Bearing fault diagnosis method based on pyramid fusion of local details and residual KAN

Abstract

The invention provides a bearing fault diagnosis method based on pyramid fusion of local details and residual KAN, relates to the technical field of deep learning and intelligent fault diagnosis, and combines a multi-scale feature fusion technology and a feature extraction network to improve the accuracy of fault diagnosis. Firstly, bearing vibration signals are collected through a preset sampling frequency and an accelerometer, and the bearing vibration signals are divided into a training set, a verification set and a test set. Then, a one-dimensional vibration signal is adopted as an input feature, and random Gaussian noise is added on the basis to improve the generalization performance of data; according to the constructed deep residual Kolmogorov-Arnod network, through a pyramid fusion strategy with local details reserved and a global attention mechanism, features are effectively extracted. Finally, the trained and optimized fault diagnosis model can quickly and accurately identify the reason of the bearing fault, and on this basis, the fault diagnosis requirements under different working conditions can be met.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning and intelligent fault diagnosis, and in particular to a bearing fault diagnosis method based on pyramid fusion of local details and residual KAN. Background Art

[0002] Bearings are called "industrial joints" and are widely used in various industrial equipment. The health monitoring and fault diagnosis of bearings are crucial to the development of industrial equipment. Bearing fault diagnosis usually adopts methods based on vibration signal analysis, mainly collecting vibration signals through sensors, and amplifying, filtering and converting these signals to obtain the vibration amplitude or vibration frequency. Early methods relied on traditional signal processing, but as the data scale increases and the failure mode becomes more complicated, these methods are no longer sufficient to meet the challenges of feature extraction.

[0003] Machine learning algorithms such as support vector machines and artificial neural networks have been introduced to extract features, but these methods are too dependent on expert knowledge and have difficulty capturing deep data features. Deep learning methods have attracted attention for their ability to automatically extract deep features, but the traditional deep learning network structure is fixed and has difficulty extracting complex nonlinear characteristics in high-dimensional data.

[0004] The Kolmogorov-Arnold network can realize complex nonlinear modeling. The connection weights are spline functions, which can describe the transformation and adjustment process of signals when they are transmitted between nodes, and can accurately realize complex nonlinear relationship modeling.

[0005] On this basis, methods such as attention mechanism and multi-scale feature fusion are introduced to improve feature extraction, but the recognition accuracy still needs to be improved. Summary of the invention

[0006] In order to solve the technical problem of low bearing fault diagnosis accuracy in the prior art, the present invention provides a bearing fault diagnosis method based on pyramid fusion of local details and residual KAN.

[0007] The technical solution provided by the present invention is as follows:

[0008] The bearing fault diagnosis method based on local detail pyramid fusion and residual KAN provided by the present invention includes:

[0009] S1. Collect bearing vibration signals at multiple time points through a preset sampling frequency and accelerometer layout to generate time series vibration data;

[0010] S2. Divide the collected vibration signal into data sets, wherein the data sets include a training set, a validation set, and a test set, each of the data sets is independent of each other, and each of the data sets contains a corresponding vibration signal and a fault type;

[0011] S3. Extract fault diagnosis features from vibration signals using a deep residual network (ResNet) Kolmogorov-Arnold network (KAN), wherein the Kolmogorov-Arnold network includes a Kolmogorov-Arnold convolutional layer using Gram polynomials to replace the B-spline function of the traditional convolutional Kolmogorov-Arnold network, and on this basis, a global attention mechanism is added to each layer of the network to realize the collection of important features;

[0012] S4, by pooling the low-level features of the network during feature extraction and fusing them with high-level features through pyramid features, weighted fusion of multi-scale features is achieved;

[0013] S5, inputting the training set into the Kolmogorov-Arnold network for training, and using the validation set to adjust the hyperparameters to optimize the model performance;

[0014] S6. Use the trained and optimized model to perform fault diagnosis on the test set to achieve automatic identification of bearing fault types.

[0015] The beneficial effects brought about by the technical solution provided by the present invention include at least:

[0016] In the present invention, by combining the local detail-preserving pyramid fusion strategy, the global attention mechanism, and the deep residual Kolmogorov-Arnold network, the complex nonlinear features of the bearing vibration signal are effectively extracted and the fault type is accurately identified. Compared with the traditional method, the present invention significantly improves the accuracy and generalization performance of bearing fault diagnosis, and can maintain a high diagnostic accuracy under noise interference and different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0018] Figure 1 A flow chart of a bearing fault diagnosis method based on pyramid fusion of local details and residual KAN provided in an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of a multi-scale feature fusion process of a bearing fault diagnosis method based on local detail pyramid fusion and residual KAN provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0020] The technical solution of the present invention is described below in conjunction with the accompanying drawings.

[0021] In the embodiments of the present invention, words such as "exemplarily" and "for example" are used to indicate examples, illustrations or explanations. Any embodiment or design described as "example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of the word "example" is intended to present the concept in a specific way. In addition, in the embodiments of the present invention, the meaning expressed by "and / or" can be both, or it can be either of the two.

[0022] In the embodiments of the present invention, "image" and "picture" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same. "of", "corresponding, relevant" and "corresponding" can sometimes be used interchangeably. It should be noted that when the difference between them is not emphasized, the meanings they intend to express are the same.

[0023] In the embodiments of the present invention, sometimes a subscript such as W1 may be mistakenly written as a non-subscript such as W1. When the difference is not emphasized, the meanings to be expressed are consistent.

[0024] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0025] Reference Manual Attached Figure 1 , shows a flow chart of a bearing fault diagnosis method based on local detail pyramid fusion and residual KAN provided in an embodiment of the present invention.

[0026] The embodiment of the present invention provides a bearing fault diagnosis method based on local detail pyramid fusion and residual KAN. The processing flow may include the following steps:

[0027] S1. Collect bearing vibration signals at multiple time points through a preset sampling frequency and accelerometer layout to generate time series vibration data;

[0028] S2. Divide the collected vibration signal into data sets, wherein the data sets include a training set, a validation set, and a test set, each of the data sets is independent of each other, and each of the data sets contains a corresponding vibration signal and a fault type;

[0029] S3. Extract fault diagnosis features from vibration signals using a deep residual network (ResNet) Kolmogorov-Arnold network (KAN), wherein the Kolmogorov-Arnold network includes a Kolmogorov-Arnold convolutional layer using Gram polynomials to replace the B-spline function of the traditional convolutional Kolmogorov-Arnold network, and on this basis, a global attention mechanism is added to each layer of the network to realize the collection of important features;

[0030] S4, by pooling the low-level features of the network during feature extraction and fusing them with high-level features through pyramid features, weighted fusion of multi-scale features is achieved;

[0031] S5, inputting the training set into the Kolmogorov-Arnold network for training, and using the validation set to adjust the hyperparameters to optimize the model performance;

[0032] S6. Use the trained and optimized model to perform fault diagnosis on the test set to achieve automatic identification of bearing fault types.

[0033] In the embodiment of the present invention, the specific implementation is as follows:

[0034] The signal is preprocessed and the relevant data set is prepared. Specifically, the original vibration signal is sampled with a sampling length of 1024, and the sampled data is randomly divided in a ratio of 7:2:1, of which 70% is used as a training set. 20% is used as a validation set to adjust the hyperparameters in the experiment and evaluate the model performance, and the remaining 10% is used as a test set to verify the generalization performance of the model. There is no overlap between the training set samples and the test set samples. At the same time, a series of data enhancement methods are used, including but not limited to random Gaussian noise, random scale and time series stretching.

[0035] Gram polynomials are used instead of B-splines in KAN convolution as the activation function, called GKAN convolution. This GKAN convolution is used to replace the standard convolution in the residual network to form a deep residual Kolmogorov-Arnold network. The specific implementation is as follows:

[0036] The divided data sets are enhanced with random Gaussian noise data in turn and used as feature input;

[0037] The KAN convolution method is:

[0038] φ=w1·spline(x)+w2·silu(x)

[0039] Among them, w1 and w2 are used to control the overall size of the activation function and can be regarded as redundant terms, while spline uses the B-spline function as a learnable activation function;

[0040] The Gram-KAN convolution method is:

[0041] The Gram-KAN convolution method is an improved KAN convolution operation that uses the Gram polynomial basis to construct an orthogonal basis of the function space, thereby replacing the B-spline function in the traditional KAN convolution.

[0042] For a given input, the polynomial basis can be generated by the following recurrence relation:

[0043]

[0044] Among them, P n represents the nth polynomial, which can be obtained by recursion. These polynomial basis functions are used to parameterize the weights of the convolution kernels.

[0045] In the Gram-KAN convolution method, the convolution kernel can be expressed as follows:

[0046]

[0047] In the formula, N is the highest order of the polynomial, is an n-degree polynomial of x, used to represent the high-dimensional features of the input data. Weight function Where δ(·) is the Dirac function, s is the displacement of the Dirac pulse, and the Dirac function can map the position of the input signal during the convolution process and efficiently represent the local information of the input features;

[0048] The GKAN convolution is used to replace the standard convolution layer in the original residual block to construct a deep residual Kolmogorov-Arnold network. The deep residual Kolmogorov-Arnold network consists of multiple stacked residual blocks, each of which contains multiple GKAN convolution modules. The input features are combined with the features processed by the GKAN convolution through skip connections. This ensures smooth propagation of gradients throughout the deep network and reduces the risk of gradient vanishing. In order to keep the input and output feature dimensions consistent in the skip connection, additional convolution operations are used to adjust the dimensions as needed.

[0049] The embodiment of the present invention also proposes a local detail-preserving pyramid fusion strategy. Compared with the traditional feature fusion method, the design of the local detail-preserving pyramid fusion strategy module fully considers the inherent connection between features at different levels and can more effectively fuse and extract deep features, such as Figure 2 As shown, the specific implementation is as follows:

[0050] The input data first goes through regular convolution, normalization layer and max pooling operation, and then enters multiple basic blocks for GKAN convolution operation. When processing one-dimensional bearing vibration signals, the lowest level convolution usually captures the local features of the signal, which are essential for identifying key patterns and abnormal changes in the signal. Directly incorporating the lowest level features into pyramid fusion will ignore local details, thus affecting the recognition ability of the model. To this end, the lowest level features are processed with adaptive max pooling and average pooling to better retain these key local information.

[0051] After entering the high-level output, the single-scale output cannot effectively capture targets of different scales, resulting in poor generalization of the model. Therefore, the feature pyramid structure is introduced to enhance the information fusion ability of the model. By integrating multi-level feature information, the complementarity of features at each level is fully utilized to capture targets of different scales. For one-dimensional vibration signal processing, the mid- and high-level convolutional features represent abstract information at different levels. The fusion of these features helps improve the performance of the model in complex scenes. The lowest-level features have retained local detail information through pooling operations, so they are directly upsampled to the second layer for fusion in the feature pyramid extraction stage.

[0052] Finally, the feature pyramid fusion output is weightedly fused with the results of the maximum pooling layer and the average pooling layer. W1, W2, and W3 correspond to the weights added after the three outputs. This method not only effectively retains the local detail features of the lowest layer, but also integrates high-level multi-scale information, thereby improving the accuracy of feature extraction and fusion. W1, W2, and W3 will redistribute their respective proportions each time the weights are updated. The formula is as follows:

[0053]

[0054] Among them, w i is the i-th element of the original weight vector w, ε is a small constant used to prevent errors caused by zero division, and n refers to the length of the weight vector. During each training process, it is automatically adjusted through the back propagation algorithm to minimize the loss function. Specifically, it is a set of learnable weight vectors whose values ​​are updated according to the loss gradient during each training iteration.

[0055] The beneficial effects brought about by the technical solution provided by the embodiment of the present invention include at least:

[0056] In the present invention, by combining the local detail-preserving pyramid fusion strategy, the global attention mechanism, and the deep residual Kolmogorov-Arnold network, the complex nonlinear features of the bearing vibration signal are effectively extracted and the fault type is accurately identified. Compared with the traditional method, the present invention significantly improves the accuracy and generalization performance of bearing fault diagnosis, and can maintain a high diagnostic accuracy under noise interference and different working conditions.

[0057] The above contents are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art who is familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

[0058] There are a few points to note:

[0059] (1) The drawings of the embodiments of the present invention only relate to the structures related to the embodiments of the present invention, and other structures may refer to the general design.

[0060] (2) For the sake of clarity, in the drawings used to describe the embodiments of the present invention, the thickness of the layers or regions is exaggerated or reduced, that is, these drawings are not drawn according to the actual scale. It is understood that when an element such as a layer, film, region or substrate is referred to as being "on" or "under" another element, the element may be "directly" "on" or "under" the other element or there may be intermediate elements.

[0061] (3) In the absence of conflict, the embodiments of the present invention and the features therein may be combined with each other to obtain new embodiments.

[0062] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. The protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A bearing fault diagnosis method based on local detail pyramid fusion and residual KAN, characterized in that: include: S1. Collect bearing vibration signals at multiple time points through a preset sampling frequency and accelerometer layout to generate time series vibration data; S2. Divide the collected vibration signal into data sets, wherein the data sets include a training set, a validation set, and a test set, each of the data sets is independent of each other, and each of the data sets contains a corresponding vibration signal and a fault type; S3. Extract fault diagnosis features from vibration signals using a deep residual network (ResNet) Kolmogorov-Arnold network (KAN), wherein the Kolmogorov-Arnold network includes a Kolmogorov-Arnold convolutional layer using Gram polynomials to replace the B-spline function of the traditional convolutional Kolmogorov-Arnold network, and on this basis, a global attention mechanism is added to each layer of the network to realize the collection of important features; S4, by pooling the low-level features of the network during feature extraction and fusing them with high-level features through pyramid features, weighted fusion of multi-scale features is achieved; S5, inputting the training set into the Kolmogorov-Arnold network for training, and using the validation set to adjust the hyperparameters to optimize the model performance; S6. Use the trained and optimized model to perform fault diagnosis on the test set to achieve automatic identification of bearing fault types.

2. The bearing fault diagnosis method based on local detail pyramid fusion and residual KAN according to claim 1 is characterized in that: The generated data set in S2 further comprises: After preprocessing the signal, the original collected vibration signal is windowed and sampled with a sampling length of 1024, and the sampled data is randomly divided in a ratio of 7:2:1, of which 70% is used as a training set, 20% is used as a validation set to adjust the hyperparameters in the experiment and evaluate the model performance, and the remaining 10% is used as a test set to verify the generalization performance of the model. There is no overlap between the training set samples and the test set samples.

3. The bearing fault diagnosis method based on local detail pyramid fusion and residual KAN according to claim 2 is characterized in that: The KAN specifically includes: Gram polynomials are used to replace B-splines in KAN convolution as the activation function to form GKAN convolution. The GKAN convolution replaces the standard convolution in the residual network to form a deep residual KAN. The specific process is as follows: The divided data sets are enhanced with random Gaussian noise data in turn and used as feature input; The KAN convolution method is: φ=w1·spline(x)+w2·silu(x) Among them, w1 and w2 are used to control the overall size of the activation function and can be regarded as redundant terms, while spline uses the B-spline function as a learnable activation function; The Gram-KAN convolution method is: The Gram polynomial basis is used to construct an orthogonal basis of the function space, thereby replacing the B-spline function in the traditional KAN convolution. For a given input x, the polynomial basis is generated by the following recursive relationship: Among them, P n Represents the nth polynomial, which is obtained by recursive relationship. The above polynomial basis function is used to parameterize the weight of the convolution kernel. In the Gram-KAN convolution method, the convolution kernel is expressed as: in, N is the highest order of the polynomial, is an n-order polynomial of x, used to represent the high-dimensional features of the input data, and the weight function Where δ(·) is the Dirac function, s is the displacement of the Dirac pulse, and the Dirac function can map the position of the input signal during the convolution process and efficiently represent the local information of the input features; GKAN convolution is used to replace the standard convolution layer in the original residual block to construct a deep residual Kolmogorov-Arnold network. The deep residual Kolmogorov-Arnold network consists of multiple stacked residual blocks, each of which contains multiple GKAN convolution modules. The input features are combined with the features processed by GKAN convolution through jump connections. In order to maintain the consistency of the input and output feature dimensions in the jump connection, additional convolution operations are used to adjust the dimensions as needed.

4. The bearing fault diagnosis method based on local detail pyramid fusion and residual KAN according to claim 3 is characterized in that: The fusion of the multi-scale features further includes: After conventional convolution, normalization and maximum pooling operations, the input data enters multiple basic blocks for GKAN convolution operations, and adaptive maximum pooling and average pooling are performed on the bottom-level features; The feature pyramid structure is introduced to enhance the information fusion capability of the model. By integrating multi-level feature information, targets of different scales are captured. The feature pyramid extraction stage is directly upsampled to the second layer for fusion. The results of the feature pyramid fusion output, the maximum pooling layer, and the average pooling layer are weighted fused. W1, W2, and W3 correspond to the weights added to the outputs of the three layers, and the weights are redistributed to their respective proportions when updated: Among them, w i is the i-th element of the original weight vector w, ε is a small constant used to prevent errors caused by division by zero, and n refers to the length of the weight vector, which is automatically adjusted by the back-propagation algorithm during each training process to minimize the loss function.

5. The bearing fault diagnosis method based on local detail pyramid fusion and residual KAN according to claim 4 is characterized in that: The weighted fusion includes: The weights are redistributed to their respective proportions when updated: Among them, w i is the i-th element of the original weight vector w, ε is a small constant used to prevent errors caused by division by zero, and n refers to the length of the weight vector, which is automatically adjusted by the back-propagation algorithm during each training process to minimize the loss function.

6. The bearing fault diagnosis method based on local detail pyramid fusion and residual KAN according to claim 1, characterized in that: The optimization model performance further includes: The global attention mechanism is used to fuse features and then input into the classifier to reduce model complexity and parameter quantity and realize fault classification. The global attention mechanism selectively focuses on important areas or features in the input data, captures the relationship between different parts of the input, and assigns different weights to them to improve the overall performance of the model. The goal of the global attention mechanism is to optimize model performance by capturing the relationship between global features.

Images