Bearing multi-task fault diagnosis method based on prior knowledge and depth feature fusion

Abstract

The invention provides a bearing multi-task fault diagnosis method based on prior knowledge and depth feature fusion, and relates to the technical field of mechanical fault diagnosis. The method comprises the following steps: 1) main-task collaborative data acquisition: generating a simulation vibration signal based on a two-degree-of-freedom dynamic model; actually measured vibration signals are collected through a sensor; 2) signal processing and feature extraction: after denoising and normalization, using a lightweight convolution-KAN hybrid network to extract four types of features; 3) performing prior guidance fusion: aggregating four types of features through an attention sub-module, performing simulation-actual measurement feature alignment, and screening first 128-dimensional high-discrimination fusion features; and 4) cooperative training diagnosis: taking actual measurement diagnosis as a main task, taking simulation classification and feature alignment as auxiliary tasks, and adopting a mixed loss function training model. According to the invention, dependence on measured data is reduced through cooperation of main and auxiliary tasks, and the method has end-to-end characteristics and strong interpretability, and is suitable for fault diagnosis of single fault and composite fault of the rolling bearing in an industrial scene.

Description

Technical Field

[0001] This invention belongs to the field of rolling bearing fault diagnosis technology, specifically involving a bearing multi-task fault diagnosis method based on the fusion of prior knowledge and deep features. Background Technology

[0002] In the field of fault diagnosis, deep learning-based intelligent methods have been widely applied to key equipment such as gears and rolling bearings. However, intelligent algorithms typically require a large amount of labeled data for model training. While single-fault data from mechanical equipment has fostered a relatively robust data-sharing ecosystem, labeled composite fault data is lacking. In practical engineering, due to cascading effects, multiple faults are more likely to occur in several parts of a component or across multiple system components. When composite faults occur, the resulting risks are usually much more severe than those caused by a single fault. As a core component of rotating machinery, rolling bearings account for more than 40% of all rotating machinery faults, and accurate early fault identification is crucial for ensuring reliable equipment operation. Existing technologies have significant limitations in adaptability to nonlinear signal processing and composite fault diagnosis. Purely physical model-driven methods construct basic dynamic equations, combine them with numerical algorithms to solve vibration responses, and then extract fault features. The core of these methods relies on theoretical modeling of bearing dynamic behavior and is suitable for ideal scenarios with constant speed and single faults. However, its limitations become apparent in complex real-world operating conditions: when faced with compound faults (such as multi-source excitation caused by the simultaneous failure of the inner ring and rolling elements), traditional models struggle to distinguish independent fault characteristics due to the inability to separate coupled vibration signals, resulting in a misjudgment rate as high as 25%. The fundamental problem lies in the fact that the simplified assumptions of the theoretical model regarding actual operating conditions deviate significantly from the real operating environment, leading to a disconnect between feature extraction and the actual fault mechanism.

[0003] Purely data-driven methods rely on deep learning models such as CNNs and LSTMs to directly learn from the original vibration signals end-to-end to achieve feature extraction and fault classification. The advantage of this approach is that it does not require pre-setting the fault mechanism and can autonomously learn feature patterns through a large amount of data, making it suitable for scenarios with abundant data and stable operating conditions. However, this method has two major bottlenecks: first, its "black box" nature leads to a lack of physical interpretability; model decisions rely on abstract features such as high-level convolutional kernels and cannot be linked to the physical evolution of the fault; second, its generalization ability is weak with small samples. When facing complex fault types, a large amount of data needs to be re-labeled to update the model, which not only increases training costs but also easily leads to overfitting in data-scarce scenarios. Its core flaw lies in its departure from the physical essence of bearing faults, relying solely on statistical data patterns, making it difficult to cope with the fluctuations in operating conditions in actual industrial environments.

[0004] Traditional feature fusion methods combine the advantages of physical models and data-driven approaches, simply concatenating theoretical features generated by the physical model with signal features extracted by data-driven approaches before inputting them into the classifier. However, this "concatenation-style" fusion has a fundamental flaw: the distribution of physical features, which are mostly sparse vectors in the frequency domain, differs significantly from that of data features, which are mostly dense vectors in the time domain. Direct concatenation leads to an explosion in the dimensionality of the feature space, which not only increases the computational burden on the classifier but also reduces discriminative power due to interference from redundant information.

[0005] In the acquisition and processing of nonlinear signals, some methods rely solely on traditional signal processing techniques such as wavelet transform and Fourier analysis, which can only extract simple time-domain or frequency-domain features of the signal, making it difficult to capture complex and deep nonlinear fault information. While some deep learning methods can mine deep features, they often employ a single network structure, such as pure convolutional networks. This invention uses a lightweight convolutional-KAN hybrid network to extract four types of features: analog signal physical information features, analog signal data features, measured signal physical information features, and measured signal data features. It separates and extracts analog signal physical information (including dynamic laws) and data features, as well as corresponding features of measured signals, achieving a deep fusion of physical priors and data-driven approaches, avoiding feature loss caused by traditional fragmented processing.

[0006] At the feature fusion and ranking level, traditional methods often simply splice together four types of features, which is prone to redundancy due to signal heterogeneity. The lack of alignment mechanisms leads to broken correlations between simulated and measured features, and they often rely on experience or simple statistical screening, easily retaining redundant and low-discriminative features. This invention dynamically assigns differentiated weights to the four types of features through channel attention, accurately aggregating key features within a single signal (such as highlighting the physical laws of simulated signals or fault pulses in measured signals). Then, cross-signal attention achieves semantic alignment between simulated and measured features, eliminating modal difference interference and solving the problem of poor feature correlation in traditional fusion methods. The SHAP method quantifies the contribution of features to the diagnostic results, accurately selecting the top 128 high-discriminative features. This reduces computational redundancy, improves model efficiency, and avoids the subjectivity of traditional screening. Furthermore, prior knowledge is integrated throughout the fusion and ranking process, ensuring that features always conform to the laws of bearing dynamics, avoiding the unreliability of fused features caused by traditional methods detaching from physical constraints, and significantly improving feature quality and subsequent diagnostic accuracy.

[0007] In the diagnostic model building and training phases, most deep learning diagnostic models are "black box" models, such as traditional deep neural networks, which cannot clearly explain the decision-making basis of diagnostic results. This contradicts the interpretability requirements of maintenance personnel in industrial scenarios, greatly limiting the application of models in practical engineering. Traditional training strategies often treat actual fault classification as a single task, failing to fully utilize the advantage of the large amount of analog signals that can be generated to design auxiliary tasks. This makes the model prone to overfitting when experimental data is scarce. Model training lacks effective constraints on physical laws and fails to ensure the consistency between analog and experimental signals in terms of dynamic characteristics, further reducing the reliability and generalization ability of diagnostic results.

[0008] To address the limitations of traditional methods, this invention proposes a bearing multi-task fault diagnosis method based on the fusion of prior knowledge and deep features, achieving a breakthrough in traditional fault diagnosis techniques through systematic improvements.

[0009] A search revealed application CN113191240B, which discloses a multi-task deep neural network method and apparatus for bearing fault diagnosis. The method includes: acquiring bearing vibration signals of different fault types and extracting the corresponding spectrum of each vibration signal; creating a bearing sample set; constructing a multi-task deep neural network with auxiliary tasks, including an auxiliary task execution module and a fault diagnosis task execution module. The auxiliary task execution module includes a feature extraction module and a spectrum generation module; the fault diagnosis task execution module includes a feature extraction module and a fault classification module; inputting the bearing sample set into the multi-task deep neural network for training; and using the trained multi-task deep neural network to diagnose the bearing under test. Through this method, this application enables the hidden layers of the neural network to learn features in the vibration signal spectrum during fault identification, thereby improving the robustness of the neural network in bearing fault diagnosis.

[0010] This patent addresses the weak generalization ability and lack of physical interpretability of small samples in complex real-world operating conditions by proposing a multi-task bearing fault diagnosis method based on the fusion of prior knowledge and deep features. First, to address the poor generalization ability of small samples under complex conditions, this patent expands the sample size and provides prior knowledge by constructing a composite fault dynamics model of the rolling bearing system, thereby enhancing the model's robustness. Second, to address the lack of physical interpretability, this patent extracts physical and data features from simulated and measured signals separately, and then fuses four types of features to achieve deep integration of physical priors and data-driven approaches, avoiding feature loss caused by traditional fragmented processing. Finally, addressing the issues of poor generalization ability and incomplete fault category coverage under complex operating conditions with small samples, this patent proposes a primary and secondary task. On the one hand, the primary task focuses on six key states in the actual operation of rolling bearings, accurately serving the real-time monitoring and fault determination of bearing operating status in production. On the other hand, secondary task 1 covers eight more detailed fault states, providing the model with richer fault sample information and helping the model learn diverse fault modes. Second, secondary task 2 minimizes the difference in feature distribution between simulated and measured signals, allowing the model to more efficiently integrate signal features from different sources. The synergistic effect of the two tasks effectively expands the model's cognitive boundaries regarding faults and enhances the model's generalization ability and diagnostic accuracy under complex operating conditions. Summary of the Invention

[0011] This invention aims to solve the problems of the prior art. It proposes a bearing multi-task fault diagnosis method based on the fusion of prior knowledge and deep features. The technical solution of this invention is as follows:

[0012] A bearing multi-task fault diagnosis method based on prior knowledge and deep feature fusion includes the following steps:

[0013] S1. Collaborative Data Acquisition of Main and Auxiliary Tasks: Construct a dynamic model of a composite fault in a rolling bearing system, i.e., build a simulation model based on a two-degree-of-freedom dynamic equation, solve the dynamic equation using the Runge-Kutta algorithm, and generate simulated vibration signals of single and composite faults; collect rolling bearing fault vibration signals under different conditions using sensors to obtain measured signal samples.

[0014] S2. Signal Processing and Feature Extraction: The simulated vibration signal and the measured vibration signal are subjected to unified preprocessing; then feature extraction is performed on them. A lightweight convolutional-KAN hybrid network with residual connections and feature normalization is used to extract the physical information features and nonlinear data features of the two signals respectively.

[0015] S3. Prior-guided feature fusion: Construct a prior-guided multi-layer attention fusion module, input the extracted physical information features of two signals and nonlinear data features, and then aggregate single signal features through channel attention, achieve simulation-measured feature alignment through cross-signal source attention, and sort and filter highly discriminative fusion features.

[0016] S4. Collaborative Training and Diagnosis of Main and Auxiliary Tasks: The main task is fault diagnosis of measured signals, while the auxiliary tasks are classification of simulated signals and alignment of multiple signal features. The KAN fault diagnosis model is trained using a hybrid loss function of "cross-entropy + contrastive + physical constraint MSE" to output fault categories and interpretability results.

[0017] Furthermore, S1 specifically includes the following steps:

[0018] ① Construction of a two-degree-of-freedom dynamic model for rolling bearings:

[0019] Single and compound faults in rolling bearings are simplified as cuboid pit defects. Single faults in rolling bearings include faults in the outer ring, inner ring, and rolling elements. Compound faults include combinations of inner ring + outer ring and inner ring + rolling elements. Defect parameters are defined as: length L, width B, and depth H. Based on geometric relationships, the maximum displacement excitation ΔH of the fault is calculated as follows:

[0020]

[0021] Where D is the diameter of the rolling element;

[0022] Based on Hertz contact theory, a two-degree-of-freedom fault model for rolling bearings is constructed as follows:

[0023]

[0024] Where m represents the bearing mass, C is the equivalent damping, and W... x W y External excitations in the X and Y directions;

[0025] in, It represents the acceleration of the bearing in the X direction, describing the rate of change of the bearing's velocity in the X direction. This represents the bearing's acceleration in the Y direction, i.e., the rate of change of the bearing's velocity in the Y direction. K is the contact stiffness, reflecting the ability of the rolling bearing's contact parts to resist deformation, and is related to factors such as the bearing's material and geometry. j δ is a coefficient related to the j-th rolling element, which may involve the influence of factors such as the rolling element's position and contact state on the contact force. j (t): Represents the contact deformation between the j-th rolling element and the inner and outer rings of the bearing. Contact force is generated only when the contact deformation is greater than 0. θj : is the contact angle of the j-th rolling element, used to describe the angular relationship between the rolling element and the contact point between the inner and outer rings, and affects the components of the contact force in the X and Y directions.

[0026] ②Simulated data generation and measured data acquisition

[0027] The dynamic equations are solved using the Runge-Kutta algorithm to generate simulated vibration signals for single and combined faults. Then, sensors are used to collect measured vibration signals of rolling bearings under different fault states, i.e., single fault + combined fault, and the fault type and operating parameters are recorded simultaneously.

[0028] Furthermore, S2 specifically includes the following steps:

[0029] ① Feature preprocessing:

[0030] Adaptive wavelet denoising: db4 wavelet basis decomposition with 6 layers, Stein unbiased risk estimation with soft thresholding in the 5th-6th layers of high frequency, and hard thresholding in the 1st-4th layers of low frequency.

[0031] Min-Max normalization:

[0032]

[0033] Where: x mim / x max To find the global maximum and minimum values ​​in the simulated and measured training sets, output a standardized analog signal X. sim Standardized measured signal X real ;

[0034] ② Four types of feature extraction: Convolutional-KAN hybrid network, with shared parameters in both domains;

[0035] Design 1D convolutional blocks to extract physical information features, and obtain the physical information features F of the analog signal. sim_phy Includes fault frequency, contact force characteristics, and physical information features of measured signals F real_phy ;

[0036] The KAN layer is designed to extract data features to obtain the analog signal data features F. sim_data Measured signal data characteristics F real_data .

[0037] Furthermore, the prior-guided multi-layer attention fusion module construction in S3 specifically includes the following steps:

[0038] ① Channel attention submodule construction and single feature aggregation:

[0039] The channel attention submodule comprises a serial structure of "global average pooling (GAP) + 2 fully connected layers + sigmoid activation", which then applies the analog signal data features F to... sim_data Physical information characteristics of the measured signal F real_phy Measured signal data characteristics F real_data Input the channel attention submodule, calculate the channel weights of each feature and perform weighted aggregation, output:

[0040] 1) Characteristics of analog signal data aggregation:

[0041] F sim_data_agg =F sim_data ⊙W ch_sim_data

[0042] Among them, W ch_sim_data The corresponding channel weights are defined in dimension [M, 64].

[0043] 2) Aggregation characteristics of physical information of measured signals:

[0044] F real_phy_agg =F real_phy ⊙W ch_real_phy

[0045] Among them, W ch_real_phy The corresponding channel weights are defined in dimension [N, 64], where N is the number of measured samples.

[0046] 3) Aggregation characteristics of measured signal data

[0047] F real_data_agg =F real_data ⊙W ch_real_data

[0048] Among them, W ch_real_data The corresponding channel weights are defined in dimension [N, 64].

[0049] Then, the weights of the prior guidance channels are optimized, and the two types of aggregated features F of the analog signal are further optimized. sim_phy_agg F sim_data_agg The fault physical parameters calculated based on S1 are the maximum fault displacement excitation ΔH and the contact force F. j Adjust the channel weights. After adjustment, the weights need to be normalized to ensure that the sum of the weights is 1.

[0050] ② Construction and feature alignment of cross-signal attention submodules, prior-guided simulation-test feature matching

[0051] (1) First, feature splicing of the two signals is performed. The aggregated features of the same signal type are spliced ​​in the order of "physical information features + data features" to obtain two types of cross-modal features: analog signal cross-modal features and measured signal cross-modal features;

[0052] (2) Then perform similarity matrix calculation for the two fused signals and design cross-signal attention weights based on cosine similarity;

[0053] (3) Then perform prior guidance weight adjustment, and dynamically adjust W according to the matching degree of physical parameters between the analog signal and the measured signal. cross ;

[0054] (4) Finally, cross-signal feature alignment and fusion are performed, based on the adjusted W. cross Align the measured signal features with the physical prior of the analog signal;

[0055] ③ Construct the SHAP feature ranking submodule and select high-discriminative features:

[0056] (1) First, the SHAP interpreter is configured, and TreeExplainer is used as the SHAP value calculation tool. The interpreter is built based on the feedforward structure of the KAN fault diagnosis model, namely the feature extractor + fusion module.

[0057] (2) Then, the SHAP value is calculated and the feature contribution is quantified, and the F align For each feature channel c (c = 1, 2, ..., 256), calculate its mean absolute SHAP value as an indicator of feature contribution.

[0058] (3) Then perform feature sorting and filtering;

[0059] (4) Finally, the results of the prior verification screening are carried out. The correlation between the 128-dimensional features and the physical parameters is verified.

[0060] Furthermore, the main-auxiliary task collaborative training and diagnosis in S4 specifically includes the following steps:

[0061] (1) Definition of main and auxiliary tasks and input feature matching

[0062] Main task: The objective is to output six states of rolling bearings, including normal state C0, inner ring single fault C1, outer ring single fault C2, rolling element single fault C3, inner ring + outer ring combined fault C4, and inner ring + rolling element combined fault C5. The input feature is the fused feature F of the measured signal output by S3. real_fusion The dimension is [N, 128], where N is the number of measured samples, which is obtained by attention fusion and SHAP sorting of the two types of features of the measured signal;

[0063] Auxiliary Task 1: The objective is to output eight types of fault states for rolling bearings, including five single faults: C0, C1, C2, C3, and cage fault C6; and three compound faults: C4, C5, and outer ring + rolling element compound fault C7. The input feature is the fusion feature F of the analog signal synchronously generated in step S3.sim_fusion The dimension is [M, 128], where M is the number of simulated samples. This feature is obtained by fusing two types of features of the simulated signal with the same attention and sorting by SHAP.

[0064] Auxiliary Task 2: The objective is to minimize the difference in feature distribution between the analog and measured signals. The input features are the cross-modal feature pairs output from step S3, i.e., the analog side F. sim_cross , measured side F real_cross All are 128-dimensional;

[0065] (2) Construction of KAN network diagnostic model

[0066] The KAN network diagnostic model described above adopts a "shared feature mapping layer + task-specific classification head" structure, with input feature F real_fusion or F sim_fusion The fusion features are then entered into the shared KAN backbone layer to achieve nonlinear feature mapping. This layer is the core feature processing unit of the model. Through a 2-layer KAN network, nonlinear compression and physical association enhancement of 128-dimensional input features are achieved. Batch normalization and regularization are added simultaneously to ensure generalization.

[0067] Task-specific category headers: This layer designs independent category units for different tasks, with parameters not shared, to ensure adaptation to the different output category requirements of main and auxiliary tasks, and constructs a main task category header and an auxiliary task 1 category header;

[0068] The diagnostic network parameters are initialized using KAN layer B-spline nodes, based on the fault physical parameters ΔH([0,0.5D]) and contact force F. j The normalization range of ([0,1e4]N) is set to ensure that the nodes cover the range of physical feature values; the weights of the fully connected layers are initialized using He normality. The bias is initialized to 0.01;

[0069] (3) Hybrid loss function calculation, i.e., multi-objective collaborative optimization:

[0070] The hybrid loss function is a weighted fusion of the primary and secondary task losses, and the formula is as follows:

[0071] L total =0.6L main +0.2L aux1 +0.2L aux2

[0072] The calculation methods for each sub-loss are as follows:

[0073] 1)L main Main task loss, L aux1 To mitigate the losses of Task 1, L aux2The loss is for auxiliary task 2, namely contrastive loss + physical constraint MSE loss;

[0074] (4) Model training optimization strategy:

[0075] 1) Optimizer configuration

[0076] The AdamW optimizer is used with the following parameters:

[0077] Initial learning rate: η0 = 1e-3

[0078] Weight decay coefficient: λ = 1e -4

[0079] Momentum parameters: β1 = 0.9, β2 = 0.999.

[0080] 2) Learning rate scheduling and early stopping strategy

[0081] Learning rate scheduling: Cosine annealing strategy is adopted;

[0082] Early stopping strategy: Monitor the accuracy of the validation set of the main task. The validation set accounts for 20% of the actual test data. If the improvement is ≤0.1% for 5 consecutive rounds, save the current best model.

[0083] 3) Parameter update method

[0084] L is calculated using automatic differentiation. total For the gradient of all network parameters (KAN layer coefficients, classifier head weights, attention weights), update the parameters according to the gradient descent rule, and iterate until the early stopping condition is met.

[0085] (5) Diagnostic results output:

[0086] 1) Core diagnostic results

[0087] The input measured signal to be diagnosed is preprocessed in S2 and fused in S3 to obtain F. real_fusion Output:

[0088] 6-dimensional probability vector: p test = [p0, p1, ..., p5], where p k The signal to be diagnosed belongs to C k Confidence level;

[0089] Fault category determination: C pred =arg max k p k If max p k If the value is less than 0.7, output "Suspected fault";

[0090] 2) Interpretability results

[0091] SHAP Feature Contribution Heatmap: Labeled F real_fusion The physical meaning of the top 10 SHAP values ​​in the data;

[0092] Physical parameter matching report: Calculate the matching degree (Match) between the impact amplitude of the signal to be diagnosed and the theoretical ΔH. If Match < 0.6, it indicates "the feature deviates significantly from the physical model".

[0093] The advantages and beneficial effects of this invention are as follows:

[0094] This invention boasts significant innovation in its technical solution, with its core advantages lying in the end-to-end optimization of data utilization, feature processing, and model training. Firstly, it breaks through the strong dependence of traditional fault diagnosis on measured data. By constructing a two-degree-of-freedom dynamic model to generate simulated vibration signals and designing a main-auxiliary task collaborative data acquisition mode, it combines simulated and measured signals. This not only supplements scarce fault samples (especially composite fault samples) but also injects prior knowledge through the physical model, addressing the pain points of difficult acquisition and high annotation costs of industrial field fault data. Secondly, it employs a lightweight convolutional-KAN hybrid network to extract four types of features, taking into account the physical and data features of simulated signals and the corresponding features of measured signals. This avoids the shortcomings of traditional single-network or signal processing techniques in capturing nonlinear fault information. Simultaneously, through a priori-guided multi-layer attention fusion module, it achieves channel attention aggregation of key features and cross-signal attention alignment of simulated and measured features. High-discriminative features are then selected using the SHAP method, solving the problems of redundancy, alignment discrepancies, and subjective selection in traditional feature fusion, significantly improving feature quality. Third, we innovated a hybrid loss function of “cross-entropy + contrastive + physical constraint MSE”, with actual diagnostics as the main task and simulated classification and feature alignment as auxiliary tasks. This not only avoids overfitting in single-task training, but also ensures that the features conform to the dynamic laws of bearings through physical constraints, thus breaking through the limitation of poor interpretability of traditional “black box” models.

[0095] The beneficial effects of this invention can be seen in both industrial applications and diagnostic performance. Regarding industrial adaptability, the lightweight network architecture reduces the number of parameters and computational costs, allowing deployment on embedded devices to meet real-time monitoring needs. Strong interpretability, achieved through the SHAP method, quantifies feature contribution, helping maintenance personnel understand the diagnostic basis, increasing trust in the diagnostic results, and solving the problem of difficult engineering implementation of traditional models. In terms of diagnostic performance, multi-task collaboration and prior-deep feature fusion enable the model to maintain high accuracy even under conditions of scarce measured data and strong noise interference. It is suitable for single-fault diagnosis of rolling bearings and can accurately identify compound faults, effectively reducing equipment downtime risks and maintenance costs. Furthermore, this method eliminates the need for manually pre-setting feature extraction rules, achieving end-to-end diagnosis and simplifying industrial operation and maintenance processes. Compared to traditional methods, it significantly improves data dependency, diagnostic accuracy, generalization ability, and engineering practicality, providing an efficient and reliable technical solution for fault diagnosis of industrial rotating machinery bearings. Attached Figure Description

[0096] Figure 1 This is a flowchart of a preferred embodiment of the method provided by the present invention;

[0097] Figure 2 This is a schematic diagram of the feature extractor structure;

[0098] Figure 3 This is a schematic diagram of a feature fusion multi-module attention structure;

[0099] Figure 4 This is a schematic diagram of the fault diagnosis network structure. Detailed Implementation

[0100] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0101] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0102] A bearing multi-task fault diagnosis method based on prior knowledge and deep feature fusion includes the following steps:

[0103] S1. Collaborative Data Acquisition of Main and Auxiliary Tasks: A dynamic model of a composite fault in a rolling bearing system is constructed. This model is based on a two-degree-of-freedom dynamic equation to build a simulation model. The dynamic equation is solved using the Runge-Kutta algorithm to generate simulated vibration signals for single and composite faults. Vibration signals of rolling bearing faults under different conditions are collected using sensors to obtain measured signal samples.

[0104] S2. Signal Processing and Feature Extraction: The simulated vibration signal and the measured vibration signal undergo unified preprocessing. Feature extraction is then performed using a lightweight convolutional-KAN hybrid network with residual connections and feature normalization to extract the physical information features and nonlinear data features of the two signals, respectively.

[0105] S3. Prior-guided feature fusion: Construct a prior-guided multi-layer attention fusion module, input the extracted physical information features of two signals and nonlinear data features, and then aggregate single signal features through channel attention, achieve simulation-measured feature alignment through cross-signal source attention, and sort and filter highly discriminative fusion features.

[0106] S4. Collaborative Training and Diagnosis of Main and Auxiliary Tasks: The main task is fault diagnosis of measured signals, while the auxiliary tasks are classification of simulated signals and alignment of multiple signal features. The KAN fault diagnosis model is trained using a hybrid loss function of "cross-entropy + contrastive + physical constraint MSE" to output fault categories and interpretability results.

[0107] Furthermore, step S1 specifically includes the following steps:

[0108] ① Construction of a two-degree-of-freedom dynamic model for rolling bearings

[0109] Single faults (outer ring, inner ring, rolling element faults) and compound faults (such as combinations of inner ring + outer ring, inner ring + rolling element, etc.) in rolling bearings are simplified into cuboid pit defects. Defect parameters are defined as: length L, width B, and depth H. The maximum displacement excitation ΔH of the fault is calculated based on geometric relationships.

[0110]

[0111] Where D is the diameter of the rolling element.

[0112] The outer ring fault displacement excitation function of the bearing is expressed as:

[0113]

[0114] Where θ j The rolling element position angle, Let Δθ be the initial angle. o The central angle of the fault area.

[0115] The inner ring fault displacement excitation function of the bearing is expressed as:

[0116]

[0117] Where Δθ i Let the central angle of the fault area be... The function of time t is expressed as:

[0118] The rolling element fault displacement excitation function of the bearing is expressed as:

[0119] Other situations

[0120] Where ω b ω is the angular velocity of the rolling element.

[0121] The normal contact deformation between the raceway and the rolling elements incorporates the influence of the bearing fault displacement excitation function. The contact deformation can be expressed as:

[0122] δ j (t)=x(t)cosθ j +y(t)sinθ j -δ0-Ho j -Hl j -Hb j

[0123] Where x and y represent displacements in the X and Y directions, respectively, and are excited by Ho containing fault displacement. j Hl j Hb j The presence or absence of [something] is used to simulate various types of failures in rolling bearings.

[0124] Based on Hertz contact theory, the contact force and deformation satisfy a 3 / 2 power nonlinear relationship. This relationship only holds if δ... j When (t)>0, the rolling element is in contact with the inner and outer rings, and the contact force is:

[0125] F j =μ j K·max(δ j (t),0) 3 / 2

[0126] Where K: equivalent contact stiffness, μ j Contact judgment coefficient (fault location μ) j =0, contact force disappears; normal position μ j =1, contact force effective);

[0127] By decomposing the contact force in the X and Y directions, its components are:

[0128] F j,x =F j cosθ j ,F j,y =F j sinθ j

[0129] Therefore, the two-degree-of-freedom fault model of a rolling bearing is:

[0130]

[0131] Where m represents the bearing mass, C is the equivalent damping, and W... x W y These are external stimuli in the X and Y directions.

[0132] ②Simulated data generation and measured data acquisition

[0133] The dynamic equations are solved using the Runge-Kutta algorithm to generate simulated vibration signals for single and combined faults.

[0134] The two-degree-of-freedom dynamic equations of the faulty bearing are transformed into a system of first-order differential equations, and the state vector is defined. The system of first-order differential equations is expressed as follows:

[0135]

[0136] Given the initial condition z(t0) = z0, and setting the time step h, solve for z(t) using the fourth-order Runge-Kutta iterative formula within the time interval [t0, T].

[0137]

[0138] Where t n =t0+nh(n=0,1,…,N-1), N=[(T-t0) / h].

[0139] Finally, x(t) and y(t) are extracted from the obtained state vector z(t) to serve as simulated vibration signals of the rolling bearing under single faults (outer ring, inner ring, and rolling element faults) and compound faults (superposition of fault displacement excitation functions).

[0140] Then, sensors are used to collect measured vibration signals of rolling bearings under different fault conditions (single fault + compound fault), and the fault type and operating parameters are recorded simultaneously.

[0141] Furthermore, in S2, a lightweight convolutional-KAN hybrid network with residual connections and feature normalization is used to extract the physical information features and nonlinear data features of the two signals, specifically:

[0142] ① Feature preprocessing

[0143] Adaptive wavelet denoising: db4 wavelet basis decomposition with 6 layers, Stein unbiased risk estimation soft thresholding for high-frequency layers (5-6 layers), and hard thresholding for low-frequency layers (1-4 layers);

[0144] Min-Max normalization:

[0145]

[0146] Where: x min / x max To find the global maximum and minimum values ​​in the simulated and measured training sets, output a standardized analog signal X. sim Standardized measured signal X real ;

[0147] ② Four types of feature extraction (convolutional-KAN hybrid network, dual-domain parameter sharing)

[0148] The design of 1D convolutional blocks for extracting physical information features involves the following steps:

[0149] Convolutional block 1: kernel=16, stride=2, output channels 32 → BatchNorm → LeakyReLU → 1×1 convolutional residual connection;

[0150] Convolutional block 2: kernel=8, stride=2, output channels 64 → BatchNorm → LeakyReLU → 1×1 convolutional residual connection;

[0151] Output physical information characteristics: Analog signal physical information characteristics F sim_phy Includes fault frequency, contact force characteristics, and physical information features of measured signals F real_phy ;

[0152] The specific steps for designing a KAN layer to extract data features are as follows:

[0153] Convolution output flattening (dimension 64×256=16384) → KAN layer 1 (input 16384 → output 128, B-spline order 3, number of nodes 10, node range [0,0.5D]) → KAN layer 2 (input 128 → output 128, L1 sparse regularization);

[0154] Global Average Pooling (GAP): Output data characteristics: Analog signal data characteristics F sim_data Measured signal data characteristics F real_data ;

[0155] Furthermore, in S3, the channel attention submodule construction and single feature aggregation method are specifically as follows:

[0156] The channel attention submodule contains a serial structure of "Global Average Pooling (GAP) + 2 fully connected layers + Sigmoid activation", wherein: the first fully connected layer has the same input dimension as the feature to be aggregated (both are 64-dimensional), the output dimension is 32-dimensional, and the activation function is ReLU; the second fully connected layer has an input dimension of 32-dimensional, an output dimension of 64-dimensional, and the activation function is Sigmoid, which is used to output channel weights;

[0157] The four types of features are then input into the channel attention submodule, where the channel weights of each feature are calculated and weighted aggregation is performed. The specific process is as follows:

[0158] (1) Physical information characteristics F of analog signals sim_phy (Dimension([M,64]), where M is the number of simulated samples), first calculate the global mean of each channel using GAP, expressed by the formula:

[0159]

[0160] Where: c = 1, 2, ..., 64 are the channel indices;

[0161] (2) GAP(F) sim_phy Input the channel attention submodule and calculate the channel weights W. ch_sim_phy , is represented as:

[0162] W ch_sim_phy =Sigmoid(W2·ReLU(W1·GAP(F)) sim_phy )+b1)+b2)

[0163] In the formula The first fully connected layer can learn weights and biases. The second fully connected layer can learn weights and biases;

[0164] (3) The aggregated physical information features of the analog signal are obtained through weighted aggregation, and are expressed as follows:

[0165] F sim_phy_agg =F sim_phy ⊙W ch_sim_phy

[0166] Where ⊙ represents element-wise multiplication, with the dimension [M,64] preserved;

[0167] (4) Repeat steps (1) to (3) above, respectively, for the analog signal data features F sim_data Physical information characteristics of the measured signal F real_phy Measured signal data characteristics F real_data Execute channel attention aggregation, output:

[0168] 1) Characteristics of Analog Signal Data Aggregation

[0169] F sim_data_agg =F sim_data ⊙W ch_sim_data

[0170] Among them, W ch_sim_data The corresponding channel weights are defined in dimension [M, 64].

[0171] 2) Aggregation characteristics of physical information of measured signals

[0172] F real_phy_agg =F real_phy ⊙W ch_real_phy

[0173] Among them, W ch_real_phy The corresponding channel weights are defined in dimension [N, 64], where N is the number of measured samples.

[0174] 3) Aggregation characteristics of measured signal data

[0175] F real_data_agg =F real_data ⊙W ch_real_data

[0176] Among them, W ch_real_data The corresponding channel weights are defined in dimension [N, 64].

[0177] Then, the weights of the prior guidance channels are optimized, and the two types of aggregated features F of the analog signal are further optimized. sim_phy_agg F sim_data_agg Based on the fault physical parameters calculated by S1 (maximum fault displacement excitation ΔH, contact force F) j Adjusting channel weights is represented as follows:

[0178] If a certain channel characteristic has a Pearson correlation coefficient Corr(F) with ΔH, then... ch If ΔH)≥0.7, then the channel weight is within the original W... ch Increase by 15% on the basis;

[0179] If a certain channel characteristic is similar to F j Pearson correlation coefficient Corr(F) ch ,F j If the value is greater than or equal to 0.7, then the channel weight is within the original W range. ch Increase by 10% on the basis;

[0180] The weights after adjustment need to be normalized: Ensure that the sum of the weights is 1.

[0181] Furthermore, in S3, the cross-signal attention submodule construction and feature alignment, and the prior-guided simulation-test feature matching method, are specifically as follows:

[0182] (1) First, feature concatenation of the two signals is performed. The aggregated features of the same signal type are concatenated in the order of "physical information features + data features" to obtain two types of cross-modal features:

[0183] 1) Cross-modal characteristics of analog signals: F sim_cross =[F sim_phy_agg ,F sim_data_aggThe dimensions are concatenated to form [M, 128] (64 dimensions + 64 dimensions);

[0184] 2) Measured signal cross-modal characteristics: F real_cross =[[F real_phy_agg ,F real_data_agg The dimensions are concatenated to [M, 128].

[0185] (2) Next, calculate the similarity matrix of the two fused signals, and design the cross-signal attention weights based on cosine similarity. The formula is as follows:

[0186]

[0187] In the formula, i = 1, 2, ..., M are the simulated sample indices, j = 1, 2, ..., N are the measured sample indices, and d k =128 represents the cross-modal feature dimension (temperature coefficient). (used to avoid similarity value saturation) This is the cross-signal attention weight matrix.

[0188] (3) Then perform prior guidance weight adjustment, and dynamically adjust W according to the matching degree of physical parameters between the analog signal and the measured signal. cross Specifically:

[0189] Calculate ΔH for simulated sample i sim (i) Impact amplitude Amp of measured sample j real (j) by F real_phy_agg The matching degree (calculated from the time-domain peak value) is expressed by the formula:

[0190]

[0191] If Match(i,j)≥0.8, then W cross (i,j)=W cross (i,j)×1.2 (Weights of enhanced physical matching samples).

[0192] (4) Finally, cross-signal feature alignment and fusion are performed, based on the adjusted W. cross Aligning the measured signal features with the physical prior of the analog signal, the formula is expressed as:

[0193] F align (i,∶)=W cross (i,∶)·F real_cross +F sim_cross (i,∶)

[0194] In the formula 256The aligned feature matrix (128-dimensional simulated cross-modal fusion features + 128-dimensional measured cross-modal fusion features) is used when the measured sample features are incomplete (e.g., insufficient data). W cross (i,j) automatically increases the proportion of simulated features and supplements physical priors.

[0195] Furthermore, in S3, the SHAP feature ranking submodule construction and high-discriminative feature selection method are specifically as follows:

[0196] (1) First, the SHAP interpreter is configured, using TreeExplainer as the SHAP value calculation tool. The interpreter is built based on the feedforward structure (feature extractor + fusion module) of the KAN fault diagnosis model, and the parameters are set as follows:

[0197] 1) A tree depth of 5 layers is used to match the non-linear fitting complexity of the KAN layer;

[0198] 2) The sample size is 1000, and F is randomly selected. align 1000 samples were used for SHAP value calculation, balancing accuracy and efficiency;

[0199] 3) The feature association method adopts a multi-classification association mode of "feature-fault category" to calculate the independent contribution of each feature to the 6 types of faults;

[0200] (2) Further calculate and quantify the SHAP value and its feature contribution: For F align For each feature channel c (c = 1, 2, ..., 256), calculate its mean absolute SHAP value as an indicator of feature contribution. The formula is as follows:

[0201]

[0202] In the formula, C = 6 represents the number of fault categories, and SHAP(F align (i,:),c,k) represents the SHAP value of the c-th feature of the i-th sample for the fault diagnosis result of the k-th class.

[0203] (3) Perform feature sorting and filtering again.

[0204] 1) Press SHAP mean (c) Sort the 256-dimensional feature channels in descending order to obtain the feature sort index, represented as: Idx = sort(SHAP mean (c),descend);

[0205] 2) Retain the first 128 features in the sorting, i.e., F fusion =F align (:,Idx(1:128)), with dimensions [M,128];

[0206] (4) Finally, the results of the prior verification screening are performed. For the 128 retained features, their correlation with the physical parameters is verified:

[0207] 1) Calculate the relationship between each retained feature and ΔH, F j The Pearson correlation coefficient requires that ≥80% of the features satisfy Corr(F) fusion (∶,c),ΔH)≥0.7 or Corr(F) fusion (∶,c),F j ≥0.7;

[0208] 2) If not satisfied, readjust the sample size of the SHAP interpreter (increase to 1500) or the weight parameters of the channel attention submodule until the verification is successful.

[0209] Furthermore, in S4, the primary-secondary task collaborative training and diagnosis method specifically includes:

[0210] (1) Definition of main and auxiliary tasks and input feature matching

[0211] 1) The specific implementation method for associating task objectives with input features is as follows:

[0212] Main task: The objective is to output six states of rolling bearings (normal state C0, inner ring single fault C1, outer ring single fault C2, rolling element single fault C3, inner ring + outer ring combined fault C4, inner ring + rolling element combined fault C5). The input feature is the fused feature F of the measured signal output by S3. real_fusion (Dimension [N, 128], where N is the number of measured samples, obtained by attention fusion and SHAP sorting of two types of features of the measured signal);

[0213] Auxiliary Task 1: The objective is to output eight types of fault states for rolling bearings (five single faults: C0, C1, C2, C3, and cage fault C6; three compound faults: C4, C5, and outer ring + rolling element compound fault C7). The input feature is the fusion feature F of the analog signal generated synchronously in step S3. sim_fusion (Dimension [M, 128], where M is the number of simulated samples), this feature is obtained by fusing two types of features of the simulated signal with the same attention + SHAP sorting;

[0214] Auxiliary Task 2: The objective is to minimize the difference in feature distribution between the analog and measured signals. The input features are the cross-modal feature pairs output from step S3 (Analog side F). sim_cross , measured side F real_cross (All are 128-dimensional).

[0215] 2) The sample pairing rules are implemented using the following method:

[0216] During the training phase, a "pairing of similar faults first" strategy is adopted: for each measured sample x resl ∈C k (k = 0, 1, ..., 5), randomly match two simulated samples x of the same type. sim_pos ∈C k (Positive sample) and 1 out-of-class simulated sample x sim_neh ∈C k′ (k′≠k, negative samples) ensures that the auxiliary task provides targeted prior guidance for the main task.

[0217] (2) Construction of KAN network diagnostic model

[0218] 1) The KAN network diagnostic model described above adopts a "shared feature mapping layer + task-specific classification head" structure, as follows:

[0219] Input feature F real_fusion or F sim_fusion 128-dimensional fusion features;

[0220] Next, we enter the shared KAN backbone layer (which implements non-linear feature mapping). This layer is the core feature processing unit of the model. It achieves non-linear compression and physical association enhancement of 128-dimensional input features through a 2-layer KAN network. Batch normalization and regularization are added simultaneously to ensure generalization. Specifically, it includes 3 sub-modules:

[0221] KAN layer 1: Input 128-dimensional → Output 64-dimensional, activation function is 3rd order B-spline function (8 nodes, node range [0,1])

[0222] BatchNorm layer: Performs batch normalization on the output of KAN layer 1.

[0223] KAN layer 2: Input 64-dimensional → Output 32-dimensional, with L2 regularization (weight decay coefficient 1e-4).

[0224] Task-specific category headers: This layer features independently designed category units for different tasks, with parameters not shared, ensuring adaptability to the different output category requirements of main and auxiliary tasks.

[0225] Main task classification header: 32-dimensional → fully connected layer (output 6-dimensional) → Softmax activation, outputting 6 types of probabilities;

[0226] Auxiliary Task 1 Classification Head: 32-dimensional → Fully connected layer (output 8-dimensional) → Softmax activation, outputting 8 class probabilities.

[0227] 2) Parameter initialization of the diagnostic network

[0228] KAN layer B-spline nodes: based on fault physical parameters ΔH([0,0.5D]) and contact force F jThe normalization range setting of ([0,1e4]N) ensures that the node covers the range of physical feature values;

[0229] Fully connected layer weights: initialized using He normal distribution The bias is initialized to 0.01.

[0230] (3) Calculation of hybrid loss function (multi-objective collaborative optimization)

[0231] The hybrid loss function is a weighted fusion of the primary and secondary task losses, and the formula is as follows:

[0232] L total =0.6L main +0.2L aux1 +0.2L aux2

[0233] The calculation methods for each sub-loss are as follows:

[0234] 1) Main task loss L main (Cross-entropy loss)

[0235]

[0236] In the formula: y i,k The one-hot label (y) of the measured sample i i,k =1 indicates that sample i belongs to C k ;p main (i,k) represents the model pair where sample i belongs to C. k The predicted probability; ∈ = 1e-10; N is the number of actual samples.

[0237] 2) Auxiliary task 1 loss L aux1 (Cross-entropy loss)

[0238]

[0239] In the formula: p represents the one-hot label of the simulated sample j; aux1 (h,k) represents auxiliary task 1, where sample j belongs to C. k The predicted probability; M is the number of simulated samples.

[0240] 3) Auxiliary Task 2 Loss L aux2 (Contrastive loss + Physically Constrained MSE loss)

[0241] L aux2 =L Contrastive +L Phys-MSE

[0242] Contrastive loss L Contrastive

[0243]

[0244] In the formula: F sim_pos For similar simulated cross-modal features, f sim_neg To simulate cross-modal features for heterogeneous models; α is the squared L2 norm; α = 0.5 is the marginal parameter.

[0245] Physical constraint MSE loss L Phys-MSE

[0246]

[0247] In the formula: (Min-Max normalization);

[0248] (4) Model training optimization strategy

[0249] 1) Optimizer configuration

[0250] The AdamW optimizer is used with the following parameters:

[0251] Initial learning rate: η0 = 1e-3

[0252] Weight decay coefficient: λ = 1e -4

[0253] Momentum parameters: β1 = 0.9, β2 = 0.999.

[0254] 2) Learning rate scheduling and early stopping strategy

[0255] Learning rate scheduling: A cosine annealing strategy is used, and the formula is as follows: (t represents the current round, T = 100 represents the maximum round);

[0256] Early stopping strategy: Monitor the accuracy of the validation set of the main task (the validation set accounts for 20% of the actual test data). If the improvement is ≤0.1% for 5 consecutive rounds, save the current best model.

[0257] 3) Parameter update method

[0258] L is calculated using automatic differentiation. total For the gradient of all network parameters (KAN layer coefficients, classifier head weights, attention weights), update the parameters according to the gradient descent rule, and iterate until the early stopping condition is met.

[0259] (5) Diagnostic results output (including interpretable information)

[0260] 1) Core diagnostic results

[0261] The input measured signal to be diagnosed (after S2 preprocessing + S3 fusion to obtain F) real_fusion Output:

[0262] 6-dimensional probability vector: p test = [p0, p1, ..., p5], where p k The signal to be diagnosed belongs to C k Confidence level;

[0263] Fault category determination: C pred =arg max k p k If max p k If the value is less than 0.7, output "Suspected fault".

[0264] 2) Interpretability results

[0265] SHAP Feature Contribution Heatmap: Labeled F real_fusion The physical meaning of the top 10 features of the SHAP value (e.g., "Channel 28: Contact force characteristics corresponding to inner ring fault ΔH = 0.3mm");

[0266] Physical parameter matching report: Calculates the degree of matching between the impulse amplitude of the signal to be diagnosed and the theoretical ΔH. If Match < 0.6, the message "Features deviate significantly from the physical model" will be displayed.

[0267] To address the challenges of fault diagnosis in rolling bearings of rotating machinery due to the mismatch between feature extraction and actual fault mechanisms, as well as the limited sample data, a multi-task fault diagnosis method for bearings based on prior knowledge and deep feature fusion is proposed. The technical solution of this invention is as follows:

[0268] A multi-task fault diagnosis method for bearings based on the fusion of prior knowledge and deep features is proposed. The process framework of this method is as follows: Figure 1 As shown, it includes the following steps:

[0269] First, a composite fault dynamic model of a rolling bearing system is constructed. A simulation model is built based on the two-degree-of-freedom dynamic equations, and the dynamic equations are solved using the Runge-Kutta algorithm to generate simulated vibration signals for single and composite faults. Then, sensors are used to collect rolling bearing fault vibration signals under different conditions, obtaining measured signal samples. The simulated and measured vibration signals are then subjected to unified preprocessing for feature extraction. A lightweight convolutional-KAN hybrid network with residual connections and feature normalization is used to extract physical information features and nonlinear data features from both types of signals. Next, a priori-guided multi-layer attention fusion module is constructed, inputting the extracted physical information features and nonlinear data features from both types of signals. Channel attention aggregates single signal features, cross-source attention aligns simulated and measured features, and feature sorting filters for highly discriminative fusion features. Finally, with measured signal fault diagnosis as the primary task and simulated signal classification and multi-signal feature alignment as auxiliary tasks, a KAN fault diagnosis model is trained using a hybrid loss function of "cross-entropy + contrastive + physical constraint MSE," outputting fault categories and interpretability results.

[0270] The specific steps include:

[0271] Step 1: Data Acquisition and Processing in a Collaborative Main-Support Task System

[0272] 1. Simulated signal generation: Based on the two-degree-of-freedom dynamic equation of rolling bearings, the equation is solved using the Runge-Kutta algorithm to generate simulated vibration signals for single and compound faults.

[0273] 2. Measured signal acquisition: Using a rolling bearing data acquisition test bench, collect rolling bearing fault vibration signals under different conditions (including normal conditions, various fault conditions, etc.) and obtain measured signal samples.

[0274] 3. Signal preprocessing: The generated simulated vibration signal and the acquired measured vibration signal are subjected to unified preprocessing operations, mainly including denoising (removing noise interference from the signal and improving signal quality) and normalization (mapping the signal values ​​to a specific range to facilitate subsequent processing and analysis), and finally the preprocessed simulated signal (s1) and measured signal (s2) are obtained.

[0275] Step 2: Feature Extraction

[0276] A lightweight convolutional-KAN hybrid network with residual connections and feature normalization is used to extract features:

[0277] 1. Perform preliminary feature extraction and feature enhancement on the input signal (analog signal (s1) or measured signal (s2)) through convolutional blocks and residual blocks to obtain intermediate features containing physical information.

[0278] 2. The features are converted into one-dimensional vector form through a flattening layer.

[0279] 3. The features are further processed through the KAN-1 and KAN-2 modules in sequence to uncover deeper feature information.

[0280] 4. The features are aggregated and dimensionality reduced using an average pooling layer to extract the physical information features (F) of the two signals respectively. phy ) and nonlinear data features (F data ).

[0281] Step 3: Prior-guided feature fusion

[0282] The two signal physical information features extracted from the input (F) phy ) and nonlinear data features (F data ), to perform multi-layered attention fusion guided by prior knowledge:

[0283] 1. Channel Attention and Single Feature Aggregation: Through the channel attention submodule, channel-level attention weighting is applied to the features (physical information features and nonlinear data features) of a single signal to aggregate the effective features of the single signal.

[0284] 2. Cross-signal attention and feature alignment: By utilizing the cross-signal source attention submodule, the alignment between the features of analog signals and measured signals is achieved, allowing the features of measured signals to align with the physical priors of analog signals, thus solving the problem of mismatch between analog and measured signal features.

[0285] 3. SHAP Feature Ranking and Filtering: Using the SHAP feature ranking submodule, the feature contribution of the aligned features is calculated and ranked, and high-discriminatory fusion features are filtered out, ultimately obtaining the fusion features (F) of the measured signal. real_fusion The fusion characteristics of analog and analog signals (F) sim_fusion ).

[0286] Step 4: Primary-Secondary Task Collaborative Training and Diagnosis

[0287] fused features of measured signals (F) real_fusion ) and analog signal fusion features (F sim_fusion Input subsequent network:

[0288] 1. The features are further processed and abstracted through KAN layer-1 and KAN layer-2.

[0289] 2. Taking the fault diagnosis of measured signals as the main task and the analog signal classification and multi-signal feature alignment as auxiliary tasks, the system outputs the 6 types of probabilities of the main task (corresponding to fault category judgment) and the 8 types of probabilities of the auxiliary tasks (corresponding to analog signal classification and other auxiliary judgments) through fully connected layers and Softmax functions, respectively, and finally obtains the fault category and interpretability results.

[0290] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0291] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0292] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A bearing multi-task fault diagnosis method based on the fusion of prior knowledge and deep features, characterized in that, Includes the following steps: S1. Collaborative Data Acquisition of Main and Auxiliary Tasks: Construct a dynamic model of a composite fault in a rolling bearing system, i.e., build a simulation model based on a two-degree-of-freedom dynamic equation, solve the dynamic equation using the Runge-Kutta algorithm, and generate simulated vibration signals of single and composite faults; collect rolling bearing fault vibration signals under different conditions using sensors to obtain measured signal samples. S2. Signal Processing and Feature Extraction: The simulated vibration signal and the measured vibration signal are subjected to unified preprocessing; then feature extraction is performed on them. A lightweight convolutional-KAN hybrid network with residual connections and feature normalization is used to extract the physical information features and nonlinear data features of the two signals respectively. S3. Prior-guided feature fusion: Construct a prior-guided multi-layer attention fusion module, input the extracted physical information features of two signals and nonlinear data features, and then aggregate single signal features through channel attention, achieve simulation-measured feature alignment through cross-signal source attention, and sort and filter highly discriminative fusion features. S4. Collaborative Training and Diagnosis of Main and Auxiliary Tasks: The main task is fault diagnosis of measured signals, while the auxiliary tasks are classification of simulated signals and alignment of multiple signal features. The KAN fault diagnosis model is trained using a hybrid loss function of "cross-entropy + contrastive + physical constraint MSE" to output fault categories and interpretability results.

2. The bearing multi-task fault diagnosis method based on prior knowledge and deep feature fusion as described in claim 1, characterized in that, S1 specifically includes the following steps: ① Construction of a two-degree-of-freedom dynamic model for rolling bearings: Single and compound faults in rolling bearings are simplified as cuboid pit defects. Single faults in rolling bearings include faults in the outer ring, inner ring, and rolling elements. Compound faults include combinations of inner ring + outer ring and inner ring + rolling elements. Defect parameters are defined as: length L, width B, and depth H. Based on geometric relationships, the maximum displacement excitation ΔH of the fault is calculated as follows: Where D is the diameter of the rolling element; Based on Hertz contact theory, a two-degree-of-freedom fault model for rolling bearings is constructed as follows: Where m represents the bearing mass, C is the equivalent damping, and W... x W y External excitations in the X and Y directions; in, This represents the acceleration of the bearing in the X direction, describing the rate of change of the bearing's velocity in the X direction. μ represents the acceleration of the bearing in the Y direction, i.e., the rate of change of the bearing's velocity in the Y direction; K is the contact stiffness, which reflects the ability of the rolling bearing's contact part to resist deformation, and is related to factors such as the bearing's material and geometry. j : is a coefficient related to the j-th rolling element, involving the influence of the rolling element's position and contact state on the contact force; δ j (t): Represents the contact deformation between the j-th rolling element and the inner and outer rings of the bearing. Contact force is generated only when the contact deformation is greater than 0; θ j : is the contact angle of the j-th rolling element, used to describe the angular relationship between the rolling element and the contact point between the inner and outer rings, and affects the components of the contact force in the X and Y directions; ②Simulated data generation and measured data acquisition The dynamic equations are solved using the Runge-Kutta algorithm to generate simulated vibration signals for single and combined faults. Then, sensors are used to collect measured vibration signals of rolling bearings under different fault states, i.e., single fault + combined fault, and the fault type and operating parameters are recorded simultaneously.

3. The bearing multi-task fault diagnosis method based on prior knowledge and deep feature fusion according to claim 1, characterized in that, S2 specifically includes the following steps: ① Feature preprocessing: Adaptive wavelet denoising: db4 wavelet basis decomposition with 6 layers, Stein unbiased risk estimation with soft thresholding in the 5th-6th layers of high frequency, and hard thresholding in the 1st-4th layers of low frequency. Min-Max normalization: Where: x min / x max To find the global maximum and minimum values ​​in the simulated and measured training sets, output a standardized analog signal X. sim Standardized measured signal X real ; ② Four types of feature extraction: Convolutional-KAN hybrid network, with shared parameters in both domains; Design 1D convolutional blocks to extract physical information features, and obtain the physical information features F of the analog signal. sim_phy Includes fault frequency, contact force characteristics, and physical information features of measured signals F real_phy ; The KAN layer is designed to extract data features to obtain the analog signal data features F. sim_data Measured signal data characteristics F real_data .

4. The bearing multi-task fault diagnosis method based on prior knowledge and deep feature fusion as described in claim 1, characterized in that, The construction of the prior-guided multi-layer attention fusion module in S3 specifically includes the following steps: ① Channel attention submodule construction and single feature aggregation: The channel attention submodule comprises a serial structure of "global average pooling (GAP) + 2 fully connected layers + sigmoid activation", which then applies the analog signal data features F to... sim_data Physical information characteristics of the measured signal F real_phy Measured signal data characteristics F real_data Input the channel attention submodule, calculate the channel weights of each feature and perform weighted aggregation, output: 1) Characteristics of analog signal data aggregation: F sim_data_agg =F sim_data ⊙W ch_sim_data Among them, W ch_sim_data The corresponding channel weights are defined in dimension [M, 64]. 2) Aggregation characteristics of physical information of measured signals: F real_phy_agg =F real_phy ⊙W ch_real_phy Among them, W ch_real_phy The corresponding channel weights are defined in dimension [N, 64], where N is the number of measured samples. 3) Aggregation characteristics of measured signal data F real_data_agg =F real_data ⊙W ch_real_data Among them, W ch_real_data The corresponding channel weights are defined in dimension [N, 64]. Then, the weights of the prior guidance channels are optimized, and the two types of aggregated features F of the analog signal are further optimized. sim_phy_agg F sim_data_agg The fault physical parameters calculated based on S1 are the maximum fault displacement excitation ΔH and the contact force F. j Adjust the channel weights. After adjustment, the weights need to be normalized to ensure that the sum of the weights is 1. ② Construction and feature alignment of cross-signal attention submodules, prior-guided simulation-test feature matching (1) First, feature splicing of the two signals is performed. The aggregated features of the same signal type are spliced ​​in the order of "physical information features + data features" to obtain two types of cross-modal features: analog signal cross-modal features and measured signal cross-modal features; (2) Then perform similarity matrix calculation for the two fused signals and design cross-signal attention weights based on cosine similarity; (3) Then perform prior guidance weight adjustment, and dynamically adjust W according to the matching degree of physical parameters between the analog signal and the measured signal. cross ; (4) Finally, cross-signal feature alignment and fusion are performed, based on the adjusted W. cross Align the measured signal features with the physical prior of the analog signal; ③ Construct the SHAP feature ranking submodule and select high-discriminative features: (1) First, the SHAP interpreter is configured, and TreeExplainer is used as the SHAP value calculation tool. The interpreter is built based on the feedforward structure of the KAN fault diagnosis model, namely the feature extractor + fusion module. (2) Then, the SHAP value is calculated and the feature contribution is quantified, and the F align For each feature channel c (c = 1, 2, ..., 256), calculate its mean absolute SHAP value as an indicator of feature contribution. (3) Then perform feature sorting and filtering; (4) Finally, the results of the prior verification screening are carried out. The correlation between the 128-dimensional features and the physical parameters is verified.

5. The bearing multi-task fault diagnosis method based on prior knowledge and deep feature fusion according to claim 1, characterized in that, The main-auxiliary task collaborative training and diagnosis in S4 specifically includes the following steps: (1) Definition of main and auxiliary tasks and input feature matching Main task: The objective is to output six states of rolling bearings, including normal state C0, inner ring single fault C1, outer ring single fault C2, rolling element single fault C3, inner ring + outer ring combined fault C4, and inner ring + rolling element combined fault C5. The input feature is the fused feature F of the measured signal output by S3. real_fusion The dimension is [N, 128], where N is the number of measured samples, which is obtained by attention fusion and SHAP sorting of the two types of features of the measured signal; Auxiliary Task 1: The objective is to output eight types of fault states for rolling bearings, including five single faults: C0, C1, C2, C3, and cage fault C6; and three compound faults: C4, C5, and outer ring + rolling element compound fault C7. The input feature is the fusion feature F of the analog signal synchronously generated in step S3. sim_fusion The dimension is [M, 128], where M is the number of simulated samples. This feature is obtained by fusing two types of features of the simulated signal with the same attention and sorting by SHAP. Auxiliary Task 2: The objective is to minimize the difference in feature distribution between the analog and measured signals. The input features are the cross-modal feature pairs output from step S3, i.e., the analog side F. sim_cross , measured side F real_cross All are 128-dimensional; (2) Construction of KAN network diagnostic model The KAN network diagnostic model described above adopts a "shared feature mapping layer + task-specific classification head" structure, with input feature F real_fusion or F sim_fusion The fusion features are then entered into the shared KAN backbone layer to achieve nonlinear feature mapping. This layer is the core feature processing unit of the model. Through a 2-layer KAN network, nonlinear compression and physical association enhancement of 128-dimensional input features are achieved. Batch normalization and regularization are added simultaneously to ensure generalization. Task-specific category headers: This layer designs independent category units for different tasks, with parameters not shared, to ensure adaptation to the different output category requirements of main and auxiliary tasks, and constructs a main task category header and an auxiliary task 1 category header; The diagnostic network parameters are initialized using KAN layer B-spline nodes, based on the fault physical parameters ΔH([0,0.5D]) and contact force F. j The normalization range of ([0,1e4]N) is set to ensure that the nodes cover the range of physical feature values; the weights of the fully connected layers are initialized using He normality. The bias is initialized to 0.01; (3) Hybrid loss function calculation, i.e., multi-objective collaborative optimization: The hybrid loss function is a weighted fusion of the primary and secondary task losses, and the formula is as follows: L total <0.6L main +0.2L aux1 +0.2L aux2 The calculation methods for each sub-loss are as follows: 1)L main Main task loss, L aux1 To mitigate the losses of Task 1, L aux2 The loss is for auxiliary task 2, namely contrastive loss + physical constraint MSE loss; (4) Model training optimization strategy: 1) Optimizer configuration The AdamW optimizer is used with the following parameters: Initial learning rate: η0 = 1e-3 Weight decay coefficient: λ = 1e -4 Momentum parameters: β1 = 0.9, β2 = 0.

999. 2) Learning rate scheduling and early stopping strategy Learning rate scheduling: Cosine annealing strategy is adopted; Early stopping strategy: Monitor the accuracy of the validation set of the main task. The validation set accounts for 20% of the actual test data. If the improvement is ≤0.1% for 5 consecutive rounds, save the current best model. 3) Parameter update method L is calculated using automatic differentiation. total For the gradient of all network parameters (KAN layer coefficients, classifier head weights, attention weights), update the parameters according to the gradient descent rule, and iterate until the early stopping condition is met. (5) Diagnostic results output: 1) Core diagnostic results The input measured signal to be diagnosed is preprocessed in S2 and fused in S3 to obtain F. real_fusion Output: 6-dimensional probability vector: p test = [p0, p1, ..., p5], where p k The signal to be diagnosed belongs to C k Confidence level; Fault category determination: C pred =argmax k p k If maxp k If the value is less than 0.7, output "Suspected fault"; 2) Interpretability results SHAP Feature Contribution Heatmap: Labeled F real_fusion The physical meaning of the top 10 SHAP values ​​in the data; Physical parameter matching report: Calculate the matching degree (Match) between the impact amplitude of the signal to be diagnosed and the theoretical ΔH. If Match < 0.6, it indicates "the feature deviates significantly from the physical model".

Images