An IPSO-Wav-KAN-based bearing fault diagnosis method for vehicle transmission system
By using the IPSO-Wav-KAN method, bearing fault diagnosis of vehicle transmission systems under complex working conditions was achieved, which solved the problems of insufficient feature extraction and parameter optimization in the existing technology, and improved the diagnostic accuracy and adaptability to portable devices.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- 63963 TROOP OF THE PLA
- Filing Date
- 2026-02-26
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to achieve adaptive feature extraction and parameter optimization for bearing faults in vehicle transmission systems under complex operating conditions, resulting in insufficient generalization ability of diagnostic models under different operating conditions and difficulty in real-time deployment on portable devices.
The method based on IPSO-Wav-KAN is adopted. Feature vectors are extracted through multi-scale wavelet decomposition and adaptive threshold denoising. The KAN network parameters are optimized by combining the IPSO algorithm with adaptive inertial weights and dynamic learning factors to achieve accurate diagnosis of bearing faults.
It improves feature recognition and diagnostic accuracy, increases convergence speed by 35% and diagnostic accuracy by 12.3%, and achieves lightweight design to adapt to portable devices, supporting rapid on-site testing.
Smart Images
Figure CN121723274B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fault diagnosis technology, and in particular to a method for diagnosing bearing faults in vehicle transmission systems based on IPSO-Wav-KAN. Background Technology
[0002] As commercial vehicles and construction machinery continue to evolve towards higher loads and higher operating speeds, the mechanical stress and dynamic load levels experienced by vehicle transmission systems during actual service are constantly increasing. Bearings, as key transmission components, operate in increasingly complex environments, subjected to alternating high loads, strong vibrations, and complex working conditions. Affected by these factors, the frequency of bearing failures is on the rise, and most failures exhibit a gradual, insidious deterioration characteristic in their initial stages. If not effectively identified early on, these failures can easily evolve into transmission system failures during operation, leading to vehicle downtime, safety accidents, and significant economic losses. Therefore, fault diagnosis methods for transmission system bearings have become an important research direction in the field of vehicle operation and maintenance and condition monitoring.
[0003] From a technological evolution perspective, bearing fault diagnosis methods have undergone a gradual development from traditional signal analysis to intelligent diagnosis. At the signal processing level, early methods primarily relied on time-domain and frequency-domain analysis. With the increasing demand for non-stationary signal analysis, wavelet transform, with its multi-scale analysis capabilities, has been gradually introduced into fault feature extraction and has been widely applied in suppressing noise interference and extracting impact components. Regarding model parameter optimization, particle swarm optimization (PSO) and genetic algorithms have been used to search for and adjust key parameters of the diagnostic model. PSO, in particular, has a certain application foundation in engineering practice due to its simple structure and low implementation cost. In terms of fault identification models, diagnostic models have gradually evolved from traditional BP neural networks to newer intelligent network forms with simpler structures and stronger nonlinear fitting capabilities, such as KAN networks. These models reduce reliance on hardware computing power to a certain extent, making them more suitable for deployment in embedded scenarios. Simultaneously, diagnostic equipment has evolved from large-scale testing systems primarily used in laboratory environments to portable maintenance assistance devices (PMAs) to meet the practical needs of rapid on-site detection and maintenance decision-making.
[0004] Despite continuous technological advancements, several shortcomings remain in engineering applications. In feature extraction, existing wavelet transform methods often rely on manual experience to set the decomposition level and denoising threshold. When vehicles are operating under varying loads or in complex noise environments, it's difficult to adaptively adjust parameter configurations according to changing conditions. This can easily lead to the mixing of fault impact signals with background noise, limiting the discriminative power of extracted features and resulting in unsatisfactory characterization of early, subtle faults. In parameter optimization, traditional particle swarm optimization algorithms typically employ fixed inertia weights and learning factors. This leads to insufficient global search capability in the early stages of iteration and limited local optimization capability in later stages, easily resulting in decreased search efficiency or getting trapped in local optima. Consequently, it affects the generalization performance of the diagnostic model under different operating conditions. Summary of the Invention
[0005] To overcome the shortcomings of insufficient generalization ability under complex working conditions and difficulty in real-time deployment on embedded portable devices, this invention provides a method for diagnosing bearing faults in vehicle transmission systems based on IPSO-Wav-KAN.
[0006] The technical implementation scheme of the present invention is: a method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN, comprising the following steps:
[0007] S1: Obtain impact vibration data of the bearing in the vehicle transmission system, perform multi-scale wavelet decomposition and denoising processing on the impact vibration data to obtain denoised wavelet coefficients, and extract feature vectors describing the bearing state based on the denoised wavelet coefficients.
[0008] S2: The particle swarm optimization algorithm is introduced based on adaptive inertia weights and dynamic learning factors to obtain an improved particle swarm optimization algorithm. The structural parameters and weight parameters of the KAN network are then optimized based on the improved particle swarm optimization algorithm to obtain the optimized network parameters.
[0009] S3: Input the obtained feature vector into the optimized KAN network to obtain the bearing fault type;
[0010] S4: Generate diagnostic results based on the output of the optimized KAN network and output diagnostic information on a portable diagnostic terminal or vehicle control unit.
[0011] Preferably, the step of acquiring impact vibration data of the vehicle transmission system bearing, performing multi-scale wavelet decomposition and denoising processing on the impact vibration data to obtain denoised wavelet coefficients, and extracting feature vectors describing the bearing state based on the denoised wavelet coefficients includes: acquiring impact vibration data of the bearing during movement from a vibration sensor on the vehicle or testing device, wherein the impact vibration data is the original vibration signal, and performing multi-scale decomposition and threshold denoising on the impact vibration data using the db4 wavelet basis function to extract fault-sensitive feature vectors; the multi-scale decomposition involves performing a fixed five-level wavelet decomposition to obtain low-frequency approximation coefficients. With high frequency detail coefficient The threshold denoising process involves soft thresholding of high-frequency detail coefficients; the multi-scale decomposition process is described as follows:
[0012] ;
[0013] In the formula, For the scale factor Translation factor The original wavelet coefficients containing noise after decomposition under the action reflect the local fault characteristics of the signal; This is a scaling factor used to control the frequency domain resolution of the signal; This is the translation factor, used to achieve point-by-point scanning in the time domain; These are db4 wavelet basis functions; The original vibration signal, specifically, the signal over time. A changing time-domain function; This is a normalization factor used to ensure that the energy of wavelets remains consistent across different scales, enabling comparability analysis of wavelet coefficients at different scales.
[0014] Preferably, the threshold denoising is soft threshold denoising of high-frequency detail coefficients, including: obtaining denoised wavelet coefficients using a soft threshold denoising formula, wherein the soft threshold denoising formula is described as follows:
[0015] ;
[0016] In the formula, These are the denoised wavelet coefficients; These are the original wavelet coefficients containing noise; This is a sign function used to preserve the positive or negative trend of the original coefficients; The threshold is determined empirically or adaptively. , The standard deviation of noise; This is the signal length.
[0017] Preferably, the extraction of fault-sensitive feature vectors includes: obtaining mean features, variance features, kurtosis features, and peak factor features based on the denoised wavelet coefficients, and constructing feature vectors based on the mean features, variance features, kurtosis features, and peak factor features. .
[0018] Preferably, the step of introducing the particle swarm optimization algorithm based on adaptive inertia weights and dynamic learning factors to obtain an improved particle swarm optimization algorithm, and then optimizing the structural parameters and weight parameters of the KAN network based on the improved particle swarm optimization algorithm to obtain optimized network parameters, includes: wherein the design description of the adaptive inertia weights is as follows:
[0019] ;
[0020] In the formula, For the first Inertia weights in the next iteration; This is the initial maximum inertia weight; The final minimum inertia weight; where is the current iteration number, satisfying... ; This represents the maximum number of iterations.
[0021] Preferably, the step of introducing adaptive inertia weights and a dynamic learning factor into the particle swarm optimization algorithm to obtain an improved particle swarm optimization algorithm includes: obtaining a dynamic learning factor through a dynamic learning formula, and updating the particle velocity and position of the IPSO based on the adaptive inertia weights and the dynamic learning factor, wherein the dynamic learning formula is:
[0022] ;
[0023] ;
[0024] In the formula, For the first The individual cognitive factor at each iteration is used to measure the degree to which a particle depends on its own historical best position. This represents the initial maximum value of the individual's cognitive factor; This represents the final minimum value of the individual's cognitive factor. For the first The social cognition factor at the next iteration is used to measure the degree to which a particle depends on the global optimal position; This represents the final maximum value of the social cognitive factor. This represents the initial minimum value of the social cognitive factor.
[0025] Preferably, updating the particle velocity and position of IPSO based on adaptive inertia weights and dynamic learning factors includes: updating the particle position according to a position update formula; and updating the particle velocity according to a velocity update formula, wherein the position update formula is:
[0026] ;
[0027] The speed update formula is:
[0028] ;
[0029] In the formula, For the first The generation The first particle Dimensional parameter values; For the first The generation The first particle Dimensional parameter values; For the first The generation The first particle Dimensional speed; For the first The generation The first particle The speed of the update after the dimension; For the first Individual cognitive factors at the next iteration; For the first Social cognitive factors at the next iteration; For the first Inertia weights in the next iteration; For the first The first particle The historical optimal position of a dimension, i.e., the individual's optimal position; This represents the historical best position for all particles as a whole, i.e., the global best position. This is a random number used to increase the randomness of the search.
[0030] Preferably, the step of updating the particle velocity and position of IPSO according to the adaptive inertia weight and dynamic learning factor includes: the preferred parameters of IPSO are 50 particles, 100 maximum iterations, and the number of hidden layer nodes of the KAN network is limited to 5 to 20, and the parameter range of position update is limited to [-1, 1]; the velocity range of velocity update is limited to [-0.5, 0.5].
[0031] Preferably, the step of inputting the obtained feature vector into the optimized KAN network to obtain the bearing fault type includes: the optimization objective of the KAN network is to minimize the diagnostic error, and the mapping formula in the optimized KAN network is:
[0032] ;
[0033] In the formula, For IPSO optimized network parameters, For the hidden layer The weights from each hidden node to the output layer; The weights from the input layer to the hidden layer, i.e., the weights of the first hidden layer. The hidden node to the first Feature mapping weights on the dimensional input; For hidden layer bias; It is a piecewise cubic polynomial activation function; The One-Hot encoded output vector is used for fast PMA parsing; For the input feature vector, the first... One component; The input feature vector; This represents the number of hidden nodes. This represents the number of input feature vectors.
[0034] Preferably, the optimization objective of the KAN network is to minimize the diagnostic error, including: the objective function is described as follows: In the formula, The objective function is... The number of samples; This is a real label; This is a predicted value; during the diagnostic process, the bearing vibration signal collected by PMA is converted from analog to digital, and the feature vector is extracted by the algorithm and input into the KAN network with optimized network parameters. The function determines the fault type, assesses the fault severity by combining the kurtosis value, and finally outputs the results on the PMA display screen.
[0035] The beneficial effects of this invention are as follows:
[0036] 1. This invention provides accurate feature extraction, corresponding to wavelet transform with 5-level multi-scale decomposition, adaptive threshold denoising, and 4 types of sensitive feature parameter extraction techniques. Traditional methods rely on experience to set the number of decomposition levels and denoising thresholds, which are difficult to adapt to the high noise conditions of vehicles. This method accurately separates fault impact signals from background noise through fixed-level decomposition of db4 wavelet basis functions and adaptive threshold processing based on noise standard deviation. The selected mean, variance, kurtosis, and peak factor comprehensively characterize the bearing state from multiple dimensions. In particular, kurtosis is sensitive to early minor faults, which greatly improves feature recognition.
[0037] 2. The present invention has highly efficient parameter optimization: The adaptive inertia weight, dynamic learning factor design, and lightweight parameter setting correspond to the IPSO algorithm. The traditional PSO algorithm suffers from optimization imbalance due to fixed parameters. This method achieves a dynamic balance between global search and local search through linearly decreasing inertia weight and dynamically adjusted learning factor, preventing premature convergence. At the same time, it reasonably controls the particle population size and the number of iterations, improving the convergence speed of the traditional PSO algorithm by more than 35%.
[0038] 3. High diagnostic accuracy: Corresponding to the strong nonlinear fitting architecture of the KAN network and the IPSO parameter optimization technique, traditional diagnostic models have limited fitting capabilities and their parameters rely on manual tuning. The KAN network in this method possesses stronger nonlinear mapping capabilities through a piecewise cubic polynomial activation function. Combined with the optimal network parameters obtained through IPSO algorithm optimization, it accurately identifies different types and severity of faults, achieving an average diagnostic accuracy of 98.9%, a 12.3% improvement over traditional methods.
[0039] 4. This invention has strong field adaptability: The entire algorithm is designed with lightweight features and hardware / software and interactive adaptation schemes for PMA. Existing high-precision algorithms are difficult to adapt to PMA embedded resources. This method uses lightweight optimization by fixing the number of wavelet decomposition layers and limiting the number of network nodes, keeping the memory usage within 512MB and the diagnostic response time less than or equal to 3s. At the same time, through hardware acceleration, software compatibility and standardized interactive design, it achieves deep integration with PMA, supporting integrated operation from field acquisition to diagnosis to output, without relying on laboratory equipment. Attached Figure Description
[0040] Figure 1 This is a flowchart of the vehicle transmission system bearing fault diagnosis method of the IPSO-Wav-KAN of the present invention;
[0041] Figure 2 This is a comparison of the convergence curves of the IPSO algorithm of this invention and the traditional PSO algorithm;
[0042] Figure 3 This is a comparison of the signals before and after wavelet transform feature extraction according to the present invention;
[0043] Figure 4 This is a flowchart of the IPSO-Wav-KAN algorithm of the present invention. Detailed Implementation
[0044] The technical solutions of 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.
[0045] Example: A method for diagnosing bearing faults in vehicle transmission systems based on IPSO-Wav-KAN, such as... Figures 1-4 As shown, it includes the following steps:
[0046] S1: Obtain impact vibration data of the bearing in the vehicle transmission system, perform multi-scale wavelet decomposition and denoising processing on the impact vibration data to obtain denoised wavelet coefficients, and extract feature vectors describing the bearing state based on the denoised wavelet coefficients.
[0047] Impact vibration data during bearing movement is acquired using vibration sensors on vehicles or testing devices. This impact vibration data is the raw vibration signal. The impact vibration data is then subjected to multi-scale decomposition and threshold denoising using the db4 wavelet basis function to extract fault-sensitive feature vectors. The multi-scale decomposition involves performing a fixed five-level wavelet decomposition to obtain low-frequency approximation coefficients. With high frequency detail coefficient The threshold denoising process involves soft thresholding of high-frequency detail coefficients; the multi-scale decomposition process is described as follows:
[0048] ;
[0049] In the formula, For the scale factor Translation factor The original wavelet coefficients containing noise after decomposition under the action reflect the local fault characteristics of the signal; This is a scaling factor used to control the frequency domain resolution of the signal; This is the translation factor, used to achieve point-by-point scanning in the time domain; These are db4 wavelet basis functions; The original vibration signal, specifically, the signal over time. A changing time-domain function; This is a normalization factor used to ensure that the energy of wavelets remains consistent across different scales, enabling comparability analysis of wavelet coefficients at different scales.
[0050] Under normal vehicle driving conditions or simulated operation of the detection device, vibration sensors are arranged at the location of the bearing in the vehicle's transmission system to collect impact vibration data generated by the bearing during operation. The vibration sensors are fixedly installed in the bearing housing or adjacent to the bearing structure, and their output signals reflect the vibration changes of the bearing under different operating conditions. The impact vibration data collected by the vibration sensors is processed as the raw vibration signal. After obtaining the raw vibration signal, multi-scale wavelet decomposition is performed on the impact vibration data. In this embodiment, the multi-scale wavelet decomposition uses the db4 wavelet basis function and a fixed five-level wavelet decomposition method. The raw vibration signal is input to the wavelet decomposition module, and decomposed sequentially based on the db4 wavelet basis function to obtain a set of low-frequency approximation coefficients and five sets of high-frequency detail coefficients. The low-frequency approximation coefficients characterize the overall trend of the bearing vibration signal, while the high-frequency detail coefficients reflect the transient impact characteristics of the bearing in different frequency bands.
[0051] The denoised wavelet coefficients are obtained using the soft-threshold denoising formula, which is described as follows:
[0052] ;
[0053] In the formula, These are the denoised wavelet coefficients; These are the original wavelet coefficients containing noise; This is a sign function used to preserve the positive or negative trend of the original coefficients; The threshold is determined empirically or adaptively. , The standard deviation of noise; This is the signal length.
[0054] After completing the fixed five-level multi-scale wavelet decomposition and obtaining the high-frequency detail coefficients, threshold denoising is performed on the high-frequency detail coefficients. In this embodiment, soft threshold denoising is used.
[0055] Impact vibration signals generated by bearing failures are easily masked by background noise, making them difficult to separate effectively using traditional methods. Therefore, multi-scale decomposition and threshold denoising of the db4 wavelet basis function are employed to extract fault-sensitive feature parameters. Specifically, soft thresholding is performed on the high-frequency detail coefficients at each scale. By weakening the noise-dominated small-amplitude wavelet coefficients, the large-amplitude wavelet coefficients reflecting the impact characteristics of bearing failures are retained, thus obtaining the denoised wavelet coefficients.
[0056] Based on the denoised wavelet coefficients, mean features, variance features, kurtosis features, and peak factor features are obtained, and a feature vector is constructed based on these features. .
[0057] Based on the denoised wavelet coefficients, feature vectors describing the bearing state are extracted. Specifically, mean feature, variance feature, kurtosis feature, and crescendo factor feature are calculated based on the denoised wavelet coefficients; among which the mean feature... The average amplitude level of the vibration signal is used to characterize the vibration signal; energy increases significantly during a fault. Variance characteristics reflect the dispersion of the vibration signal; fluctuations intensify during a fault. Kujicic characteristics reflect the sharpness of the impact component; they are extremely sensitive to impact signals and are normal under normal conditions. Approximately equal to 3, during a fault. A value greater than 5 indicates a positive correlation between the severity of the fault; the peak factor characteristic describes the relationship between the signal peak value and the overall energy distribution, reflecting the ratio of the peak value to the effective value. During a fault, the peak value is prominent. The mean feature, variance feature, kurtosis feature, and peak factor feature are combined in a predetermined order to construct a feature vector describing the bearing's operating state. This embodiment demonstrates the method for obtaining each feature: ; The signal length; It is a characteristic of the mean; ; Variance characteristics; ; This is a kurtosis feature; ; Characterized by peak factor;
[0058] In the formula, That is, described as .
[0059] S2: An improved particle swarm optimization algorithm is obtained by introducing adaptive inertia weights and a dynamic learning factor. The structural and weight parameters of the KAN network are then optimized based on this improved algorithm to obtain the optimized network parameters. The design description of the adaptive inertia weights is as follows:
[0060] ;
[0061] In the formula, For the first Inertia weights in the next iteration; This is the initial maximum inertia weight; This represents the final minimum inertia weight; For the current iteration number, satisfying ; This represents the maximum number of iterations.
[0062] Traditional PSO algorithms use fixed inertial weights and learning factors, leading to an imbalance between global and local optimization and premature convergence. Furthermore, the computational cost is too high, making them unsuitable for PMA hardware. To address this, we introduce adaptive inertial weights and dynamic learning factors, while simultaneously optimizing the algorithm's parameter scale, to construct a lightweight IPSO algorithm specifically optimized for key parameters of the KAN network. The maximum number of iterations is used to limit the computational scale and convergence process of the particle swarm optimization algorithm. This is the initial maximum inertia weight, which is set to a large value in the early stages of the algorithm to enhance the global search capability of the particles. The final minimum inertia weight is set to a small value in the later stages of the algorithm to improve the local search accuracy of particles near the optimal solution. For the first The inertia weight in each iteration is used to control the proportion of the particle's current velocity that retains the velocity from the previous moment. Through this linearly decreasing adaptive inertia weight design, the improved particle swarm optimization algorithm maintains strong global search capabilities in the early stages of iteration, reducing the likelihood of getting trapped in local optima. In the later stages of iteration, the search step size is gradually reduced, allowing for a finer search within the global optimum region, effectively improving the stability and convergence accuracy of the KAN network parameter optimization. In this embodiment... ; ; The search strategy is dynamically adjusted by linearly decreasing the search level.
[0063] The dynamic learning factor is obtained through a dynamic learning formula. Based on the adaptive inertia weight and the dynamic learning factor, the particle velocity and position of IPSO are updated. The dynamic learning formula is as follows:
[0064] ;
[0065] ;
[0066] In the formula, For the first The individual cognitive factor at each iteration is used to measure the degree to which a particle depends on its own historical best position. This represents the initial maximum value of the individual's cognitive factor; This represents the final minimum value of the individual's cognitive factor. For the first The social cognition factor at the next iteration is used to measure the degree to which a particle depends on the global optimal position; This represents the final maximum value of the social cognitive factor. This represents the initial minimum value of the social cognitive factor.
[0067] This is the initial maximum value of the individual cognitive factor, used to enhance the individual exploration ability of particles in the early stages of the algorithm; This is the final minimum value of the individual cognitive factor, used to reduce the excessive reliance of particles on their own experience in the later stages of the algorithm; This is the final maximum value of the social cognition factor, used to enhance particle convergence towards the global optimal solution in the later stages of the algorithm; The initial minimum value of the social cognition factor is used to prevent the algorithm from prematurely converging towards the global optimum in the early stages. Through the design of the dynamic learning factor, the improved particle swarm optimization algorithm focuses on individual exploration in the initial stage, enhancing the search diversity of the parameter space. In the later stages of iteration, the social collaborative search capability is gradually enhanced, prompting the particle swarm to converge towards the global optimum KAN network parameter combination, thereby reducing the premature convergence problem common in traditional particle swarm optimization algorithms. In this embodiment, the individual cognition factor is reduced from 2.5 to 1.0 to encourage autonomous exploration of particles in the early stage; the social cognition factor is increased from 1.0 to 2.5 to encourage particles to move towards the group optimum in the later stage, balancing exploration and utilization capabilities and reducing ineffective iterations.
[0068] Update particle position according to the position update formula; update particle velocity according to the velocity update formula, where the position update formula is:
[0069] ;
[0070] The speed update formula is:
[0071] ;
[0072] In the formula, For the first The generation The first particle Dimensional parameter values; For the first The generation The first particle Dimensional parameter values; For the first The generation The first particle Dimensional speed; For the first The generation The first particle The speed of the update after the dimension; For the first Individual cognitive factors at the next iteration; For the first Social cognitive factors at the next iteration; For the first Inertia weights in the next iteration; For the first The first particle The historical optimal position of a dimension, i.e., the individual's optimal position; This represents the historical best position for all particles as a whole, i.e., the global best position. This is a random number used to increase the randomness of the search.
[0073] Position updates are performed by adjusting the corresponding KAN parameters, with the parameter range limited to [-1, 1] to reduce computational complexity; speed updates are performed by controlling the step size of parameter adjustments, with the speed range limited to [-0.5, 0.5] to reduce over-search. For the first The inertial weight in the next iteration is used to adjust the degree to which particles inherit historical velocities; For the first The generation The first particle The velocity dimension is used to represent the direction and magnitude of parameter changes. Through the above-mentioned velocity and position update methods, IPSO can maintain the continuity of particle search while taking into account individual experience and group collaborative information, thereby achieving efficient search of the parameter space of the KAN network.
[0074] The preferred parameters for IPSO are: 50 particles, 100 maximum iterations, 5 to 20 hidden layer nodes in the KAN network, and the parameter range for position updates is limited to [-1, 1]; the speed range for speed updates is limited to [-0.5, 0.5].
[0075] The number of particles is preferably set to 50 to ensure diversity in the parameter space search while preventing excessive computational complexity due to an excessive number of particles. The maximum number of iterations is preferably set to 100 to achieve a balance between diagnostic accuracy and computation time. The number of hidden layer nodes in the KAN network is limited to between 5 and 20 to prevent overfitting of the model or excessive consumption of embedded computing resources due to an excessive number of hidden nodes. The parameter range for particle position updates is limited to [-1, 1] to constrain the value range of the KAN network weight parameters and prevent abnormal parameter divergence. The speed range for particle velocity updates is limited to [-0.5, 0.5] to control the magnitude of parameter adjustment and make the optimization process more stable. By limiting the above parameter ranges, IPSO has good numerical stability in actual operation and can effectively optimize the parameters of the KAN network under limited computing resources.
[0076] S3: Input the obtained feature vector into the optimized KAN network to obtain the bearing fault type;
[0077] The optimization objective of the KAN network is to minimize the diagnostic error. The mapping formula in the optimized KAN network is:
[0078] ;
[0079] In the formula, For IPSO optimized network parameters, For the hidden layer The weights from each hidden node to the output layer; The weights from the input layer to the hidden layer, i.e., the weights of the first hidden layer. The hidden node to the first Feature mapping weights on the dimensional input; For hidden layer bias; It is a piecewise cubic polynomial activation function; The One-Hot encoded output vector is used for fast PMA parsing; For the input feature vector, the first... One component; The input feature vector; This represents the number of hidden nodes. This represents the number of input feature vectors.
[0080] Compared to BP neural networks, KAN networks, based on Kolmogorov's theorem, possess stronger nonlinear fitting capabilities through piecewise multinomial activation functions. Furthermore, their network structure is simple, requiring no deep layer stacking and consuming less computation, making them the optimal choice for PMA (Problem Assessment). The lightweight design of the KAN network structure includes: an input layer with 4 nodes corresponding to feature vectors; the number of hidden layer nodes (5-20) determined by IPSO optimization; and an output layer with 4 nodes corresponding to normal, inner race fault, outer race fault, and rolling element fault. In this embodiment, the IPSO algorithm continuously adjusts the KAN network parameters to gradually reduce the objective function, thereby obtaining the network parameter configuration with the minimum diagnostic error, ensuring that the optimized KAN network possesses high fault identification accuracy and stability.
[0081] The objective function is described as follows: In the formula, The objective function is... The number of samples; This is a real label; This is a predicted value; during the diagnostic process, the bearing vibration signal collected by PMA is converted from analog to digital, and the feature vector is extracted by the algorithm and input into the KAN network with optimized network parameters. The function determines the fault type, assesses the fault severity by combining the kurtosis value, and finally outputs the results on the PMA display screen.
[0082] This embodiment uses a PMA (Polarization and Analysis System) as a field diagnostic device to perform real-time monitoring of bearings in a vehicle's transmission system. The specific process includes: first, the PMA collects vibration signals during bearing operation and performs analog-to-digital conversion on these signals; then, the digitized vibration signals undergo the aforementioned multi-scale wavelet decomposition, threshold denoising, and feature extraction processing to obtain a feature vector describing the bearing's operating state. This feature vector is then input into a KAN network with optimized network parameters to obtain an output vector in One-Hot encoded form. Based on this, the argmax(y) function is applied to the output vector to determine the bearing fault type corresponding to the largest component. The severity of the fault is assessed by combining the kurtosis values (5 < K < 7 for mild faults, 7 ≤ K < 9 for moderate faults, and K ≥ 9 for severe faults). Finally, the results are output on the PMA display screen, thus determining the current fault state of the bearing.
[0083] S4: Generate diagnostic results based on the output of the optimized KAN network and output diagnostic information on a portable diagnostic terminal or vehicle control unit.
[0084] At the hardware level: The algorithm is embedded in the ARM Cortex-A9 core processor of the PMA, and the floating-point unit of the processor accelerates the operation. The PMA collects signals through vibration and temperature sensors, and performs analog-to-digital conversion through a 16-bit ADC to provide the algorithm with input data that meets the accuracy requirements. At the software level: The algorithm is optimized in C language and compiled into an executable file adapted to the PMA embedded system. The sampling rate and iteration number parameters are configured through the PMA software interface, the diagnostic process is started and the algorithm processing results are received. At the interaction level: The diagnostic results of the fault type and severity output by the algorithm are displayed on the PMA display unit in Chinese.
[0085] The above embodiments are only for illustrating the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.
Claims
1. A method for diagnosing bearing faults in vehicle transmission systems based on IPSO-Wav-KAN, characterized in that, Includes the following steps: S1: Obtain impact vibration data of the bearing in the vehicle transmission system, perform multi-scale wavelet decomposition and denoising processing on the impact vibration data to obtain denoised wavelet coefficients, and extract feature vectors describing the bearing state based on the denoised wavelet coefficients. Extracting fault-sensitive feature vectors includes: obtaining mean features, variance features, kurtosis features, and peak factor features based on the denoised wavelet coefficients, and constructing feature vectors based on the mean features, variance features, kurtosis features, and peak factor features. ; S2: The particle swarm optimization algorithm is introduced based on adaptive inertia weights and dynamic learning factors to obtain an improved particle swarm optimization algorithm. The structural parameters and weight parameters of the KAN network are then optimized based on the improved particle swarm optimization algorithm to obtain the optimized network parameters. S3: Input the obtained feature vectors into the optimized KAN network to obtain the bearing fault type, including: The optimization objective of the KAN network is to minimize the diagnostic error, and the mapping formula in the optimized KAN network is: ； In the formula, For IPSO optimized network parameters, For the hidden layer The weights from each hidden node to the output layer; The weights from the input layer to the hidden layer, i.e., the weights of the first hidden layer. The hidden node to the first Feature mapping weights on the dimensional input; For hidden layer bias; It is a piecewise cubic polynomial activation function; The One-Hot encoded output vector is used for fast PMA parsing; For the input feature vector, the first... One component; The input feature vector; This represents the number of hidden nodes. The number of input feature vectors; The output layer has 4 nodes, corresponding to normal, inner ring fault, outer ring fault, and rolling element fault; S4: Generate diagnostic results based on the output of the optimized KAN network and output diagnostic information on a portable diagnostic terminal or vehicle control unit.
2. The method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 1, characterized in that, The process involves acquiring impact vibration data of the vehicle transmission system bearing, performing multi-scale wavelet decomposition and denoising on the impact vibration data to obtain denoised wavelet coefficients, and extracting feature vectors describing the bearing state based on the denoised wavelet coefficients. This includes: acquiring impact vibration data of the bearing during movement from vibration sensors on the vehicle or testing device, where the impact vibration data is the original vibration signal; performing multi-scale decomposition and threshold denoising on the impact vibration data using the db4 wavelet basis function; and extracting fault-sensitive feature vectors. The multi-scale decomposition involves performing a fixed five-level wavelet decomposition to obtain low-frequency approximation coefficients. With high-frequency detail coefficient The threshold denoising process involves soft thresholding of high-frequency detail coefficients; the multi-scale decomposition process is described as follows: ； In the formula, For the scale factor Translation factor The original wavelet coefficients containing noise after decomposition under the action reflect the local fault characteristics of the signal; This is a scaling factor used to control the frequency domain resolution of the signal; This is the translation factor, used to achieve point-by-point scanning in the time domain; These are db4 wavelet basis functions; The original vibration signal, specifically, the signal over time. A changing time-domain function; This is a normalization factor used to ensure that the energy of wavelets remains consistent across different scales, enabling comparability analysis of wavelet coefficients at different scales.
3. A method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 2, characterized in that, The threshold denoising involves performing soft threshold denoising on high-frequency detail coefficients, including: obtaining the denoised wavelet coefficients using a soft threshold denoising formula, wherein the soft threshold denoising formula is described as follows: ； In the formula, These are the denoised wavelet coefficients; These are the original wavelet coefficients containing noise; This is a sign function used to preserve the positive or negative trend of the original coefficients; The threshold is determined empirically or adaptively. , The standard deviation of noise; This is the signal length.
4. A method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 1, characterized in that, The process involves introducing adaptive inertia weights and a dynamic learning factor into the particle swarm optimization algorithm to obtain an improved particle swarm optimization algorithm. The structural and weight parameters of the KAN network are then optimized based on this improved algorithm to obtain the optimized network parameters, including: The design description of the adaptive inertia weights is as follows: ； In the formula, For the first Inertia weights in the next iteration; This is the initial maximum inertia weight; This represents the final minimum inertia weight; For the current iteration number, satisfying ; This represents the maximum number of iterations.
5. A method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 4, characterized in that, The improvement of the particle swarm optimization algorithm by introducing adaptive inertia weights and dynamic learning factors includes: obtaining a dynamic learning factor through a dynamic learning formula; and updating the particle velocity and position of the IPSO based on the adaptive inertia weights and the dynamic learning factor, wherein the dynamic learning formula is: ； ； In the formula, For the first The individual cognitive factor at each iteration is used to measure the degree to which a particle depends on its own historical best position. This represents the initial maximum value of the individual's cognitive factor; This represents the final minimum value of the individual's cognitive factor. For the first The social cognition factor at the next iteration is used to measure the degree to which a particle depends on the global optimal position; This represents the final maximum value of the social cognitive factor. This represents the initial minimum value of the social cognitive factor.
6. A method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 5, characterized in that, The step of updating the particle velocity and position of IPSO based on adaptive inertia weights and dynamic learning factors includes: updating the particle position according to the position update formula; and updating the particle velocity according to the velocity update formula, wherein the position update formula is: ； The speed update formula is: ； In the formula, For the first The generation The first particle Dimensional parameter values; For the first The generation The first particle Dimensional parameter values; For the first The generation The first particle Dimensional speed; For the first The generation The first particle The speed of the update after the dimension; For the first Individual cognitive factors at the next iteration; For the first Social cognitive factors at the next iteration; For the first Inertia weights in the next iteration; For the first The first particle The historical optimal position of a dimension, i.e., the individual's optimal position; This represents the historical best position for all particles as a whole, i.e., the global best position. This is a random number used to increase the randomness of the search.
7. A method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 6, characterized in that, The step of updating the particle velocity and position of IPSO based on adaptive inertia weights and dynamic learning factors includes: the preferred parameters of IPSO are 50 particles, 100 maximum iterations, and the number of hidden layer nodes of the KAN network is limited to 5 to 20, and the parameter range for position update is limited to [-1, 1]; the velocity range for velocity update is limited to [-0.5, 0.5].
8. A method for diagnosing bearing faults in a vehicle transmission system based on IPSO-Wav-KAN as described in claim 1, characterized in that, The optimization objective of the KAN network is to minimize the diagnostic error, including: the objective function is described as follows: In the formula, The objective function is... The number of samples; This is a real label; This is a predicted value; during the diagnostic process, the bearing vibration signal collected by PMA is converted from analog to digital, and the feature vector is extracted by the algorithm and input into the KAN network with optimized network parameters. The function determines the fault type, assesses the fault severity by combining the kurtosis value, and finally outputs the results on the PMA display screen.