A KAN-based method for fault diagnosis of rotating machinery

By using the KAN diagnostic model and symbolic mapping technology, an explicit mathematical expression for the fault diagnosis model of rotating machinery is established, which solves the problem of the inability to interpret existing models and achieves high-precision and high-reliability fault diagnosis.

CN123115852APending Publication Date: 2026-11-17CHONGQING TSINGSHAN IND
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611261881.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing fault diagnosis models for rotating machinery struggle to establish explicit mathematical relationships between data characteristics and physical fault mechanisms, resulting in uninterpretable and unreliable diagnostic results.

Method used

A fault diagnosis method based on Kolmogorov-Arnold network (KAN) is adopted, which combines symbolic mapping technology to transform the trained model parameters into explicit mathematical expressions. Nonlinear features are dynamically captured through the learnable activation function of the KAN diagnostic model, and dual-path extraction of multi-scale time-frequency domain physical features and two-dimensional time-frequency image abstract features is provided in the feature extraction stage.

Benefits of technology

This approach achieves a direct correlation between fault diagnosis results and physical fault mechanisms, improving the reliability and interpretability of the diagnosis results, reducing computational resource consumption, and enhancing the robustness and generalization ability of the model under different operating conditions and noise interference.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115852A_ABST
    Figure CN123115852A_ABST
Patent Text Reader

Abstract

This invention discloses a KAN-based fault diagnosis method for rotating machinery, comprising: acquiring vibration signals during the operation of the rotating machinery, preprocessing and extracting features from the vibration signals to obtain a target feature set; constructing a KAN diagnostic model based on a Kolmogorov-Arnold network, wherein activation functions are set on the network connection edges of the KAN diagnostic model; training the KAN diagnostic model using the target feature set to obtain a trained KAN diagnostic model; acquiring the vibration signal of the rotating mechanism under test and inputting it into the trained KAN diagnostic model for fault diagnosis, outputting the fault diagnosis result of the rotating mechanism under test; and using symbolic mapping technology to transform the parameters in the trained KAN diagnostic model into explicit mathematical expressions to establish the correlation between the fault diagnosis result and the physical fault mechanism. This invention solves the technical problems in existing rotating machinery fault diagnosis caused by the "black box" characteristic of deep learning models, which makes the decision-making process uninterpretable and makes it difficult to establish the correlation between the data-driven model and the physical mechanism.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of mechanical fault diagnosis and artificial intelligence, and in particular to a method for diagnosing rotating machinery faults based on KAN. Background Technology

[0002] As a core architecture in the field of artificial intelligence, the Multi-Layer Perceptron (MLP) exhibits powerful data relationship modeling capabilities through its feedforward hierarchical structure and superimposed nonlinear transformations of weight connections. However, the inherent paradigm of fixed activation functions in its hidden layers limits model interpretability, making it difficult to trace the causal relationship between feature mappings and decision logic. To overcome this bottleneck, the Kolmogorov-Arnold Network (KAN) was developed. Based on the Kolmogorov-Arnold representation theorem, it reconstructs the network architecture, replacing the linear weight-activation separation mode of traditional MLPs with dynamically parameterized nonlinear basis functions. KAN utilizes the learnable properties of edge weight functions to transform implicit features into explicit mathematical expressions, thereby constructing hierarchically interpretable mapping relationships.

[0003] Currently, fault diagnosis of rotating machinery (such as bearings and gearboxes) mainly relies on data-driven deep learning methods, such as convolutional neural networks (CNNs) and multilayer perceptrons (MLPs). However, these models are often considered "black boxes," as their core architecture is based on a paradigm that separates fixed linear weights from nonlinear activation functions. This paradigm makes it difficult for the model to establish an explicit causal relationship between input features and output decisions. In practical engineering applications, when the model identifies a fault, technicians cannot know which physical mechanism (such as amplitude modulation or frequency coupling) in the vibration signal the decision corresponds to, severely limiting the reliability and interpretability of the diagnostic results.

[0004] The core problem with existing technologies lies in the difficulty of establishing an explicit mathematical connection between data features and physical fault mechanisms. For example, while CNN models can improve the ability to extract periodic fault features in noisy environments, it remains unclear how the model uses these extracted features to determine the fault in the bearing's inner race. Although multi-attention feature aggregation networks enhance model robustness, their decision-making logic is completely independent of the physical characteristics of the vibration signal (such as amplitude modulation and frequency coupling). Existing models cannot convert trained parameters into explicit physical formulas, making it difficult to trace diagnostic results back to specific fault features, severely limiting their reliability and application potential in engineering scenarios.

[0005] Therefore, there is an urgent need for a method for diagnosing rotating machinery faults that can maintain high diagnostic accuracy and establish an explicit correlation between physical characteristics and fault categories. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention provides a KAN-based method for diagnosing rotating machinery faults. By employing a KAN diagnostic model based on the Kolmogorov-Arnold network and introducing symbolic mapping technology to transform the trained basis functions into explicit mathematical expressions, this invention solves the technical problems in existing rotating machinery fault diagnosis caused by the "black box" characteristics of deep learning models, which result in uninterpretable decision-making processes and difficulty in establishing the correlation between data-driven models and physical mechanisms.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] A KAN-based method for diagnosing faults in rotating machinery includes the following steps:

[0009] S1. Collect vibration signals during the operation of rotating machinery, and preprocess and extract features from the vibration signals to obtain a target feature set;

[0010] S2. Construct a KAN diagnostic model based on the Kolmogorov-Arnold network, wherein activation functions are set on the network connection edges of the KAN diagnostic model;

[0011] S3. Train the KAN diagnostic model using the target feature set to obtain the trained KAN diagnostic model;

[0012] S4. Acquire the vibration signal of the rotating mechanism under test and input it into the trained KAN diagnostic model to perform fault diagnosis and output the fault diagnosis result of the rotating mechanism under test.

[0013] S5. Using symbolic mapping technology, the parameters in the trained KAN diagnostic model are transformed into explicit mathematical expressions to establish the correlation between fault diagnosis results and physical fault mechanisms.

[0014] Furthermore, in step S1, the preprocessing and feature extraction of the vibration signal includes:

[0015] According to a preset time window, the vibration signal is divided into multiple sub-samples of the same length;

[0016] For each of the sub-samples, a feature extraction operation is performed to obtain the corresponding feature vector as the target feature set.

[0017] Furthermore, for each of the sub-samples, a feature extraction operation is performed, constructing a target feature set using any of the following paths:

[0018] Path 1: Extract multi-scale time-frequency domain physical features from the vibration signal using wavelet transform, and construct the target feature set. The time-frequency domain physical features include mean, kurtosis, variance, root mean square, skewness, and peak factor.

[0019] Path 2: The vibration signal is transformed into a two-dimensional time-frequency image by wavelet transform, and the abstract features of the two-dimensional time-frequency image are extracted using a preset convolutional neural network to construct the target feature set.

[0020] Furthermore, in step S2, the KAN diagnostic model is constructed based on the Kolmogorov-Arnold representation theorem, and its mathematical expression is:

[0021] ;

[0022] In the formula: For KAN diagnostic model; These are the features in the target feature set; The number of features; It is an outer unary function; It is an inner unary function; For the input vector of the th One component; The index for traversing the input features; This is an index for traversing the number of neurons in the KAN diagnostic model.

[0023] Furthermore, in step S2, the outer unary function in the KAN diagnostic model With inner univariate function All are learnable activation functions on the network connection edges. These activation functions are parameterized by spline functions composed of piecewise polynomials, and are expressed as follows:

[0024] ;

[0025] In the formula: It is a spline function; For the first basis functions These are the coefficients of the corresponding basis functions.

[0026] Furthermore, in step S3, the KAN diagnostic model is trained through the following process:

[0027] The target feature set is divided into a training set and a validation set according to a certain ratio; the training set is input into the KAN diagnostic model, the prediction output is calculated through forward propagation, and the prediction error is calculated based on a preset loss function; the parameters of the piecewise polynomial spline are updated using the backpropagation algorithm until the prediction error meets the preset convergence condition, and the trained KAN diagnostic model is obtained.

[0028] Furthermore, in step S4, the fault diagnosis and output of the fault diagnosis result of the rotating mechanism under test specifically includes:

[0029] The trained KAN diagnostic model is used to calculate the vibration signal of the input rotating mechanism under test, and a continuous fault probability value is obtained based on the output of the single neuron output layer. The fault probability value is mapped and compared with a preset fault threshold to determine the fault category of the rotating mechanism under test, which is used as the fault diagnosis result.

[0030] Furthermore, in step S5, the step of using symbolic mapping technology to convert the parameters in the trained KAN diagnostic model into explicit mathematical expressions specifically includes:

[0031] Extract the piecewise polynomial spline parameters of each network connection edge in the trained KAN diagnostic model after training convergence; use the symbolic regression algorithm to transform the piecewise polynomial spline parameters into a closed mathematical combination formula; based on the mathematical combination formula, analyze the explicit mathematical correlation between the fault diagnosis result and the input time-frequency domain physical features or abstract features.

[0032] Compared with the prior art, the present invention has the following technical effects:

[0033] 1. This invention employs a KAN diagnostic model based on the Kolmogorov-Arnold representation theorem. It utilizes symbolic mapping technology to transform model parameters into explicit mathematical expressions. Unlike traditional convolutional neural networks or fully connected neural networks, this invention not only outputs fault diagnosis results but also establishes a direct correlation between fault diagnosis results and physical fault mechanisms. This makes the diagnostic process no longer a "black box" but allows for a clear explanation of the physical causes of faults through specific mathematical formulas. This increases technicians' confidence in the intelligent diagnostic results and provides a traceable theoretical basis for subsequent maintenance decisions.

[0034] 2. This invention sets activation functions on the network connection edges of the KAN diagnostic model and uses piecewise polynomial splines for parameterization. This transforms the traditional "fixed activation function + linear weight" mode into a "learnable activation function" mode. Due to the smoothness and fitting ability of spline functions, the KAN diagnostic model can fit the complex nonlinear characteristics in the vibration signal of rotating machinery more accurately with fewer parameters. This not only reduces the training difficulty of the model, but also effectively improves the robustness and generalization ability of the model when facing different working conditions or noise interference, thereby improving the accuracy of fault diagnosis.

[0035] 3. This invention provides two paths in the feature extraction stage. Path one is to extract multi-scale time-frequency domain physical features through wavelet transform, which preserves the original physical properties of the signal and enhances the interpretability of the final diagnostic results. Path two is to use convolutional neural networks to extract abstract features from two-dimensional time-frequency images, which ensures both physical clarity in simple scenes and high-precision recognition in complex scenes.

[0036] 4. Compared to traditional deep learning models that often require a large number of parameters to improve performance, this invention uses a Kolmogorov-Arnold theorem-based architecture to approximate multivariate functions through the superposition of univariate functions. This allows for the achievement of the same fitting effect with fewer parameters, reducing the computational resource consumption during model training and inference.

[0037] 5. This invention divides continuous vibration signals into multiple sub-samples of equal length, which can effectively solve the problem of dilution of local transient fault features in long time-series signals, thereby improving the fine capture of local transient features of the signal. Attached Figure Description

[0038] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0039] Figure 1 This is a flowchart of a rotating machinery fault diagnosis method based on KAN disclosed in this invention;

[0040] Figure 2 This is a flowchart illustrating the diagnostic process based on the KAN diagnostic model in an embodiment of the present invention.

[0041] Figure 3 This is a comparison chart of classification accuracy under various loads on the CWRU dataset for different implementation paths in this embodiment of the invention. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0044] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the figures, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, the terms "horizontal," "vertical," etc., do not indicate that the component is required to be absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] Example:

[0046] In the existing field of rotating machinery fault diagnosis, although deep learning technologies, represented by convolutional neural networks, have been widely used, they are essentially "black box" models. These models can only provide a mapping between input (vibration signals) and output (fault category), but cannot elucidate the underlying physical mechanisms of fault occurrence. In industrial settings, technicians not only focus on "what fault occurred," but also urgently need to understand "why the fault occurred" through mathematical logic or physical indicators.

[0047] To address this technical problem, this invention proposes a KAN-based method for diagnosing rotating machinery faults. This method utilizes a KAN diagnostic model based on the Kolmogorov-Arnold representation theorem as the core diagnostic model. It replaces the fixed activation function and linear weight separation mode of traditional MLP with learnable univariate functions (such as B-spline parameterization) on the network connection edges. This preserves the nonlinear fitting of deep learning. More importantly, the activation functions on the edges can be directly transformed into explicit mathematical expressions after training through symbolic mapping technology. This allows each decision of the model to be traced back to the specific input physical features and their computational relationships, thereby enabling the nonlinear relationship between fault features and diagnostic results to be analyzed from a mathematical structure perspective. This significantly improves the reliability and interpretability of the diagnostic results.

[0048] Specifically, the KAN-based rotating machinery fault diagnosis method proposed in this invention, such as... Figure 1 As shown, it includes the following steps:

[0049] S1. Collect vibration signals during the operation of rotating machinery, and preprocess and extract features from the vibration signals to obtain a target feature set;

[0050] In this embodiment, the preprocessing and feature extraction of the acquired vibration signal specifically includes: dividing the vibration signal into multiple sub-samples of length 1024 according to a preset time window; performing feature extraction operation on each sub-sample to obtain the corresponding feature vector as the target feature set.

[0051] In this embodiment, the target feature set adopts the following approach: Figure 2 The following can be constructed using any of the paths shown:

[0052] Path 1: Extract multi-scale time-frequency domain physical features from the vibration signal using wavelet transform, and construct a target feature set with clear physical meaning, suitable for scenarios requiring mechanism analysis; wherein, the time-frequency domain physical features include mean, kurtosis, variance, root mean square, skewness, and peak factor;

[0053] Path Two: The vibration signal is transformed into a two-dimensional time-frequency image using wavelet transform. A pre-built convolutional neural network is then used to extract abstract features from the two-dimensional time-frequency image, constructing the target feature set. This approach is suitable for scenarios requiring high diagnostic accuracy. In this path, a convolutional autoencoder is constructed for unsupervised training, reconstructing the time-frequency image to learn feature representations. After training, the abstract features of the time-frequency image are extracted for training and classification of the KAN diagnostic model.

[0054] This embodiment provides two paths in the feature extraction stage. Path one is to extract multi-scale time-frequency domain physical features through wavelet transform, which preserves the original physical properties of the signal and enhances the interpretability of the final diagnostic results. Path two is to use convolutional neural networks to extract abstract features from two-dimensional time-frequency images, which ensures both physical clarity in simple scenes and high-precision recognition in complex scenes.

[0055] S2. Construct a KAN diagnostic model based on the Kolmogorov-Arnold network, wherein activation functions are set on the network connection edges of the KAN diagnostic model;

[0056] In this embodiment, the KAN diagnostic model is constructed based on the Kolmogorov-Arnold representation theorem, and its mathematical expression is:

[0057] ;

[0058] In the formula: For KAN diagnostic model; These are the features in the target feature set; The number of features; It is an outer unary function; It is an inner unary function; For the input vector of the th One component; The index for traversing the input features; This is an index for traversing the number of neurons in the KAN diagnostic model.

[0059] Meanwhile, abandoning the paradigm of separating fixed activation functions from linear weights, we instead use univariate functions as learnable activation functions on network edges. Nodes perform only simple summation operations, thus making the function on each connection edge in the network... It can be parameterized as an adaptive univariate function, thus dynamically capturing the nonlinear characteristics and smooth transition properties of the input data. Therefore, the activation function on the network connection edges of the KAN diagnostic model is parameterized by piecewise polynomial splines. The spline function defines nodes in the input domain and uses low-order polynomials for interpolation between adjacent nodes, thereby achieving flexible approximation of complex curves. (Based on input features...) For example, its corresponding spline function can be expressed as:

[0060] ;

[0061] In the formula: It is a spline function; For the first basis functions These are the coefficients of the corresponding basis functions.

[0062] In this embodiment, activation functions are set on the network connection edges of the KAN diagnostic model, and piecewise polynomial splines are used for parameterization. This transforms the traditional "fixed activation function + linear weight" mode into a "learnable activation function" mode. Due to the smoothness and fitting ability of spline functions, the KAN diagnostic model can fit the complex nonlinear characteristics of rotating machinery vibration signals more accurately with fewer parameters. This not only reduces the training difficulty of the model, but also effectively improves the robustness and generalization ability of the model when facing different working conditions or noise interference, thereby improving the accuracy of fault diagnosis. Compared with traditional deep learning models, which often require a large number of parameters to exchange for performance, the Kolmogorov-Arnold theorem-based architecture approximates multivariate functions through the superposition of univariate functions, thus achieving the same fitting effect with fewer parameters and reducing the computational resource consumption of the model training and inference process.

[0063] S3. Train the KAN diagnostic model using the target feature set to obtain the trained KAN diagnostic model;

[0064] In this embodiment, the KAN diagnostic model is trained through the following process:

[0065] The target feature set is divided into a training set and a validation set in an 8:2 ratio; the training set is input into the KAN diagnostic model, the prediction output is calculated through forward propagation, and the prediction error is calculated based on a preset loss function; the parameters of the piecewise polynomial spline are updated using the backpropagation algorithm until the prediction error meets the preset convergence condition, and the trained KAN diagnostic model is obtained.

[0066] S4. Acquire the vibration signal of the rotating mechanism under test and input it into the trained KAN diagnostic model to perform fault diagnosis and output the fault diagnosis result of the rotating mechanism under test.

[0067] In this embodiment, the trained KAN diagnostic model is used to calculate the vibration signal of the input rotating mechanism under test, and a continuous fault probability value is obtained based on the output of the single neuron output layer. The fault probability value is mapped and compared with a preset fault threshold to determine the fault category of the rotating mechanism under test, which is used as the fault diagnosis result.

[0068] In this embodiment, the output layer of the KAN diagnostic model is designed to retain only one neuron, outputting continuous values ​​and mapping them to multiple fault categories through a preset threshold. This greatly simplifies the structure of the output layer, avoids parameter redundancy and the assumption of independence between categories brought about by traditional multi-neuron softmax output. The single-neuron output combined with threshold mapping is suitable for continuous quantitative assessment of fault severity and is easier to interface with mathematical expressions generated by symbolic mapping.

[0069] S5. Using symbolic mapping technology, the parameters in the trained KAN diagnostic model are transformed into explicit mathematical expressions to establish the correlation between fault diagnosis results and physical fault mechanisms.

[0070] In this embodiment, piecewise polynomial spline parameters are extracted from each network connection edge of the trained KAN diagnostic model after training convergence; the piecewise polynomial spline parameters are transformed into a closed mathematical combination formula using a symbolic regression algorithm; based on the mathematical combination formula, the explicit mathematical correlation between the fault diagnosis result and the input time-frequency domain physical features or abstract features is analyzed (e.g., "outer ring fault decision = 0.3 × kurtosis + 1.7 × frequency modulation amplitude").

[0071] In practice, regression analysis simplifies the spline basis functions into closed-form mathematical combinations. For example, the mapping from the input layer to the first layer may learn the squared term of the vibration signal amplitude. The second layer may capture the periodic modulation of frequency components. This study traces the model's decision-making process back to the characteristics of vibration signals, providing quantifiable evidence for fault mechanism analysis. A KAN diagnostic model with only one neuron in the output layer is constructed, outputting continuous values. (C is the number of fault categories), which is mapped to fault categories through a threshold or other means.

[0072] This embodiment utilizes symbolic mapping technology to transform model parameters into explicit mathematical expressions. Unlike traditional convolutional neural networks or fully connected neural networks, this invention can not only output fault diagnosis results but also establish a direct correlation between fault diagnosis results and physical fault mechanisms. This makes the diagnosis process no longer a "black box" but can clearly explain the physical causes of faults through specific mathematical formulas, thereby increasing technicians' trust in intelligent diagnosis results and providing a traceable theoretical basis for subsequent maintenance decisions.

[0073] To better illustrate the effectiveness of the technical solution of this invention, this embodiment provides experimental test results for two different implementation paths, focusing on the interpretability of the model and the diagnostic accuracy, respectively. Figure 3 As shown.

[0074] In Path One, which emphasizes interpretability, this embodiment employs a pure KAN diagnostic strategy based on wavelet transform time-frequency features. In the data preprocessing stage, this path divides the data into subsamples according to requirement S1, extracts time-frequency domain physical features via path S2, and concatenates them into feature vectors, which are then input into the KAN diagnostic model trained according to S3. Test results show that the diagnostic accuracy of this scheme under four load conditions (0HP, 1HP, 2HP, and 3HP) is 81.6%, 85.0%, 79.8%, and 83.4%, respectively. Importantly, compared to the traditional multilayer perceptron, this model reduces the number of parameters by 66.3%. Through symbolic mapping, this path can generate explicit mathematical expressions, directly establishing a mathematical relationship between physical features and fault types, thereby achieving transparent analysis of the fault mechanism.

[0075] In Path Two, which focuses on high accuracy, this embodiment employs a hybrid diagnostic method based on two-dimensional time-frequency images and CNN-KAN. This path also divides subsamples, but transforms them into two-dimensional time-frequency images using wavelet transform. It then uses a convolutional autoencoder to extract deep abstract features along the S2 path, subsequently configuring and training a KAN diagnostic model. Performance tests show that this hybrid method achieves a diagnostic accuracy exceeding 98% under all load conditions, demonstrating extremely high recognition capabilities. However, because the convolutional neural network extracts high-dimensional abstract features, these features lack a direct correspondence with the physical parameters of the original signal, making it difficult to trace the symbol mapping results back to the specific physical mechanism.

[0076] In summary, this embodiment provides a KAN-based fault diagnosis method for rotating machinery, aiming to address the technical problem of the lack of interpretability in traditional "black box" models. The method first collects vibration signals from the rotating machinery and divides them into equal-length subsamples. Second, it employs a dual-path feature construction strategy: one is to extract interpretable time-frequency domain physical features such as mean, kurtosis, and root mean square through wavelet transform; the other is to generate a two-dimensional time-frequency graph through wavelet transform and extract high-dimensional abstract features using a convolutional autoencoder, adapting to the interpretability or accuracy requirements of different engineering scenarios. Based on this, a KAN diagnostic model based on the Kolmogorov-Arnold representation theorem is constructed, introducing B-spline parameterized learnable activation functions on the network connection edges to enhance the model's dynamic capture capability of nonlinear fault features. Subsequently, the KAN model is trained using a feature set, and fault classification is achieved through single-neuron output layer mapping to a threshold. Furthermore, this invention introduces symbolic mapping technology to transform the trained spline basis functions into explicit mathematical expressions (e.g., "outer ring fault decision = 0.3 × kurtosis + 1.7 × frequency modulation amplitude"), thereby establishing a physical correlation between features and faults and enabling the decision-making process to be interpretable and traceable. Through the above methods, this embodiment significantly improves the transparency and reliability of the model while ensuring high diagnostic accuracy, providing reliable technical support for the intelligent operation and maintenance of rotating machinery.

[0077] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A method for diagnosing rotating machinery faults based on KAN, characterized in that, Includes the following steps: S1. Collect vibration signals during the operation of rotating machinery, and preprocess and extract features from the vibration signals to obtain a target feature set; S2. Construct a KAN diagnostic model based on the Kolmogorov-Arnold network, wherein activation functions are set on the network connection edges of the KAN diagnostic model; S3. Train the KAN diagnostic model using the target feature set to obtain the trained KAN diagnostic model; S4. Acquire the vibration signal of the rotating mechanism under test and input it into the trained KAN diagnostic model to perform fault diagnosis and output the fault diagnosis result of the rotating mechanism under test. S5. Using symbolic mapping technology, the parameters in the trained KAN diagnostic model are transformed into explicit mathematical expressions to establish the correlation between fault diagnosis results and physical fault mechanisms.

2. The KAN-based rotating machinery fault diagnosis method according to claim 1, characterized in that, In step S1, the preprocessing and feature extraction of the vibration signal includes: According to a preset time window, the vibration signal is divided into multiple sub-samples of the same length; For each of the sub-samples, a feature extraction operation is performed to obtain the corresponding feature vector as the target feature set.

3. The KAN-based rotating machinery fault diagnosis method according to claim 2, characterized in that, For each of the sub-samples, a feature extraction operation is performed, and the target feature set is constructed using any of the following paths: Path 1: Extract multi-scale time-frequency domain physical features from the vibration signal using wavelet transform, and construct the target feature set. The time-frequency domain physical features include mean, kurtosis, variance, root mean square, skewness, and peak factor. Path 2: The vibration signal is transformed into a two-dimensional time-frequency image by wavelet transform, and the abstract features of the two-dimensional time-frequency image are extracted using a preset convolutional neural network to construct the target feature set.

4. The KAN-based rotating machinery fault diagnosis method according to claim 1, characterized in that, In step S2, the KAN diagnostic model is constructed based on the Kolmogorov-Arnold representation theorem, and its mathematical expression is: ; In the formula: For KAN diagnostic model; These are the features in the target feature set; The number of features; It is an outer unary function; It is an inner unary function; For the input vector of the th One component; The index for traversing the input features; This is an index for traversing the number of neurons in the KAN diagnostic model.

5. The KAN-based rotating machinery fault diagnosis method according to claim 4, characterized in that, In step S2, the outer unary function in the KAN diagnostic model With inner univariate function All are learnable activation functions on the network connection edges. These activation functions are parameterized by spline functions composed of piecewise polynomials, and are expressed as follows: ; In the formula: It is a spline function; For the first basis functions These are the coefficients of the corresponding basis functions.

6. The KAN-based rotating machinery fault diagnosis method according to claim 5, characterized in that, In step S3, the KAN diagnostic model is trained through the following process: The target feature set is divided into a training set and a validation set according to a certain ratio; the training set is input into the KAN diagnostic model, the prediction output is calculated through forward propagation, and the prediction error is calculated based on a preset loss function; the parameters of the piecewise polynomial spline are updated using the backpropagation algorithm until the prediction error meets the preset convergence condition, and the trained KAN diagnostic model is obtained.

7. The KAN-based rotating machinery fault diagnosis method according to claim 1, characterized in that, In step S4, the fault diagnosis is performed, and the fault diagnosis result of the rotating mechanism under test is output, specifically including: The trained KAN diagnostic model is used to calculate the vibration signal of the input rotating mechanism under test, and a continuous fault probability value is obtained based on the output of the single neuron output layer. The fault probability value is mapped and compared with a preset fault threshold to determine the fault category of the rotating mechanism under test, which is used as the fault diagnosis result.

8. The KAN-based rotating machinery fault diagnosis method according to claim 1 or 6, characterized in that, In step S5, the step of converting the parameters in the trained KAN diagnostic model into explicit mathematical expressions using symbolic mapping technology specifically includes: Extract the piecewise polynomial spline parameters of each network connection edge in the trained KAN diagnostic model after training convergence; use the symbolic regression algorithm to transform the piecewise polynomial spline parameters into a closed mathematical combination formula; based on the mathematical combination formula, analyze the explicit mathematical correlation between the fault diagnosis result and the input time-frequency domain physical features or abstract features.