Deep groove ball bearing fault diagnosis method based on fusion of frequency domain features and improved residual network

Abstract

The present invention discloses a deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network, comprising the following steps: step one, collecting the vibration signal of the vibration acceleration sensor of the deep groove ball bearing installed at the driving end; step two, using the inverse cosine function to perform numerical coding conversion; step three, embedding the multi-scale attention module into the deep residual neural network to obtain the multi-scale attention residual network fault diagnosis model; step four, using the multi-scale attention residual network fault diagnosis model obtained in step three to perform fault classification on the encoded Gram angle field feature map obtained in step two, so as to realize the fault diagnosis of the rolling bearing. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network of the present invention can effectively realize the fault diagnosis of the rolling bearing through the settings of steps one to four.

Description

Technical Field

[0001] The present invention relates to a bearing fault diagnosis method, and more specifically to a deep groove ball bearing fault diagnosis method integrating frequency domain features with an improved residual network. Background Art

[0002] Robust identification of the operating status of deep groove ball bearings is of great significance for ensuring the service performance of industrial equipment. However, in practical applications, the collected vibration signals are usually affected by complex transmission paths and component coupling, resulting in interference from irrelevant noise, which poses a challenge to the accurate diagnosis of deep groove ball bearing faults. Due to long-term high-load operation and complex terrain conditions, bearings are prone to various faults such as wear, cracks, and spalling. If these faults are not diagnosed and handled in a timely manner, they may cause serious equipment damage or even accidents.

[0003] Traditional deep groove ball bearing fault diagnosis methods have the following main problems: First, they rely on the experience and expertise of engineers and cannot make full use of modern technologies such as big data and deep learning; second, manual feature extraction leads to incomplete information extraction, which easily ignores some important fault characteristics and affects the accuracy of diagnosis; in addition, traditional methods are sensitive to environmental conditions, and changes in factors such as temperature and humidity may interfere with the diagnosis results; at the same time, due to human factors and subjective judgment, traditional methods may not be able to detect some hidden or early signs of faults, resulting in missed diagnosis or misjudgment; finally, traditional methods require a lot of manpower and time costs, including data collection, feature extraction, analysis and interpretation, which are costly and inefficient, and are difficult to adapt to the needs of large-scale industrial applications. Summary of the invention

[0004] In view of the deficiencies in the prior art, the object of the present invention is to provide a deep groove ball bearing fault diagnosis method that integrates frequency domain features and an improved residual network.

[0005] To achieve the above object, the present invention provides the following technical solution: a deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network, characterized in that it comprises the following steps:

[0006] Step 1: collect the vibration signal of the vibration acceleration sensor of the deep groove ball bearing installed at the driving end, with a sampling frequency of 12kHz, and process the vibration signal using the Fourier transform algorithm to obtain the non-redundant spectrum of the vibration signal, and normalize the spectrum value to the range of [-1,1] through linear mapping;

[0007] Step 2: Use the arccosine function to perform numerical coding conversion, map the normalized non-redundant spectrum to polar coordinates, obtain the polar coordinate representation of each frequency component, and use the mapped polar coordinate data to establish a Gram matrix to describe the relationship between different signal points. The Gram matrix is ​​constructed by calculating the inner product between the signal points to generate the encoded Gram angular field feature map;

[0008] Step 3: embed the multi-scale attention module into the deep residual neural network to obtain a multi-scale attention residual network fault diagnosis model;

[0009] Step 4: Use the multi-scale attention residual network fault diagnosis model obtained in step 3 to classify the encoded Gram angular field feature map obtained in step 2 to achieve rolling bearing fault diagnosis.

[0010] As a further improvement of the present invention, the non-redundant spectrum of the vibration signal obtained in step one is obtained in the following manner: a threshold is set, the frequency of the frequency domain signal of the vibration signal processed by the Fourier transform algorithm is compared with the threshold, and all frequency components higher than the threshold are removed, thereby removing high-frequency redundant information to obtain a non-redundant spectrum.

[0011] As a further improvement of the present invention, the specific steps of normalizing the spectrum value to the range of [-1, 1] by linear mapping in step 1 are as follows: the non-redundant spectrum F=f is obtained by using the maximum and minimum normalization method. 1 , f 2 , ..., f n Perform linear mapping to normalize the spectrum value to the range of [-1,1], the formula is:

[0012] As a further improvement of the present invention, the polar coordinate representation formula of each frequency component obtained in step 2 is as follows:

[0013]

[0014] Where r is the polar diameter in polar coordinates, is the polar angle in polar coordinates, is the normalized non-redundant spectrum, ti is the spectrum number, and L is the spectrum length.

[0015] As a further improvement of the present invention, the Gram angle field feature map in step 2 includes a feature map encoded by a cosine function and a feature map encoded by a sine function. The feature map encoded by the cosine function is called a Gram angle sum field, and the feature map encoded by the sine function is called a Gram angle difference field. Therefore, for a non-redundant spectrum f of a given signal, its corresponding Gram angle sum field can be defined as:

[0016]

[0017] For a given signal's non-redundant spectrum f, its Gram angle difference field can be expressed as:

[0018]

[0019] Wherein, subscripts i and j represent different spectrum sampling points respectively.

[0020] As a further improvement of the present invention, in step 3, the multi-scale attention module is embedded into the deep residual neural network, and the specific steps of obtaining the multi-scale attention residual network fault diagnosis model are as follows:

[0021] Step 31: introduce a multi-scale feature extraction module, compress the channel dimension of the input Gram angular field feature map through the multi-scale feature extraction module, and divide a feature map X with a channel number C into S branches through a compression operation, and the channel number of each branch is C′=C / S;

[0022] Step 32: splice the S branch outputs divided in step 31 to form a feature map that integrates multi-scale information;

[0023] Step 33: Introduce the channel attention module and use one-dimensional convolution to replace the fully connected layer in the channel attention module;

[0024] Step 34: Based on the multi-scale feature extraction module of step 331 and the channel attention module of step 33, a multi-scale attention module is proposed, and then the multi-scale attention module is integrated into the residual mechanism. The convolution layer with a step size of 4 and a convolution kernel of 4×4 and the layer normalization layer are used as the input layer of the multi-scale attention residual network fault diagnosis model. The four residual mechanisms with multi-scale attention modules are used as the feature extraction layer of the multi-scale attention residual network fault diagnosis model. The global average pooling layer, layer normalization layer and linear layer constitute the output layer of the multi-scale attention residual network fault diagnosis model, completing the construction of the multi-scale attention residual network fault diagnosis model.

[0025] As a further improvement of the present invention, in each branch divided in step 31, the input Gram angular field feature map is processed by convolution layers with different kernel sizes to capture spatial information of different scales. The i-th branch uses a kernel size of k. i The convolution kernel is expressed as follows:

[0026] k i =2×(i+1)+1;

[0027] And apply group convolution in each branch. The operation of group convolution can be expressed by the following formula:

[0028]

[0029] Among them, K represents the convolution kernel size, and G represents the number of groups;

[0030] Then the multi-scale feature map generation function formula is obtained as follows:

[0031]

[0032] in, Indicates the number of groups is G i , the kernel size is k i The grouped convolution operation of , X is the input feature map;

[0033] Based on the multi-scale feature map generated above, the feature fusion operation is performed using the following formula:

[0034] F=Concat([F 0 ,F 1 , ..., F S-1 ])

[0035] Among them, F is the comprehensive feature map after splicing, F i is the output feature map of the i-th branch.

[0036] As a further improvement of the present invention, the specific steps of introducing the channel attention module in step three are as follows: first, global average pooling is performed to extract the global information of each channel to prepare for the subsequent channel attention allocation. The formula is as follows:

[0037]

[0038] Where X∈R H×W×C is the output feature map of the convolutional layer, H and W are the height and width of the feature map, respectively, and C is the number of channels;

[0039] Then, one-dimensional convolution is used to replace the fully connected layer in the traditional attention module to model the relationship between channels. The formula is as follows:

[0040] ω=σ(Conv 1D (g))

[0041] Among them, g is the output of the global average pooling operation, Conv 1D is a one-dimensional convolution operation, σ is the Sigmoid activation function, ω∈R C is the generated channel attention weight vector;

[0042] In the above steps, an adaptive module is introduced to determine the kernel size of the one-dimensional convolution. Based on the nonlinear mapping of the number of channels, the kernel size is dynamically adjusted to adapt to different network structures and numbers of channels. The formula is as follows:

[0043]

[0044] Among them, C is the number of channels, k is the kernel size of the one-dimensional convolution, γ and b are adjustable parameters, and the kernel size of the one-dimensional convolution is adaptively determined according to the number of channels.

[0045] Beneficial effects of the present invention:

[0046] 1. Through the combination of discrete Fourier transform and Gram angle field, the present invention effectively improves the anti-noise performance of the rolling bearing fault diagnosis model. Discrete Fourier transform can convert the rolling bearing vibration signal into the frequency domain, while the Gram angle field reconstructs the frequency domain components into a feature map in the form of an image. This process reduces the sensitivity to noise and effectively improves the anti-noise performance of the rolling bearing fault diagnosis model, making it more suitable for applications in complex industrial environments.

[0047] 2. The present invention combines a multi-scale feature extraction module and a channel attention module, which significantly improves the accuracy of the rolling bearing fault diagnosis model. The multi-scale feature extraction module can capture the characteristic information of the rolling bearing vibration signal from different scales, so that the model can understand the characteristics of the signal more comprehensively, thereby improving the recognition accuracy and generalization ability of the model. The channel attention module can automatically learn and emphasize the feature channels that are critical to the fault diagnosis task, so that the model can automatically identify and focus on important areas of the image, which not only improves the speed of feature extraction, but also enhances the recognition accuracy of signal feature-intensive areas in the feature image. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is the overall flow chart of the present invention;

[0049] FIG2 is a schematic diagram of the Gram angle field encoding process in an embodiment of the present invention, wherein (a) corresponds to a normal rolling bearing, (b) corresponds to a rolling bearing with a rolling element damage fault, (c) corresponds to a rolling bearing with an inner ring damage fault, and (d) corresponds to a rolling bearing with an outer ring damage fault;

[0050] Figure 3 Schematic diagram of a multi-scale feature extraction module in an embodiment of the present invention;

[0051] Figure 4 Schematic diagram of a channel attention module in an embodiment of the present invention;

[0052] Figure 5 Schematic diagram of a multi-scale attention module in an embodiment of the present invention;

[0053] Figure 6 Schematic diagram of a residual mechanism in an embodiment of the present invention;

[0054] FIG7 is a curve of recognition accuracy and a curve of loss function in the model training process according to an embodiment of the present invention, wherein (a) corresponds to a curve of recognition accuracy changing with the number of iterations, and (b) corresponds to a curve of loss function changing with the number of iterations;

[0055] FIG8 is a confusion matrix under different signal-to-noise ratios in an embodiment of the invention, wherein (a) corresponds to a 0Db confusion matrix, (b) corresponds to a -2dB confusion matrix, (c) corresponds to a -4Db confusion matrix, and (d) corresponds to a -6Db confusion matrix. DETAILED DESCRIPTION

[0056] The present invention will be further described below in detail with reference to the embodiments shown in the accompanying drawings.

[0057] Reference Figure 1 As shown in FIG. 8 , a deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network in this embodiment is specifically implemented as follows:

[0058] Step 1: Collect the vibration signal from the vibration accelerometer with a sampling frequency of 12kHz from the deep groove ball bearing installed at the drive end. The outer ring, inner ring and rolling element of the bearing have artificial faults with different diameters (0.18mm, 0.36mm and 0.54mm), as shown in Table 1, including the vibration signal of the healthy bearing, covering a total of ten typical fault categories, each of which contains four different load conditions.

[0059] Table 1 Description of bearing fault types

[0060]

[0061]

[0062] The sample length is set to 1024 data points. By performing overlapping sampling in continuous time windows with a step size of 512, the model can better understand the contextual information in the sequence. Each load condition is divided into 200 samples, and each load condition contains 200×10=2000 samples in total. The entire data set covers a total of 8000 samples, which are divided into training set and test set in a ratio of 7:3.

[0063] After collecting the vibration signal through the above steps, the rolling bearing vibration signal is converted into a frequency domain signal using the fast Fourier transform algorithm. In the frequency domain signal, the high-frequency part may contain noise or irrelevant information, which may not be important for bearing fault diagnosis. Therefore, by setting a threshold, all frequency components above the threshold are removed, thereby removing high-frequency redundant information. Corresponding to the removal of high-frequency information, the low-frequency components are retained because these components usually contain useful information related to the health status of the bearing. The maximum and minimum normalization method is used to obtain the non-redundant spectrum F = f 1 , f 2 , ..., f n Perform linear mapping to normalize the spectrum value to the range of [-1,1], the formula is:

[0064]

[0065] Step 2: Use the arccosine function to perform numerical encoding conversion, map the normalized non-redundant spectrum to polar coordinates, and obtain the polar coordinate representation of each frequency component. After the non-redundant spectrum is mapped to polar coordinates, the polar coordinate representation of each frequency component is obtained. These polar coordinates contain the amplitude and phase information of the signal, so the internal characteristics of the signal can be characterized based on the obtained polar coordinates. The formula is:

[0066]

[0067] Where r is the polar diameter in polar coordinates, is the polar angle in polar coordinates, is the normalized non-redundant spectrum, ti is the spectrum number, and L is the spectrum length.

[0068] The mapped polar coordinate data is then used to build a Gram matrix to describe the relationship between different signal points. The inner product between the signal points is calculated to construct the Gram matrix and generate the encoded Gram angular field feature map. The relationship between different sampling points can be established using sine and cosine functions. Among them, the feature map encoded by the cosine function is called Gramian Angular Summation Fields (GASF), while the feature map encoded by the sine function is called Gramian Angular Difference Fields (GADF).

[0069] Therefore, based on the above reasons, for a given signal's non-redundant spectrum f, its corresponding GASF can be defined as:

[0070]

[0071] Similarly, for a given signal’s non-redundant spectrum f, its GADF can be expressed as:

[0072]

[0073] Wherein, subscripts i and j represent different spectrum sampling points respectively.

[0074] The Gram angle field encoding process of the rolling bearing vibration signal is shown in Figure 2, which includes the time domain waveform, spectrum, GADF and GASF images from left to right.

[0075] Step 3: The target objects in the GAF image (Gram Angle Field) may appear in different sizes and shapes, which requires the model to capture all-round information from local details to global context. By introducing the Multi-scale Feature Extraction Module (MFEM) into the model, the main structure is as follows Figure 3 As shown, the model can better understand the image content, thereby improving the recognition accuracy and generalization ability of the model.

[0076] In this embodiment, the multi-scale feature extraction module is introduced as follows: the channel dimension of the input feature map is compressed to facilitate feature extraction and fusion in subsequent steps. For a feature map X with a channel number of C, it is divided into S branches through compression operation, and the number of channels in each branch is C′=C / S. In each branch, the input feature map is processed by convolution layers with different kernel sizes to capture spatial information of different scales. The i-th branch uses a kernel size of k. i The convolution kernel is expressed as follows:

[0077] k i =2×(i+1)+1

[0078] In order to further optimize the computational efficiency, group convolution is applied in each branch to reduce the number of parameters and reduce the computational complexity. The operation of group convolution can be expressed by the following formula:

[0079]

[0080] Among them, K represents the convolution kernel size and G represents the number of groups. The formula for generating the multi-scale feature map is as follows:

[0081]

[0082] in, Indicates the number of groups is G i , the kernel size is k i The grouped convolution operation is performed, and X is the input feature map.

[0083] The outputs of all branches are concatenated to form a feature map that integrates multi-scale information. The feature fusion operation can be expressed by the following formula:

[0084] F=Concat([F 0 ,F 1 , ..., , F S-1 ])

[0085] Among them, F is the comprehensive feature map after splicing, F i is the output feature map of the i-th branch.

[0086] The encoding method of GAF images causes the signal features to be linearly distributed along a specific direction, which provides a new direction for the adaptive attention of the model. Different from the traditional attention module, the channel attention module (CAM) adopted has the following main structure: Figure 4 As shown in the figure, by selecting one-dimensional convolution and adaptive kernel size, the modeling of inter-channel dependencies is achieved, so that the model can automatically identify and focus on important areas of the image, which not only improves the speed of feature extraction, but also enhances the recognition accuracy of signal feature-dense areas in GAF images.

[0087] In this embodiment, the channel attention module is introduced as follows: first, global average pooling is performed to extract the global information of each channel to prepare for the subsequent channel attention allocation. The formula is as follows:

[0088]

[0089] Where X∈R H×W×C is the output feature map of the convolutional layer, H and W are the height and width of the feature map, respectively, and C is the number of channels. The global average pooling operation compresses the feature map of each channel into a single numerical vector.

[0090] One-dimensional convolution is used to replace the fully connected layer in the traditional attention module. This design allows the model to model the relationship between channels while maintaining spatial information. The formula is as follows:

[0091] ω=σ(Conv 1D (g))

[0092] Among them, g is the output of the global average pooling operation, Conv 1D is a one-dimensional convolution operation, σ is the Sigmoid activation function, ω∈R C is the generated channel attention weight vector.

[0093] The kernel size of the one-dimensional convolution is a key parameter that determines the range of channels considered when generating channel attention weights. An adaptive module is introduced to determine the kernel size of the one-dimensional convolution. Based on the nonlinear mapping of the number of channels, the kernel size is dynamically adjusted to adapt to different network structures and numbers of channels. In this way, more accurate attention weights can be generated for each channel while avoiding the need to manually adjust parameters. The formula is as follows:

[0094]

[0095] Among them, C is the number of channels, k is the kernel size of the one-dimensional convolution, γ and b are adjustable parameters, and the kernel size of the one-dimensional convolution is adaptively determined according to the number of channels.

[0096] Then, based on the above multi-scale feature extraction module and channel attention module, a multi-scale attention module (Multi-Scale Attention Module, Msa module) is proposed. The main structure is as follows Figure 5 As shown in the figure, the diversity of feature representation is enhanced by comprehensively utilizing spatial information at different scales. The input feature map is processed by the MFEM module, which can generate multi-scale feature representation in the channel dimension. This step captures feature information at different levels through convolution kernels of different sizes and merges this information in the channel dimension to obtain a feature map containing rich multi-scale information. The CAM module is used to analyze each branch in the multi-scale feature map and extract the channel attention vector of each branch. The CAM module can effectively identify which channels are more important for the task, thereby providing a basis for subsequent feature recalibration. The original attention vector output by the CAM module is Softmax processed to achieve normalization and calibration of the attention weight. This step ensures that the model can more accurately evaluate the importance of feature channels and provide weight guidance for feature fusion. The above output is multiplied element by element with the multi-scale feature map, which enables the model to strengthen important features while suppressing unimportant features according to the attention weights, forming a feature map containing rich multi-scale feature information, which can provide richer and more discriminative feature representations for subsequent fault diagnosis tasks.

[0097] As the core component of ResNet, the residual mechanism has proven its effectiveness in building deep networks. However, in order to further improve the performance of the model, especially in the rolling bearing fault diagnosis scenario, the residual mechanism needs to be improved. The Msa module is integrated into the residual mechanism to enhance the model's ability to extract GAF image features and diagnostic accuracy. Figure 5As shown in the figure. First, the order of the convolutional layers is adjusted. The Msa module is used followed by two 1x1 convolutional modules. This helps the model to perform more refined feature selection and channel number adjustment through 1x1 convolution after the initial feature extraction. Secondly, the GELU activation function is introduced to replace the traditional ReLU, which can provide a smoother nonlinear transformation for the model and help improve the learning ability and generalization performance of the model. Finally, a 2*2 convolution with a step size of 2 is used to achieve downsampling, which can reduce information loss. At the same time, a layer normalization layer is added before and after downsampling to help maintain the stability of the model during training.

[0098] In the multi-scale attention residual network model, the input layer contains a convolution layer with a stride of 4 and a convolution kernel of 4×4 and a layer normalization layer. The feature extraction layer contains four residual mechanisms with multi-scale attention modules, with a structure of (3, 3, 9, 3). The output layer contains a global average pooling layer, a layer normalization layer, and a linear layer. Through the above steps, not only the model's ability to capture multi-scale features is enhanced, but also the feature expression ability and the model's diagnostic accuracy are improved through the effective channel attention module. It has shown excellent performance in complex tasks such as rolling bearing fault diagnosis, providing a new perspective and method for the application of deep learning models in the industrial field.

[0099] Step 4: Use bearing vibration signals under four different load conditions, convert them into GADF images according to the signal data set preprocessing method and input them into the model. During the training process, as the number of iterations increases, the model gradually learns and optimizes itself until it reaches a convergence state. Figure 7a and Figure 7b The fault recognition accuracy and loss function curves under four load conditions are shown in the figure. After about 5-10 iterations, the fault recognition accuracy stabilizes at a level close to 100%, and the loss function curve continues to decline, indicating that the model is not overfitting, verifying the feasibility of the model.

[0100] In order to verify that the model has better diagnostic efficiency in the field of bearing fault diagnosis, it is compared with one-dimensional convolutional neural network (1D-CNN) and long short-term memory network (LSTM). The experimental results are shown in Table 2.

[0101] Table 2 Comparison of accuracy of different models

[0102]

[0103]

[0104] Compared with one-dimensional neural networks, the proposed model has better diagnostic efficiency, and achieves higher accuracy and more stable performance in fault diagnosis tasks, which further proves its advantages and effectiveness over one-dimensional neural networks.

[0105] The model proposed in this paper is analyzed for its noise resistance. In actual engineering applications, due to the presence of factors such as environmental noise, the timing signal is often interfered to a certain extent. Therefore, it is of great significance to evaluate the performance of the model in a noisy environment. Under four different load conditions, noise with a certain level of signal-to-noise ratio is artificially added, and the diagnostic accuracy of the model under different noise levels is recorded. The experimental results show that the model proposed in this paper exhibits strong anti-interference ability when facing different load conditions and noise levels. Figure 8a to Figure 8d It is the confusion matrix generated by the experimental results under different signal-to-noise ratios. Although the accuracy of the model may decrease in a high-noise environment, it can still maintain a high diagnostic accuracy compared with other models. Therefore, the proposed model shows good performance in the model noise resistance analysis, has strong anti-interference ability and stability, and provides reliable support for practical engineering applications.

[0106] The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network, characterized in that: The steps include: Step 1: collect the vibration signal of the vibration acceleration sensor of the deep groove ball bearing installed at the driving end, with a sampling frequency of 12kHz, and process the vibration signal using the Fourier transform algorithm to obtain the non-redundant spectrum of the vibration signal, and normalize the spectrum value to the range of [-1,1] through linear mapping; Step 2: Use the arccosine function to perform numerical coding conversion, map the normalized non-redundant spectrum to polar coordinates, obtain the polar coordinate representation of each frequency component, and use the mapped polar coordinate data to establish a Gram matrix to describe the relationship between different signal points. The Gram matrix is ​​constructed by calculating the inner product between the signal points to generate the encoded Gram angular field feature map; Step 3: embed the multi-scale attention module into the deep residual neural network to obtain a multi-scale attention residual network fault diagnosis model; Step 4: Use the multi-scale attention residual network fault diagnosis model obtained in step 3 to classify the encoded Gram angular field feature map obtained in step 2 to achieve rolling bearing fault diagnosis.

2. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 1 is characterized in that: The non-redundant spectrum of the vibration signal obtained in step 1 is obtained in the following manner: a threshold is set, the frequency of the frequency domain signal of the vibration signal processed by the Fourier transform algorithm is compared with the threshold, and all frequency components higher than the threshold are removed, thereby removing high-frequency redundant information to obtain a non-redundant spectrum.

3. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 2 is characterized in that: The specific steps of normalizing the spectrum value to the range of [-1, 1] by linear mapping in step 1 are as follows: using the maximum and minimum normalization method to obtain the non-redundant spectrum F = f1, f2, ..., f n Perform linear mapping to normalize the spectrum value to the range of [-1,1], the formula is:

4. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 1, 2 or 3, characterized in that: The polar coordinate representation formula of each frequency component obtained in step 2 is as follows: Where r is the polar diameter in polar coordinates, φ is the polar angle in polar coordinates, is the normalized non-redundant spectrum, ti is the spectrum number, and L is the spectrum length.

5. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 4 is characterized in that: The Gram angle field feature map in step 2 includes a feature map encoded by a cosine function and a feature map encoded by a sine function. The feature map encoded by a cosine function is called a Gram angle sum field, and the feature map encoded by a sine function is called a Gram angle difference field. Therefore, for a non-redundant spectrum f of a given signal, its corresponding Gram angle sum field can be defined as: For a given signal's non-redundant spectrum f, its Gram angle difference field can be expressed as: GADF=sin(φ i -f j ) Wherein, subscripts i and j represent different spectrum sampling points respectively.

6. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 1, 2 or 3, characterized in that: In step 3, the multi-scale attention module is embedded into the deep residual neural network, and the specific steps of obtaining the multi-scale attention residual network fault diagnosis model are as follows: Step 31: introduce a multi-scale feature extraction module, compress the channel dimension of the input Gram angular field feature map through the multi-scale feature extraction module, and divide a feature map X with a channel number C into S branches through a compression operation, and the channel number of each branch is C'=C / S; Step 32: splice the S branch outputs divided in step 31 to form a feature map that integrates multi-scale information; Step 33: Introduce the channel attention module and use one-dimensional convolution to replace the fully connected layer in the channel attention module; Step 34: Based on the multi-scale feature extraction module of step 331 and the channel attention module of step 33, a multi-scale attention module is proposed, and then the multi-scale attention module is integrated into the residual mechanism. The convolution layer with a step size of 4 and a convolution kernel of 4×4 and the layer normalization layer are used as the input layer of the multi-scale attention residual network fault diagnosis model. The four residual mechanisms with multi-scale attention modules are used as the feature extraction layer of the multi-scale attention residual network fault diagnosis model. The global average pooling layer, layer normalization layer and linear layer constitute the output layer of the multi-scale attention residual network fault diagnosis model, completing the construction of the multi-scale attention residual network fault diagnosis model.

7. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 6 is characterized in that: In each branch divided in step 3, the input Gram angular field feature map is processed by convolutional layers with different kernel sizes to capture spatial information of different scales. The i-th branch uses a kernel size of k. i The convolution kernel is expressed as follows: k i =2×i+1+1; And apply group convolution in each branch. The operation of group convolution can be expressed by the following formula: Among them, K represents the convolution kernel size, and G represents the number of groups; Then the multi-scale feature map generation function formula is obtained as follows: in, Indicates the number of groups is G i , the kernel size is k i The grouped convolution operation of , X is the input feature map; Based on the multi-scale feature map generated above, the feature fusion operation is performed using the following formula: F=ConcatF0,F1,…,F S-1 Among them, F is the comprehensive feature map after splicing, F i is the output feature map of the i-th branch.

8. The deep groove ball bearing fault diagnosis method integrating frequency domain features and improved residual network according to claim 7 is characterized in that: The specific steps of introducing the channel attention module in step 33 are as follows: First, global average pooling is performed to extract the global information of each channel to prepare for the subsequent channel attention allocation. The formula is as follows: Where X∈R H×W×C is the output feature map of the convolutional layer, H and W are the height and width of the feature map, respectively, and C is the number of channels; Then, one-dimensional convolution is used to replace the fully connected layer in the traditional attention module to model the relationship between channels. The formula is as follows: ω=σ(Conv 1D (g)) Among them, g is the output of the global average pooling operation, Conv 1D is a one-dimensional convolution operation, σ is the Sigmoid activation function, ω∈R C is the generated channel attention weight vector; In the above steps, an adaptive module is introduced to determine the kernel size of the one-dimensional convolution. Based on the nonlinear mapping of the number of channels, the kernel size is dynamically adjusted to adapt to different network structures and numbers of channels. The formula is as follows: Among them, C is the number of channels, k is the kernel size of the one-dimensional convolution, γ and b are adjustable parameters, and the kernel size of the one-dimensional convolution is adaptively determined according to the number of channels.

Images