A rapid diagnosis method for rolling bearing fault based on RegNet-SES

Abstract

The application discloses a kind of based on RegNet-SES's quick diagnosis rolling bearing fault method, its specific steps include: collecting healthy, different fault types and the one-dimensional vibration acceleration signal data of the degree of fault of rolling bearing;Using continuous wavelet transform converts one-dimensional vibration signal into the two-dimensional time-frequency chart of obvious feature information, and is made into rolling bearing mixed fault data set;Build RegNet-SES network model and complete network parameter initialization, using the weight parameter of pre-training model is transferred to RegNet-SES network using depth migration learning;Rolling bearing mixed fault data set is used for RegNet-SES network training, and the mixed fault diagnosis model for rolling bearing is obtained.The application overcomes the problem that the existing model is not good for rolling bearing mixed fault diagnosis effect and the high model complexity leads to large calculation cost, improves the model diagnosis precision and improves the speed of model diagnosis fault.

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology, specifically relating to a method for rapid diagnosis of rolling bearing faults based on RegNet-SES. Background Technology

[0002] As key components of rotating machinery, rolling bearings often operate in harsh environments characterized by wear, high temperatures, and heavy loads, making them highly susceptible to damage. A failure in a rolling bearing can damage the entire mechanical system, causing economic losses and even serious safety accidents. Therefore, real-time monitoring and fault diagnosis of rolling bearing operation are crucial to ensuring the normal operation of rotating machinery.

[0003] The introduction of convolutional neural networks (CNNs) has opened up a new direction for rolling bearing fault diagnosis. Unlike traditional intelligent fault diagnosis methods, deep CNN-based fault diagnosis methods emphasize model structure, enabling the model to learn deeper features representing samples and integrating feature extraction and fault classification. This avoids the loss of original sample feature information and makes fault type clustering more accurate. Some researchers have used the original bearing vibration signal as the initial input to a one-dimensional CNN, supplementing it with spectral analysis data incrementally, effectively achieving fault diagnosis of rolling bearings under strong noise conditions. Others have optimized a two-dimensional CNN, converting the original signal into a two-dimensional grayscale image, and the neural network extracts image features to achieve rolling bearing fault diagnosis. Some researchers have used SDP technology to fuse fault features collected by multiple sensors and combined the resulting SDP feature map with a VGG16 network, achieving higher diagnostic accuracy than a single sensor. Still others have applied an improved residual network to achieve fault diagnosis of rolling bearings under complex working conditions, improving the noise resistance and generalization ability of the fault diagnosis model. The diagnostic model proposed by the above method has a certain degree of robustness and high accuracy. However, to achieve the above results, it is necessary to train the model in multiple layers using a large sample fault dataset, which results in a huge time, labor and computing cost.

[0004] It's worth noting that the large datasets used for training are only one factor contributing to the significant time and computational cost of model training; another factor is the model design strategy. Traditional convolutional neural network models such as VGG, GoogLeNet, and ResNet use standard convolutional operations. While these models have strong feature representation capabilities, their high parameter and computational requirements lead to problems such as excessive complexity and slow inference speed. To address these issues, the field of artificial intelligence has focused on researching lightweight and fast inference models in recent years. Lightweight models can effectively reduce the computational cost and parameter count, improving computational efficiency and inference speed, making them more suitable for real-time computing and embedded devices. For example, the MobileNet network proposed by Google in 2017 uses strategies such as depthwise separable convolutions and adjustable network width to effectively reduce the model's computational cost. In 2020, Facebook proposed a highly adjustable, modular convolutional neural network architecture, Regular Network (RegNet). This architecture, based on the combination design of basic modules and an adaptive network width design strategy, allows for the construction of RegNet networks of varying complexity to adapt to different tasks, thus providing greater flexibility for different scenarios. However, lightweight models also have some limitations. They suffer from decreased accuracy, especially when dealing with larger and more complex datasets and tasks. Furthermore, while lightweight models have been successfully applied to some practical tasks, their use in rolling bearing fault diagnosis is limited.

[0005] Besides optimizing model design strategies, employing deep transfer learning is also a common strategy to improve model training efficiency. Deep transfer learning, based on a pre-trained model trained on a large dataset, adjusts the model according to the specific task, achieving better performance with minimal training time compared to an initial training model. Some researchers have combined transfer learning with ResNet-50 to effectively diagnose various rolling bearing faults under different operating conditions. Others have used GAF ​​image encoding technology combined with a ResNet34 model based on transfer learning to classify single rolling bearing faults under varying speeds. Still others have proposed a fast rolling bearing fault diagnosis method based on transfer learning-based ResNet18, improving upon the problems of large training parameters and long training times in existing deep neural network rolling bearing fault diagnosis models.

[0006] It should be noted that the above methods are mainly used to identify and analyze single faults in rolling bearings. However, in the actual use of rotating machinery, with the passage of time and changes in the environment, rolling bearing faults are often no longer singular, but rather the result of a mixture of multiple fault modes. Applying the single fault diagnosis methods described above to identify mixed faults can easily lead to misjudgments. Therefore, based on the single fault diagnosis algorithm for rolling bearings, new methods and models need to be proposed to diagnose mixed faults in rolling bearings. Summary of the Invention

[0007] 1. The technical problem that the invention aims to solve

[0008] Given the poor diagnostic performance of existing solutions when dealing with mixed faults in the inner and outer rings of rolling bearings, and the high computational cost due to the high complexity of the model algorithms, this invention provides a fast method for diagnosing rolling bearing faults based on RegNet-SES. Compared with existing VGG16, GoogLeNet, ResNet34, and ResNet50 networks, this invention not only has lower model complexity, reducing the time and computational costs required for training the model, but also exhibits superior performance in the task of diagnosing mixed faults in rolling bearings, demonstrating significant practical application value.

[0009] 2. Technical Solution

[0010] To achieve the above objectives, the technical solution provided by the present invention is as follows:

[0011] The present invention provides a method for rapid diagnosis of rolling bearing faults based on RegNet-SES, comprising:

[0012] S1: Collect one-dimensional vibration acceleration signal data of rolling bearings in good condition, with different fault types and fault degrees;

[0013] S2: Sample the acquired one-dimensional vibration acceleration signal data, and use continuous wavelet transform to generate a two-dimensional time-frequency diagram from the sampled data;

[0014] S3: Create a mixed fault dataset for rolling bearings and divide it into training and test sets;

[0015] S4: Build the RegNet-SES network and use transfer learning to transfer the weight parameters of the pre-trained model to the completed RegNet-SES network;

[0016] S5: Train the RegNet-SES network using the mixed fault dataset of rolling bearings and update the network weight parameters. Train the RegNet-SES network until it converges, save the optimal model and use it as the mixed fault diagnosis model to perform rolling bearing fault diagnosis.

[0017] Furthermore, the failure types of rolling bearings include: inner ring failure, outer ring failure, and mixed inner and outer ring failure. For inner ring failure, outer ring failure, and mixed inner and outer ring failure, two levels of damage are collected respectively. The damage level is divided into level 1 and level 2. Level 1 represents a damage diameter of less than 2 mm, and level 2 represents a damage diameter of more than 2 mm but less than 4.5 mm.

[0018] Furthermore, step S2 employs a continuous wavelet transform method with Cmor as the wavelet basis function to perform continuous wavelet transform on the sampled data to generate a two-dimensional time-frequency diagram.

[0019] Furthermore, the RegNet-SES network includes a network input layer, a network body, and a network output layer. Transfer learning transfers the weight parameters of the pre-trained model's input layer and network body to the initialized RegNet-SES network.

[0020] Furthermore, the network body comprises four stages, each with a very similar structure, but differing in the number of SES attention group convolutional residual modules it contains. The advantage of this structure is that the network depth and width can be adjusted, thereby effectively improving model performance and reducing computational costs. The four stages are as follows:

[0021] Stage 1: Consists of a SES attention group convolutional residual module with downsampling;

[0022] Stage 2 consists of one SES attention group convolutional residual module with downsampling and two SES attention group convolutional residual modules;

[0023] Stage 3 consists of one SES attention group convolutional residual module with downsampling and five SES attention group convolutional residual modules.

[0024] Stage 4 consists of one SES attention group convolutional residual module with downsampling and five SES attention group convolutional residual modules.

[0025] Furthermore, the SES attention group convolutional residual module includes a standard convolutional layer with a kernel size of 1×1, a grouped convolutional layer with a kernel size of 3×3, an SES attention mechanism module, and a standard convolutional layer with a kernel size of 1×1. The SES attention group convolutional residual module constructs the residual structure by adding cross-layer connections at the input and output.

[0026] Furthermore, the SES attention group convolutional residual module with downsampling adds a standard convolutional layer with a kernel size of 1×1 at the cross-layer connection. The SES attention group convolutional residual module with downsampling performs a downsampling operation to ensure the connection with the previous layer.

[0027] Furthermore, the SES attention mechanism module is a hybrid attention mechanism module that combines the SE attention mechanism and the spatial attention mechanism. This module can better focus on the information of input features in different channels and different spatial locations, thereby improving the model's representation ability.

[0028] Furthermore, the network input layer includes a standard convolutional layer, a batch normalization (BN) layer, and a ReLU activation function layer. The BN layer normalizes the input feature tensor, and the ReLU activation function layer is used to introduce nonlinear factors into the neural network. The network output layer includes an adaptive average pooling layer, a dropout layer, and a fully connected layer.

[0029] Furthermore, when training the RegNet-SES network, the SGD optimizer is selected to update the model parameters, and the cosine annealing strategy is used to adjust the learning rate.

[0030] 3. Beneficial effects

[0031] Compared with existing known technologies, the technical solution provided by this invention has the following significant advantages:

[0032] (1) The present invention provides a method for rapid diagnosis of rolling bearing faults based on RegNet-SES. It uses continuous wavelet transform with Cmor as the wavelet basis function to transform the one-dimensional vibration acceleration signal data of rolling bearings into a two-dimensional time-frequency diagram with obvious feature information. This method is more in line with the RegNet-SES network model and is beneficial for the network to extract features and identify faults.

[0033] (2) The main body of the network is an important component of the RegNet-SES network. The construction of the main body of the network is directly related to the complexity and feature extraction capability of the entire network model. The fast rolling bearing fault diagnosis method based on RegNet-SES of this invention combines the SES attention mechanism with the group convolution residual structure to ensure the lightweight nature of the model, achieve high computational efficiency and fast inference speed, and improve the feature extraction capability of the model.

[0034] (3) In order to achieve faster effective extraction of input sample features by the model, the present invention proposes a fast rolling bearing fault diagnosis method based on RegNet-SES, which introduces transfer learning to reduce the time and computational costs consumed by the model to achieve high-precision rolling bearing hybrid fault diagnosis.

[0035] (4) The fast rolling bearing fault diagnosis method based on RegNet-SES of the present invention has lower model complexity and reduces the time and computation costs required for training the model compared with the existing VGG16 network, GoogLeNet network, ResNet34 network and ResNet50 network. It also has better performance in the rolling bearing hybrid fault diagnosis task and has very good practical application value. Attached Figure Description

[0036] Figure 1 This is a flowchart of a method for rapid diagnosis of rolling bearing faults based on RegNet-SES according to the present invention.

[0037] Figure 2 This is a schematic diagram of the time-domain vibration signals of normal, inner ring level 1 fault, inner ring level 2 fault, outer ring level 1 fault, outer ring level 2 fault, mixed inner and outer ring level 1 fault, and mixed inner and outer ring level 2 fault according to an embodiment of the present invention.

[0038] Figure 3 This is a schematic diagram of the continuous wavelet transform process.

[0039] Figure 4 This is a schematic diagram of the SES attention group convolutional residual module.

[0040] Figure 5 This is a schematic diagram of the SE attention mechanism module.

[0041] Figure 6 This is a schematic diagram of the spatial attention mechanism module.

[0042] Figure 7 This is a schematic diagram of the RegNet-SES network structure.

[0043] Figure 8 A diagram illustrating the transfer of weights from a pre-trained model for transfer learning. Detailed Implementation

[0044] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.

[0045] Example 1

[0046] like Figure 1 As shown in this embodiment, a method for rapid diagnosis of rolling bearing faults based on RegNet-SES includes the following steps:

[0047] S1: Collect one-dimensional vibration acceleration signal data of rolling bearings in good condition, with different fault types and fault degrees.

[0048] like Figure 2As shown, the different condition types of the collected rolling bearings include: normal, inner ring level 1 fault, inner ring level 2 fault, outer ring level 1 fault, outer ring level 2 fault, mixed inner and outer ring level 1 fault, and mixed inner and outer ring level 2 fault. Level 1 represents a damage diameter of less than 2 mm, and Level 2 represents a damage diameter of more than 2 mm but less than 4.5 mm.

[0049] S2: Sample the acquired one-dimensional vibration acceleration signal data, and use continuous wavelet transform to generate a two-dimensional time-frequency diagram from the sampled data.

[0050] Specifically, step S1 collects seven types of faults, saving each type of fault to a folder with filenames from 0 to 6. The corresponding rolling bearing vibration acceleration signal data is obtained by accessing the folder path, and the data is sampled. In this embodiment, the sampling frequency is 64kHz, and the sample length is 1024. Seven new folders are created, with filenames from 0 to 6. A continuous wavelet transform method using Cmor as the wavelet basis function is applied to each type of sample, generating time-frequency graph samples of size 236×236×3. These time-frequency graph samples are then saved to the corresponding folders. The transformation process is as follows: Figure 3 As shown, the final seven folders named 0, 1, 2, 3, 4, 5, and 6 contain seven types of two-dimensional time-frequency images: normal, inner circle level 1 fault, inner circle level 2 fault, outer circle level 1 fault, outer circle level 2 fault, mixed inner and outer circle level 1 fault, and mixed inner and outer circle level 2 fault.

[0051] S3: In this embodiment, the 7 types of fault two-dimensional time-frequency images obtained in step S2 are used as a mixed fault dataset for rolling bearings. The total number of samples in the dataset is 6994, of which the number of training set samples is 6300 and the number of test set samples is 694. The specific sample distribution is shown in Table 1.

[0052] Table 1. Distribution of mixed failure samples in rolling bearings

[0053]

[0054] S4: Build the RegNet-SES network and use transfer learning to transfer the weight parameters of the pre-trained model to the fully initialized RegNet-SES network; specifically:

[0055] like Figure 4 As shown, this embodiment constructs an SES attention group convolutional residual module, wherein the group convolutional residual module is the main module of the SES attention group convolutional residual module. The SES attention mechanism module is combined with the group convolutional residual module in an embedded manner, with the aim of improving the ability of the group convolutional residual module to acquire input feature information. The RegNet-SES network composed of this module has the advantages of lightweight and strong feature extraction capability.

[0056] The SES attention group convolutional residual module mainly includes a standard convolutional layer with a kernel size of 1×1, a grouped convolutional layer with a kernel size of 3×3, the SES attention mechanism module, and a standard convolutional layer with a kernel size of 1×1. The SES attention group convolutional residual module constructs the residual structure by adding cross-layer connections at the input and output.

[0057] The grouped convolutional residual module differs from the general residual module in that it replaces the standard convolutional layer with a kernel size of 3×3 in the general residual module with a grouped convolutional layer. This grouping operation can reduce the number of parameters and computation. Specifically, under the condition that the input feature size, convolutional kernel size and output feature size are the same, the number of parameters and computation required for the standard convolution operation is g times that of the grouped convolution (divided into g groups). See Table 2 for a detailed comparison. In addition, using the grouped convolution operation can also improve the computational efficiency of the model, especially when accelerating computation on a GPU.

[0058] Table 2 Comparison of Standard Convolution and Grouped Convolution Parameters

[0059]

[0060] The SES attention mechanism module combines the advantages of the Squeeze-and-Excitation (SE) attention mechanism and the Spatial Attention mechanism. Its specific implementation is as follows: Figure 5 As shown, given an input feature X, a series of convolution operations are performed to obtain a feature U with height H, width W, and number of feature channels C. Feature U is then input into the SE attention mechanism.

[0061] The SE attention mechanism is based on two key ideas: Squeeze and Excitation. First, the Squeeze stage uses global average pooling to process the two-dimensional features u of each channel of the feature U. i (i∈[1,C]) are compressed to reduce the features of this channel to a scalar Z. i The mathematical definition of the Squeeze operation is as follows:

[0062]

[0063] In the formula, i represents the channel index, u i This represents the global statistics of the i-th channel, which is a form of feature compression in spatial dimension because the scalar Z... i It is based on two-dimensional feature u iAll values ​​are calculated, so it has a global receptive field to some extent. After performing the Squeeze operation on feature U, we obtain the global information Z of feature U, Z = [z1, z2, ..., z]. c ]∈R C R C Let C represent a vector space consisting of C real numbers.

[0064] The excitation phase uses a two-layer fully connected network to learn the importance weights for each channel. The first fully connected layer compresses the Z obtained from the previous Squeeze operation to a smaller dimension, i.e.:

[0065] f(z i )=W1z i +b1

[0066] In the formula, W1∈R C×C / r and b1∈R C / r These are the weights and biases of the first fully connected layer, R. C×C / r Let r represent a real matrix of size C / r×C, where r is an adjustable scaling factor.

[0067] The first fully connected layer compresses C channels into C / r channels to reduce computation; then, ReLU is used as the activation function to obtain:

[0068] a i =Relu(f(z) i ))

[0069] Then, the channel is restored to C-dimensionality through a second fully connected layer, i.e.:

[0070] s i =W2a i +b2

[0071] In the formula, W2∈R C×C / r and b2∈R C These are the weights and biases of the second fully connected layer, respectively. The function of these two fully connected layers is to fuse the feature information s from each channel. i , for s i Perform sigmoid normalization to obtain the attention weights S for all channels, S = [s1, s2, ..., s]. c ] T Finally, the attention weight S is weighted (Fscaled) with the input feature U to output feature X'.

[0072] Combination Figure 6This embodiment uses spatial attention to weight the feature map X' output after the channel attention mechanism in the preceding text in a spatial dimension. The aim is to allow the model to better focus on the information of the input feature X in different channels and spatial locations, thereby improving the model's representational ability. The specific implementation consists of two steps: first, max pooling and average pooling operations are performed on the feature map X' in the spatial dimension, mathematically defined as follows:

[0073] P = [p] max (X'),p avg (X')]

[0074] In the formula, P max (X') and P avg( X') are two statistical features obtained from max pooling and average pooling.

[0075] P max (X') and P avg( X') are concatenated to obtain feature P, and then feature P is passed through a fully connected layer with a kernel size of 7×7 to obtain a scalar weight M, that is:

[0076] M = W3P + b3

[0077] In the formula, W3∈R 2 b3∈R are the weights and biases of the fully connected layer, respectively. Then, the weights Ms are obtained by normalizing M using the Sigmoid function, i.e.:

[0078]

[0079] Finally, the obtained weights Ms are weighted and averaged with the input features X' to obtain the final output features F. Features F contain the feature information of the input features X in different channels and different spatial locations. Then, features F are input into the second standard convolutional layer with a kernel size of 1×1 in the group convolutional residual module for the next step of feature extraction.

[0080] The RegNet-SES network constructed in this embodiment has the following structure: Figure 7 As shown, it includes the network input layer, the network body, and the network output layer.

[0081] The network input layer consists of a standard convolutional layer with a kernel size of 3×3 and 32 kernels, a Batch Normalization (BN) layer, and a ReLU activation function layer. Figure 7The diagram only shows the most important standard convolutional layer in the input layer; the BN and ReLU layers are not shown. When a sample of size 236×236×3 is input into the input layer, a feature of size 118×118×32 will be obtained, and then the feature of size 118×118×32 will be fed into the main body of the network.

[0082] The Batch Normalization (BN) layers used in the RegNet-SES network constructed in this invention normalize the input feature tensor to avoid gradient vanishing or exploding in the neural network. It normalizes the mean and variance of each channel along the batch dimension, making the mean of the input data 0 and the variance 1, and adjusts the data distribution through scaling factors and offset terms to enhance the network's generalization ability. Specific parameter settings include:

[0083] 1) eps: Used to prevent numerical instability caused by a denominator that is too small. It is generally set to 1e-5, i.e., 1×10 -5 ;

[0084] 2) momentum: The momentum term used in batch normalization, typically 0.1, is used to smooth the updates of the mean and variance;

[0085] 3) affine: Whether to use affine transformation, usually True. If True, then the scaling parameters gamma and beta also need to be learned.

[0086] 4) track_running_stats: Whether to track global mean and variance running statistics. The default is True.

[0087] ReLU (Modified Linear Unit) is a commonly used activation function to introduce nonlinearity into neural networks. The ReLU function has the form f(x) = max(0,x), meaning it outputs 0 when the input x is less than 0, and outputs x itself when x is greater than 0.

[0088] This invention constructs a RegNet-SES network, including a main network body. The RegNet-SES network body consists of four stages, each with a very similar structure, but differing in the number of SES attention group convolutional residual modules it contains. The advantage of this structure is that the network depth and width can be adjusted, thereby effectively improving model performance and reducing computational costs. Each stage specifically comprises:

[0089] Stage 1 consists of a SES attention group convolutional residual module with downsampling. Specifically, it includes: a standard convolutional layer with a 1×1 kernel and 48 kernels; a grouped convolutional layer with a 3×3 kernel, 48 kernels, and 6 groups; the SES attention mechanism module; and the standard convolutional layer with a 1×1 kernel and 48 kernels. The SES attention group convolutional residual module with downsampling adds a standard convolutional layer with a 1×1 kernel and 48 kernels at the cross-layer connection of the SES attention group convolutional residual module to downsample the input features, allowing the input features to be added to the output features. The 118×118×32 features are fed into Stage 1 of the network body, resulting in features of size 59×59×48, which are then passed to Stage 2.

[0090] Stage 2 consists of one SES attention group convolutional residual module with downsampling and two SES attention group convolutional residual modules, which are connected in series.

[0091] The SES attention group convolutional residual module with downsampling includes: a standard convolutional layer with a kernel size of 1×1 and 104 kernels; a grouped convolutional layer with a kernel size of 3×3, 104 kernels, and 13 groups; the SES attention mechanism module and the standard convolutional layer with a kernel size of 1×1 and 104 kernels. The SES attention group convolutional residual module with downsampling adds a standard convolutional layer with a kernel size of 1×1 and 104 kernels at the cross-layer connection of the SES attention group convolutional residual module to downsample the input features, so that the input features can be added to the output features.

[0092] Both SES attention group convolutional residual modules have the same configuration, including: a standard convolutional layer with a kernel size of 1×1 and 104 kernels; a grouped convolutional layer with a kernel size of 3×3, 104 kernels, and 13 groups; the SES attention mechanism module and the standard convolutional layer with a kernel size of 1×1 and 104 kernels, which add the input and output features by adding cross-layer connections. The 59×59×48 feature is fed into Stage 2 of the network body, resulting in a feature size of 30×30×104, which is then fed into Stage 3.

[0093] Stage 3 consists of one SES attention group convolutional residual module with downsampling and five SES attention group convolutional residual modules, which are connected in series.

[0094] The SES attention group convolutional residual module with downsampling includes: a standard convolutional layer with a kernel size of 1×1 and 208 kernels; a grouped convolutional layer with a kernel size of 3×3, 208 kernels, and 26 groups; the SES attention mechanism module and the standard convolutional layer with a kernel size of 1×1 and 208 kernels. The SES attention group convolutional residual module with downsampling adds a standard convolutional layer with a kernel size of 1×1 and 208 kernels at the cross-layer connection of the SES attention group convolutional residual module to downsample the input features, so that the input features can be added to the output features.

[0095] The five SES attention group convolutional residual modules are configured identically: a standard convolutional layer with a kernel size of 1×1 and 208 kernels; a grouped convolutional layer with a kernel size of 3×3, 208 kernels, and 26 groups; and the SES attention mechanism module and the standard convolutional layer with a kernel size of 1×1 and 208 kernels, which add input and output features by adding cross-layer connections. The 30×30×104 features are fed into Stage 3 of the network body, resulting in features of size 15×15×208, which are then passed to Stage 4.

[0096] Stage 4 consists of one SES attention group convolutional residual module with downsampling and five SES attention group convolutional residual modules, which are connected in series.

[0097] The SES attention group convolutional residual module with downsampling includes: a standard convolutional layer with a kernel size of 1×1 and 440 kernels; a grouped convolutional layer with a kernel size of 3×3, 440 kernels, and 55 groups; the SES attention mechanism module and the standard convolutional layer with a kernel size of 1×1 and 440 kernels. The SES attention group convolutional residual module with downsampling adds a standard convolutional layer with a kernel size of 1×1 and 440 kernels at the cross-layer connection of the SES attention group convolutional residual module to downsample the input features, so that the input features can be added to the output features.

[0098] The five SES attention group convolutional residual modules are configured identically, each including: a standard convolutional layer with a kernel size of 1×1 and 440 kernels; a grouped convolutional layer with a kernel size of 3×3, 440 kernels, and 55 groups; and the SES attention mechanism module and the standard convolutional layer with a kernel size of 1×1 and 440 kernels, which add input and output features by adding cross-layer connections. The 15×15×208 features are fed into Stage 3 of the network body, resulting in features of size 7×7×440, which are then passed to the output layer.

[0099] This invention constructs a RegNet-SES network, including a network output layer, which comprises an adaptive average pooling (Avgpool) layer, a dropout layer, and a fully connected (FC) classification layer.

[0100] The adaptive average pooling layer adaptively scales the input features of size 7×7×440 to the size of Output_size, and then performs an averaging operation on each scaled region to obtain output features of the same size as Output_size. The number of feature channels is not changed during the operation. In this invention, Output_size is set to (1, 1), and the output feature size is 1×1×440.

[0101] In neural networks, the dropout operation is a regularization method that randomly sets some elements of the input data to zero, aiming to reduce overfitting during model training. Depending on the dropout rate p, the dropout layer will randomly set each element of the input tensor to zero with probability p. In this invention, the dropout rate p is set to 0.25.

[0102] The fully connected classification layer connects each element in the input data tensor (because it is a dense layer) to each element in the output data tensor, thereby mapping the information from the input layer to the output layer. An activation function is then applied to the output data tensor to obtain a set of non-linear feature representations for subsequent classification tasks. In this invention, the input tensor is set to 440, and the output data tensor to 7.

[0103] like Figure 8 As shown, transfer learning is used to transfer the pre-trained model weight parameters to the RegNet-SES rolling bearing hybrid fault diagnosis model. This process includes the following steps:

[0104] 1) Initialize the RegNet-SES rolling bearing hybrid fault diagnosis model: It is necessary to initialize a RegNet-SES rolling bearing hybrid fault diagnosis model so that the pre-trained weight parameters can be transferred to this model.

[0105] 2) Transfer weight parameters: Transfer the weight parameters of the pre-trained model obtained from the ImageNet dataset to the initialized RegNet-SES rolling bearing hybrid fault diagnosis model.

[0106] 3) Freeze model layers: After transferring the weight parameters, freeze the weight parameters of the input layer and the main body of the network, and do not participate in the training. Only update the weight parameters of the output layer.

[0107] 4) Train the RegNet-SES network using the rolling bearing mixed fault dataset and update the network weight parameters. Train the RegNet-SES network to learn how to identify mixed faults in rolling bearings from the rolling bearing mixed fault dataset.

[0108] Before training the network, a deep learning environment needs to be configured. This requires a Windows 10 operating system with an Intel Core i3-12100f processor, 16GB of RAM, an NVIDIA GTX 3060ti 8GB GPU, CUDA version 11.6, PyTorch as the deep learning framework, Python programming language, and PyCharm compiler. Since GPUs perform better with batch sizes that are powers of 2, a batch size of 32 was set during training. The model was trained for 30 iterations, using the SGD optimizer to update model parameters, and cosine annealing to adjust the learning rate η. t The formula is shown below:

[0109]

[0110] In the formula η max and η min This represents the range of the learning rate, x is the x-th epoch of model training, and T is the total number of training iterations. This embodiment sets the parameter η... max and η min The values ​​are 1 and 0.01 respectively, and T is 30.

[0111] After configuration, training begins. After 30 iterations, the model parameter weights from each iteration are saved to a designated folder. The fault identification performance of each model is verified using a mixed test set of rolling bearings. Finally, the model with the best identification performance is selected as the mixed fault diagnosis model for rolling bearings. The number of parameters required for training the RegNet-SES model of this invention is 3,907,815 params, and the model complexity is 487.41 MFlops. The prediction results and recognition accuracy of the optimal model for each type of fault on the mixed test set of rolling bearings are shown in Table 3. The results show that the RegNet-SES mixed fault diagnosis model for rolling bearings of this invention can achieve a 100% recognition rate for whether a rolling bearing has a fault. It can achieve a 100% recognition rate for five types of faults: healthy, outer ring 1, mixed 1, mixed 2, and inner ring 1. For the outer ring 2 and inner ring 2 types, only one sample is mispredicted. This indicates that the RegNet-SES network with a complexity of 400 MFlops can effectively distinguish between single faults and mixed faults of different damage levels in rolling bearings.

[0112] Table 3. Prediction results and recognition accuracy of single-class samples on the test set.

[0113]

[0114] This invention is applied to the identification and classification of mixed faults with different damage levels in rolling bearings. The rolling bearing dataset (PU dataset) from the University of Paderborn, Germany, was used for experimental verification. The PU dataset consists of data from healthy bearings and faulty bearings. The faulty bearing data includes data from bearings with actual damage and those with artificial damage. The PU dataset collects experimental data from a 6203 deep groove ball bearing, with a sampling frequency of 64kHz. This invention primarily uses the vibration signal data of bearings with actual damage from this dataset to verify the effectiveness and superiority of the proposed method. The experimental data were collected under the conditions of a rotational speed N = 900 r / min, a load torque M = 0.7 Nm, and a radial force F = 1000 N on the bearing.

[0115] Compared with existing VGG16, GoogLeNet, ResNet34 and ResNet50 networks (see Table 4), this invention not only has lower model complexity and reduces the number of parameters required to train the model, effectively reducing the time and computational costs of training the model, but also has higher recognition accuracy in the mixed fault diagnosis task of rolling bearings, and has good practical application value.

[0116] Table 4 Comparison of different network models

[0117]

[0118] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A fast diagnosis method for rolling bearing fault based on RegNet-SES, characterized in that, The application relates to a rolling bearing fault diagnosis method based on RegNet-SES network. The method comprises the following steps: S1: collecting one-dimensional vibration acceleration signal data of healthy rolling bearings, rolling bearings with different fault types and rolling bearings with different fault degrees; S2: sampling the collected one-dimensional vibration acceleration signal data, and using continuous wavelet transformation to generate two-dimensional time-frequency diagrams from the sampled data samples; S3: preparing a rolling bearing mixed fault data set and dividing the data set into a training set and a test set; S4: building a RegNet-SES network, and using transfer learning to transfer pre-trained model weight parameters to the initialized RegNet-SES network; the RegNet-SES network comprises a network input layer, a network main body and a network output layer, and the transfer learning is used to transfer the weight parameters of a pre-trained model input layer and the network main body to the initialized RegNet-SES network; The network main body comprises four stages, and the four stages are as follows: Stage 1: composed of one SES attention group convolution residual module with downsampling; Stage 2: composed of one SES attention group convolution residual module with downsampling and two SES attention group convolution residual modules; Stage 3: composed of one SES attention group convolution residual module with downsampling and five SES attention group convolution residual modules; Stage 4: composed of one SES attention group convolution residual module with downsampling and five SES attention group convolution residual modules; The SES attention group convolution residual module comprises a standard convolution layer with a convolution kernel size of 1*1, a group convolution layer with a convolution kernel size of 3*3, an SES attention mechanism module and a standard convolution layer with a convolution kernel size of 1*1, and the SES attention group convolution residual module is composed of a residual structure by adding a cross-layer connection at the input and the output; The SES attention group convolution residual module with downsampling is provided with a standard convolution layer with a convolution kernel size of 1*1 at the cross-layer connection, and the SES attention group convolution residual module with downsampling performs a downsampling operation to ensure the connection with the previous layer; The SES attention mechanism module is a hybrid attention mechanism module combining an SE attention mechanism and a spatial attention mechanism; 2. The fast diagnosis method of rolling bearing fault based on RegNet-SES according to claim 1, characterized in that: S5: training the RegNet-SES network by using the rolling bearing mixed fault data set and updating the network weight parameters, training the RegNet-SES network to convergence, saving the optimal model as a mixed fault diagnosis model and performing rolling bearing fault diagnosis.

3. The fast diagnosis method of rolling bearing fault based on RegNet-SES according to claim 2, characterized in that: The fault types of the rolling bearings include inner ring faults, outer ring faults and inner-outer ring mixed faults, wherein two fault damage degrees are collected for the inner ring faults, the outer ring faults and the inner-outer ring mixed faults, and the damage degrees are divided into level 1 and level 2.

4. The fast diagnosis method of rolling bearing fault based on RegNet-SES according to claim 3, characterized in that: The continuous wavelet transformation method with Cmor as a wavelet base function is adopted in step S2 to generate two-dimensional time-frequency diagrams from the sampled data samples. The network input layer comprises a standard convolution layer, a BN layer and an activation function ReLU layer, the BN layer normalizes the input feature tensor, and the activation function ReLU layer is used to introduce a nonlinear factor in the neural network. The network output layer includes an adaptive average pooling layer, a Dropout layer and a fully connected layer.

5. The fast diagnosis method of rolling bearing fault based on RegNet-SES according to claim 4, characterized in that: When training the RegNet-SES network, an SGD optimizer is selected to update the model parameters, and a cosine annealing strategy is used to adjust the learning rate.

Images