A multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis

Abstract

The present invention proposes a multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis, comprising the following steps: first, by designing a multi-scale feature extraction module, one-dimensional vibration signal features are extracted using convolution kernels of different sizes; second, time-frequency images are generated using short-time Fourier transform and pseudo-color processing techniques; finally, the time-frequency images are segmented and rearranged, flattened into one-dimensional vectors and spliced ​​together as input sequences, and the advantages of the self-attention mechanism and encoder are utilized to input them into the Vision-Transformer model for rolling bearing fault diagnosis. The present invention can not only simultaneously solve the problem of low diagnostic accuracy inherent in mainstream deep learning-based fault diagnosis models, but also can be applied to the fault diagnosis of other rotating machinery to improve their diagnostic accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault diagnosis methods, and in particular relates to a multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis. Background Art

[0002] Rolling bearings are crucial components in modern industry, directly impacting production efficiency and equipment life. They are widely used in aviation, aerospace, shipbuilding, and automotive industries. However, due to multiple factors such as vibration, friction, and wear, rolling bearings are susceptible to damage, and the risk of failure is increasing. Therefore, timely and accurate fault diagnosis is crucial to maintaining production efficiency, extending equipment life, and reducing maintenance costs.

[0003] Bearing fault diagnosis has always been a complex and challenging task. First, traditional fault diagnosis techniques often require manual feature design and specialized expertise, resulting in significant limitations. Second, RNNs and their variants, such as LSTMs, are not suitable for parallel computing, resulting in very low training efficiency on large datasets and unsuitable for real-world operating conditions. RNNs and their variants also suffer from the inability to establish effective connections between long-distance sequences. CNN models also have some shortcomings, such as a failure to capture relationships between objects, treating all pixels equally and lacking specificity. The local receptive field of convolutional kernels necessitates stacking many convolutional layers to obtain global information. To address this, Vaswani et al. proposed a new network architecture, the Transformer. This architecture abandons the traditional encoder-decoder model's inherent requirement for combining CNNs or RNNs, relying solely on attention. They also proposed two new attention mechanisms, namely Scaled Dot-Product Attention and Multi-Head Attention. However, these models process single-scale data. Using multi-scale data fusion to improve fault diagnosis accuracy is a primary challenge in intelligent bearing fault diagnosis. Summary of the Invention

[0004] To address the challenges of the prior art, this paper proposes a multi-scale feature fusion Vision-Transformer model for rolling bearing fault diagnosis. This method uses vibration signals as input and uses multiple convolution kernels of varying sizes to learn features at different scales in parallel. This model then fuses these features using convolutional channels. This method also leverages the advantages of the Vision-Transformer's self-attention mechanism and encoder to enhance the connections between key features and improve the model's classification performance.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis includes the following steps:

[0007] Step 1: Perform multi-scale feature fusion on the one-dimensional vibration signal;

[0008] Step 2: Generate a three-channel time-frequency color image using short-time Fourier transform and pseudo-color processing technology;

[0009] Step 3: Build the Vision-Transformer model;

[0010] Step 4: Perform position encoding on the time-frequency image to calculate the relative position, then segment and rearrange the time-frequency image and input it into the Vision-Transformer for diagnosis, and output the diagnosis result.

[0011] Furthermore, the formula for multi-scale feature fusion used in step 1 is:

[0012]

[0013] Among them, X′ represents the processed feature matrix, X represents the original input data, is a convolutional layer with a convolution kernel size of n1×1 and the number of convolution kernels C, and σ(·) is the ReLU activation function.

[0014] Furthermore, the short-time Fourier transform formula used in step 2 is:

[0015]

[0016] Among them, STFT x (ω,τ) is the result of short-time Fourier transform, ω,τ are frequency variables and time variables respectively, ω(t-τ) represents the window function, which is usually a localized function. The user and the intercepted signal x(t) are a small one-dimensional vibration signal, e -jωt is a complex exponential function, which is the core part of Fourier transform and is used to convert signals from time domain to frequency domain. Where: j, t represent imaginary units and time respectively;

[0017] Specific formula based on pseudo color processing technology:

[0018]

[0019] Among them, x h is the pixel value of the grayscale image, R, G, B are the pixel values ​​of the red, green, and blue channels of the generated color image, respectively, and fR ,f G ,f B It is a function that maps grayscale values ​​to corresponding channels.

[0020] Furthermore, the Vision-Transformer model in step 3 includes a multi-head attention module, a position encoding module, layer normalization, and a residual connection; the first module includes a self-attention layer and a normalization layer, and the other module includes two convolutional layers and a feedforward layer; the feedforward layer module formula is as follows:

[0021]

[0022] Among them, LayerNorm represents the layer normalization operation, which standardizes the features after residual connection. Input is the input feature. SA (Input) represents the application of self-attention mechanism to the input feature to obtain the self-attention feature. Output SA Output represents the output after self-attention mechanism (Self-Attention) and layer normalization (Layer Normalization); FF Represents the output after the feedforward network and layer normalization, Conv1(Output SA ) indicates Output SA The first convolution operation applied, ReLU(·) means applying the ReLU activation function to the result of the first convolution operation, Conv2(·) means applying the second convolution operation to the features after the ReLU activation function, Output SA +Conv2(ReLU(Conv1(Output SA ))) represents the residual connection, which adds the features after self-attention to the output features of the feedforward network.

[0023] Furthermore, the position encoding formula used in step 4 is:

[0024]

[0025] Among them, PE represents position encoding, pos represents the position of the element in the vector; 2i, 2i+1 represent the dimensions for encoding even positions and odd positions respectively, and d model Represents the dimension of the model, the dimension encoding the even positions, the dimension encoding the odd positions, sin, cos are trigonometric functions;

[0026] Specific split-based rearrangement formula:

[0027] patches=reshape(X,N,P×P×C p )

[0028] Among them, X, N, P, C p are the input image, the number of tiles, the side length of each tile, and the number of channels of the image.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] (1) The method of extracting features using multiple convolution kernels in the present invention can capture characteristic information of signals from different scales and angles. Multi-source information, such as vibration signals and current signals, may have different frequencies, amplitudes, and phases. In contrast, using only vibration signals or current signals may be limited to signals within certain frequency ranges.

[0031] (2) The Vision-Transformer in this invention can simultaneously consider the information of the entire image at each layer, rather than focusing only on local features. This consideration of global information enables the network to understand the image more comprehensively and helps better capture the semantic and contextual information in the image. Secondly, because ViT uses a self-attention mechanism, the network output can be interpreted as the importance of each position to the classification result. This makes the network's prediction results more interpretable and helps to understand the network's decision-making process.

[0032] (3) The present invention supports intelligent diagnosis of rotating machinery bearing faults, comprehensively monitors equipment status, and accurately diagnoses problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is the overall architecture diagram of the rolling bearing intelligent fault diagnosis of the present invention.

[0034] Figure 2 This is the multi-scale feature fusion layer structure of the present invention.

[0035] Figure 3 This is the Vision-Transformer model method constructed by the present invention.

[0036] Figure 4 This is a comparative experiment of the algorithm of the present invention.

[0037] Figure 5 This is the t-SNE visualization result display of the present invention.

[0038] Figure 6 This figure compares the effects of setting the number of encoders in the network model of the present invention. DETAILED DESCRIPTION

[0039] The present invention will be further described below in conjunction with the embodiments.

[0040] Example 1

[0041] A multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis includes the following steps:

[0042] Step 1: Perform multi-scale feature fusion on the one-dimensional vibration signal;

[0043] The formula for multi-scale feature fusion used in step 1 is:

[0044]

[0045] Among them, X′ represents the processed feature matrix, X represents the original input data, is a convolutional layer with a convolution kernel size of n1×1 and the number of convolution kernels C. Similarly, σ(·) is the ReLU activation function;

[0046] Step 2: Generate a three-channel time-frequency color image using short-time Fourier transform and pseudo-color processing technology;

[0047] The short-time Fourier transform formula used in step 2 is:

[0048]

[0049] Among them, STFT x (ω,τ) is the result of short-time Fourier transform, ω,τ are frequency variables and time variables respectively, ω(t-τ) represents the window function, which is usually a localized function. The user and the intercepted signal x(t) are a small one-dimensional vibration signal, e -jωt It is a complex exponential function, which is the core part of Fourier transform and is used to convert the signal from time domain to frequency domain. In the formula, j and t represent the imaginary unit and time respectively. The specific formula based on pseudo color processing technology is:

[0050]

[0051] Among them, x h is the pixel value of the grayscale image, R, G, B are the pixel values ​​of the red, green, and blue channels of the generated color image, respectively, and f R ,f G ,f B It is a function that maps grayscale values ​​to corresponding channels;

[0052] Step 3: Build the Vision-Transformer model;

[0053] The Vision-Transformer model in step 3 includes a multi-head attention module, a position encoding module, layer normalization, and a residual connection; the first module includes a self-attention layer and a normalization layer, and the other module includes two convolutional layers and a feedforward layer; the feedforward layer module formula is as follows:

[0054]

[0055] Among them, LayerNorm represents the layer normalization operation, which standardizes the features after residual connection. Input is the input feature. SA (Input) represents the application of self-attention mechanism to the input feature to obtain the self-attention feature. Output SA Output represents the output after self-attention mechanism (Self-Attention) and layer normalization (Layer Normalization); FF Represents the output after the feedforward network and layer normalization, Conv1(Output SA ) indicates Output SA The first convolution operation applied, ReLU(·) means applying the ReLU activation function to the result of the first convolution operation, Conv2(·) means applying the second convolution operation to the features after the ReLU activation function, Output SA +Conv2(ReLU(Conv1(Output SA ))) represents the residual connection, which adds the features after self-attention to the output features of the feedforward network;

[0056] Step 4: Position encode the time-frequency image to calculate the relative position, then segment and rearrange the time-frequency image and input it into Vision-Transformer for diagnosis, and output the diagnosis result;

[0057] The position encoding formula used in step 4 is:

[0058]

[0059] Among them, PE represents position encoding, pos represents the position of the element in the vector; 2i, 2i+1 represent the dimensions for encoding even positions and odd positions respectively, and d model Represents the dimension of the model, the dimension encoding the even positions, the dimension encoding the odd positions, sin, cos are trigonometric functions;

[0060] And the specific split-based rearrangement formula:

[0061] patches=reshape(X,N,P×P×C p )

[0062] Among them, X, N, P, C p are the input image, the number of tiles, the side length of each tile, and the number of channels of the image.

[0063] Example 2

[0064] like Figure 1 As shown in Figure 1, a multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis includes a multi-scale feature fusion module, an image segmentation and reordering module, and a fault classification module. The specific implementation steps are as follows:

[0065] Step 1: Follow the conventional training process based on the deep learning target detection model and process the data set first. The present invention is not targeted at a specific data set, so it can be applied to public rolling bearing data sets represented by CWRU, PU, ​​etc. In particular, when using a private data set collected by oneself, the data set is usually divided into a training set, a validation set, and a test set in a ratio of 7:2:1. Specifically, the vibration signal of the bearing under running state can be collected by an acceleration sensor to produce a data set, and the collected samples can be divided into a training set, a validation set, and a test set;

[0066] Step 1-1: Taking a private dataset collected by yourself as an example, for the training set and validation set divided from the dataset, first read the specified file into memory through the Python language interface and process it into the commonly used Numpy data type for subsequent processing;

[0067] Step 1-2: Then, convert the Numpy data type into a Pytorch-specific tensor data type and process it through the dataloader interface to output only one training tensor of the specified batch size each time.

[0068] Step 2: Specific implementation of the multi-scale feature fusion layer based on deep learning. The processed vibration signal is subjected to feature fusion using three convolution kernels of different sizes, and then short-time Fourier transform and pseudo-color processing techniques are used to generate time-frequency image samples. Figure 2 The specific process is described as follows:

[0069] Step 2-1: Implementation of the multi-scale feature fusion layer. Three convolution kernels of different sizes were selected, namely 101, 201 and 301, covering the feature ranges of different scales. Then, the convolution layer was defined in the Conv class, in which the convolution kernel sizes were 101, 201 and 301, respectively, to extract features of specific scales from the input signal. Then, after extracting features through each convolution kernel, three feature vectors of different scales were obtained, which were fused together through splicing operations to form the final output that integrates multi-scale features. Subsequently, the entire multi-scale feature fusion layer was used as a module to participate in the training process of the entire fault diagnosis model, where the number of feature channels was set to 4, corresponding to the number of feature channels generated by each convolution kernel. The formula is as follows:

[0070]

[0071] Among them, X′ represents the processed feature matrix, X represents the original input data, is a convolution layer with a convolution kernel size of n1×1 and the number of convolution kernels C. Similarly, σ(·) is the ReLU activation function.

[0072] Step 2-2: Time-frequency color image sample generation. This step mainly performs data preprocessing on the vibration signal after multi-scale feature fusion to generate a data format that can be input into the Vision-Transformer model for fault diagnosis. First, a short-time Fourier transform (STFT) is used to generate a time-frequency diagram. STFT is based on the discrete Fourier transform (DFT), which divides the signal into multiple time periods and processes each time period using a window function, followed by an STFT. This method allows the observation of the frequency changes of the signal in different time periods, and the superposition of the results of each section can obtain a frequency domain representation of the entire signal. Converting the signal into a time-frequency diagram aims to fully capture the changes in the signal in time and frequency in a two-dimensional representation, which can more accurately understand the changes in the signal's spectral structure over time. Compared with one-dimensional signals, time-frequency images can more intuitively show the time-frequency characteristics of the signal. In addition, STFT is chosen as the signal conversion method because it can provide balanced resolution in the time and frequency domains, has good local feature extraction capabilities, and is easy to understand and calculate, which is conducive to subsequent feature extraction work. STFT is defined as:

[0073] STFT x (ω,τ)=∫ R x(t)ω(t-τ)e -jωt dt

[0074] Among them, STFTx (ω,τ) is the result of short-time Fourier transform, ω,τ are frequency variables and time variables respectively, ω(t-τ) represents the window function, which is usually a localized function. The user and the intercepted signal x(t) are a small one-dimensional vibration signal, e -jωt is a complex exponential function, the core of the Fourier transform, used to convert signals from the time domain to the frequency domain; where j and t represent imaginary units and time, respectively. In actual use, the signal is generally a discrete sampled signal, and the formula is as follows:

[0075]

[0076] Among them, Y[m,k] is the transformation result in the frequency domain, which represents the complex amplitude of the frequency component with frequency k at the time of m sampling times, x[n] is the input discrete sequence, which represents the signal in the time domain, and ω[nk] is the complex exponential term in the frequency domain, which is used to represent the frequency k. is the rotation factor used to rotate the signal at discrete frequencies for frequency domain transformation, L is the length of the input signal, representing the number of samples of the input x[n], M is the number of frequency components, m is the frequency index, and k is the window position index. The window function selects the Hanning window, and its expression is as follows:

[0077]

[0078] Among them, ω Hann (n t -m c ) is the Hanning window function, which means that the window function is relative to the center position m c The offset, n t ,m c Represents the time index and the center index of the window function respectively, π is pi, cos is the cosine function, is the definition of the Hanning window, indicating that at position n t -m c The Hamming window value at n t ,m cfunction. Since the image generated by STFT is a grayscale image, it cannot be directly input into the deep learning model. In order to improve the model's ability to distinguish image details, the pseudo-color processing method in OpenCV is used. This method converts the original single-channel grayscale image into a three-channel color time-frequency image, increases the number of image channels, and thus improves the information richness. In image processing, color images usually contain more information than grayscale images because they have more color details. By mapping the grayscale image to the color space, the visual expression and recognition accuracy of the image are improved. This chapter chooses to evenly map the grayscale values ​​to the three channels of red (Red), green (Green), and blue (Blue) to ensure that the colors corresponding to different grayscale levels have good distinction. This mapping retains the brightness information of the original image and introduces color information, making the image more vivid and diverse. In summary, converting grayscale images into color time-frequency images is an effective image enhancement method that can improve the deep learning model's understanding and learning ability of image data, and can also improve the visual quality and visualization effect of the image. The specific calculation formula is as follows:

[0079] (R,G,B)=ColorMap(C gray )

[0080] Among them, C gray Represents the pixel value in a grayscale image. ColorMap is a function that returns the corresponding color value (R, G, B) based on the grayscale. A grayscale image is an image that only contains brightness information but no color information. The grayscale value of each pixel represents the brightness level of the pixel, usually in the range [0, 255].

[0081] Step 3: Implementation of the multi-scale fusion Vision-Transformer model method. Figure 1 The detailed implementation steps are as follows:

[0082] Step 3-1: Implement the Embeddings module. First, this module is responsible for encoding the input image data into sequence data. It includes image segmentation, position encoding, and adding classifier tokens.

[0083] Step 3-1-1: Segment and rearrange the image as follows

[0084] x p =Conv2d(x)

[0085] Among them, x p It is the feature representation of the input image data after being processed by the two-dimensional convolution layer.

[0086] Step 3-1-2: Position encode the divided image blocks as follows:

[0087] embeddings=x p +position_embeddings

[0088] Among them, position_embeddings and embeddings are the position encoding information of the image block and the embedded representation of the image block respectively.

[0089] Step 3-1-3: Classifier token, add a special token to represent the information of the entire image.

[0090] Step 3-2: Implementation of the multi-head self-attention module. The self-attention mechanism first processes the input sequence matrix and calculates the Q, K, and V matrices as follows:

[0091]

[0092] In the above formula, attention represents the calculation of self-attention, softmax is the activation function, Q, K, and V represent query, key, and value respectively. These three are input matrices, which represent different forms of input expression. Usually, they are obtained by linear transformation of the input signal. K T represents the transpose of the matrix K, d is the dimension of query and key, By taking the square root of the dimension, the attention score can be scaled to control the size of the score. The dot product of Q and K is to find the cosine similarity between the two, that is, to calculate the similarity of each Q and K, and then pass it through the softmax function.

[0093] Step 3-2-1: The multi-head attention mechanism refers to the process of using multiple self-attention mechanisms to calculate attention separately and finally merge them together. The multi-head attention mechanism allows the model to focus on different positions, providing multiple self-attentions to represent multiple subspaces, thereby focusing on different positions of the input sequence at different times and under different circumstances, so that attention can provide multiple possibilities. The multi-head attention mechanism is to execute the above single attention mechanism H z times, such as Figure 3 As shown, the calculation steps are as follows:

[0094]

[0095]

[0096] In the formula, head H represents the output of the Hth attention head, each attention head performs the attention mechanism calculation independently, It means connecting the outputs of all attention heads to form a large matrix, W ORepresents the linear transformation matrix after the multi-head attention mechanism, which is used to linearly transform the connected results. There are four learnable parameter matrices, H z is the number of self-attention mechanism heads. Represents the linear transformation of query matrix Q, key matrix K, and value matrix V respectively

[0097] Step 3-2-2: The role of layer normalization is to map the passing vector value to between 0 and 1, speeding up the convergence of the model. Given an input x, it is calculated as follows:

[0098]

[0099] Where: LayerNorm(x) is the output after the normalization layer; γ and β are adjustment parameters, μ x ,σ x are the mean and standard deviation of the input x, respectively. The x in layer normalization refers to the patch feature vector after embedding and position encoding.

[0100] Step 3-3: Implementation of the position encoder module. The self-attention mechanism uses parallel operation, which will result in the input being out of order. Therefore, the present invention proposes to use sine and cosine functions to implement position encoding, the formula is as follows:

[0101]

[0102] Among them, PE represents position encoding, pos represents the position of the element in the vector; 2i, 2i+1 represent the dimensions for encoding even positions and odd positions respectively, d model The dimension of the representation model encodes the dimensions for even positions and the dimensions for odd positions, with sin and cos being trigonometric functions. The advantages of this positional encoding are that each position has a unique positional encoding, it can accommodate longer input sequences, and it makes computation easier.

[0103] Step 4: Model testing. Input the validation set into the fault diagnosis network model trained in Step 3 to evaluate the model's performance and fine-tune the model's parameters based on the experimental results.

[0104] Step 5: Finally, use the test set to perform performance testing on the trained model, calculate the evaluation indicators, and output the results.

[0105] The method of the present invention can collect the vibration signals of the bearing under the running state through the acceleration sensor to produce a data set, and divide the collected samples into a training set, a verification set and a test set; build a multi-scale feature fusion Vision-Transformer model method, first subject the vibration signals in the training set to multi-scale feature fusion, and then convert them into time-frequency color images through short-time Fourier transform and pseudo-color processing technology; the time-frequency image is segmented and rearranged, and through position encoding technology, it is input into the encoder part of the network model to output classification.

[0106] The present invention proposes a multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis, comprising the following steps: first, by designing a multi-scale feature extraction module, one-dimensional vibration signal features are extracted using convolution kernels of different sizes; second, time-frequency images are generated using short-time Fourier transform and pseudo-color processing technology; finally, the time-frequency images are segmented and rearranged, flattened into one-dimensional vectors and spliced ​​together as input sequences, and the advantages of the self-attention mechanism and encoder are utilized to input them into the Vision-Transformer model for rolling bearing fault diagnosis; the present invention can not only solve the problem of low diagnostic accuracy inherent in mainstream deep learning-based fault diagnosis models, but also can be applied to fault diagnosis of other rotating machinery to improve their diagnostic accuracy.

[0107] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis, characterized by: The following steps are involved: Step 1: Perform multi-scale feature fusion on the one-dimensional vibration signal; Step 2: Generate a three-channel time-frequency color image using short-time Fourier transform and pseudo-color processing technology; Step 3: Build the Vision-Transformer model; Step 4: Position encode the time-frequency image to calculate the relative position, then segment and rearrange the time-frequency image and input it into Vision-Transformer for diagnosis, and output the diagnosis result; The Vision-Transformer model in step 3 includes a multi-head attention module, a position encoding module, layer normalization, and a residual connection; the first module includes a self-attention layer and a normalization layer, and the other module includes two convolutional layers and a feedforward layer; the feedforward layer module formula is as follows: Among them, LayerNorm represents the layer normalization operation, which standardizes the features after residual connection. Input is the input feature. SA (Input) represents the application of self-attention mechanism to the input feature to obtain the self-attention feature. Output SA Output represents the output after the self-attention mechanism and layer normalization; FF Represents the output after the feedforward network and layer normalization, Conv1(Output SA ) indicates Output SA The first convolution operation applied, ReLU(·) means applying the ReLU activation function to the result of the first convolution operation, Conv2(·) means applying the second convolution operation to the features after the ReLU activation function, Output SA +Conv2(ReLU(Conv1(Output SA ))) represents the residual connection, which adds the features after self-attention to the output features of the feedforward network.

2. The multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis according to claim 1 is characterized in that: The formula for multi-scale feature fusion used in step 1 is: Among them, X′ represents the processed feature matrix, X represents the original input data, is a convolutional layer with a convolution kernel size of n1×1 and the number of convolution kernels C, and σ(·) is the ReLU activation function.

3. The multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis according to claim 1 is characterized in that: The short-time Fourier transform formula used in step 2 is: Among them, STFT x (ω,τ) is the result of short-time Fourier transform, ω,τ are frequency variables and time variables respectively, ω(t-τ) represents the window function, which is usually a localized function. The user and the intercepted signal x(t) are a small one-dimensional vibration signal, e -jωt is a complex exponential function, which is the core part of Fourier transform and is used to convert signals from time domain to frequency domain. Where: j, t represent imaginary units and time respectively; Specific formula based on pseudo color processing technology: Among them, x h is the pixel value of the grayscale image, R, G, B are the pixel values ​​of the red, green, and blue channels of the generated color image, respectively, and f R ,f G ,f B It is a function that maps grayscale values ​​to corresponding channels.

4. The multi-scale feature fusion Vision-Transformer model method for rolling bearing fault diagnosis according to claim 1 is characterized in that: The position encoding formula used in step 4 is: Among them, PE represents position encoding, pos represents the position of the element in the vector; 2i, 2i+1 represent the dimensions for encoding even positions and odd positions respectively, and d model Represents the dimension of the model, the dimension encoding the even positions, the dimension encoding the odd positions, sin, cos are trigonometric functions; Specific split-based rearrangement formula: patches=reshape(X,N,P×P×C p ) Among them, X, N, P, C p are the input image, the number of tiles, the side length of each tile, and the number of channels of the image.

Images