Rolling bearing service life prediction method based on Swin Transformer model

By adopting the Swin Transformer model in the bearing life prediction, combining the convolutional neural network and the fully connected layer, and using the mobile window self-attention mechanism, the problem of low accuracy of the traditional model is solved, and more efficient and accurate bearing life is achieved. Forecast improves reliability and reduces operation and maintenance costs.

CN118940625APending Publication Date: 2024-11-12山东浪潮智能生产技术有限公司
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410990506.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing bearing residual service life prediction model based on deep learning is not very accurate, and traditional convolutional neural networks show weak generalization ability and low accuracy in bearing life prediction.

Method used

The rolling bearing life prediction method based on the Swin Transformer model is adopted. By Z-score normalizing the original vibration data, the Swin Transformer model is constructed, combining the convolutional neural network and the fully connected layer, and the mobile window self-attention mechanism and multi-head Self-attention mechanism improves the model's feature extraction and global perception ability, and conducts training and prediction.

Benefits of technology

It improves the accuracy of bearing residual life prediction, improves the reliability and safety of bearings, reduces operation and maintenance costs, and shows that the prediction results are more efficient and accurate than traditional convolutional neural networks and fully connected networks.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940625A_ABST
    Figure CN118940625A_ABST
Patent Text Reader

Abstract

A rolling bearing life prediction method based on a Swin Transform model comprises the following steps: dividing original vibration data into a training set and a test set, carrying out Z-score normalization processing on the original vibration data, constructing the Swin Transform model, importing training set data and labels into the Swin Transform model for training, and training the Swin Transform model until convergence; and finally, importing the test set and the tag into the trained Swin Transformer model to obtain the residual life of the rolling bearing. The method can predict the service life of the rolling bearing more effectively and accurately.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of bearing life prediction, and in particular to a rolling bearing life prediction method based on a Swin Transformer model. Background Art

[0002] Rolling bearings are one of the most important parts in machinery, and are also the most basic parts that are prone to failure. They are widely used in various industrial fields. Once a failure occurs, it will affect the normal operation and working efficiency of the machine. In severe cases, it will cause mechanical damage, resulting in economic losses and casualties. With the gradual development of the machinery industry, people often need to complete tasks more efficiently based on limited machine resources, so it is very important to reduce the failure rate of equipment. Therefore, it is of great significance to accurately and effectively predict the remaining useful life (RUL) of rolling bearings.

[0003] Remaining service life prediction can be roughly divided into two categories, namely model-based methods and data-driven methods. Model-based methods rely heavily on a large amount of professional knowledge, and the models usually show weak generalization ability and low accuracy. Based on data-driven methods, with the development of artificial intelligence, deep learning has been widely used in bearing life prediction due to its powerful feature extraction and generalization capabilities, and has improved the accuracy of predictions. However, at present, most of the remaining service life predictions of bearings based on deep learning use traditional convolutional neural networks to build models for remaining service life prediction. However, the prediction results of the models built by this network are not accurate. Summary of the invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a rolling bearing life prediction method based on the SwinTransformer model, which can effectively improve the accuracy of the prediction of the remaining service life of the bearing, improve the reliability and safety of the bearing, and reduce the operation and maintenance costs.

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

[0006] A rolling bearing life prediction method based on a Swin Transformer model comprises the following steps:

[0007] Step 1: Divide the original vibration data into a training set and a test set, perform Z-score normalization on the original vibration data, and convert the data into a standard normal distribution with a mean of 0 and a variance of 1;

[0008] Step 2: Build a Swin Transformer model. In the Swin Transformer model, a convolutional neural network is first used to process the input data, and then data segmentation is performed. Convolutional embedding replaces the ordinary fully connected layer embedding in the Swin Transformer module. Then it enters the Swin Transformer network composed of more than two Swin Transformer modules, and then average pooling is performed, and then it is flattened through the Flatten layer, and finally the prediction result is output through the fully connected layer;

[0009] Step 3: Import the training set data and labels into the Swin Transformer model for training. Train the Swin Transformer model until convergence. The Adam optimization algorithm is used during training, and the loss function is the MSE loss function.

[0010] Step 4: Import the test set and labels into the trained Swin Transformer model to obtain the remaining life of the rolling bearing.

[0011] In step 1, the Z-score normalization process is performed on the original vibration data, and the calculation is expressed as:

[0012]

[0013] Where mean is the mean, std is the standard deviation, x is the original vibration data, and y is the data after the original vibration data is normalized by Z-score. The labels are also normalized, and the life labels are converted between [0,1]. The operation of normalizing the life labels is expressed as follows:

[0014]

[0015] Where y c is the current remaining life, y t is the total lifespan, is the normalized lifespan label;

[0016] The structure of the Swin Transformer module described in step 2 is first a LN (LayerNormalization) layer, through which normalization can be performed, and then through the W-MSA (Window-based Multi-head Self-Attention) module, which divides the window, performs self-attention calculation on each window and generates a weighted vector, and then splices the weighted vectors calculated by multiple windows together. The spliced ​​vectors are then transformed through a linear transformation and an LN layer, and finally through the MLP (Multilayer Perceptron) layer and then input into the next Swin Transformer module after linear transformation. In the MLP layer, it first passes through a fully connected layer, then uses the GELU activation function, then passes through a dropout layer, then passes through a fully connected layer without using an activation function, and finally passes through a dropout layer; in the next Swin Transformer module, the W-MSA (Window-based Multi-head Self-Attention) module is replaced with the SW-MSA (Shifted Window-based Multi-head Through the window displacement operation, SW-MSA can enhance the information interaction between windows, thereby improving the performance of the model. The other components are the same as the first Swin Transformer module.

[0017] Convolutional neural network model and fully connected network model were constructed, and the same training set was input into the convolutional neural network model and the fully connected network model for training, and the test set and label were input into the two trained convolutional neural network models and the fully connected network model, and the remaining life of rolling bearings under the convolutional neural network model and the fully connected network model were obtained respectively; the bearing prediction results under the Swin Transformer model, convolutional neural network model and fully connected network model were compared, and the root mean square error (RMSE) was used to evaluate the prediction results. The calculation formula of RMSE is:

[0018]

[0019] In the formula, is the predicted value, y i is the actual value, V is the test sample length, and the accuracy of the model is evaluated by comparing RMSE.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] The rolling bearing life prediction method based on the Swin Transformer model provided by the present invention innovatively integrates the Swin Transformer module with the bearing life prediction. Since the Swin Transformer module has a moving window self-attention mechanism, the computational efficiency is improved. When the convolution embedding replaces the ordinary fully connected layer embedding in the Swin Transformer module before the data is input into the Swin Transformer module, the data features can be more fully retained. Compared with traditional convolutional neural networks and fully connected neural networks, the Swin Transformer model has powerful feature extraction capabilities and global perception capabilities, and can predict the remaining life of the bearing more efficiently and accurately. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 It is a flow chart of the present invention.

[0023] Figure 2 It is a network framework diagram of the Swin Transformer model of an embodiment of the present invention.

[0024] Figure 3 It is a schematic diagram of the Swin Transformer module of an embodiment of the present invention.

[0025] Figure 4 4 is a bearing life prediction diagram of an embodiment of the present invention.

[0026] Figure 5 This is a comparative diagram of bearing life prediction based on convolutional neural network.

[0027] Figure 6 This is a comparative bearing life prediction diagram based on a fully connected neural network. DETAILED DESCRIPTION

[0028] In order to better understand the features, advantages and technical contents of the present invention, the present invention is described in detail below in conjunction with embodiments and drawings.

[0029] Reference Figure 1 , a rolling bearing life prediction method based on the Swin Transformer model, comprising the following steps:

[0030] Step 1: Using the Xi'an Jiaotong University Bearing Open Source Dataset (XJTU-SY), select the data with the working condition of 2100rpm (35Hz) and 12KN as the original vibration data, and divide the original vibration data into a training set and a test set; perform Z-score normalization on the original vibration data, normalize the labels, and convert the life labels to between [0,1]; the Z-score normalization of the original vibration data is expressed as:

[0031]

[0032] Where mean is the mean, std is the standard deviation, x is the original vibration data, and y is the data after the original vibration data is normalized by Z-score; this method can convert the original vibration data into a standard normal distribution with a mean of 0 and a variance of 1;

[0033] The normalization operation for the lifespan label is expressed as:

[0034]

[0035] Where y c is the current remaining life, y t is the total lifespan, is the normalized lifespan label;

[0036] Step 2: Build the Swin Transformer model, such as Figure 2 As shown in the figure, in the Swin Transformer model, a convolutional neural network is first used to process the input data, and then data segmentation is performed. Convolutional embedding replaces the ordinary fully connected layer embedding in the Swin Transformer module, and then enters the Swin Transformer network composed of 6 Swin Transformer modules, followed by average pooling, and then flattening through the Flatten layer, and finally the prediction result is output through the fully connected layer;

[0037] The structure of the Swin Transformer module is as follows Figure 3As shown in the figure, first there is an LN (LayerNormalization) layer, through which normalization can be performed, and then through the W-MSA (Window-based Multi-head Self-Attention) module, which divides the window, performs self-attention calculation on each window and generates a weighted vector, and then concatenates the weighted vectors calculated by multiple windows, and the concatenated vectors pass through a linear transformation and an LN layer, and finally pass through the MLP (Multilayer Perceptron) layer and then through a linear transformation and input into the next Swin Transformer module. In the MLP layer, it first passes through a fully connected layer, then uses a GELU activation function, then passes through a dropout layer, then passes through a fully connected layer without using an activation function, and finally passes through a dropout layer; in the next Swin Transformer module, the W-MSA (Window-based Multi-head Self-Attention) module is replaced with a SW-MSA (Shifted Window-based Multi-head Self-Attention) module. Through the displacement operation of the window, SW-MSA can enhance the information interaction between windows, thereby improving the performance of the model. The other components are the same as the first Swin The Transformer module is the same;

[0038] Step 3: Import the training set data and labels into the Swin Transformer model for training. Train the Swin Transformer model until convergence. The Adam optimization algorithm is used during training, and the loss function is the MSE loss function.

[0039] Step 4: Import the test set and labels into the trained Swin Transformer model to obtain the remaining life of the rolling bearing.

[0040] Comparative example: A convolutional neural network model and a fully connected network model are constructed respectively, and the same training set is input into the convolutional neural network model and the fully connected network model for training respectively, and the test set and the label are input into the two trained convolutional neural network models and the fully connected network model, and the remaining life of the rolling bearing under the convolutional neural network model and the fully connected network model is obtained respectively;

[0041] The bearing prediction results under the Swin Transformer model, convolutional neural network model, and fully connected network model are compared, such as Figure 4 , Figure 5 , Figure 6As shown in the figure, it can be seen that the prediction results of the Swin Transformer model are more accurate than those of the convolutional neural network model and the fully connected network model. The root mean square error (RMSE) is used to evaluate the prediction results. The calculation formula of RMSE is:

[0042]

[0043] In the formula, is the predicted value, y i is the actual value, V is the test sample length, and the accuracy of the model is evaluated by comparing RMSE. The specific results are shown in Table 1;

[0044] Table 1 RMSE values ​​corresponding to the prediction results of different methods

[0045]

[0046] By comparing the RMSE values ​​of the three networks, it was found that the Swin Transformer network had the highest prediction accuracy.

Claims

1. A rolling bearing life prediction method based on Swin Transformer model, characterized in that: The following steps are involved: Step 1: Divide the original vibration data into a training set and a test set, perform Z-score normalization on the original vibration data, and convert the data into a standard normal distribution with a mean of 0 and a variance of 1; Step 2: Build a Swin Transformer model. In the Swin Transformer model, a convolutional neural network is first used to process the input data, and then data segmentation is performed. Convolutional embedding replaces the ordinary fully connected layer embedding in the Swin Transformer module. Then it enters the Swin Transformer network composed of more than two Swin Transformer modules, and then average pooling is performed, and then it is flattened through the Flatten layer, and finally the prediction result is output through the fully connected layer; Step 3: Import the training set data and labels into the Swin Transformer model for training. Train the Swin Transformer model until convergence. The Adam optimization algorithm is used during training, and the loss function is the MSE loss function. Step 4: Import the test set and labels into the trained Swin Transformer model to obtain the remaining life of the rolling bearing.

2. The rolling bearing life prediction method based on the Swin Transformer model according to claim 1 is characterized in that: In step 1, the Z-score normalization process is performed on the original vibration data, and the calculation is expressed as: Where mean is the mean, std is the standard deviation, x is the original vibration data, and y is the data after the original vibration data is normalized by Z-score. The labels are also normalized, and the life labels are converted between [0,1]. The operation of normalizing the life labels is expressed as follows: Where y c is the current remaining lifespan, y t is the total lifespan, is the normalized lifespan label.

3. The rolling bearing life prediction method based on the Swin Transformer model according to claim 1 is characterized in that: The structure of the Swin Transformer module described in step 2 is first a LN (LayerNormalization) layer, through which normalization can be performed, and then through the W-MSA (Window-based Multi-head Self-Attention) module, which divides the window, performs self-attention calculation on each window and generates a weighted vector, and then splices the weighted vectors calculated by multiple windows together. The spliced ​​vectors are then transformed through a linear transformation and an LN layer, and finally through the MLP (Multilayer Perceptron) layer and then input into the next Swin Transformer module after linear transformation. In the MLP layer, it first passes through a fully connected layer, then uses the GELU activation function, then passes through a dropout layer, then passes through a fully connected layer without using an activation function, and finally passes through a dropout layer; in the next Swin Transformer module, the W-MSA (Window-based Multi-head Self-Attention) module is replaced with the SW-MSA (Shifted Window-based Multi-head Through the window displacement operation, SW-MSA can enhance the information interaction between windows, thereby improving the performance of the model. The other components are the same as the first Swin Transformer module.

4. The rolling bearing life prediction method based on the Swin Transformer model according to claim 1 is characterized in that: Convolutional neural network model and fully connected network model were constructed, and the same training set was input into the convolutional neural network model and the fully connected network model for training, and the test set and label were input into the two trained convolutional neural network models and the fully connected network model, and the remaining life of rolling bearings under the convolutional neural network model and the fully connected network model were obtained respectively; the bearing prediction results under the Swin Transformer model, convolutional neural network model and fully connected network model were compared, and the root mean square error (RMSE) was used to evaluate the prediction results. The calculation formula of RMSE is: In the formula, is the predicted value, y i is the actual value, V is the test sample length, and the accuracy of the model is evaluated by comparing RMSE.