Bearing life prediction method based on combination of pooling-multi-scale convolutional neural network and Transform

Abstract

The invention provides a bearing life prediction method based on combination of a variable pooling-multi-scale convolutional neural network and a Transformer, and relates to the technical field of bearing prediction, a VPMCNN and the Transformer are combined, the ability of capturing global and detail features in the VPMCNN is reserved, meanwhile, the Transformer learns abundant expressions of sequences before and after input signals through a self-attention mechanism, and the prediction accuracy of the bearing life is improved. The residual life of the bearing can be better predicted through the input acoustic emission signal of the rolling bearing; the disadvantage that the Transform cannot capture detail features is solved, meanwhile, a self-attention mechanism in a Transform model can capture a long-distance dependency relationship in an input signal, and the problem of memory deterioration of a prediction sequence is effectively solved. In order to reduce the excessive confidence of the model for correct labels, avoid overfitting and improve the generalization ability of the model, a Lable-Smoothing regularization technology is embedded when a training data set is labeled, smooth processing of the labels is carried out, and hard labels are changed into soft labels during training.

Description

Technical Field

[0001] The present invention relates to the relevant fields of bearing life, and in particular to a bearing life prediction method based on the combination of variable pooling-multi-scale convolutional neural network and Transformer. Background Art

[0002] Rolling bearings are key components of modern industrial production and equipment operation. Continuous wear during operation will reduce the service life of the equipment. Therefore, predicting the life of rolling bearings is of great significance and reference value for predicting the remaining service life of bearings in components such as engines.

[0003] like Figure 1 As shown, in the prior art, the method of establishing MSCNN (1D-multiscale-CNN convolutional neural network) is usually used to predict the life of rolling bearings, which consists of four convolutional pooling layers, a fully connected layer and a softmax layer. The multiscale layer (Multiscale, MS) is placed behind the convolutional pooling layer. It aims to capture the different scale features of the input data through convolution kernels of different sizes. Smaller convolution kernels can capture detailed features, while larger convolution kernels can capture overall features, thereby predicting the life of rolling bearings. It can be seen that in the above prediction method, the pooling layer in MSCNN only performs downsampling operations to reduce the dimensionality of the input original acoustic emission data; however, this unmodified dimensionality reduction may lose some features with relatively large weights and are more important to the prediction task. At the same time, the original MSCNN also has the problem of prediction sequence degradation, and the model is overconfident about the correct label, and there is a risk of overfitting. Summary of the invention

[0004] The purpose of an embodiment of the present invention is to provide a bearing life prediction method based on a variable pooling-multi-scale convolutional neural network combined with a Transformer, so as to alleviate the problem that the pooling layer in the MSCNN in the prior art only performs a downsampling operation and reduces the dimension of the input original acoustic emission data; however, this unmodified dimensionality reduction may lose some features with relatively large weights and are relatively important for the prediction task. At the same time, the original MSCNN also has a prediction sequence degradation problem, and the model is overconfident about the correct label and has a technical problem of overfitting risk.

[0005] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0006] A first aspect of the present invention provides a life prediction method for bearing acoustic emission signals, comprising the following steps:

[0007] S10. Acoustic emission signal x of the original bearing *={x1,x2,…,x n} to collect;

[0008] Where x* represents the normalized data;

[0009] S20. Perform data enhancement on the original bearing acoustic emission signal, expand the data sample, and divide it into a training data set and a test data set;

[0010] S30.,establish a neural network that is sequentially provided with a convolutional layer, a variable pooling layer, a multi-scale layer, and a Transformer model, and outputs the prediction result of the one-dimensional sequence through a fully connected layer and a Softmax function;

[0011] S40. Input the features of the multi-scale layer output in the variable pooling-multi-scale convolutional neural network into the Transformer model; after encoding and decoding, select the mean square error as the loss function; after the full connection layer and softmax prediction output, output the predicted value. The variable pooling layer changes the proportion of the bearing signal features by assigning different weight coefficients to the channel domain features of the pooling layer. The variable pooling layer calculates the weight coefficient of the pooling layer channel, and then combines the weight coefficient with the features of the pooling layer channel. By assigning weight coefficients to the pooling layer channels, feature information with a large predictive effect can be enhanced, and feature information with a small predictive effect can be weakened.

[0012] Furthermore, in the multi-scale layer in step S30, a convolution kernel 1×1 is added before the convolution kernel 5×1, and a 3×1 maximum pooling is set to perform feature extraction on the data.

[0013] Furthermore, the steps of establishing the variable pooling layer in step S30 are as follows:

[0014] S31. The features generated by the convolutional layer are combined and converted into a feature vector, which is expressed as:

[0015]

[0016] Where H is the length of the pooling layer, u c represents the jth neuron in the cth pooling channel field of the pooling layer, z c Represents the average feature information of the c-th pooled channel domain;

[0017] S32. Calculate the weight coefficient of the channel domain feature of the pooling layer, the expression is:

[0018] S c =σ(W2δ(W1z c ))

[0019] In the formula, δ is the Relu function, W1 is the weight of the Relu function; σ is the Sigmoid function, W2 is the weight of the Sigmoid function; the role of the Sigmoid function is to map the weight coefficient to between 0 and 1;

[0020] S33: Allocation of channel domain weight coefficients, using the weight coefficients s learned in step S20 c , multiplied by the channel domain features of the pooling layer, and different weight coefficients are assigned to the pooling channel domain features; the distribution expression of the weight coefficient is:

[0021] u c ′=s c *z c

[0022] Among them, u c ′ is the channel domain of the updated pooling layer.

[0023] Furthermore, during the prediction phase, the test data set is imported into the trained variable pooling-multiscale convolutional neural network, and the predicted remaining service life curve is compared with the actual bearing life curve;

[0024] The real bearing life curve is reflected by the reliability of the bearing, and the reliability factor is then constructed as the remaining service life curve; the reliability factor formula is as follows:

[0025]

[0026] Where R(i) is the reliability factor of the i-th sampling point, T i is the running time of the bearing at the i-th sampling point, T1 is the running time of the bearing at the initial fault point, T e is the total running time of the bearing.

[0027] Furthermore, in step S20, the data label of the bearing device before the initial fault point is 0; the data label of the bearing device after the initial fault point is 1, and the above data is used as label data of the training data set and the test data.

[0028] Furthermore, in step S20, the label-smoothing regularization technique is applied in the model training stage to add a "noise" to the real life label; that is, a small part of ε=0.1 is taken out from the data with the bearing label of 1, and then assigned to the category with the label of 0 to reduce the overfitting of the data.

[0029] Furthermore, the application of the Lable-smoothing regularization technology in the model training stage includes:

[0030] Convert hard labels into soft labels to make network optimization smoother; the original hard labels are

[0031]

[0032] Among them, the initial fault point is the sample point where the bearing detects the fault for the first time; after using the LabelSmoothing regularization technology, the updated label vector is as follows:

[0033]

[0034] Where K is the total number of categories, and K = 2, α is a small hyperparameter, generally 0.2.

[0035] The present invention proposes a method for predicting the life of bearings, combining VPMCNN and Transformer, which not only retains the ability of VPMCNN to capture global and detailed features, but also learns rich representations of the sequence before and after the input signal through the self-attention mechanism of Transformer, and can better predict the remaining life of the bearing through the input rolling bearing acoustic emission signal; it solves the disadvantage of Transformer's inability to capture detailed features, and at the same time, the self-attention mechanism in the Transformer model can capture long-distance dependencies in the input signal, effectively overcoming the memory degradation problem of the prediction sequence. In order to reduce the model's overconfidence in the correct label, avoid overfitting, and improve the generalization ability of the model, this patent embeds the Label-Smoothing regularization technology when labeling the training data to smooth the labels. The specific effects are:

[0036] 1. In terms of feature extraction: a variable pooling-multi-scale convolutional neural network is proposed. Compared with the ordinary one-dimensional convolutional neural network, the variable pooling layer can give higher weights to the features that are more meaningful for life prediction, thereby better extracting features that are beneficial to bearing prediction and achieving better prediction results.

[0037] 2. The present invention uses the Transformer model to make improvements. First, the acoustic emission signal features extracted by VPMCNN are sent to the Transformer as input, and the multi-head attention mechanism is used to capture the dependencies in the acoustic emission signal sequence, which to a certain extent avoids the speed of calculating the quadratic increase when the sequence length increases. Then the prediction output is performed through the fully connected layer and Softmax. During the training process, the batch size is selected as 128, and the optimizer selects the Adam optimizer.

[0038] 3. To improve the generalization ability of model prediction and adapt to the noise in the data, when labeling the life span of training samples, the Label-Smoothing regularization technology is embedded to convert the "hard" label, i.e. one-hot encoded label, into the "soft" label, i.e. soft label, during training. Label Smoothing softens the label to make the model smoother during prediction and not overly dependent on a certain feature or label. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required to be used in the embodiments are briefly introduced below.

[0040] Figure 1 A flow chart of the MSCNN method in the prior art provided by the present invention;

[0041] Figure 2 A method flow chart of a bearing life prediction method based on a variable pooling-multi-scale convolutional neural network combined with Transformer provided by the present invention;

[0042] Figure 3 A structural principle diagram of the convolutional pooling layer in a bearing life prediction method based on a variable pooling-multi-scale convolutional neural network combined with Transformer provided by the present invention;

[0043] Figure 4 A schematic diagram of the principle of Transformer in a bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer provided by the present invention;

[0044] Figure 5 A schematic diagram of the principle of the reliability factor in a bearing life prediction method based on a variable pooling-multi-scale convolutional neural network combined with Transformer provided by the present invention;

[0045] Figure 6 Provide a schematic diagram of the LSTM structure in Example 2 of the present invention;

[0046] Figure 7 The VMD-AE flowchart of Example 2 of the bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer provided by the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0049] Embodiment 1

[0050] like Figure 2 As shown in the figure, a bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer specifically includes the following steps:

[0051] S10. Acoustic emission signal x of the original bearing * ={x1,x2,…,x n} to collect;

[0052] Where x* represents the normalized data;

[0053] S20. Perform data enhancement on the original bearing acoustic emission signal, expand the data samples, and divide them into training data set, validation data set, and test data set; apply label-smoothing regularization technology in the model training stage to add a "noise" to the real life label. That is, take a small part of ε=0.1 from the data after the initial failure point of the bearing (labeled as 1), and then assign it to the category labeled as 0 (that is, the data before the initial failure point) to reduce data overfitting. In the test data set, we directly use the output of the model for the test data set as the prediction result.

[0054] S30. Set the training epochs to N, and establish a variable pooling-multiscale convolutional neural network that is sequentially provided with a convolutional pooling layer, a multi-scale layer, a Transformer model, a VP layer, a GAP layer, a softmax layer, and a Lable-Smoothing model;

[0055] S40. Input the features of the multi-scale layer output in the variable pooling-multi-scale convolutional neural network into the Transformer model; after encoding and decoding, select the mean square error as the loss function; and output through the fully connected layer and softmax prediction;

[0056] In the multi-scale structure, in the same layer, there are convolution kernels of different sizes, 1*1, 3*1, and 5*1, which are used to extract the features of bearing signals at different scales. Finally, the extracted features of different scales are fused to achieve the extraction of multi-scale features. When using convolution and pooling operations, padding is used to ensure that the output features are of the same size. The advantages of multi-scale blocks are as follows:

[0057] (1) Increased the depth and width of the network;

[0058] (2) Multi-scale reduces the parameters of the training network;

[0059] (3) Use convolution kernels of different sizes, 1*1, 3*1, and 5*1, to extract the multi-scale features of the original acoustic emission signal.

[0060] Finally, the multi-scale module extracts multiple features that can capture complementary and rich prediction information at different scales. This greatly improves the feature learning ability and thus improves the prediction performance.

[0061] S50. Embed Lable-Smoothing to perform label smoothing.

[0062] Combining VPMCNN and Transformer not only retains the ability of VPMCNN to capture global and detailed features, but also Transformer learns rich representations of the sequence before and after the input signal through the self-attention mechanism, which can better predict the remaining life of the bearing through the input rolling bearing acoustic emission signal; it solves the disadvantage of Transformer's inability to capture detailed features, and at the same time, the self-attention mechanism in the Transformer model can capture long-distance dependencies in the input signal, effectively overcoming the memory degradation problem of the prediction sequence. In order to reduce the model's overconfidence in the correct label, avoid overfitting, and improve the generalization ability of the model, this paper embeds Lable-Smoothing after the VPMCNN-Transformer model to smooth the labels. The specific effects are as follows:

[0063] In terms of feature extraction:

[0064] 1. A multiscale layer is introduced into the traditional 1D-CNN to extract local and full-view features;

[0065] 2. A variable pooling layer is proposed, which can give higher weights to features that are more meaningful for life prediction, thereby achieving better prediction results.

[0066] The present invention uses the Transformer model to make improvements. The acoustic emission signal features extracted by VPMCNN are first sent to the Transformer as input, which avoids the speed of calculating the quadratic increase when the sequence length increases to a certain extent. In addition, the Label-Smoothing algorithm is embedded in the training data to convert the "hard" label into a "soft" label during training.

[0067] It should be noted that the Transformer model is mainly composed of two parts: the encoder and the decoder. The encoder takes the features extracted by the variable pooling-multi-scale CNN as input and passes it into the self-attention layer in the Transformer, then enters the feedforward neural network, and then outputs it to the next encoder. The encoder starts processing the input sequence. The output of the top encoder is then transformed into a set of attention vectors Key and Value. The attention vector will be used in the "encoder-decoder attention" layer of each decoder. If the acoustic emission data is directly input into the Transformer model without processing, the large number of samples will cause the model parameters to surge accordingly, and network training will be very difficult. This patent uses variable pooling-multi-scale CNN to first extract the time series feature vectors in the long-term degradation sequence of rolling bearing acoustic emission to reduce the dimension of the input data.

[0068] Inputting the features output by VPMCNN into the Transformer model can help find the relationship between time series in the acoustic emission signals of rolling bearings better and faster, and extract information related to life prediction. In addition, during model training, Label-Smoothing is applied to the labeling of training data sets to smooth the output labels, so that the extreme values ​​min and max at both ends are moved to the middle, increasing generalization ability.

[0069] Furthermore, if Figure 3 As shown, in the multi-scale layer in step S10, a convolution kernel 1×1 is added before the convolution kernel 5×1, and a 3×1 maximum pooling is set to extract features from the data.

[0070] The convolution kernel 5×1 will increase the computational complexity of the model, so a convolution kernel 1×1 is added before the convolution kernel 5×1 to reduce the parameters of the model, and a 3×1 maximum pooling is introduced to further extract features from the data.

[0071] Furthermore, the steps of establishing the variable pooling layer in step S10 are as follows:

[0072] Step 1: Pooling layer channel domain feature combination

[0073] The features generated by the convolutional layer are combined and converted into a feature vector. The feature combination process is described as shown in Formula 1:

[0074]

[0075] In formula (1), H is the length of the pooling layer, u c represents the jth neuron in the cth pooling channel field of the pooling layer, z c Represents the average feature information of the c-th pooled channel domain.

[0076] Step 2: Calculate the weight coefficient of the channel domain feature of the pooling layer

[0077] The calculation process of the weight coefficient of the channel domain feature is shown in formula (2):

[0078] S c =σ(W2δ(W1z c )) (2)

[0079] In formula (2), δ is the Relu function, W1 is the weight of the Relu function, σ is the Sigmoid function, and W2 is the weight of the Sigmoid function. The role of the Sigmoid function is to map the weight coefficient to between 0 and 1.

[0080] Step 3: Assignment of channel domain weight coefficients

[0081] Use the weight coefficient s learned in step 2 c , multiplied by the channel domain features of the pooling layer, and different weight coefficients are assigned to the pooled channel domain features. The distribution of weight coefficients is shown in formula (3).

[0082] u c ′=s c *z c (3)

[0083] In formula (3), u c ′ is the channel domain of the updated pooling layer.

[0084] Furthermore, during the prediction stage, the test data set is imported into the trained variable pooling-multiscale convolutional neural network, and the predicted RUL (remaining useful life) curve is compared with the actual bearing life curve.

[0085] like Figure 5 As shown, the current real bearing life curve is generally reflected by the reliability of the bearing, and then the reliability factor is constructed as RUL. The reliability factor formula and graphical representation are as follows:

[0086]

[0087] Where R(i) is the reliability factor of the i-th sampling point, T i is the running time of the bearing at the i-th sampling point, T1 is the running time of the bearing at the initial fault point, T e is the total running time of the bearing.

[0088] Before the acoustic emission signal is sent to the Transformer model for life prediction, it is firstly extracted in VPMCNN to give more weight to the variables that are meaningful for prediction. In addition, the multi-scale layer can extract both local and global information of the data, which is conducive to life prediction. Secondly, in order to increase the generalization ability of the prediction model, the Label-Smoothing algorithm is embedded after softmax when training the model to smooth the labels and increase the generalization ability of the model. The specific steps are as follows:

[0089] (1) Offline modeling stage:

[0090] Step 1: Original bearing AE signal x * ={x1,x2,…,x n}, where x* represents the normalized data.

[0091] Step 2: Perform data enhancement on the original signal, expand the data samples, and divide them into training data set and test data set.

[0092] Step 3: To increase the generalization ability of the model, the Label-Smoothing regularization technology is embedded when labeling the training data set, which can prevent the model from being overconfident about the correct labels and alleviate overfitting.

[0093] Step 4: Set the number of training iterations to N, the loss of the training variable pooling-multi-scale convolutional neural network model to M, build the model, and perform feature extraction.

[0094] Step 5: Input the data processed by the variable pooling-multi-scale convolutional neural network model into the Transformer model, encode it first, then decode it, and select the root mean squared error (RMSE) as the loss function. Then pass through the fully connected layer and softmax to predict the output.

[0095] (2) Online prediction stage:

[0096] Step 1: Import the test data set into the trained variable pooling multi-scale-Transformer prediction model, and compare the predicted RUL curve with the true bearing life curve. The true bearing life curve is the reliability factor in the figure above. In this patent, the evaluation indicator root mean square error (RMSE) is used to evaluate the quality of the prediction results. The smaller the evaluation indicator value, the better the prediction effect. The root mean square error is the square root of the mean of the sum of squared errors. In order to reduce the error, the model will inevitably find a way to prioritize the samples with the largest deviation as close to the true value as possible. In other words, the greater the deviation of the sample, the greater the impact on the model. If the sample is noise, then this will have a significant negative impact on the accuracy of the model.

[0097]

[0098] Where the sampling time is N, i represents the i-th sampling point, RUL(i) pre represents the predicted RUL of the i-th sampling point, and RUL(i) real Represents the real RUL.

[0099] In order to better illustrate the algorithm adopted by the present invention, a brief supplementary explanation is given as follows:

[0100] like Figure 4 As shown, the Transformer model is a pre-trained model launched by GOOGLE in 2021. Some people have improved it and used it in life prediction and achieved good results. However, the computational complexity of the Transformer model is relatively large, which is proportional to the square of the length. Therefore, corresponding processing can be performed before the data enters the Transformer. In the present invention, the feature extraction of VPMCNN is adopted, which can retain the multi-scale features of the data. At the same time, the proposed VP layer can assign different weights to channel variables, making the prediction results more accurate. The model is embedded in Lable-Smoothing, which increases the generalization performance of the model.

[0101] Label-Smoothing is introduced as follows:

[0102] Label smoothing is a regularization method in the field of machine learning. It is usually used to label training data in diagnosis and prediction. The purpose is to prevent the model from overconfidently predicting labels during training and improve the problem of poor generalization ability. It converts hard labels into soft labels to make network optimization smoother. The original hard label is

[0103]

[0104] The initial fault point is the sample point where the bearing is first detected to have a fault. After using the LabelSmoothing regularization technique, the updated label vector is as follows:

[0105]

[0106] Where K is the total number of classification categories. In this patent, K=2, and α is a smaller hyperparameter (generally 0.2).

[0107] Embodiment 2

[0108] The method of this application analyzes and studies bearing fatigue acoustic emission signals. To further illustrate the effectiveness of this method, the proposed method is compared with the method (VMD-AE) that extracts degradation features through traditional VMD denoising and then puts them into the autoencoder network (AE). The algorithm proposed in this patent is used to extract features from the original signals collected from the bearings in the data set, and the experimental results are better than those of VMD, EMD, EEMD and EWT. It avoids the difficulties of large data dimensions and high computational complexity.

[0109] The experiment groups the bearing data with similar failure modes into one group (similar failure modes refer to the similar degradation trends of bearing data, and such data sets are mostly caused by the same type of faults). Data with different failure modes are grouped into different groups (different failure modes refer to the similarity of the degradation trends of bearing data is very low or even completely different, and such data sets are mostly caused by different types and degrees of faults). The VMD-AE flow chart is shown in Figure 7:

[0110] The method of this application is also compared with LSTM prediction. The LSTM structure is as follows Figure 6 As shown in Figure 2, the prediction effect is better than that of LSTM, and the root mean square error of the proposed method is much smaller than the corresponding value of LSTM.

[0111] The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. A bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer, characterized in that: The method comprises the following steps: S10. Acoustic emission signal x of the original bearing * ={x1,x2,…,x n } to collect; Where x* represents the normalized data; S20. Perform data enhancement on the original bearing acoustic emission signal, expand the data sample, and divide it into a training data set and a test data set; S30.,establish a neural network that is sequentially provided with a convolutional layer, a variable pooling layer, a multi-scale layer, and a Transformer model, and outputs the prediction result of the one-dimensional sequence through a fully connected layer and a Softmax function; S40. Input the features of the multi-scale layer output in the variable pooling-multi-scale convolutional neural network into the Transformer model; after encoding and decoding, select the mean square error as the loss function; after the fully connected layer and softmax prediction output, output the predicted value.

2. According to claim 1, the bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer is characterized in that: In the multi-scale layer in step S30, a convolution kernel 1×1 is added before the convolution kernel 5×1, and a 3×1 maximum pooling is set to perform feature extraction on the data.

3. The bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer according to claim 1 is characterized in that: The steps of establishing the variable pooling layer in step S30 are as follows: The variable pooling layer changes the weight of the bearing signal features by assigning different weight coefficients to the channel domain features of the pooling layer; the variable pooling layer calculates the weight coefficient of the pooling layer channel, and then combines the weight coefficient with the features of the pooling layer channel; by assigning weight coefficients to the pooling layer channel, it can enhance the feature information with great predictive effect and weaken the feature information with small predictive effect; S31. The features generated by the convolutional layer are combined and converted into a feature vector, which is expressed as: Where H is the length of the pooling layer, u c represents the jth neuron in the cth pooling channel field of the pooling layer, z c Represents the average feature information of the c-th pooled channel domain; S32. Calculate the weight coefficient of the channel domain feature of the pooling layer, the expression is: S c =σ(W2δ(W1z c )) In the formula, δ is the Relu function, W1 is the weight of the Relu function; σ is the Sigmoid function, W2 is the weight of the Sigmoid function; the role of the Sigmoid function is to map the weight coefficient to between 0 and 1; S33: Allocation of channel domain weight coefficients, using the weight coefficients s learned in step S20 c , multiplied by the channel domain features of the pooling layer, and different weight coefficients are assigned to the pooling channel domain features; the distribution expression of the weight coefficient is: u c ′=s c *z c Among them, u c ′ is the channel domain of the updated pooling layer.

4. The bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer according to claim 1 is characterized in that: During the prediction phase, the test data set is imported into the trained variable pooling-multi-scale convolutional neural network, and the predicted remaining service life curve is compared with the actual bearing life curve; The real bearing life curve is reflected by the reliability of the bearing, and the reliability factor is then constructed as the remaining service life curve; the reliability factor formula is as follows: Where R(i) is the reliability factor of the i-th sampling point, T i is the running time of the bearing at the i-th sampling point, T1 is the running time of the bearing at the initial fault point, T e is the total running time of the bearing.

5. The bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer according to claim 1 is characterized in that: In the step S20, the data label of the bearing device before the initial fault point is 0; the data label of the bearing device after the initial fault point is 1, and the above data is used as the label data of the training data set and the test data.

6. The bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer according to claim 5 is characterized in that: In step S20, the label-smoothing regularization technology is applied in the model training stage to add a "noise" to the real life label; that is, a small part of ε=0.1 is taken out from the data with the bearing label of 1, and then assigned to the category with the label of 0 to reduce the overfitting of the data.

7. The bearing life prediction method based on variable pooling-multi-scale convolutional neural network combined with Transformer according to claim 6 is characterized in that: The label-smoothing regularization technique applied in the model training stage includes: Convert hard labels into soft labels to make network optimization smoother; the original hard labels are Among them, the initial fault point is the sample point where the bearing detects the fault for the first time; after using the Label Smoothing regularization technology, the updated label vector is as follows: Where K is the total number of categories, and K = 2, α is a small hyperparameter, generally 0.2.

Images