking mechanism to mask the enhanced frequency domain signals, and use a high-order Laplace renormalization algorithm (HoLa Signal Renorm) to dynamically fuse the masked frequency domain signals from multiple sensors to obtain fused frequency domain signals; In a preferred embodiment, the step of acquiring a one-dimensional time-series signal reflecting the operating status of mechanical equipment based on multiple sensors pre-deployed on different rotating machines; performing a fast Fourier transform and data enhancement processing on the one-dimensional time-series signal to obtain an enhanced frequency domain signal; masking the enhanced frequency domain signal using a random masking mechanism; and performing multi-sensor dynamic fusion of the masked frequency domain signal using a high-order Laplace renormalization algorithm to obtain a fused frequency domain signal includes the following steps: S11. Perform a one-dimensional fast Fourier transform on the one-dimensional time series signal collected by each sensor; take the modulus of the one-dimensional fast Fourier transform result to obtain the amplitude spectrum, and perform normalization and truncation processing to obtain the initial frequency domain signal. Specifically, assuming the original one-dimensional time series sample segment captured by multiple sensors (such as X, Y, and Z axes) is... Where B is the batch size and C=3 is the number of sensor channels. This represents the original number of sampling points. First, a one-dimensional Fast Fourier Transform (1D-FFT) is performed, calculated using the following formula: ; In the formula, For the first The amplitude corresponding to each frequency index This represents the original number of sampling points. For discrete sampling point index, For frequency domain indexing, The base of the natural logarithm, The imaginary unit, This is a mold taking operation.
[0026] Based on the Nyquist sampling theorem and spectral symmetry, the first half is extracted to form the fundamental frequency domain signal. In the formula, This represents the truncated initial frequency domain signal tensor (i.e., the initial frequency domain signal). Represents the set of real numbers; Indicates the batch size when inputting data from the network. This indicates the number of channels for multiple sensors (e.g., a triaxial sensor). ); This indicates the length of the effective frequency domain sequence retained after truncating the first half. This represents the number of sampling points of the input one-dimensional original time series, and satisfies... .
[0027] S12. Use a sliding window to traverse the initial frequency domain signal, perform data augmentation processing including frequency domain elastic deformation and amplitude jitter, and generate an enhanced frequency domain signal. Specifically, during the training phase, frequency-domain elastic deformation is performed to simulate fluctuations in machine rotational speed. A uniformly distributed frequency scaling factor is generated. In the formula, This is the frequency scaling factor; Mathematical symbols representing a specific probability distribution; Indicates a continuous and uniform distribution; To control the fluctuation hyperparameter of frequency domain deformation intensity (preferred) ); and These represent the lower and upper limits of the values for the continuous uniform distribution, respectively.
[0028] Then, the length after deformation is calculated. For the initial frequency domain signal Perform linear interpolation mapping function along the length dimension : ; In the formula, This is the frequency domain vector after deformation. It is a linear interpolation mapping function. The initial frequency domain signal, The length of the deformed sequence.
[0029] like Then add to the end of the sequence. A zero; if Then cut off the front Each point. Then, amplitude jitter is performed to generate a perturbation matrix. Amplified signal obtained : ; In the formula, This is the random disturbance matrix simulating amplitude fluctuations; This indicates that the elements in the matrix follow a continuous uniform distribution with a lower limit of 0.5 and an upper limit of 1.5; Represents the set of real numbers; Indicates batch size; Represents the broadcast coefficient in terms of sequence channels and length; The enhanced frequency domain signal matrix; This is an element-wise multiplication operation.
[0030] S13. Use the Random Masking (MAE) mechanism to mask the enhanced frequency domain signal to obtain the masked frequency domain signal; In a preferred embodiment, the step of masking the enhanced frequency domain signal using a random masking mechanism to obtain a masked frequency domain signal includes the following steps: S131. For the input enhanced frequency domain signal, generate a standard pure random noise matrix corresponding to the length of the frequency domain signal sequence; S132. Sort the standard pure random noise matrix in ascending order to obtain the corresponding position index, and truncate the position index of the first proportion as the occlusion area, and the remaining position index as the reserved area. S133. Using a Boolean substitution algorithm, the occluded region is assigned a value of zero, and the retained region is assigned a value of one to generate a binary mask matrix. The binary mask matrix is then subjected to a Hadamard product operation with the enhanced frequency domain signal to obtain the mask frequency domain signal.
[0031] Specifically, such as Figure 2 The diagram shown illustrates the principle of a high-order Laplace renormalization and random masking multi-sensor fusion mechanism according to an embodiment of the present invention. Unlike traditional high-energy filtering-based methods, this invention aims to completely eliminate ill-conditioned dependencies and generate a pure random noise matrix that follows a standard uniform distribution. And set In the formula, The total number of spectral blocks after the frequency domain sequence is equally divided; The effective frequency domain sequence length retained after the aforementioned truncation; The size of the number of sequence points contained in a single spectral block is a preset value. This is the generated pure random floating-point noise matrix; Represents the set of real numbers; This is the batch size. Subsequently, the pure random noise matrix is sorted in ascending order along the sequence dimension. Get Index : ; In the formula, This is the index matrix of the corresponding positions obtained in ascending order. It is a pure random floating-point noise matrix. This is a function for ascending order sorting.
[0032] Set occlusion rate , cut off Construct a binary mask matrix using indexes. In the formula, The calculated number of spectrum blocks that need to be forcibly blocked; This is the preset mask occlusion rate parameter; This indicates a round-down operation; It is a binary mask matrix consisting of 0s and 1s; This represents the number of sensor channels. The area covered by the selected index (i.e., the occluded area) is assigned a value of 0, while the remaining areas (i.e., the reserved areas) are assigned a value of 1. This yields the mask frequency domain signal. : ; In the formula, To force the frequency domain representation after partial feature loss (i.e., masked frequency domain signal). For the enhanced signal, It is a binary mask matrix consisting of 0s and 1s.
[0033] S14. Construct a causal diffusion matrix based on the high-order Laplace matrix of the multi-sensor topology and the learnable diffusion weights. Use the causal diffusion matrix to perform causal diffusion modulation on the mask frequency domain signal and output the fused frequency domain signal.
[0034] In a preferred embodiment, the construction of a causal diffusion matrix based on a high-order Laplacian matrix and learnable diffusion weights using a multi-sensor topology, and the causal diffusion modulation of the masked frequency domain signal using the causal diffusion matrix to output the fused frequency domain signal, includes the following steps: S141. To address the coupling failure mechanism of multiple sensors, a high-order Laplace matrix is constructed. S142. Introduce the diffusion coefficient that the network can learn dynamically, and combine it with the identity matrix and the higher-order Laplacian matrix to calculate the causal diffusion matrix. S143. Multiply the transpose of the mask frequency domain signal with the causal diffusion matrix to obtain the smoothed intermediate features; S144. Perform one-dimensional convolution embedding processing on the smoothed intermediate features, expand the number of channels of the smoothed intermediate features to the preset feature dimension, and sequentially perform instance normalization and corrected linear unit nonlinear activation processing to output the fused frequency domain signal.
[0035] Specifically, to address coupling failures between sensors, a physical topological Laplace matrix is constructed. Since the three-axis sensors are in the same mechanical physical field, the initial state matrix is defined as follows: ; In the formula, This is a high-order Laplace matrix describing the physical topology of a three-node sensor.
[0036] Introducing a dynamically learnable diffusion coefficient for the network (Initialize to 0.5), calculate the causal diffusion matrix of the multidimensional signal. : ; In the formula, The calculated causal diffusion matrix, It is a unit diagonal matrix. The diffusion coefficient represents the dynamic learning of the network. It is a higher-order Laplace matrix.
[0037] The mask signal is rearranged in dimensionality and multiplied in feature space with the diffusion matrix to achieve physical smoothing and coupling. ; In the formula, To achieve multidimensional physical smoothing and the intermediate features after coupling, For mask signal, This is a matrix transpose operation.
[0038] Finally, a one-dimensional convolutional embedding operation is used to expand the number of channels of the input features to the preset model feature dimension. Subsequently, instance normalization and nonlinear activation processing are performed to obtain the fused deep multi-sensor features. The calculation formula is: ; In the formula, To extract the hidden layer feature channel dimension (i.e. model feature dimension) in the main architecture of the network for saliency mixing. The fused features (i.e., the fused frequency domain signal) are processed. These are the parameters of a one-dimensional convolution kernel; This is the convolution operator; These are intermediate features after the aforementioned physical smoothing and coupling are completed; and These are the mean and variance of the instance sequence along the channel dimension, respectively. and For network-learnable affine transformation scaling and translation parameters; To prevent the denominator from being divided by zero by tiny floating-point constants; To correct the nonlinear activation function of the linear unit.
[0039] S2. The fused frequency domain signal is input into a saliency hybrid extraction network containing a state-space model to extract local and global deep features of the sequence, resulting in global dense features and unpooled features. Based on the global dense features and unpooled features, a class prototype for few-shot tasks is constructed, and the manifold probability distribution of the query sample for each fault category is calculated using a manifold causal routing mechanism. The unpooled features are amplified and mapped using a lightweight decoder, and the reconstructed frequency domain signal is obtained by decoding and restoration. In a preferred embodiment, the step of inputting the fused frequency domain signal into a saliency fusion extraction network containing a state-space model to extract local and global deep features of the sequence, and obtaining global dense features and unpooled features, includes the following steps: Local texture features of the fused frequency domain signal are extracted by local convolution and activation functions; adaptive average pooling is applied to the local texture features, and the pooling result is input into a two-layer fully connected network to generate channel attention weight vectors; the channel attention weight vectors are multiplied element-wise with the local texture features to obtain weighted local features. Specifically, the saliency hybrid extraction network consists of alternating cascades of a base downsampling layer and multiple saliency hybrid blocks. For features input to the hybrid blocks... First, local textures are obtained through a local convolution module. : ; In the formula, For the extracted local texture features, Given the input feature sequence, For depthwise separable convolution weights, For batch normalization operations, This is the activation function.
[0040] Furthermore, through channel attention mechanisms (including adaptive average pooling) With two-layer perceptron Generate dynamic weight vectors And weighted to obtain : ; ; In the formula, For weighted local features, For the generated channel dynamic weight vector, and For activation function, and For the perceptron weights, For adaptive average pooling, This is an element-wise multiplication operation.
[0041] Calculate the local energy scalar of the fused frequency domain signal along the channel dimension at each time step, and obtain the reordering index matrix by sorting them in descending order of energy; use the reordering index matrix to reorder the fused frequency domain signal in time steps to obtain the reordered signal with high energy feature priority. Specifically, such as Figure 3 The diagram illustrates a saliency mixing block and state-space sequence energy reordering mechanism according to an embodiment of the present invention. To address the sequential forgetting defect of the state-space model (Mamba), this invention presents a unique sequence energy reordering mechanism. The mechanism calculates the energy reordering at each time step in the sequence. Local energy scalar along the channel dimension : ; In the formula, For the first The local energy scalar at each time step The input values are the specific dimensions of the feature tensor.
[0042] Obtain the index matrix by sorting in descending order of energy. Using this index, a rearrangement mapping is performed on the original input along the sequence dimension to obtain... : ; In the formula, It is the rearranged high-energy front sequence (i.e., the rearranged signal with high-energy features as the priority). This is the reordered index matrix obtained by sorting in descending order of energy.
[0043] The rearranged signal is sequentially input into the layer normalization and Mamba state space model to extract global long-range dependency features; the global long-range dependency features are restored in time step by time using the inverse mapping of the rearranged index matrix; the global features restored in time step are added with the fused frequency domain signal and the weighted local features by residual addition to obtain the enhanced output features. The enhanced output features are adaptively pooled and flattened to obtain a globally dense feature vector; and the unpooled 3D sequence features in the last layer are taken as unpooled features.
[0044] Specifically, the rearranged high-energy front sequence is fed into the layer for normalization. Extracting global long-range dependency features using the Mamba model. : ; In the formula, These are long-range causal dependency features extracted from the reordered state. For state-space model operation functions, This is a layer normalization operation.
[0045] Finally, the inverse mapping array of descending indices is used. By restoring the time steps, we obtain the global features. And perform residual summation to obtain : ; ; In the formula, To obtain the long-range causal global feature matrix after time step restoration, A function to obtain the inverse mapped array, This is to perform the output feature after adding the residuals (i.e., the enhanced output feature).
[0046] After deep stacking, the final output sequence is adaptively pooled and flattened to obtain a dense global feature vector for few-sample classification. (For query sets) and (For the support set). Simultaneously, the unpooled final layer 3D sequence features are denoted as... .
[0047] In addition, to meet the requirement of algorithm structure transparency, the specific architecture, inter-layer connections and core parameters of the saliency extraction network (Backbone) that deeply integrates local CNN and global Mamba model of this invention are fully disclosed. The details are now explained in Table 1 (taking an input sequence of length 1024 as an example).
[0048] Table 1: Saliency Hybrid Extraction Network Architecture and Parameter Configuration Table As a preferred implementation, the step of constructing a class prototype for a few-sample task based on global dense features and unpooled features, and calculating the manifold probability distribution of query samples for each fault category using a manifold causal routing mechanism includes the following steps: For each fault category, aggregate the global dense features of the corresponding support set and calculate the class prototype of the current fault category; use exponential moving average to fuse the class prototype with the corresponding historical round information to obtain the updated global class prototype. Calculate the Euclidean distance matrix between the global dense features of the query sample and the updated global class prototype, and transform the Euclidean distance matrix into the initial similarity distribution through negative exponential Softmax mapping; Calculate the Euclidean distance matrix between each global class prototype and introduce an infinite diagonal mask matrix to construct the prototype manifold adjacency matrix. The initial similarity distribution is multiplied by the adjacency matrix of the prototype manifold to obtain the modulated similarity distribution. The modulated similarity distribution and the initial similarity distribution are then weighted and summed using hyperparameters to obtain the manifold probability distribution.
[0049] Specifically, in small sample scenarios, each fault category is calculated separately. The current supported set of average prototypes : ; In the formula, Current fault category Support set average prototype (i.e., current fault category) (class prototype) For the sample size, To support set features.
[0050] By incorporating historical round information using exponential moving average (EMA), the global class prototype maintained by the system is smoothed and updated. : ; In the formula, To ensure a smooth update of the global class prototype, The hyperparameter is the translational slip (preferably 0.9).
[0051] Calculate query sample features Euclidean distance matrix between all class prototypes The initial Euclidean similarity distribution matrix is obtained through negative softmax mapping. : ; In the formula, The initial similarity distribution probability matrix, To query the distance matrix between features and prototypes, This represents the total number of categories.
[0052] To explore nonlinear geometric manifolds in the feature space, the internal Euclidean distance matrix between class prototypes is calculated. Introducing an infinite diagonal mask matrix. (Diagonal elements are 1e9, the rest are 0) To mask self-similarity interference, construct the prototypical manifold adjacency matrix. : ; In the formula, For the prototypical manifold adjacency matrix, This is the internal distance matrix between class prototypes. It is an infinite diagonal mask matrix. This is the feature dimension scaling factor.
[0053] By combining the initial distribution and the manifold topology, a final manifold probability distribution matrix with strong generalization is generated by multiplying the probability matrices. Furthermore, a small extremum clamp of 1e-4 is added to prevent logarithmic overflow: ; In the formula, For manifold probability distribution, and This is the fusion adjustment coefficient.
[0054] Simultaneously, a decoder network containing extremely lightweight one-dimensional convolution and upsampling operators is used to process the unpooled data. Feature amplification mapping is performed to decode and reconstruct the frequency domain signal with the same size as the input. .
[0055] S3. Using a combined loss function consisting of manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss constructed based on the reconstructed frequency domain signal, the manifold probability distribution of the saliency hybrid extraction network is jointly optimized; the final manifold probability distribution is calculated using the jointly optimized saliency hybrid extraction network, and the fault diagnosis results of the target machine are output.
[0056] In a preferred embodiment, the construction of the combined loss function includes the following steps: The manifold routing loss is obtained by calculating the negative log-likelihood between the manifold probability distribution and the true label. Specifically, it is calculated based on the standard cross-entropy formula. Approaching real labels Degree: ; In the formula, For manifold routing loss, This is a one-hot encoding indicator. This refers to the batch size.
[0057] The initial similarity distribution of the query samples to each type of prototype is multiplied by the prototype manifold adjacency matrix between the prototypes to generate the causal intervention probability distribution. The classification cross-entropy is calculated based on the causal intervention probability distribution and the true label to obtain the causal intervention loss. Specifically, the aforementioned query distance prototype distribution (i.e., the initial similarity distribution of various prototypes) is utilized. The internal distribution of prototypes (i.e., the adjacency matrix of prototype manifolds between class prototypes). The matrix multiplication of the two is used to obtain the probability distribution of causal intervention. Then, the negative log-likelihood loss for classification is calculated: ; In the formula, For causal intervention loss, This is the probability matrix of the causal intervention distribution.
[0058] L2 normalization is performed on the global dense features of the query sample and the updated global class prototype, respectively. The cosine similarity matrix with temperature coefficient between the global dense features of the query sample and the updated global class prototype is calculated, and the cosine similarity matrix is transformed into feature discrimination loss. Specifically, for and Perform L2 normalization on the channel dimension separately. Calculate the temperature coefficients for both. dot product cosine similarity matrix This forces the query sample features to cluster tightly around their respective prototype cluster: ; In the formula, For feature discrimination loss, The dot product cosine similarity matrix is... This is the normalized temperature coefficient.
[0059] Based on the manifold probability distribution, the prediction entropy of each sample prediction result and the global marginal distribution entropy of all categories are calculated, and the sum of the prediction entropy and the global marginal distribution entropy is used as the confidence entropy loss. Specifically, to enhance the determinism of cross-machine sample decision boundaries, the sum of the prediction entropy of a single sample and the global marginal distribution entropy is calculated: ; In the formula, For confidence entropy loss, This represents the global marginal distribution probability.
[0060] For the positions marked as occlusion areas in the mask matrix, the corresponding enhanced frequency domain signals and the reconstructed frequency domain signals output by the lightweight decoder are extracted, normalized, and then the mean square error is calculated to obtain the MAE reconstruction penalty loss. Specifically, extract the mask matrix Boolean index set marked as 0 (obscured) To prevent the sharp extrema generated by the Fourier transform itself from overwhelming the backpropagation gradient of the network, the real augmentation signal is... and reconstructed signals Execute separately Normalization is performed, followed by calculation of the mean squared error: ; In the formula, Penalty loss for signal mask reconstruction, The set of Boolean indices marked as 0 in the mask matrix. To reconstruct the signal (i.e., reconstruct the frequency domain signal). This is the actual enhanced signal (i.e., the corresponding enhanced frequency domain signal). This is a data normalization function.
[0061] The combined loss function is obtained by weighting and summing the manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss according to preset weights.
[0062] Specifically, the penalty terms with five distinct physical mechanisms are linearly superimposed and fused with hyperparameter coefficients to obtain the final scalar total loss equation used for gradient backpropagation. : ; In the formula, This is the scalar total loss equation. It is particularly important to emphasize that the reconstruction loss weight is constrained to 0.2 to treat it as a low-intensity regularization penalty mechanism. This both forces the backbone network to maintain its global frequency band structure despite memory gaps and avoids gradient interference from the reconstruction task on the mainstream classification task.
[0063] like Figure 4 The diagram shows the confusion matrix and T-SNE visualization results of the cross-machine migration task according to an embodiment of the present invention. After training, for an unknown cross-domain target machine, the network parameters are frozen and all random masking mechanisms and data augmentation operations are stopped. The multi-sensor signals of the target machine are processed by FFT and input into the backbone network to extract feature vectors. The manifold probability is calculated directly using the class prototype solidified in step S22, and the final manifold probability distribution is obtained. The neuron index corresponding to the largest median value is used as the predictive output for cross-machine fault types, thereby achieving high-precision real-time diagnosis with few samples across machines.
[0064] It should be noted that, in combination Figure 4 The visualization results are illustrated in the following figures. On the left, the vertical axis of the confusion matrix represents the true class of the sample (True), and the horizontal axis represents the predicted class (Predicted). The values on the diagonal reflect the distribution of samples that were correctly classified. On the right, the T-SNE feature distribution visualization shows the relative positional components of the high-dimensional deep physical features after nonlinear dimensionality reduction mapping on a two-dimensional plane. These coordinate values have no specific physical dimensions and are only used to intuitively represent the relative distances and topological aggregation structures of different sample classes in the feature manifold space. The category labels 0 to 5 correspond to the six true health states of the target machine's gearbox: Category 0 represents a healthy state; Category 1 represents gear pitting failure; Category 2 represents gear wear failure; Category 3 represents gear tooth loss failure; Category 4 represents gear tooth cracking failure; and Category 5 represents gear tooth breakage failure.
[0065] According to a second embodiment of the present invention, a cross-machine fault diagnosis system based on Laplace renormalization and manifold causal Mamba is provided, the system comprising: The frequency domain masking and Laplace topology fusion module is used to perform fast Fourier transform and data enhancement processing on the acquired one-dimensional time series signal to obtain the enhanced frequency domain signal; the enhanced frequency domain signal is masked, and a high-order Laplace renormalization algorithm is used to perform multi-sensor dynamic fusion on the masked frequency domain signal to obtain the fused frequency domain signal; The energy rearrangement Mamba and manifold causal inference modules are used to input the fused frequency domain signal into a saliency hybrid extraction network containing a state-space model to extract local and global deep features of the sequence, obtaining global dense features and unpooled features. Based on the global dense features and unpooled features, a class prototype for few-shot tasks is constructed, and the manifold probability distribution of query samples for each fault category is calculated using a manifold causal routing mechanism. Finally, a lightweight decoder is used to amplify and map the unpooled features, and the reconstructed frequency domain signal is obtained by decoding and restoring them. The five-element manifold regularization loss optimization module is used to jointly optimize the manifold probability distribution of the saliency hybrid extraction network using a combined loss function consisting of manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss constructed based on the reconstructed frequency domain signal. The module then uses the jointly optimized saliency hybrid extraction network to calculate the final manifold probability distribution and outputs the fault diagnosis results of the target machine.
[0066] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores static and dynamic information data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements the steps in the above method embodiments.
[0067] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0068] In addition, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.
[0069] In addition, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.
[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.
[0071] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba, characterized in that, The method includes: S1. Perform Fast Fourier Transform and data augmentation on the acquired one-dimensional time series signal to obtain the enhanced frequency domain signal; perform masking on the enhanced frequency domain signal, and use a high-order Laplace renormalization algorithm to perform multi-sensor dynamic fusion on the masked frequency domain signal to obtain the fused frequency domain signal. S2. The fused frequency domain signal is input into a saliency hybrid extraction network containing a state-space model to extract local and global deep features of the sequence, resulting in global dense features and unpooled features. Based on the global dense features and unpooled features, a class prototype for few-shot tasks is constructed. The manifold causal routing mechanism is used to calculate the manifold probability distribution of the query samples for each fault category. The unpooled features are then amplified and mapped using a lightweight decoder to decode and restore the reconstructed frequency domain signal. S3. Using a combined loss function consisting of manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss constructed based on the reconstructed frequency domain signal, the manifold probability distribution of the saliency hybrid extraction network is jointly optimized; the final manifold probability distribution is calculated using the jointly optimized saliency hybrid extraction network, and the fault diagnosis results of the target machine are output.
2. The cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in claim 1, characterized in that, The acquired one-dimensional time series signal is subjected to Fast Fourier Transform and data enhancement processing to obtain an enhanced frequency domain signal; The enhanced frequency domain signal is masked, and a high-order Laplace renormalization algorithm is used to perform multi-sensor dynamic fusion of the masked frequency domain signal to obtain the fused frequency domain signal. The steps include: S11. Perform a one-dimensional fast Fourier transform on the one-dimensional time series signal collected by each sensor. The amplitude spectrum is obtained by taking the modulus of the one-dimensional fast Fourier transform result, and then normalized and truncated to obtain the initial frequency domain signal. S12. Use a sliding window to traverse the initial frequency domain signal, perform data augmentation processing including frequency domain elastic deformation and amplitude jitter, and generate an enhanced frequency domain signal. S13. Use a random masking mechanism to mask the enhanced frequency domain signal to obtain the masked frequency domain signal; S14. Construct a causal diffusion matrix based on the high-order Laplace matrix of the multi-sensor topology and the learnable diffusion weights. Use the causal diffusion matrix to perform causal diffusion modulation on the mask frequency domain signal and output the fused frequency domain signal.
3. The cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in claim 2, characterized in that, The process of masking the enhanced frequency domain signal using a random masking mechanism to obtain the masked frequency domain signal includes the following steps: S131. For the input enhanced frequency domain signal, generate a standard pure random noise matrix corresponding to the length of the frequency domain signal sequence; S132. Sort the standard pure random noise matrix in ascending order to obtain the corresponding position index, and truncate the position index of the first proportion as the occlusion area, and the remaining position index as the reserved area. S133. Using a Boolean substitution algorithm, the occluded region is assigned a value of zero, and the retained region is assigned a value of one to generate a binary mask matrix. The binary mask matrix is then subjected to a Hadamard product operation with the enhanced frequency domain signal to obtain the mask frequency domain signal.
4. The cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in claim 2, characterized in that, The construction of a causal diffusion matrix based on a high-order Laplace matrix and learnable diffusion weights using a multi-sensor topology, and the causal diffusion modulation of the masked frequency domain signal using the causal diffusion matrix to output the fused frequency domain signal, includes the following steps: S141. To address the coupling failure mechanism of multiple sensors, a high-order Laplace matrix is constructed. S142. Introduce the diffusion coefficient that the network can learn dynamically, and combine it with the identity matrix and the higher-order Laplacian matrix to calculate the causal diffusion matrix. S143. Multiply the transpose of the mask frequency domain signal with the causal diffusion matrix to obtain the smoothed intermediate features; S144. Perform one-dimensional convolution embedding processing on the smoothed intermediate features, expand the number of channels of the smoothed intermediate features to the preset feature dimension, and sequentially perform instance normalization and corrected linear unit nonlinear activation processing to output the fused frequency domain signal.
5. The cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in claim 1, characterized in that, The process of inputting the fused frequency domain signal into a saliency fusion extraction network containing a state-space model to extract local and global deep features of the sequence, and obtaining global dense features and unpooled features, includes the following steps: Local texture features of the fused frequency domain signal are extracted by local convolution and activation functions; adaptive average pooling is applied to the local texture features, and the pooling result is input into a two-layer fully connected network to generate channel attention weight vectors; the channel attention weight vectors are multiplied element-wise with the local texture features to obtain weighted local features. Calculate the local energy scalar of the fused frequency domain signal along the channel dimension at each time step, and obtain the reordering index matrix by sorting them in descending order of energy; use the reordering index matrix to reorder the fused frequency domain signal in time steps to obtain the reordered signal with high energy feature priority. The rearranged signal is sequentially input into the layer normalization and Mamba state space model to extract global long-range dependency features; the global long-range dependency features are restored in time step by time using the inverse mapping of the rearranged index matrix; the global features restored in time step are added with the fused frequency domain signal and the weighted local features by residual addition to obtain the enhanced output features. The enhanced output features are adaptively pooled and flattened to obtain a globally dense feature vector; and the unpooled 3D sequence features in the last layer are taken as unpooled features.
6. The cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in claim 1, characterized in that, The process of constructing a class prototype for few-shot tasks based on globally dense features and unpooled features, and calculating the manifold probability distribution of query samples for each fault category using a manifold causal routing mechanism, includes the following steps: For each fault category, aggregate the global dense features of the corresponding support set and calculate the class prototype of the current fault category; use exponential moving average to fuse the class prototype with the corresponding historical round information to obtain the updated global class prototype. Calculate the Euclidean distance matrix between the global dense features of the query sample and the updated global class prototype, and transform the Euclidean distance matrix into the initial similarity distribution through negative exponential Softmax mapping; Calculate the Euclidean distance matrix between each global class prototype and introduce an infinite diagonal mask matrix to construct the prototype manifold adjacency matrix. The initial similarity distribution is multiplied by the adjacency matrix of the prototype manifold to obtain the modulated similarity distribution. The modulated similarity distribution and the initial similarity distribution are then weighted and summed using hyperparameters to obtain the manifold probability distribution.
7. The cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in claim 6, characterized in that, The construction of the combined loss function includes the following steps: The manifold routing loss is obtained by calculating the negative log-likelihood between the manifold probability distribution and the true label. The initial similarity distribution of the query samples to each type of prototype is multiplied by the prototype manifold adjacency matrix between the prototypes to generate the causal intervention probability distribution. The classification cross-entropy is calculated based on the causal intervention probability distribution and the true label to obtain the causal intervention loss. L2 normalization is performed on the global dense features of the query sample and the updated global class prototype, respectively. The cosine similarity matrix with temperature coefficient between the global dense features of the query sample and the updated global class prototype is calculated, and the cosine similarity matrix is transformed into feature discrimination loss. Based on the manifold probability distribution, the prediction entropy of each sample prediction result and the global marginal distribution entropy of all categories are calculated, and the sum of the prediction entropy and the global marginal distribution entropy is used as the confidence entropy loss. For the positions marked as occlusion areas in the mask matrix, the corresponding enhanced frequency domain signals and the reconstructed frequency domain signals output by the lightweight decoder are extracted, normalized, and then the mean square error is calculated to obtain the MAE reconstruction penalty loss. The combined loss function is obtained by weighting and summing the manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss according to preset weights.
8. A cross-machine fault diagnosis system based on Laplace renormalization and manifold causal Mamba, used to implement the cross-machine fault diagnosis method based on Laplace renormalization and manifold causal Mamba as described in any one of claims 1-7, characterized in that, The system includes: The frequency domain masking and Laplace topology fusion module is used to perform fast Fourier transform and data enhancement processing on the acquired one-dimensional time series signal to obtain the enhanced frequency domain signal; the enhanced frequency domain signal is masked, and a high-order Laplace renormalization algorithm is used to perform multi-sensor dynamic fusion on the masked frequency domain signal to obtain the fused frequency domain signal; The energy rearrangement Mamba and manifold causal inference modules are used to input the fused frequency domain signal into a saliency hybrid extraction network containing a state-space model to extract local and global deep features of the sequence, obtaining global dense features and unpooled features. Based on the global dense features and unpooled features, a class prototype for few-shot tasks is constructed, and the manifold probability distribution of query samples for each fault category is calculated using a manifold causal routing mechanism. Finally, a lightweight decoder is used to amplify and map the unpooled features, and the reconstructed frequency domain signal is obtained by decoding and restoring them. The five-element manifold regularization loss optimization module is used to jointly optimize the manifold probability distribution of the saliency hybrid extraction network using a combined loss function consisting of manifold routing loss, causal intervention loss, feature discrimination loss, confidence entropy loss, and MAE reconstruction penalty loss constructed based on the reconstructed frequency domain signal. The module then uses the jointly optimized saliency hybrid extraction network to calculate the final manifold probability distribution and outputs the fault diagnosis results of the target machine.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.