Bearing fault diagnosis method based on physical perception KAM network

Abstract

The invention relates to the field of rotating machinery fault diagnosis, and discloses a bearing fault diagnosis method based on a physical perception KAM network, and the method comprises the steps: carrying out the discretization of a bearing vibration signal through a Gabor filter group based on physical prior initialization, and generating modal feature lexical elements with physical frequency band meanings; and inputting the lexical elements into a PC-KAM backbone network, calculating a hidden state vector by using a state space model branch, and dynamically adjusting the position of a primary function node of a Kolmogov-Arnod network branch to realize collaborative dynamic feature extraction. In the training stage, an orthogonal subspace constraint and physical perception low-rank adaptation fine tuning mechanism is introduced. And finally, searching a historical fault case, performing multi-modal fusion on the historical fault case and the deep feature sequence, mapping a fusion representation into a soft prompt, and inputting the soft prompt into a large language model to generate a diagnosis report. According to the method, the problems of poor physical interpretability of characteristics and few-sample diagnosis under variable working conditions are effectively solved, and the generalization and decision-making ability of a diagnosis system are improved.

Description

Technical Field

[0001] This invention relates to the field of rotating machinery fault diagnosis technology, specifically a bearing fault diagnosis method based on a physical sensing KAM network. Background Technology

[0002] As a key component in rotating machinery, the operating condition of rolling bearings directly affects the safety and reliability of the entire mechanical system. In industrial settings such as wind turbine generators, high-speed trains, and precision manufacturing equipment, real-time condition monitoring and fault diagnosis of bearings are crucial for ensuring stable equipment operation and reducing maintenance costs.

[0003] With the development of industrial internet technology, data-driven methods based on deep learning have become the mainstream research direction in the field of bearing fault diagnosis. Traditional signal processing methods rely on expert experience for manual feature extraction, while deep learning models such as convolutional neural networks, recurrent neural networks, and Transformer architectures can automatically learn potential fault feature representations from massive monitoring data through multi-layer nonlinear transformations. Currently, various improved deep neural network models have been widely used to process vibration acceleration signals, achieving automatic identification and classification of different fault types and damage degrees through end-to-end training.

[0004] However, existing deep neural network models typically use random initialization to set the parameters of the first-layer convolutional kernel or the feature extraction weights when processing bearing vibration signals. This purely data-driven initialization strategy lacks clear support from the physical mechanism of signal processing, resulting in the feature maps extracted by the first layer of the network often lacking clear frequency band physical meaning. It is difficult to ensure that the model pays attention to the key frequency bands containing fault impact components in the early stage of training. Especially in the context of strong noise, the randomly initialized feature extraction layer is easily affected by interference signals, causing the model to lose weak fault impact information in the feature extraction stage, which limits the interpretability of the diagnostic model and the physical accuracy of feature extraction. Summary of the Invention

[0005] This invention aims to address the problems of poor physical interpretability, low accuracy under limited sample conditions, and insufficient utilization of cross-modal knowledge in existing bearing fault diagnosis technologies.

[0006] The first aspect of this invention provides a bearing fault diagnosis method based on a physical sensing KAM network, comprising the following steps:

[0007] Obtain the original vibration acceleration sequence and synchronous speed information of the bearing to be diagnosed.

[0008] The physical sensing signal is discretized. Edge computing nodes use stored physical sensing parameters to construct a Gabor filter bank and perform one-dimensional convolution on the original vibration acceleration sequence, outputting a modal feature word sequence. The physical sensing parameters include the center frequency, statistically determined based on variational mode decomposition results from historical device data. The Gabor filter bank defines the convolution kernel weights based on the product of a Gaussian window function and a cosine function. The scale parameter of the Gaussian window function is set to be inversely proportional to the center frequency, and the frequency parameter of the cosine function is initialized to the center frequency. During training, the center frequency parameter, scale parameter, and phase offset parameter of the Gabor filter bank are treated as learnable variables and adjusted based on gradient updates.

[0009] Cooperative dynamic feature extraction is performed. The modal feature word sequence is input into the PC-KAM backbone network, which includes a parallel state-space model branch and a Kolmogorov-Arnold network branch. The state-space model branch performs a selective scan of the input sequence to calculate the hidden state vector at the current time step, reflecting the signal phase evolution. This hidden state vector is used as a control variable, and the grid node vectors required by the Kolmogorov-Arnold network branch are output through the grid prediction head. The grid node vectors are then sorted to satisfy a monotonically increasing constraint. The Kolmogorov-Arnold network branch uses the grid node vectors to construct a time-varying B-spline basis function to perform nonlinear fitting on the input sequence. The fitting process includes: calculating the response value of the B-spline basis function at the input values, using learnable control coefficients to perform a weighted summation of the response values, and multiplying the summation result by the input values ​​processed by the activation function.

[0010] During model training, a hybrid loss function with orthogonal subspace constraints is constructed. Orthogonal regularization loss is calculated, which constrains the row vectors of the classification weight matrix to remain orthogonal by minimizing the difference between the autocorrelation matrix and the identity matrix of the classification weight matrix, thus projecting different types of fault features into mutually perpendicular feature subspaces. Furthermore, a physically-aware low-rank adaptation mechanism is used to fine-tune the backbone network. While freezing the backbone parameters, a low-rank adaptive bypass containing both dimensionality-reduced and dimensionality-increasing matrices is constructed. When updating the low-rank adaptive bypass parameters, the spectral radius of the new system matrix resulting from the combination of backbone and low-rank parameters is limited to less than a preset threshold, thus constraining the dynamic behavior of the fine-tuned system.

[0011] The process involves performing retrieval-enhanced diagnostic generation. Based on the deep feature sequence output by the PC-KAM backbone network, similar historical fault cases are retrieved from the vector database. The text features of the retrieved historical fault cases are then fused with the deep feature sequence to generate multimodal information. The fusion process includes: using a linear projection layer to map the deep feature sequence and text features into projected signal features and projected text features, respectively; using a multi-head cross-attention mechanism, with the projected text features as the query vector and the projected signal features as the key and value vectors, calculating and aggregating interaction features; and calculating adaptive fusion coefficients through a gating network, using these coefficients to perform a weighted summation of the interaction features and projected text features to output a multimodal fusion representation.

[0012] Generate a diagnostic report. A linear projection layer is used to map the multimodal fusion representation into a soft cue feature sequence within the word embedding space of a large language model. A text cue word sequence is constructed, including role definitions, knowledge context, reasoning tasks, and output constraints. The soft cue feature sequence and the text cue word sequence are concatenated to form the input sequence, and a diagnostic report containing fault conclusions and maintenance suggestions is generated using the large language model.

[0013] A second aspect of the present invention provides a bearing fault diagnosis system based on a physical sensing KAM network, comprising:

[0014] The on-site data acquisition subsystem is configured to acquire the original vibration acceleration sequence and synchronous speed information of the bearing to be diagnosed.

[0015] Edge computing nodes are configured to call stored physical sensing parameters to construct a Gabor filter bank, and use this filter bank to perform a one-dimensional convolution operation on the original vibration acceleration sequence to generate a modal feature word sequence containing frequency band physical meaning.

[0016] The cloud-based analysis and diagnostic platform is configured to load and run the PC-KAM backbone network. Modal feature lexical sequences are input into the PC-KAM backbone network, which utilizes its internal parallel state-space model branches to calculate hidden state vectors. These hidden state vectors are then used to dynamically adjust the basis function node position parameters of the Kolmogorov-Arnold network branches, outputting a deep feature sequence. The platform is also configured to perform retrieval-enhanced diagnostic generation. It retrieves similar historical fault cases and fuses them with the deep feature sequences to generate a multimodal fusion representation. A linear projection layer maps this multimodal fusion representation into a soft-cue feature sequence, which is then input into a large language model to output a diagnostic report.

[0017] This invention provides a bearing fault diagnosis method based on a physical sensing KAM network. It has the following beneficial effects:

[0018] 1. This invention replaces traditional convolutional layers with Gabor filter banks initialized based on variational mode decomposition results of historical equipment data, and sets the filter center frequency and scale parameters as learnable variables. This gives the feature extraction operation of the first layer of the network a clear physical meaning of bandpass filtering, avoiding the problem of missing physical meaning of features caused by random initialization of convolutional kernels. At the same time, the gradient update mechanism is used to adaptively fine-tune the filter frequency band parameters, ensuring that the discretized modal feature word sequence can accurately cover the fault feature frequency band, thereby improving the model's ability to capture early weak fault impact signals and the interpretability of features.

[0019] 2. This invention constructs a backbone architecture that includes a parallel state-space model and a Kolmogorov-Arnold network. By dynamically adjusting the position of the basis function grid nodes of the Kolmogorov-Arnold branch using the hidden state vectors extracted from the state-space model, a dynamic mapping between signal phase evolution and network topology is established. This solves the problem that fixed structures are difficult to adapt to non-stationary signals under variable speed conditions. In conjunction with orthogonal subspace constraints and a physical perception low-rank adaptation mechanism, the fault features are forced to be orthogonally decoupled and the radius of the fine-tuning parameter spectrum is limited, ensuring the generalization performance of the model and the stability of system dynamics in scenarios with few samples.

[0020] 3. This invention uses retrieval enhancement and multimodal soft cue mapping technology to retrieve similar historical fault cases in the vector space and uses a cross-attention mechanism to perform physical and semantic fusion with current signal features. This transforms numerical deep features into soft cue word sequences that can be understood by a large language model, realizing a leap from simple fault classification labels to structured diagnostic reports that include fault mechanism deduction and maintenance suggestions. It makes full use of expert experience in the historical operation and maintenance knowledge base to assist current diagnostic reasoning, improving the logical interpretability and human-computer interaction efficiency of the intelligent diagnostic system. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example:

[0024] Please see the appendix Figure 1This invention provides the hardware architecture and logical composition of a bearing fault diagnosis system based on a physical sensing KAM network and enhanced retrieval. The system adopts an edge-cloud collaborative architecture, including a field data acquisition subsystem, edge computing nodes, and a cloud-based analysis and diagnosis platform.

[0025] The field data acquisition subsystem is mounted on the bearing housing or casing surface of the rotating machinery. This subsystem includes a vibration sensor assembly and a rotational speed measurement assembly. The vibration sensor assembly uses a single-axis or triaxial piezoelectric accelerometer with a frequency response range covering 0Hz to 20kHz. The vibration sensor is connected to the data acquisition card via a shielded cable. The data acquisition card has a built-in analog-to-digital converter (ADC) configured to convert analog vibration signals into digital vibration acceleration sequences at a sampling frequency of at least 51.2kHz. The speed measurement component uses a photoelectric encoder, which is installed at the end of the rotating shaft and configured to collect the instantaneous speed pulse signal of the shaft for subsequent calculation of the theoretical fault characteristic frequency of the bearing under varying operating conditions.

[0026] Edge computing nodes are connected to the field data acquisition subsystem via an industrial fieldbus. Each edge computing node consists of an industrial control computer equipped with an embedded processor. The edge computing node's memory stores pre-calculated physical sensing parameters, including the center frequency determined based on variational mode decomposition results of historical device data. Scale parameters and phase shift The edge computing node is configured to perform discretization operations on the physical sensing signals. Specifically, it calls the physical sensing parameters in memory to construct a Gabor convolution kernel, and uses this kernel to process the real-time acquired digital vibration acceleration sequence. Perform a one-dimensional convolution operation to generate a sequence of modal feature terms that contain the physical meaning of frequency bands. Edge computing nodes will use feature word sequences Synchronous speed information is transmitted to a cloud-based analysis and diagnostic platform via network protocol.

[0027] The cloud-based analytics and diagnostics platform is deployed in a server cluster and is equipped with a graphics processing unit (GPU) acceleration unit. Logically, the platform is divided into a collaborative dynamic feature extraction unit, a physical constraint training unit, and a retrieval-enhanced diagnostic generation unit.

[0028] The Cooperative Dynamic Feature Extraction Unit (CDRSU) is configured to load and run the PC-KAM backbone network. The PC-KAM backbone network contains parallel state-space model branches and Kolmogorov-Arnold Network (KAN) branches. The CDRSU utilizes the GPU to perform forward inference computation, specifically including: processing feature word sequences... The input state-space model branch is used to calculate the hidden state vector that reflects the phase evolution of the signal; this hidden state vector is used as a control variable to dynamically map and adjust the position parameters of the B-spline basis function nodes in the KAN branch; the adjusted KAN branch is used to fit the deterministic interference components in the input sequence; and differential operations are performed to separate the residual feature vector containing fault impacts.

[0029] The physical constraint training unit is configured to calculate theoretical fault characteristic frequencies based on bearing geometry parameters and real-time rotational speed. The physical constraint training unit stores a bearing parameter library, including ball diameters. Pitch circle diameter Number of balls and contact angle The physical constraint training unit is based on rotational speed information. Calculate the frequency of peripheral faults and inner ring failure frequency The physical consistency loss is calculated by combining the power spectral density of the network output features, and the low-rank fitness matrix parameters in the network are updated by the backpropagation algorithm.

[0030] The retrieval-enhanced diagnostic generation unit comprises a vector database server and a large language model inference service. The vector database server stores a high-dimensional feature vector library of historical fault cases. The retrieval-enhanced diagnostic generation unit is configured to perform similarity matching operations, calculating the distance metric between the currently extracted fault feature vector and historical feature vectors in the library, and selecting the most similar historical cases based on the distance metric. The retrieval-enhanced diagnostic generation unit inputs the retrieved case maintenance records and current physical spectrum feature parameters into the large language model. The large language model outputs a diagnostic report containing the fault type, confidence level, and maintenance recommendations. The diagnostic report is presented to the user via a display terminal.

[0031] The bearing fault diagnosis method based on state-guided dynamic KAM network and orthogonal subspace decoupling in this invention includes the following steps:

[0032] Step S100: Discretization of physical sensing signals based on variational mode decomposition prior

[0033] Edge computing nodes acquire the raw vibration acceleration sequence and synchronous rotational speed information of the bearing to be diagnosed. The edge computing nodes then access stored physical sensing parameters, which are statistically derived from variational mode decomposition results based on typical historical data of the equipment, including multiple modal center frequencies. The edge computing nodes initialize the weight parameters of the Gabor filter bank using these modal center frequencies, ensuring each filter corresponds to a specific physical frequency band. Finally, the edge computing nodes perform a one-dimensional convolution operation on the raw vibration acceleration sequence using this physical sensing filter bank, outputting a modal feature word sequence with clearly defined frequency band physical meaning.

[0034] Step S200: State-guided collaborative dynamic feature extraction and interference stripping

[0035] The collaborative dynamic feature extraction unit inputs the modal feature word sequence into the PC-KAM backbone network. The PC-KAM backbone network contains a mutually cooperating state-space model branch and a Kolmogorov-Arnold network (KAN) branch.

[0036] The collaborative dynamic feature extraction unit uses the selective scanning mechanism of the state-space model branch to process the input sequence, capture the temporal evolution of the sequence, and calculate the hidden state vector at the current time.

[0037] The collaborative dynamic feature extraction unit uses the hidden state vector as a control variable and dynamically predicts the B-spline node position vectors required for the KAN branch at the next time step through linear projection and normalization operations. This step establishes a dynamic mapping relationship between the system phase state and the fitted network topology.

[0038] The collaborative dynamic feature extraction unit uses a KAN branch to fit the deterministic interference components in the input sequence. Specifically, a time-varying B-spline basis function is constructed using dynamically generated node position vectors, and regression fitting is performed on the input sequence to obtain the interference estimate.

[0039] The collaborative dynamic feature extraction unit performs a difference operation, subtracting the interference estimate from the input features to generate a residual feature vector that has been freed from strong background noise.

[0040] Step S300: Orthogonal subspace constraints and low-rank adaptation training of physical perception

[0041] During the model training phase, the physical constraint training unit performs orthogonal subspace projection constraints and physical perception low-rank adaptive fine-tuning.

[0042] The physical constraint training unit calculates the orthogonal subspace projection loss. Specifically, it calculates the cosine similarity between the deterministic feature vector and the residual feature vector, and optimizes by minimizing this similarity, thereby forcing the network to achieve orthogonal decoupling between the deterministic components and the random impulse components in the feature space.

[0043] The physical constraint training unit freezes the backbone network parameters and updates only the injected low-rank matrix parameters. It calculates the theoretical fault characteristic frequencies based on real-time rotational speed and bearing geometry parameters, and also calculates the power spectral density of the residual eigenvectors. The physical constraint training unit constructs a physical consistency loss, aiming to maximize the power spectral density response at the theoretical fault frequencies, and updates the network parameters using a backpropagation algorithm.

[0044] Step S400: Generation of fault samples based on the retrieval-enhanced diffusion model

[0045] When the number of training samples falls below a preset threshold, the retrieval-enhanced diagnostic generation unit performs data augmentation. The unit uses the currently extracted residual feature vector as a query condition to retrieve the most similar feature vectors from the historical fault case database. The retrieved feature vectors are then used as conditional guidance information and input into a pre-trained Denoising Diffusion Probability Model (DDPM). The DDPM performs a reverse denoising process to generate synthetic fault sample data that conforms to the physical distribution characteristics of the current operating conditions.

[0046] Step S500: Output intelligent diagnostic report based on retrieval enhancement

[0047] The retrieval-enhanced diagnostic generation unit extracts the fault feature vector of the signal to be diagnosed and retrieves the most similar historical maintenance cases from the vector database. The unit then concatenates the retrieved case descriptions, the physical characteristic parameters of the current signal, and the classification confidence scores output by the model into structured prompts. These structured prompts are then input into a large language model, which outputs a diagnostic report containing qualitative fault analysis, physical mechanism tracing, and maintenance recommendations.

[0048] The offline processing module extracts the center frequencies of key modes from historical data in advance, specifically through steps S211 to S214:

[0049] Step S211: Construction of a set of typical historical signals

[0050] The offline processing module extracts historical typical signal sets from industrial field databases or test bench data. These historical typical signal sets contain multiple vibration acceleration sequences under normal operating conditions, inner race fault conditions, outer race fault conditions, and rolling element fault conditions. The historical typical signal sets cover low-speed, medium-speed, and high-speed operating conditions within the equipment's permissible speed range to ensure that the extracted frequency features are representative of all operating conditions.

[0051] Step S212: Construction of the variational mode decomposition problem

[0052] For each vibration acceleration sequence in the historical typical signal set, the offline processing module uses the variational mode decomposition algorithm to decompose the vibration acceleration sequence into... A set of finite-bandwidth eigenmode functions around a center frequency. An offline processing module constructs a constrained variational problem whose optimization objective is to minimize the sum of the estimated bandwidths of each mode component, ensuring that the sum of the modal components can reconstruct the original input signal. The modal decomposition number... The number of frequency bands to be considered is determined based on the spectral complexity of the bearing vibration signal and the preset number of frequency bands of interest.

[0053] Step S213: The offline processing module for frequency iteration based on the alternating direction multiplier method uses the alternating direction multiplier method to iteratively solve the constrained variational problem, alternately updating the modal components and center frequencies. During the iteration process, the offline processing module updates the center frequency of each mode to the centroid position of the current power spectrum of that mode. The offline processing module continues to execute the iteration process until the convergence condition is met, that is, the modal component error between two consecutive iterations is less than a preset threshold, thereby obtaining the vibration acceleration sequence corresponding to the... The modal center frequency values.

[0054] Step S214: Statistical aggregation of global modal center frequencies

[0055] The offline processing module performs the above decomposition process on all samples in the historical typical signal set to obtain a set of center frequencies containing multiple samples. The offline processing module then uses the K-means clustering algorithm to statistically aggregate the center frequency values ​​extracted from all samples. Specifically, the offline processing module maps the frequency values ​​obtained from the decomposition of all samples to a one-dimensional frequency space, setting the number of clusters to be equal to the number of mode decompositions. The offline processing module then performs iterative clustering operations. The cluster center values ​​are determined as the prior set of VMD center frequencies. The VMD center frequency prior set is stored on the edge computing nodes. It is used for initializing the Gabor convolution kernel parameters in subsequent steps.

[0056] The physical sensing signal discretization module constructs a learnable Gabor filter bank based on the prior set of VMD center frequencies. The specific construction process includes steps S221 to S224:

[0057] Step S221: Define the function for parameterizing the Gabor convolution kernel

[0058] The physical sensing signal discretization module establishes a parameterized Gabor function model, which serves as the weight generation function for the first layer of the neural network's convolution. The convolution kernel weights of the k-th filter... Defined as: in, For discrete-time indexing; This is a normalization constant; For the center frequency parameter, Determines the frequency band location of the filter response; The scaling parameter of the Gaussian window, Determines the time-frequency resolution of the filter; This refers to the phase offset parameter. In the above formula... , and Defined as a learnable parameter in a neural network that can be updated with gradients, enabling the network to dynamically adjust the physical frequency band it focuses on during training.

[0059] Step S222: Initialization of physical parameters based on VMD priors

[0060] The physical sensing signal discretization module utilizes the VMD center frequency prior set. The convolutional kernel parameters are initialized with physical consistency to give the network initial physical awareness. The specific initialization logic is as follows:

[0061] The physical sensing signal discretization module discretizes the center frequency parameter The initial value is directly set to the corresponding VMD modal center frequency. .

[0062] The physical sensing signal discretization module discretizes the scale parameters. The initial value is set to be inversely proportional to the center frequency, and the calculation formula is: in The preset bandwidth control coefficient has a value range set to [value range missing]. This initialization strategy follows the uncertainty principle, ensuring that the filter bank has high frequency resolution in the low-frequency band to distinguish dense RF harmonics and high time resolution in the high-frequency band to capture transient fault impacts.

[0063] The physical sensing signal discretization module discretizes the phase parameters Initialize to zero.

[0064] Step S223: Parameter Constraints and Gradient Update Mechanism Configuration

[0065] To ensure the validity of physical parameters during training (e.g., scale parameters must be positive), the physical sensing signal discretization module configures parameter constraint transformations in the computational graph. Specifically, the physical sensing signal discretization module introduces unconstrained intermediate variables and maps them to positive scale parameters using the Softplus activation function. During the backpropagation phase of the network, the optimizer calculates the gradient of the loss function with respect to these learnable parameters, thereby enabling adaptive fine-tuning of the center frequency and bandwidth.

[0066] Step S224: Multi-channel convolutional feature map generation. The physical sensing signal discretization module utilizes the generated... Gabor convolution kernels For the input original vibration acceleration sequence Perform a one-dimensional convolution operation. The output of the convolution operation includes... Modal feature sequence of each channel Modal feature sequences Each channel in the system corresponds to a specific physical frequency band component, thus completing the feature mapping of the signal from the time domain to the physical time-frequency domain at the data input end. The physical sensing signal discretization module discretizes the modal feature sequence. It is then transmitted to the subsequent lexical segmentation unit.

[0067] The lexical segmentation unit converts the multi-channel modal feature sequence output by the physical sensing signal discretization module into a discrete lexical sequence suitable for deep neural network processing. The specific processing steps include S231 to S233:

[0068] Step S231: The temporal block segmentation unit of the multi-channel modal features obtains the multi-channel modal feature sequence, which contains Each physical frequency band channel. The lexical segmentation unit uses a sliding window mechanism to segment the multi-channel modal feature sequence along the time dimension, with a window length set to [value missing]. And the step size is For segments of the sequence shorter than a window length at the end, the lexical segmentation unit uses zero-padding to fill in the gaps. The lexical segmentation unit divides the continuous sequence into multiple overlapping local feature blocks. Each local feature block is a dimensional unit. A two-dimensional matrix that completely preserves all data within a specific time window. Vibration response information of each physical frequency band and their instantaneous coupling relationship.

[0069] Step S232: Linear Embedding Projection and Dimension Mapping. The lexical segmentation unit performs flattening and linear projection transformation on each local feature block, mapping the physical feature space to the semantic latent space. Specifically, the lexical segmentation unit first flattens the physical feature space to the semantic latent space. The local feature block is flattened to a length of The first one-dimensional vector is then transformed into a vector of dimension 1 through a learnable linear projection layer. The embedding vectors. This process compresses the original high-dimensional physical features into a unified hidden layer dimension representation within the model, while preserving the semantic information of the local time-frequency structure.

[0070] Step S233: Positional Encoding Injection and Sequence Generation. The lexical segmentation unit superimposes positional encoding information onto the embedding vector to mark the temporal position of local feature blocks in the original signal. The lexical segmentation unit calls a preset positional encoding vector, the dimension of which is the same as the dimension of the embedding vector. To maintain consistency, the lexical segmentation unit performs element-wise addition of the positional encoding vector and the embedding vector to generate the final modal feature lexical sequence. This modal feature lexical sequence integrates local physical frequency band energy distribution and global temporal positional information, and serves as input data for the subsequent collaborative dynamic feature extraction network.

[0071] The PC-KAM basic coding block adopts a dual-branch cooperative architecture, utilizing a state-space model to handle temporal dependencies and a Kolmogorov-Arnold network to handle nonlinear features. The specific processing flow includes steps S311 to S315:

[0072] Step S311: Input normalization and branching

[0073] The PC-KAM basic coding block receives the modal feature word sequence output from the previous layer. The PC-KAM basic coding block uses layer normalization operations to standardize the modal feature word sequence, generating a normalized sequence. The PC-KAM basic coding block normalizes the sequence. The data flow is split into two parallel data paths: the first path inputs to the state-space model branch for extracting the temporal context, and the second path inputs to the Kolmogorov-Arnold network branch for feature transformation.

[0074] Step S312: State-space modeling based on selective scanning. The state-space model branch uses a linear projection layer to normalize the sequence. Mapped to input vector Forget gate parameters and state control parameters and The state-space model branch discretizes the parameters of the continuous system using the zero-order preserve method and performs a selective scan operation. The selective scan operation recursively calculates the hidden state vector at the current time step. Calculation formula: in, and The time-varying system matrix after discretization. This is the hidden state vector. Aggregates up to the current moment The historical time sequence information is used as a representation of the system's current phase state, providing context for the subsequent generation of the dynamic mesh. The final output of the state-space model branch contains a sequence containing temporal context information.

[0075] Step S313: State-guided dynamic mesh generation

[0076] PC-KAM basic coding blocks utilize hidden state vectors The topology of the KAN branch is dynamically adjusted. Specifically, the PC-KAM basic coding block uses the hidden state vector calculated in step S312. The input is fed into the GridPredictor. The GridPredictor, composed of a multilayer sensing mechanism, outputs the grid node vectors required for the B-Spline basis function at the next time step. The formula for generating grid node vectors is: in, and These are learnable weights and biases. The key operations included in the above formulas are defined as follows:

[0077] Sigmoid: Maps the unbounded output of a neural network to the (0, 1) interval;

[0078] Sort: Performs an ascending sort operation on the output vector to ensure that the generated mesh nodes satisfy... The monotonically increasing constraint is a prerequisite for constructing an effective B-spline basis function;

[0079] Scale: Linearly maps normalized numerical values ​​to the input features of the KAN branch. Within the actual numerical distribution range (e.g.) ).

[0080] Step S314: Nonlinear fitting KAN calculation branch based on dynamic spline basis function utilizes the dynamic mesh node vector generated in step S313. Build A set of B-spline basis functions. The KAN computation branch uses this set of basis functions to normalize the sequence of the second path input. (This is recorded as input) Perform a nonlinear transformation. The output of the KAN branch. The calculation is as follows: in, Indicates dynamic mesh The first The spline basis functions are at the input The response value at the location; These are learnable control coefficients. This is achieved by introducing a dynamic mesh. The KAN computation branch can be based on the real-time status of the system. The shape and density of the activation function are adaptively adjusted, and more basis function nodes are automatically allocated in regions of drastic signal fluctuations (corresponding to moments of drastic changes in hidden states) to improve local fitting accuracy.

[0081] Step S315: Gating fusion and residual connection

[0082] The PC-KAM basic coding block performs a gated fusion operation, using the contextual features output from the state-space model branch as a gate signal and multiplying them element-wise with the fitted features output from the KAN computation branch. The PC-KAM basic coding block then restores the fused feature sequence to its original dimension through a linear projection layer and performs Dropout regularization. Finally, the PC-KAM basic coding block combines the regularized feature sequence with the original input. Perform addition operations and generate the final coded block output sequence through residual concatenation. .

[0083] The collaborative dynamic feature extraction unit (CMU) is composed of multiple stacked PC-KAM basic coding blocks connected in series, forming a deep feature extraction network. The CMU processes the temporal dependence and nonlinear characteristics of vibration signals at different abstraction levels through a hierarchical state-guided mechanism. The specific construction and operation process includes steps S321 to S323:

[0084] Step S321: Hierarchical stacking construction of deep networks

[0085] The number of collaborative dynamic feature extraction units configured is 1. The PC-KAM basic coding blocks are concatenated sequentially according to the sequence processing order. PC-KAM basic coding blocks (of which ) Receive the output feature sequence from the (l-1)th coded block As input, it outputs the feature sequence $Z_l$ of the current layer. For the first coding block, the input is the modal feature lexical sequence output from the signal lexicalization process. The propagation of feature sequences between layers follows a recursive relationship: in, Indicates the first The computational mapping function for the basic coded block of layer PC-KAM; Indicates the first The layer-level coding block contains an independent set of learnable parameters, including the state-space system matrix, grid prediction head weights, and control coefficients of the KAN branch. This recursive structure ensures that the collaborative dynamic feature extraction unit can abstract features layer by layer: the parameters of the lower-level coding blocks. The parameters of higher-level coding blocks tend to capture microscopic physical states such as high-frequency impacts. It tends to capture macroscopic semantic states such as fault evolution trends.

[0086] Step S322: Cooperative dynamic control flow at the entire network level

[0087] The collaborative dynamic feature extraction unit independently performs the mesh deformation process at each layer. Each layer's state-space model branch independently maintains the hidden state vector for that layer. Hidden state vector Dynamically update based on the input features of the current level, and directly drive the grid node vectors of the KAN branch at the same level. Generation. This mechanism establishes an independent state-aware and mesh deformation link at every level of the network's full depth. Dynamic meshes at different levels. By making differentiated adaptive adjustments based on the signal characteristics of interest at each level, the network topology was matched with the physical characteristics of bearing faults across the entire time and space scale.

[0088] Step S323: Final Normalization and Feature Output

[0089] The collaborative dynamic feature extraction unit performs a final-layer normalization operation at the output of the last PC-KAM basic coding block. Deep feature sequence output by the layer Statistical standardization is performed to achieve a mean of 0 and a variance of 1. Subsequently, the collaborative dynamic feature extraction unit utilizes learnable scaling parameters. Translation parameters Affine transformation is applied to the standardized features to restore the expressive power of the feature distribution. The deep feature sequence is then normalized. It has a uniform distribution characteristic and is transmitted to the subsequent fault classification module for decoding and discrimination.

[0090] The model optimization unit introduces an orthogonal subspace projection mechanism to map deep features to mutually independent fault category subspaces, and performs supervised learning updates on the deep neural network. The specific training process includes steps S411 to S413:

[0091] Step S411: Global aggregation of temporal features

[0092] The model optimization unit receives the deep feature sequence output by the collaborative dynamic feature extraction unit. The model optimization unit performs a global average pooling operation on the deep feature sequence. The global average pooling operation compresses the deep feature sequence along the time dimension, generating a global feature vector. This global feature vector aggregates fault mode information over the entire time window and serves as input data for subsequent classification and projection.

[0093] Step S412: Calculation of the logarithmic odds of classification

[0094] The model optimization unit uses a fully connected classification layer to project the global feature vectors into the fault category space. The model optimization unit defines the weight matrix of the classification layer as follows: ,in This represents the total number of fault categories. The feature dimension is used. The model optimization unit calculates the input global feature vector and weight matrix. The matrix product yields the predicted log odds vector. During this process, the weight matrix... Each row vector is considered as a basis vector for the corresponding fault category. The classification calculation is essentially a measure of the degree of matching between the input features and each fault basis vector.

[0095] Step S413: Construct a hybrid loss function with orthogonal constraints

[0096] The model optimization unit constructs a hybrid loss function with orthogonal subspace constraints to optimize network parameters.

[0097] The model optimization unit first calculates the classification principal loss. It then compares the predicted log-odds vector with the true fault labels and calculates the classification error using the cross-entropy loss function. Subsequently, the model optimization unit calculates the orthogonal regularization loss. Forced classification weight matrix The row vectors are kept orthogonal. The formula for calculating the orthogonal regularization loss is: in, for The identity matrix, Describing the Frobenius norm, This is the transpose of the weight matrix. Orthogonal regularization loss. constraint The row vectors satisfy autocorrelation of 1 and cross-correlation of 0. This constraint forces the network to project different types of fault features into mutually perpendicular feature subspaces, thereby eliminating feature coupling and aliasing between different fault modes.

[0098] Finally, the model optimization unit performs a weighted summation of the classification main loss and the orthogonal regularization loss to calculate the total loss function, and updates all learnable parameters of the entire network through the backpropagation algorithm.

[0099] The model optimization unit utilizes a physically-aware low-rank adaptation mechanism to efficiently fine-tune the parameters of the pre-trained PC-KAM model, adapting to scenarios involving changing operating conditions or the emergence of new fault categories. This mechanism updates the physical dynamic features by injecting a learnable low-rank matrix while freezing the core parameters of the model. The specific implementation process includes steps S421 to S423:

[0100] Step S421: Main parameters frozen and bypass initialization

[0101] The model optimization unit locks the pre-trained weight matrix in the collaborative dynamic feature extraction unit. The parameter values ​​are set to keep the pre-trained weight matrix constant during fine-tuning. The model optimization unit constructs a low-rank adaptive bypass along the computation path of the pre-trained weight matrix. The model optimization unit initializes two low-rank matrices: a dimensionality reduction matrix and a dimensionality reduction matrix. and the increasing dimension matrix .in, The set rank, and The value is much smaller than the input dimension. and output dimensions The model optimization unit uses a normal Gaussian distribution to reduce the dimensionality of the matrix. Perform random initialization and use an all-zero matrix to pair the increased-dimensional matrix. Initialization is performed. This initialization strategy ensures that the low-rank adaptive bypass output is a zero vector at the start of fine-tuning, guaranteeing that the initial state of the model is completely consistent with the physical behavior of the pre-trained model.

[0102] Step S422: Low-rank update computation model optimization unit for physical dynamic parameters on input feature vector Perform parallel computation. Input feature vector. Simultaneously, a frozen main path and a learnable low-rank adaptive bypass are used. The main path computes the original output. .

[0103] Low-rank adaptive bypass computation incremental output The calculation formula is: in This refers to the scaling hyperparameter. In the formula... The term is used to adjust the weight ratio of adaptive features. The model optimization unit decomposes the update of the model weights into the product of two low-rank matrices. This significantly compresses the parameter space to be optimized, enabling the model to quickly converge to a new operating condition distribution using a small number of fault samples.

[0104] Step S423: Physical stability constraints of the state-space system

[0105] The model optimization unit outputs the main path. With low-rank adaptive bypass output Element-wise addition is performed to obtain the final output. For the parameter matrix representing the dynamic evolution of the system in the State-Space Model (SSM) branch, the model optimization unit updates the matrix... and Physical stability constraints are introduced at this time. Specifically, the model optimization unit monitors the new system matrix after combination. The spectral radius. The model optimization unit restricts the magnitude of all eigenvalues ​​of the new system matrix to be strictly less than 1 through constraints. This constraint forces the fine-tuned model's dynamic behavior to conform to the energy dissipation and boundedness principles of physical systems, preventing numerical divergence in discrete systems caused by parameter overfitting and ensuring that the generated fault evolution trajectory is physically plausible.

[0106] The RAG diagnostic module transforms unstructured historical maintenance records into computer-computable spatial vectors, constructing a high-dimensional structured historical case vector library to provide domain knowledge support for subsequent diagnostic reasoning. The specific construction process includes steps S511 to S513:

[0107] Step S511: Structured cleaning of historical maintenance data

[0108] The RAG diagnostic module extracts historical fault case data from the maintenance log database. It then performs cleaning and field structuring operations on the extracted raw data to generate standardized metadata objects. Standardized metadata objects It includes four core fields: fault description, root cause of the fault, repair and handling plan, and list of replacement parts. The RAG diagnostic module uses a preset industrial stop word list to remove redundant information in the text and a thesaurus to replace the colloquial descriptions of the same fault phenomenon by different technicians with standard fault terms, thus eliminating semantic noise in the data source.

[0109] Step S512: Vectorized encoding of semantic features

[0110] The RAG diagnostic module utilizes a pre-trained semantic encoding model to process standardized metadata objects. Feature extraction is performed. The RAG diagnostic module standardizes metadata objects. The fault description field and the root cause field are concatenated to form the input text sequence. The RAG diagnostic module uses a semantic encoding model to process the input text sequence. Mapped to dense feature vectors of fixed dimensions The calculation formula is: in, This represents a bidirectional text encoder based on the Transformer architecture. Feature vectors. Historical cases are represented in a high-dimensional semantic space. The fault mechanism characteristics. In this semantic space, the Euclidean distance between the feature vectors corresponding to cases with similar fault mechanisms is small, while the Euclidean distance between the feature vectors corresponding to cases with significantly different fault mechanisms is large.

[0111] Step S513: Index building and storage based on HNSW

[0112] The RAG diagnostic module constructs a key-value mapping storage structure. The RAG diagnostic module stores feature vectors... As a search key, a standardized metadata object containing the complete repair solution text will be used. As a stored value.

[0113] The RAG diagnostic module employs a hierarchical navigable small-world graph algorithm to construct a vector index. The RAG diagnostic module builds a multi-level nearest neighbor graph structure in the vector space. The bottom-level graph contains the connections between all data nodes, while the upper-level graph contains only the connections between sparsely selected routing nodes. During the retrieval process, the RAG diagnostic module utilizes a greedy traversal strategy, first quickly locating the approximate region of the query vector in the upper-level graph, and then converging downwards layer by layer to the precise neighborhood of the bottom-level graph. The HNSW index structure supports approximate nearest neighbor search in logarithmic time complexity within a large-scale vector library, meeting the timeliness requirements of real-time diagnostics. The completed vector library is stored in a non-relational database.

[0114] The RAG diagnostic module combines the features of the current fault with retrieved historical cases to generate a structured fault diagnosis report using a conditional diffusion model. This process performs denoising operations to restore the target text embedding sequence from Gaussian noise in a continuous latent semantic space. The specific implementation process includes steps S521 to S523:

[0115] Step S521: Case retrieval based on feature similarity

[0116] The RAG diagnostic module uses the global feature vector of the current signal extracted in the previous steps as the query vector. The RAG diagnostic module performs an approximate nearest neighbor search in the historical case vector database, retrieving the nearest neighbor based on Euclidean distance. A set of historical case metadata objects, denoted as the retrieval collection. Retrieval Set Historical cases provide historical handling experience data that are most similar to the physical mechanisms of current failures, and this experience data constitutes the knowledge background for subsequent generation processes.

[0117] Step S522: Context encoding of the guiding conditions

[0118] The RAG diagnostic module constructs a conditional vector to guide diffusion generation. The RAG diagnostic module combines the fault category labels output by the current classification network with the retrieval set. The maintenance plan text is concatenated to form a comprehensive context sequence. The RAG diagnostic module uses a text encoder to map the comprehensive context sequence into a fixed-dimensional conditional embedding vector. Conditional embedding vector It integrates current objective diagnostic results with historical confirmatory knowledge, which are then input as external control signals into the subsequent generative model.

[0119] Step S523: Iterative denoising generation in the latent space

[0120] The RAG diagnostic module performs a reverse diffusion process to generate a characteristic sequence for the diagnostic report. The RAG diagnostic module first starts from a standard normal distribution... Mid-sampling yields a pure noise vector sequence ,in The preset total number of diffusion steps. The RAG diagnostic module uses a denoising neural network to perform... The denoising operation is performed in each iteration. (from Decrease to 1), the denoising neural network receives the noisy state at the current moment. Current time step and conditional embedding vectors As input, the denoising neural network uses the conditional embedding vector through its internal cross-attention layer. The noise component in the current state is predicted by integrating it into the network feature stream. The RAG diagnostic module then calculates the potential state from the previous time step based on the predicted noise component. The updated formula is: in, The function mapping representing the denoising neural network, and These are the preset noise scheduling parameters. It is random Gaussian noise. (The formula contains...) The terms reflect the conditional embedding vector Explicit guidance on denoising trajectories. The RAG diagnostic module, through... Step iteration, transforming the initial pure noise vector Gradually restored to a text feature sequence with a clear semantic structure Finally, the RAG diagnostic module uses a text decoder to process the text feature sequence. The final diagnostic report is mapped into natural language form.

[0121] The RAG diagnostic module performs feature alignment and deep interaction on heterogeneous data, mapping physical sensor signal features and semantic textual knowledge features to a unified representation space. The specific implementation process includes steps S531 to S533:

[0122] Step S531: Projection of common subspace of heterogeneous features

[0123] The RAG diagnostic module receives signal feature sequences from the Co-dynamic Feature Extraction Unit. and text feature sequences from the text encoder Because the original dimensions and distribution spaces of signal features and text features are inconsistent, the RAG diagnostic module constructs a shared common interactive subspace. The RAG diagnostic module utilizes a linear projection layer to transform the features. The RAG diagnostic module uses a weight matrix... Signal feature sequence Mapped to projected signal features And through the weight matrix Text feature sequence Mapped to projected text features Linear projection operation makes the characteristics of the projected signal... With projected text features By maintaining consistency in feature dimension values, a mathematical correspondence between signal waveform patterns and text semantic terms was established, aligning the data specifications for subsequent matrix operations.

[0124] Step S532: Feature interaction based on cross-attention

[0125] The RAG diagnostic module utilizes a multi-head cross-attention mechanism to calculate the attention level of text features to signal features. The RAG diagnostic module projects text features... As a query vector, the projected signal features Simultaneously serving as both a key vector and a value vector. The RAG diagnostic module calculates the... The output of each attention head is calculated using the following formula: in, For the first The learnable weight matrix for each attention head. The scaling factor is used. The Softmax term in the formula calculates the similarity weight between each text term and each signal time step. Based on this, the RAG diagnostic module searches for and aggregates waveform features semantically related to the text description (such as "impact" or "wear") in the time-series signal, generating interactive features containing physical verification information. .

[0126] Step S533: Adaptive Gated Fusion of Physical and Semantic Features

[0127] The RAG diagnostic module introduces an adaptive gating unit to integrate interactive features. Features of the original projected text The RAG diagnostic module first calculates adaptive fusion coefficients based on the current feature context. The calculation formula is: in, This indicates a feature concatenation operation. and The weights and biases of the gating network are defined. Subsequently, the RAG diagnostic module utilizes the fusion coefficients. The features are weighted and summed to generate the final multimodal fusion representation. The formula is: The adaptive gating unit can dynamically adjust the weights according to the characteristics of the input data. When signal features contribute significantly to the diagnostic conclusion, the network automatically increases in size. The value is determined by the condition; otherwise, more original text information is retained. The final generated fusion representation... It combines the theoretical semantics of the fault with the actual physical state, serving as a direct feature input for generating subsequent diagnostic reports.

[0128] The RAG diagnostic module utilizes the reasoning capabilities of a large language model to transform the multimodal fusion representation obtained in previous steps and the retrieved historical knowledge into a diagnostic report in natural language. This process is based on soft cue alignment technology and structured instruction templates. The specific implementation process includes steps S541 to S543:

[0129] Step S541: Soft cue space mapping of multimodal features

[0130] The RAG diagnostic module receives the multimodal fusion characterization output in step S533. Due to multimodal fusion representation As continuous numerical vectors, they cannot be directly input into large language models as discrete text terminology indices. The RAG diagnostic module utilizes a learnable linear projection layer to fuse multimodal representations. Mapping to the word embedding space of a large language model to generate soft cue feature sequences The mapping calculation formula is: in, Let be the projection weight matrix. For bias vectors, This sets the preset word embedding dimension for the large language model. Through linear projection operations, the RAG diagnostic module transforms the signal features from the physical space into continuous vector representations within the semantic space of the large language model. Soft cue feature sequence. In the input sequence, it acts as a virtual word, directly participating in subsequent attention calculation and text generation processes.

[0131] Step S542: Assembling structured mind chain cue words

[0132] The RAG diagnostic module is constructed by including a sequence of text prompts with explicit instructions and contextual constraints. The RAG diagnostic module converts the set of similar historical cases retrieved in step S513 into text paragraphs and assembles them using a preset system instruction template. (Text prompt word sequence) It includes the following structured fields: Role Definition field: Specifies that the model is an industrial equipment diagnostic expert; Knowledge Context field: Contains the fault phenomena and maintenance solutions text of retrieved historical cases; Inference Task field: Requires the model to combine soft-hint feature sequences. Analyze the physical mechanism and root cause of the fault; output constraint fields: specify the output format as a standard JSON structure, and it must contain three key-value pairs: "Fault Conclusion", "Mechanism Analysis", and "Suggested Solution". The RAG diagnostic module will display the soft-suggestion feature sequence. With text prompt word sequence The corresponding word embedding vectors are concatenated to form the complete input sequence of the large language model. .

[0133] Step S543: Autoregressive generation and format validation

[0134] The RAG diagnostic module will input the complete sequence. The input is fed into a large language model. The large language model generates the output word sequence of the target diagnostic report one by one based on an autoregressive mechanism. The large language model then outputs the complete input sequence. As a constraint, the probability distribution of each word in the vocabulary in the current context is calculated, and the output word at the current moment is selected according to a preset sampling strategy (such as greedy search or kernel sampling). This generation process continues until the model outputs an end symbol or reaches the maximum length limit. After generation, the RAG diagnostic module uses a JSON parser to perform format validation on the output word sequence. The RAG diagnostic module verifies whether the output content conforms to the predefined JSON key-value structure. If the validation passes, the RAG diagnostic module extracts the content from the "suggested solution" field to generate the final diagnostic report; if the validation fails, the RAG diagnostic module triggers a regeneration command.

[0135] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A bearing fault diagnosis method based on a physical sensing KAM network, characterized in that, Includes the following steps: Obtain the original vibration acceleration sequence and synchronous speed information of the bearing to be diagnosed; perform the physical sensing signal discretization step: call the stored physical sensing parameters to construct a Gabor filter bank, and use the Gabor filter bank to perform a one-dimensional convolution operation on the original vibration acceleration sequence to generate a modal feature word sequence containing frequency band physical meaning; Perform the collaborative dynamic feature extraction step: input the modal feature word sequence into the PC-KAM backbone network, use the parallel state space model branches in the PC-KAM backbone network to calculate the hidden state vector that reflects the signal phase evolution law, and use the hidden state vector as a control variable to dynamically adjust the basis function node position parameters of the Kolmogorov-Arnold network branch to output the deep feature sequence. The enhanced diagnostic generation step involves: retrieving similar historical fault cases from the vector database based on the deep feature sequence; fusing the retrieved historical fault cases with the deep feature sequence to generate a multimodal fusion representation; The multimodal fusion representation is mapped to a soft cue feature sequence using a linear projection layer, and the soft cue feature sequence is input into a large language model to output a diagnostic report containing fault conclusions and maintenance suggestions.

2. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 1, characterized in that, In the step of performing physical sensing signal discretization, the physical sensing parameters include the center frequency determined based on the variational mode decomposition results of historical device data; The construction strategy of the Gabor filter bank is as follows: the weight of the convolution kernel is defined by the product of the Gaussian window function and the cosine function, wherein the scaling parameter of the Gaussian window function is inversely proportional to the center frequency, and the frequency parameter of the cosine function is initialized to the center frequency. The center frequency parameter, scale parameter, and phase offset parameter in the Gabor filter bank are all learnable parameters, configured to be adaptively fine-tuned based on gradient updates during training.

3. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 1, characterized in that, The specific steps of performing collaborative dynamic feature extraction include: The state space model branch is used to perform a selective scanning operation on the input modal feature word sequence, and the hidden state vector at the current time is recursively calculated. The hidden state vector is input into the grid prediction head, the grid prediction head is used to output the grid node vector at the next time step, and the grid node vector is sorted to satisfy the monotonically increasing constraint. A time-varying B-spline basis function is constructed using the grid node vectors, and the input sequence is nonlinearly fitted using the Kolmogorov-Arnold network branch.

4. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 3, characterized in that, The nonlinear fitting of the input sequence using the Kolmogorov-Arnold network branch specifically includes: Calculate the response value of the B-spline basis function at the input value based on the grid node vector; The response values ​​are weighted and summed using learnable control coefficients, and the weighted sum is multiplied by the input values ​​processed by the activation function to obtain the fitted output.

5. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 1, characterized in that, During the training phase of the PC-KAM backbone network, a hybrid loss function containing orthogonal subspace constraints is constructed. The orthogonal subspace constraint is used to ensure that the row vectors of the classification weight matrix remain orthogonal. By minimizing the difference between the autocorrelation matrix and the identity matrix of the classification weight matrix, the network is forced to project different types of fault features into mutually perpendicular feature subspaces to eliminate feature coupling between fault modes.

6. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 1, characterized in that, It also includes fine-tuning the PC-KAM backbone network using a physical-aware low-rank adaptation mechanism: Freeze the backbone parameters of the PC-KAM backbone network and construct low-rank adaptive bypasses containing dimensionality reduction and dimensionality increase matrices on the sides of the backbone paths; The incremental output is calculated using the low-rank adaptive bypass, and then the incremental output is added to the original output calculated using the trunk parameters. Physical stability constraints are introduced when updating the reduced-dimensional matrix and the increased-dimensional matrix. By limiting the spectral radius of the new system matrix after combining the main parameters and low-rank parameters to be less than a preset threshold, the convergence of the dynamic behavior of the fine-tuned system is ensured.

7. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 1, characterized in that, The step of fusing the retrieved historical fault cases with the deep feature sequence using multimodal information specifically includes: The deep feature sequence is mapped to projected signal features using a linear projection layer, and the text features of the historical fault cases are mapped to projected text features. The interaction features are calculated using a multi-head cross-attention mechanism, wherein the projected text features serve as the query vector and the projected signal features serve as both the key vector and the value vector. Calculate the similarity weight between the query vector and the key vector, and use the similarity weight to perform weighted aggregation on the value vector to generate the interaction feature containing physical verification information.

8. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 7, characterized in that, The generation of multimodal fusion representations also includes adaptive gating fusion of physical and semantic features: The interactive features are concatenated with the projected text features, and an adaptive fusion coefficient is calculated using a gating network. The interactive features and the projected text features are weighted and summed using the adaptive fusion coefficients to generate the multimodal fusion representation.

9. The bearing fault diagnosis method based on a physical sensing KAM network according to claim 1, characterized in that, The process of mapping the multimodal fusion representation into a soft cue feature sequence using a linear projection layer specifically includes: The multimodal fusion representation in the physical space is mapped to the word embedding space of a large language model using a learnable linear projection matrix, generating the soft cue feature sequence as virtual lexical units; Construct a sequence of text prompts that includes role definition fields, knowledge context fields, reasoning task fields, and output constraint fields; The soft cue feature sequence is concatenated with the word embedding vectors corresponding to the text cue word sequence to form the complete input sequence of the large language model.

10. A bearing fault diagnosis system based on a physical sensing KAM network, comprising implementing the bearing fault diagnosis method based on a physical sensing KAM network as described in any one of claims 1-9, characterized in that, include: The on-site data acquisition subsystem is configured to acquire the original vibration acceleration sequence and synchronous speed information of the bearing to be diagnosed. The edge computing node is configured to call the stored physical sensing parameters to construct a Gabor filter bank, and use the Gabor filter bank to perform a one-dimensional convolution operation on the original vibration acceleration sequence to generate a modal feature word sequence containing frequency band physical meaning; The cloud-based analysis and diagnostic platform is configured to load and run the PC-KAM backbone network. The modal feature word sequence is input into the PC-KAM backbone network, the hidden state vector is calculated using its internal parallel state space model branches, and the basis function node position parameters of the Kolmogorov-Arnold network branches are dynamically adjusted using the hidden state vectors to output the deep feature sequence. The cloud-based analysis and diagnostic platform is also configured to perform retrieval-enhanced diagnostic generation, retrieve similar historical fault cases and fuse them with the deep feature sequence to generate a multimodal fusion representation, map the multimodal fusion representation into a soft cue feature sequence using a linear projection layer, input it into a large language model, and output a diagnostic report.

Images