Improved bearing fault diagnosis method based on ghost feature enhancement and light-weight mobilevit

Abstract

The application relates to a MobileVit light-weight bearing fault diagnosis method based on a Ghost feature enhancement improvement, and belongs to the technical field of mechanical fault diagnosis, and comprises the following steps: S1, multi-dimensional feature extraction is carried out by using a three-channel preprocessing module FFE; S2, features extracted from each channel are fused to generate a fused feature dataset; S3, the fused feature dataset is processed by a Ghost enhanced feature attention network GEFA-Net, feature representation is enhanced, additional feature mapping is generated, and a fault diagnosis result is generated. The application has strong robustness and high efficiency, can effectively extract key features of a vibration signal, optimizes feature expression, and has light weight while ensuring accurate identification.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis technology, and relates to a MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement. Background Technology

[0002] Rolling bearings are key components in rotating machinery and are widely used in various industrial equipment. However, due to long-term operation, rolling bearings are prone to wear, fatigue, and other failures. Their early fault characteristics are weak and easily masked by noise, posing a significant challenge to fault diagnosis. Traditional fault diagnosis methods mainly fall into two categories: signal processing methods and machine learning-based techniques.

[0003] Signal processing methods mainly include three types: Mode Decomposition (EMD), Wavelet Transform (WT), and Variational Mode Decomposition (VMD). Mode Decomposition breaks down complex signals into simpler signals of different scales for preliminary fault analysis. Wavelet Transform can capture the local characteristics of a signal and is suitable for analyzing non-stationary signals. Variational Mode Decomposition improves upon EMD and WT, enabling more accurate separation of different modal components in a signal.

[0004] In addition, machine learning techniques mainly include Principal Component Analysis (PCA), Support Vector Machines (SVM), and Artificial Neural Networks (ANN). PCA reduces data complexity through dimensionality reduction, improving the efficiency of fault diagnosis. Support Vector Machines, based on statistical learning theory, are suitable for fault classification with small sample sizes. Artificial Neural Networks, by simulating biological neural networks, enable the learning and classification of complex nonlinear fault features.

[0005] Although the above methods improve the accuracy of fault diagnosis to some extent, their high requirements for engineer experience, high demand for computing resources, and complex parameter settings limit their widespread application in practice. In addition, these methods are computationally complex and their diagnostic accuracy and speed are insufficient. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A lightweight bearing fault diagnosis method based on Ghost feature enhancement and improvement includes the following steps:

[0009] S1: Multidimensional feature extraction is performed using the three-channel preprocessing module FFE;

[0010] S2: The features extracted from each channel are fused to generate a fused feature dataset;

[0011] S3: The fused feature dataset is processed by the Ghost Enhanced Feature Attention Network GEFA-Net to enhance feature representation and generate additional feature maps, thus generating fault diagnosis results.

[0012] Furthermore, step S1, which involves multidimensional feature extraction using the three-channel preprocessing module FFE, specifically includes:

[0013] The first channel uses Fast Fourier Transform (FFT) to analyze the periodicity and frequency distribution of the vibration signal and capture the basic structure of the fault signal.

[0014] The second channel utilizes the frequency-weighted energy operator (FWEO) to enhance the frequency characteristics and energy distribution of the signal, thereby isolating key fault features.

[0015] The third channel uses signal envelope technology to extract the transient and nonlinear characteristics of vibration signals, comprehensively capturing subtle changes in the fault.

[0016] Furthermore, the first channel employs Fast Fourier Transform (FFT) to analyze the periodicity and frequency distribution of the vibration signal, capturing the basic structure of the fault signal, specifically including:

[0017] First, the vibration signal is segmented.

[0018] Then, the FFT algorithm is used to transform each signal segment to obtain the spectral information. The calculation formula is as follows:

[0019]

[0020] Where X(f) is the result of the Discrete Fourier Transform, i.e., the frequency domain signal; x(n) is the discrete time domain signal, representing the signal value at time n; f is the frequency exponent; N is the total number of sampling points of the signal, i.e., the length of the signal; j is the imaginary unit; and n is the time index, with a value range of 0, 1, 2…N-1.

[0021] Finally, the spectral amplitude was analyzed to identify fault characteristic frequencies.

[0022] Furthermore, the second channel utilizes the frequency-weighted energy operator (FWEO) to enhance the frequency characteristics and energy distribution of the signal, isolating key fault features, specifically including:

[0023] The first and second derivatives of the signal are calculated using FWEO, and the fault characteristics are extracted by combining the instantaneous frequency and energy distribution. The calculation formula is as follows:

[0024]

[0025] Where x(t) represents the signal or function at time t, which represents the state of the system at time t; Ω(x(t)) represents the characteristic quantity of the signal x(t), through which fault characteristics can be extracted; t is a time variable.

[0026] Furthermore, the third channel employs signal envelope technology to extract the transient and nonlinear characteristics of the vibration signal, comprehensively capturing subtle changes in the fault, specifically including:

[0027] First, perform a Hilbert transform on the signal to obtain a complex analytic signal;

[0028] Then calculate the amplitude of the complex analytic signal to obtain the signal envelope;

[0029] Finally, the envelope of the signal is analyzed to extract the impact fault characteristics of the rotating machinery.

[0030] Furthermore, the GEFA-Net combines the Ghost Module and the efficient Pyramid Squared Attention (EPSA) mechanism.

[0031] The Ghost Module first performs standard convolution operations to generate a subset of feature maps; then it performs a series of linear transformations on these feature maps to generate additional feature maps.

[0032] The EPSA mechanism introduces a multi-scale feature extraction mechanism, which processes features at each scale independently, and then performs weighted fusion through an attention mechanism to model the features at multiple scales.

[0033] Furthermore, the feature map is subjected to a series of linear transformations, including weighted summation and nonlinear activation functions.

[0034] Furthermore, the specific process of the EPSA mechanism includes:

[0035] The input feature map is decomposed into sub-feature maps of different scales;

[0036] Each sub-feature map is processed independently to extract local features;

[0037] An enhanced feature representation is generated by weighted fusion of features at different scales through an attention mechanism.

[0038] The beneficial effects of this invention are as follows: this method exhibits high diagnostic accuracy and low computational complexity under complex working conditions, demonstrating strong robustness and high efficiency. The lightweight and efficient bearing fault diagnosis method proposed in this invention can effectively extract key features of vibration signals and optimize feature representation, ensuring accurate identification while maintaining lightweight design, thus possessing broad industrial application prospects.

[0039] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0040] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0041] Figure 1 This is a flowchart illustrating the workflow of the three-channel feature extraction module described in this invention.

[0042] Figure 2 This is a schematic diagram of the lightweight GEFA-Net network structure described in this invention;

[0043] Figure 3 This is a linear program structure diagram of the Ghost Module described in this invention;

[0044] Figure 4 This is a structural diagram of the EPSA described in this invention;

[0045] Figure 5 This is a schematic diagram of the Context Broadcasting module (CB) structure described in this invention;

[0046] Figure 6 This is a schematic diagram of the CB-TransformerEncoder structure described in this invention;

[0047] Figure 7 An experimental platform built for the experimental testing described in this invention;

[0048] Figure 8 The fault types of the dataset tested in the experiment described in this invention;

[0049] Figure 9 The image is an RGB image after three-channel feature extraction as described in this invention; where (a) is an image processed by FFT, (b) is an image converted by FWEO, and (c) is an image processed by envelope detection.

[0050] Figure 10 This paper compares the experimental results of Ottawa data processed by the three preprocessing methods (original signal, CWT, FFE) described in this invention in different models.

[0051] Figure 11For the accuracy comparison of the Ottawa dataset described in this invention under different models; (a) accuracy of each model after 50 training iterations; (b) training loss of each model after 50 training iterations;

[0052] Figure 12 The following are the confusion matrices obtained from training the Ottawa dataset according to the present invention: (a) is the confusion matrix obtained after training 50 iterations of GEFA-Net with the Ottawa dataset; (b) is the confusion matrix obtained after training 50 iterations of MobileNetV3 with the Ottawa dataset; (c) is the confusion matrix obtained after training 50 iterations of ResNet with the Ottawa dataset; (d) is the confusion matrix obtained after training 50 iterations of Vision transformer with the Ottawa dataset; (e) is the confusion matrix obtained after training 50 iterations of MobileVit with the Ottawa dataset; (f) is the confusion matrix obtained after training 50 iterations of Swin transformer with the Ottawa dataset.

[0053] Figure 13 This invention compares experimental results of the three preprocessing methods (original signal, CWT, FFE) used to process self-built data in different models.

[0054] Figure 14 This invention provides a comparison of the accuracy of the self-built dataset under different models; (a) the accuracy after 50 training iterations for each model; (b) the training loss after 50 training iterations for each model.

[0055] Figure 15 The following are the confusion matrices obtained from training the self-built dataset described in this invention: (a) is the confusion matrix obtained after training GEFA-Net for 50 iterations using the Ottawa dataset; (b) is the confusion matrix obtained after training MobileNetV3 for 50 iterations using the self-built dataset; (c) is the confusion matrix obtained after training ResNet for 50 iterations using the self-built dataset; (d) is the confusion matrix obtained after training Vision transformer for 50 iterations using the self-built dataset; (e) is the confusion matrix obtained after training MobileVit for 50 iterations using the self-built dataset; and (f) is the confusion matrix obtained after training Swin transformer for 50 iterations using the self-built dataset.

[0056] Figure 16 The results of the ablation experiment of the GEFA-Net network model described in this invention are shown in (a) and (b) are the Ottawa dataset and self-built dataset. Detailed Implementation

[0057] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0058] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0059] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present 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, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0060] like Figure 1-6 As shown, this invention provides a lightweight bearing fault diagnosis method based on Ghost feature enhancement and improved MobileVit, comprising the following steps: First, a three-channel preprocessing module (FFE) is used to acquire the time-domain, frequency-domain, and time-frequency-domain feature data of the vibration signal, and multi-dimensional feature extraction is performed; then, the feature data is converted into two-dimensional image information, and finally, channel fusion technology is used to convert it into three channels of an RGB image. We construct a lightweight Ghost-enhanced Feature Attention Network (GEFA-Net), and take the RGB image as input. We utilize the Ghost module and the efficient squared pyramid attention (EPSA) mechanism to enhance the feature representation, generating additional feature maps through linear operations, thereby reducing computational requirements. GEFA-Net combines advanced structures such as the Ghost module and the EPSA module, optimizing feature extraction and model architecture, improving the accuracy of fault diagnosis while reducing the number of parameters and computational complexity.

[0061] The three-channel feature extraction module (FFE) is one of the core components of this invention, designed to extract key features from vibration signals under complex operating conditions. The specific steps are as follows:

[0062] (1) Fast Fourier Transform (FFT):

[0063] FFT is used to analyze the periodicity and frequency distribution of vibration signals, capturing the basic structure of fault signals. By converting the time-domain signal to the frequency-domain signal, the frequency components become more intuitive. Calculation formula:

[0064]

[0065] In practice, the vibration signal is first segmented. Then, the FFT algorithm is used to transform each segment to obtain spectral information. By analyzing the spectral amplitude, fault characteristic frequencies can be identified. The FFT results can be used for spectral analysis to extract the main frequency components. By observing the amplitude and phase of the frequency components, it can be determined whether fault characteristics exist in the vibration signal.

[0066] (2) Frequency-weighted energy operator (FWEO):

[0067] FWEO effectively isolates key fault characteristics by emphasizing the frequency characteristics and energy distribution of the signal. By combining the signal's frequency and energy information, FWEO can more accurately detect fault characteristics in noisy environments. Calculation formula:

[0068]

[0069] In practical applications, FWEO effectively extracts fault features by calculating the first and second derivatives of a signal and combining them with its instantaneous frequency and energy distribution. This method exhibits robustness to noise and maintains high fault identification accuracy even in high-noise environments. Under complex operating conditions, FWEO can better separate fault signals from noise, improving the reliability and stability of fault diagnosis.

[0070] (3) Signal envelope analysis:

[0071] Envelope analysis is used to examine the transient and nonlinear characteristics of vibration signals, enabling a comprehensive capture of subtle changes in faults. By analyzing the envelope of a signal, envelope analysis can extract its transient features.

[0072] Calculation steps: First, perform a Hilbert transform on the signal to obtain a complex analytic signal. Then, calculate the amplitude of the complex analytic signal to obtain its envelope. Envelope analysis is particularly suitable for processing non-stationary signals. By analyzing the signal envelope, it is possible to effectively extract the impact fault characteristics of rotating machinery, such as faults in the inner or outer ring of a bearing, or in the rolling elements.

[0073] Features extracted from each channel are used to generate a fused feature dataset through a feature fusion strategy, reducing the interference of different operating conditions on fault signal diagnosis. The feature fusion strategy integrates different features extracted from the three channels to establish a comprehensive feature dataset, significantly improving the model's diagnostic efficiency and accuracy under complex operating conditions.

[0074] GEFA-Net is a lightweight neural network architecture proposed in this invention. By combining the Ghost Module and EPSA mechanism, and through linear operations and efficient attention mechanisms, it enhances feature representation, reduces computational requirements, and optimizes the model's structural framework. GEFA-Net includes a lightweight convolutional neural network architecture, combined with MobileVit and Context Broadcasting (CB) modules, to optimize the feature extraction process and improve the speed and accuracy of fault diagnosis.

[0075] The Ghost Module generates additional feature maps through linear operations, significantly reducing computational requirements. Compared to traditional convolutional operations, the Ghost Module reduces computational complexity while maintaining feature representation capabilities. Specifically, it first performs standard convolution operations to generate a subset of feature maps. Then, it applies a series of linear transformations to these feature maps to generate additional feature maps. These linear transformations include simple operations such as weighted summation and non-linear activation functions. The advantage of the Ghost Module lies in generating more feature maps through simple linear operations, thereby reducing the computational cost of convolutional operations. This significantly reduces the model's computational complexity and number of parameters while maintaining feature extraction capabilities.

[0076] The EPSA mechanism enhances feature representation capabilities by modeling features at multiple scales. EPSA effectively captures feature information at different scales, improving the diagnostic accuracy of the network. The EPSA module introduces a multi-scale feature extraction mechanism, processing features at each scale independently, and then weighted and fused them using an attention mechanism. The specific process includes:

[0077] The input feature map is decomposed into sub-feature maps of different scales.

[0078] Each sub-feature map is processed independently to extract local features.

[0079] An enhanced feature representation is generated by weighted fusion of features at different scales through an attention mechanism.

[0080] The EPSA module, through multi-scale feature modeling, can capture feature information at different scales, improving the richness and robustness of feature representation. In fault diagnosis, it can better identify different types of fault features.

[0081] The diagnostic framework is the overall architecture for intelligent fault diagnosis in this invention, including steps such as data acquisition, preprocessing, feature extraction, feature fusion, network training, and fault diagnosis. The specific steps are as follows:

[0082] Data Acquisition and Preprocessing:

[0083] Key fault features were extracted from vibration signals using FFT, FWEO, and envelope analysis.

[0084] Vibration signals are acquired by sensors and then denoised and normalized to improve signal quality and availability.

[0085] The extracted features are converted into RGB images to establish a comprehensive feature dataset. In practice, different features are mapped to different channels (R, G, B) of the image to form a multi-dimensional feature fusion image representation.

[0086] In the data preprocessing stage, the vibration signal is first filtered to remove high-frequency noise and low-frequency interference. Then, the signal is normalized to ensure that the signal amplitude is within a uniform range, reducing the impact of amplitude differences on feature extraction.

[0087] Feature fusion and network training:

[0088] Feature representation is optimized by fusing features extracted from different channels. Convolutional Neural Networks (CNNs) are used to process images and extract high-level features.

[0089] GEFA-Net was used for network training to optimize network parameters and improve fault diagnosis performance. During training, the cross-entropy loss function and Adam optimizer were used to improve the model's convergence speed and diagnostic accuracy.

[0090] In the feature fusion process, high-level features are extracted from the feature image step by step through convolutional and pooling layers. Multiple convolutional operations capture feature information at different scales and depths. Finally, fully connected layers fuse the extracted features to generate the final feature representation.

[0091] Fault diagnosis and assessment:

[0092] Fault diagnosis is performed using a pre-trained GEFA-Net model. A new vibration signal is input, and through preprocessing and feature extraction steps, a corresponding feature image is generated.

[0093] The GEFA-Net model is used to classify the feature images and output the fault type and confidence score.

[0094] Experimental verification ensures the accuracy and speed of diagnosis under complex operating conditions. Specific verification includes testing under different loads, speeds, and environmental noise conditions, followed by comparative analysis with existing methods.

[0095] In the fault diagnosis process, the input vibration signal is first preprocessed and its features are extracted to generate a feature image. Then, the feature image is input into the GEFA-Net model for feature extraction and classification. By analyzing the model's output, the fault type and confidence level are determined.

[0096] To verify the effectiveness of this invention, a dataset from the University of Ottawa and a self-built dataset were used for verification. For example... Figure 7-16 As shown, experimental results demonstrate that the method of this invention has high diagnostic accuracy and low computational complexity under complex working conditions, exhibiting strong robustness and high efficiency.

[0097] (1) Experimental dataset:

[0098] The University of Ottawa dataset contains vibration signal data under three operating conditions: normal, inner ring fault, and outer ring fault. The sampling frequency is 200kHz, and the duration of each sample is 10 seconds.

[0099] Self-built dataset: Includes vibration signal data under four operating conditions: normal, inner ring fault, outer ring fault, and rolling element fault. The sampling frequency is 25.6kHz, and the duration of each sample is 5 seconds.

[0100] (2) Experimental setup:

[0101] Preprocessing: FFT, FWEO, and envelope analysis were performed on all samples to extract key features. The features were then converted into RGB images and used as input to GEFA-Net.

[0102] Network training: The model was trained using GEFA-Net, and its performance was evaluated using cross-validation. Training parameters included a learning rate of 0.001, a batch size of 32, and 50 training epochs.

[0103] In the experiment, all datasets were first divided into training, validation, and test sets in a ratio of 6:2:2. Then, data augmentation was performed on the training set, such as random pruning, flipping, and rotation, to increase the diversity and robustness of the samples.

[0104] (3) Experimental results:

[0105] On the University of Ottawa dataset, GEFA-Net achieved a fault diagnosis accuracy of 98.53%, outperforming traditional convolutional neural networks and other lightweight models. Specific comparison results are shown in Table 1.

[0106] Table 1

[0107]

[0108] On the self-built dataset, GEFA-Net achieved a fault diagnosis accuracy of 99.98%, as shown in Table 2. Furthermore, it significantly reduced the number of model parameters and computational complexity; the comparison results for each parameter are shown in Table 3.

[0109] Table 2

[0110]

[0111] Table 3

[0112]

[0113]

[0114] Experimental results show that GEFA-Net exhibits high diagnostic accuracy and stability under different load and speed conditions. Compared with traditional fault diagnosis methods, GEFA-Net can better adapt to complex operating conditions, improving the robustness and reliability of diagnosis.

[0115] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium. When the program is executed, it can implement the steps of the method. The storage medium may be, for example, ROM / RAM, magnetic disk, optical disk, etc.

[0116] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A lightweight bearing fault diagnosis method based on Ghost feature enhancement and improvement, characterized in that: Includes the following steps: S1: Multidimensional feature extraction is performed using the three-channel preprocessing module FFE; S2: The features extracted from each channel are fused to generate a fused feature dataset; S3: The fused feature dataset is processed by the Ghost Enhanced Feature Attention Network GEFA-Net to enhance feature representation and generate additional feature maps, thus generating fault diagnosis results; The GEFA-Net combines the Ghost Module and the efficient pyramid squared attention mechanism EPSA; The Ghost Module first performs standard convolution operations to generate a subset of feature maps; then it performs a series of linear transformations on these feature maps to generate additional feature maps. The EPSA mechanism introduces a multi-scale feature extraction mechanism, which processes the features at each scale independently, and then performs weighted fusion through an attention mechanism to model the features at multiple scales. The specific process of the EPSA mechanism includes: The input feature map is decomposed into sub-feature maps of different scales; Each sub-feature map is processed independently to extract local features; An enhanced feature representation is generated by weighted fusion of features at different scales through an attention mechanism.

2. The MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement as described in claim 1, characterized in that: Step S1, which involves multidimensional feature extraction using the three-channel preprocessing module FFE, specifically includes: The first channel uses Fast Fourier Transform (FFT) to analyze the periodicity and frequency distribution of the vibration signal and capture the basic structure of the fault signal. The second channel utilizes the frequency-weighted energy operator (FWEO) to enhance the frequency characteristics and energy distribution of the signal, thereby isolating key fault features. The third channel uses signal envelope technology to extract the transient and nonlinear characteristics of vibration signals, comprehensively capturing subtle changes in the fault.

3. The MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement as described in claim 2, characterized in that: The first channel uses Fast Fourier Transform (FFT) to analyze the periodicity and frequency distribution of the vibration signal, capturing the basic structure of the fault signal, specifically including: First, the vibration signal is segmented and processed. Then, the FFT algorithm is used to transform each signal segment to obtain the spectral information. The calculation formula is as follows: in, It is the result of the Discrete Fourier Transform, i.e., the frequency domain signal; It is a discrete signal in the time domain, representing the signal value at time n; f is the frequency exponent; N is the total number of sampling points of the signal, i.e., the length of the signal; is the imaginary unit; n is the time index, with a value range of 0, 1, 2…N-1; Finally, the spectral amplitude was analyzed to identify fault characteristic frequencies.

4. The MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement as described in claim 2, characterized in that: The second channel utilizes the frequency-weighted energy operator (FWEO) to enhance the frequency characteristics and energy distribution of the signal, isolating key fault features, specifically including: The first and second derivatives of the signal are calculated using FWEO, and the fault characteristics are extracted by combining the instantaneous frequency and energy distribution. The calculation formula is as follows: in, A signal or function at time t represents the state of the system at time t; The characteristic quantity representing the signal x(t) is used to extract fault characteristics; t is a time variable.

5. The MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement according to claim 2, characterized in that: The third channel employs signal envelope technology to extract the transient and nonlinear characteristics of the vibration signal, comprehensively capturing subtle changes in the fault, specifically including: First, perform a Hilbert transform on the signal to obtain a complex analytic signal; Then calculate the amplitude of the complex analytic signal to obtain the signal envelope; Finally, the envelope of the signal is analyzed to extract the impact fault characteristics of the rotating machinery.

6. The MobileVit lightweight bearing fault diagnosis method based on Ghost feature enhancement improvement according to claim 1, characterized in that: The feature map is subjected to a series of linear transformations, including weighted summation and nonlinear activation functions.

Images