Deep learning bearing life prediction method based on CBAM attention

Abstract

The invention relates to a bearing life prediction method based on deep learning of CBAM attention. The method comprises the steps of collecting a to-be-predicted bearing vibration signal; inputting the to-be-predicted bearing vibration signal into a preset bearing life prediction model, and predicting the service life of the bearing; wherein the bearing life prediction model is a CBAM attention mechanism-based parallel time convolutional network and long and short term memory network model; the training set comprises the bearing vibration signals of the whole life cycle of the normal operation stage and the fault stage of the bearing. According to the technical scheme, richer features can be extracted, and the service life of the bearing can be predicted more accurately and reliably.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing life prediction, and particularly to a bearing life prediction method based on deep learning with CBAM attention. Background Art

[0002] Bearings are key components of rotating machinery and are constantly subjected to cyclic loads and friction, making them vulnerable to damage. Bearing failures account for 45% - 55% of all rotating machinery failure cases. Therefore, predicting the remaining useful life (RUL) of bearings has become a prominent research hotspot.

[0003] In recent years, deep learning technology has made breakthrough progress in artificial intelligence fields such as image recognition and natural language processing. Due to its capabilities of end-to-end learning and automatic feature learning, deep learning technology has been widely applied to many other fields. In the URL field, deep learning technology also has a large number of applications, mainly concentrated in four types of deep learning structures: Auto-encoder (AE), Deep Belief Network (DBN), Convolutional Neural Network (CNN), and Recurrent Neural Network (RNN). AE and DBN can improve the data expression ability and representation ability, and are generally used in the data preprocessing or feature extraction part. Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are the most commonly used deep learning methods. CNN is used more in the field of fault diagnosis due to its powerful feature extraction ability and classification ability. RNN has received more attention and achieved more results in the research of RUL prediction due to its adaptability to time series data. However, RNN has certain limitations in dealing with long-term dependencies, which may lead to problems such as gradient disappearance or gradient explosion. Existing technologies use Long short-term memory (LSTM) to solve this problem. Although LSTM can capture the dependencies in time series data, their complex chain structure leads to an extended training time. Bai et al. introduced the Temporal Convolutional Network (TCN) by combining causal convolution and dilated convolution. TCN uses a dilated causal convolutional network structure, which can parallelly process and analyze local features in the sequence, thereby accelerating the speed of model training and inference. Hu et al. observed that TCN has a lower prediction accuracy on short time scales, while it has a higher prediction accuracy on long time scales, and the opposite is true for LSTM. The dilated causal convolutional structure enables TCN to have good feature extraction ability, which can fuse the original features to obtain high-dimensional abstract features, enhance the mining of feature information, and extract long-term time relationships and higher-level spatial features from historical data. LSTM has advantages in nonlinear fitting and time series prediction and can capture correlations from time series data.

[0004] In summary, TCN has powerful and efficient feature extraction ability, and LSTM has advantages in nonlinear fitting and time series prediction. Therefore, the present invention combines the advantages of TCN and LSTM to propose a bearing remaining life prediction method based on a parallel TCN network and LSTM network with CBAM attention. Summary of the Invention

[0005] The object of the present invention is to solve the problems existing in the above-mentioned prior art. Compared with ordinary deep learning prediction methods, the technical solution of the present invention can extract richer features and can predict the bearing life more accurately and reliably.

[0006] To achieve the above object, a bearing life prediction method based on deep learning with CBAM attention is proposed, and the present invention provides the following solutions:

[0007] A bearing life prediction method based on deep learning with CBAM attention, comprising:

[0008] Collect the bearing vibration signals to be predicted;

[0009] Input the bearing vibration signals to be predicted into a preset bearing life prediction model to predict the bearing life; wherein, the bearing life prediction model is a parallel time convolutional network and long short-term memory network model based on the CBAM attention mechanism; and is obtained by training through a constructed training set, and the training set includes bearing vibration signals in the full life cycle of the normal operation stage and the fault stage of the bearing.

[0010] Optionally, constructing the training set includes:

[0011] Collect the bearing vibration signals in the full life cycle in both the horizontal and vertical directions of the bearing simultaneously;

[0012] Fuse the horizontal vibration signals and vertical vibration signals of the bearing and perform normalization processing to construct the training set.

[0013] Optionally, the bearing life prediction model is constructed using a long short-term memory network module, a time convolutional network module, and a CBAM attention mechanism.

[0014] Optionally, the bearing life prediction model predicting the bearing life includes:

[0015] Perform preliminary feature extraction on the input bearing vibration signals using one-dimensional convolution;

[0016] Respectively use the long short-term memory network module and the time convolutional network module to perform deep feature extraction on the preliminarily extracted features and fuse the extracted deep features;

[0017] Use the CBAM attention mechanism to perform weighted processing on the fused features in different dimensions, enhance the weight distribution of preset important features, and enter the fully connected layer to make a result prediction.

[0018] Optionally, the time convolutional network module is constructed in a way of stacking 5 layers of time convolutional networks and two one-dimensional convolutions;

[0019] The method for the temporal convolutional network module to perform deep feature extraction on the preliminarily extracted features is as follows:

[0020]

[0021] Among them, is the network output corresponding to the input ; is the filtering operation on the th input; is the filter window size; ; is the dilation coefficient.

[0022] Optionally, the long short-term memory network module is constructed by stacking two layers of long short-term memory networks and two one-dimensional convolutions;

[0023] The method for the long short-term memory network module to perform deep feature extraction on the preliminarily extracted features is as follows:

[0024]

[0025] Among them, represents the forget gate, represents the input gate, represents the output gate, represents the sigmoid activation function, represents the hidden layer information at time t-1, represents the information input at time t, represents the offset of the three gates.

[0026] Optionally, the CBAM attention mechanism includes: a channel attention part and a spatial attention part;

[0027] The operation of the channel attention part is expressed as:

[0028]

[0029] = {

[0030]

[0031] Among them, represents the channel attention feature map; represents the Sigmoid function; represents the multi-layer perceptron; represents average pooling; represents max pooling; Represents the input features of the channel attention part; Feature; Feature; Represents the element-wise multiplication operation; Represents the output features of the channel attention.

[0032] The operations of the spatial attention part are represented as:

[0033]

[0034]

[0035]

[0036] Among them, Represents the spatial attention feature map; Represents the Sigmoid function; Represents the multi-layer perceptron; Represents average pooling; Represents max pooling; Represents the output features of the channel attention; Feature; Feature; Represents the element-wise multiplication operation; Represents the output features of the channel attention; Represents a convolution operation with a convolution kernel of The convolution operation.

[0037] The beneficial effects of the present invention are as follows:

[0038] The present invention provides a method for predicting the remaining life of a bearing based on a parallel TCN network and an LSTM network with CBAM attention, including: collecting the vibration signal of the bearing to be predicted; inputting the vibration signal of the bearing to be predicted into a preset bearing life prediction model to predict the bearing life; wherein, the bearing life prediction model is constructed by using a long short-term memory network module, a temporal convolutional network module and a CBAM attention mechanism, and is pre-trained through a constructed data set, and the data set includes the vibration signals of the bearing in the full life cycle of the normal operation stage and the fault stage of the bearing. Compared with the ordinary deep learning prediction method, the technical solution of the present invention can extract richer features and can predict the bearing life more accurately and reliably. Description of the Drawings

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required for use in the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other accompanying drawings can also be obtained based on these drawings.

[0040] Figure 1 Schematic diagram of a bearing life prediction method based on CBAM attention in deep learning according to an embodiment of the present invention;

[0041] Figure 2 Schematic flow chart of a bearing remaining life prediction method based on a parallel TCN network and an LSTM network with CBAM attention according to an embodiment of the present invention;

[0042] Figure 3 Schematic diagram of the dilated convolution of TCN according to an embodiment of the present invention;

[0043] Figure 4 Schematic diagram of LSTM according to an embodiment of the present invention;

[0044] Figure 5 Schematic diagram of CBAM according to an embodiment of the present invention;

[0045] Figure 6 General network structure diagram of a bearing remaining life prediction method based on a parallel TCN network and an LSTM network with CBAM attention according to an embodiment of the present invention;

[0046] Figure 7 Bearing 1_4 life prediction diagram according to an embodiment of the present invention;

[0047] Figure 8 Bearing 2_4 life prediction diagram according to an embodiment of the present invention;

[0048] Figure 9 Comparison of MAE values of experimental bearings in different methods according to an embodiment of the present invention;

[0049] Figure 10 Comparison of average MAE values of experimental bearings in different methods according to an embodiment of the present invention;

[0050] Figure 11 Comparison of RMSE values of experimental bearings in different methods according to an embodiment of the present invention;

[0051] Figure 12 Comparison of average RMSE values of experimental bearings in different methods according to an embodiment of the present invention. Detailed implementation manners

[0052] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0053] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0054] As Figure 1 - Figure 2 shown, this embodiment proposes a bearing life prediction method based on deep learning with CBAM attention, including:

[0055] Collect the bearing vibration signals to be predicted;

[0056] Input the bearing vibration signals to be predicted into a preset bearing life prediction model to predict the bearing life; wherein, the bearing life prediction model is a parallel time convolutional network and long short-term memory network model based on the CBAM attention mechanism; and it is obtained by training through a constructed training set, and the training set includes the bearing vibration signals in the full life cycle of the normal operation stage and the fault stage of the bearing.

[0057] Specifically, in this embodiment, S1~Obtain the bearing vibration signals in the full life cycle, perform data preprocessing on the signals, fuse the horizontal vibration signals and vertical vibration signals of the bearing, and perform normalization processing;

[0058] S2~Divide the processed data into two parts of data, a training data set and a test data set. The training data set is used for model training, and the test data set is used for life prediction;

[0059] S3~Construct a parallel TCN network and LSTM network model with CBAM attention, use the training data set to train the model, and input the training data set into the trained model for bearing life prediction.

[0060] Furthermore, constructing the data set includes:

[0061] Collect the bearing vibration signals in the full life cycle in both the horizontal and vertical directions of the bearing at the same time;

[0062] Fuse the horizontal vibration signals and vertical vibration signals of the bearing, and perform normalization processing to construct the training set.

[0063] Specifically, in this embodiment, S101 - Build a bearing experiment platform, install sensors in the horizontal and vertical directions, and simultaneously collect the bearing vibration signals in the horizontal and vertical directions throughout the entire life cycle of the bearing;

[0064] S102 - Fusion of the horizontal vibration signal and the vertical vibration signal of the bearing, and normalization processing to make each group of values approach 1.

[0065] The normalization operation is expressed as:

[0066]

[0067] is the i-th data point in the original data set, is the mean of the original data, is the standard deviation of the original data.

[0068] Furthermore, the bearing life prediction model is a parallel time convolutional network and long short-term memory network model based on the CBAM attention mechanism;

[0069] Inputting the bearing vibration signal to be predicted into the preset bearing life prediction model includes:

[0070] Using one-dimensional convolution to perform preliminary feature extraction on the input bearing vibration signal;

[0071] Respectively using the long short-term memory network module and the time convolutional network module to perform deep feature extraction on the preliminarily extracted features, and fusing the extracted deep features;

[0072] Using the CBAM attention mechanism to perform weighted processing on the fused features in different dimensions, enhancing the weight distribution of the preset important features, and entering the fully connected layer to make a result prediction.

[0073] Specifically, in this embodiment, S201 - Select one bearing signal as the test set data, and the remaining bearing signals as the training set data.

[0074] S301 - Establish a parallel TCN network and LSTM network model with CBAM attention, and the model includes a long short-term memory network module, a time convolutional network module, and a CBAM attention mechanism;

[0075] The time convolutional network module includes:

[0076] Adopt a method of stacking 5 layers of time convolutional networks and two one-dimensional convolutions. The number of convolutional kernels of the 5 layers of time convolutional networks are 12, 10, 8, 6, 4 respectively, the dilation factors are 1, 2, 4, 8, 16 respectively, the kernel width is 3, and the number of convolutional kernels of the two one-dimensional convolutions are both 16, and the kernel width is 12.

[0077] The TCN can effectively retain time-related information by using dilated convolutions. The dilated convolution operation is expressed as:

[0078]

[0079] In the formula: is the input The corresponding network output; is the filtering operation on the th input; is the filter window size; ; is the dilation coefficient. The TCN dilated convolution is as Figure 3 shown.

[0080] The long short-term memory network module includes:

[0081] Adopt a method of stacking two layers of long short-term memory networks and two one-dimensional convolutions. The number of neurons in both layers of long short-term memory networks is 4, and the number of convolution kernels in both one-dimensional convolutions is 16, with a kernel width of 12.

[0082] The core concept of LSTM lies in the cell state and the "gate" structure. The cell state is equivalent to the path of information transmission, allowing information to be passed on in the sequence chain, including the forget gate, input gate, and output gate.

[0083] The forget gate operation is expressed as:

[0084]

[0085] In the formula, represents the forget gate, represents the sigmoid activation function, represents the weight vector from the input layer to the forget gate, represents the hidden layer information at time t-1, refers to the information input at time t, represents the offset of the forget gate.

[0086] The input gate operation is expressed as:

[0087]

[0088] In the formula, represents the input gate, represents the sigmoid activation function, represents the weight vector from the input layer to the input gate, represents the hidden layer information at time t-1, refers to the information input at time t, represents the offset of the input gate;

[0089]

[0090] In the formula, represents the candidate cell state at time t, and Tanh is the hyperbolic tangent activation function, represents the weight vector from the input layer to the cell state, represents the hidden layer information at time t-1, refers to the information input at time t, represents the offset of the cell state;

[0091]

[0092] In the formula, represents the new cell state, represents the forget gate, represents the cell state information at time t-1, represents the input gate, represents the candidate cell state at time t.

[0093] The output gate operation is expressed as:

[0094]

[0095]

[0096] In the formula, represents the output gate, represents the sigmoid activation function, represents the weight vector from the input layer to the output gate, represents the hidden layer information at time t-1, represents the hidden layer information at time t, refers to the information input at time t, refers to the offset of the output gate, and Tanh is the hyperbolic tangent activation function, represents the new cell state. The LSTM schematic diagram is as Figure 4 shown.

[0097] CBAM attention includes a channel attention part and a spatial attention part. The CBAM schematic diagram is as Figure 5 shown.

[0098] The operation of the channel attention part is expressed as:

[0099]

[0100] = {

[0101]

[0102] In the formula, represents the channel attention feature map; represents the Sigmoid function; represents the multi-layer perceptron; represents average pooling; represents max pooling; represents the input feature of the channel attention part; feature; feature; represents the element-wise multiplication operation; represents the output feature of the channel attention.

[0103] The operations of the spatial attention part are expressed as:

[0104]

[0105]

[0106]

[0107] In the formula, represents the spatial attention feature map; represents the Sigmoid function; represents the multi-layer perceptron; represents average pooling; represents max pooling; represents the output feature of the channel attention; feature; feature; represents the element-wise multiplication operation; represents the output feature of the channel attention; represents the convolution operation with a convolution kernel of .

[0108] S302 - First, perform preliminary feature extraction on the input signal using one-dimensional convolution;

[0109] S303 - Then, use the long short-term memory network module and the temporal convolutional network module respectively to perform deep feature extraction on the preliminarily processed signal and fuse them;

[0110] S304 - Next, use the CBAM attention mechanism to perform weighted processing on the features in different dimensions, assign more weights to important features, and enter the fully connected layer to make result predictions;

[0111] S305 - Finally, use the training data set to train the model and save the parameters, and input the training data set into the trained model to predict the bearing life.

[0112] Further, using the dataset to pre-train the bearing life prediction model includes:

[0113] Taking one bearing signal in the dataset as the test dataset, and the remaining bearing signals as the training dataset, and each bearing signal takes turns as the test dataset.

[0114] This embodiment also proposes a bearing life prediction system based on deep learning with CBAM attention, including:

[0115] A data preprocessing unit, which is used to preprocess the collected bearing vibration data, fuse the data, and perform normalization processing. The collected bearing vibration data includes data in the normal operation stage and the fault stage;

[0116] A model training unit, which trains the model with the divided training set data, and through multiple trainings, saves the parameters of the model when the training effect is good;

[0117] A life prediction unit, which tests the model with the divided test set data and verifies the test results.

[0118] More specifically, the following introduces the example application of this embodiment;

[0119] According to steps S1~, obtain the bearing vibration signals in the whole life cycle, preprocess the signals, fuse the horizontal vibration signal and the vertical vibration signal of the bearing, and perform normalization processing;

[0120] The dataset used is the IEEE PHM 2012 Challenge dataset, and the bearing data is collected from the PRONOSTIA experimental platform. The PRONOSTIA experimental platform is specifically used to test and verify bearing fault detection, diagnosis and prediction algorithms. Three working conditions are set respectively, and the whole life vibration data of 17 bearings are collected by using the platform. The sampling rate is 25.6 kHz, and the recording time is 0.1 second, that is, 2560 data points are collected each time. For safety reasons, when the amplitude of the vibration data exceeds 20 g, the experiment is stopped. Preprocess the signals, fuse the horizontal vibration signal and the vertical vibration signal of the bearing, and perform normalization processing;

[0121] According to steps S2~, divide the processed data into two parts of data, the training dataset and the test dataset. The training dataset is used for model training, and the test dataset is used for life prediction; 7 experiments need to be carried out for each working condition. The specific method is to select one bearing as the test dataset, and the other six bearings' labeled data as the training dataset, and each bearing takes turns as the test dataset.

[0122] Construct a parallel TCN network and an LSTM network model with CBAM attention according to steps S3~. The network structure of the bearing remaining life prediction method based on the parallel TCN network and LSTM network with CBAM attention is as Figure 6 shown. Use the training data set to train the model, and input the training data set into the trained model for bearing life prediction.

[0123] In order to intuitively reflect the prediction effect, this embodiment uses two metrics, mean absolute error (MAE) and root mean square error (RMSE), to evaluate the bearing RUL prediction model. The evaluation formula is expressed as:

[0124]

[0125]

[0126] where is the remaining life at the t-th sampling point, and q is the total number of samplings;

[0127] Through experimental data, a comparative experiment is carried out. It is found through the comparative experiment that the present invention can improve the accuracy of the prediction result. Among them, the life prediction of bearing 1_4 and the life prediction of bearing 2_4 are as Figure 7 and Figure 8 shown.

[0128] The model prediction index values of several comparative experiments are shown in Table 1:

[0129] Table 1 Model prediction index values of comparative experiments

[0130]

[0131] The comparative models selected in Table 1 are LSTM, TCN, parallel LSTM and TCN, and the method of the present invention. From the comparison results, the method proposed by the present invention, the bearing remaining life prediction method based on the parallel TCN network and LSTM network with CBAM attention, has the smallest MAE and RMSE values in most cases, indicating that the method proposed by the present invention is better than other comparative models.

[0132] Figure 9 What is shown is the MAE values of 4 prediction models. Among them, the method of the present invention obtains the lowest MAE value for 12 out of 14 bearings. Figure 10 What is shown is the average MAE value of 4 prediction models. The average value of the MAE error of the method in this article is significantly lower than that of other methods. Figure 11The RMAE values are shown. Among the 14 bearings, the method of the present invention has better effects in 13 bearings than other models. Figure 12 The average RMAE value, i.e., the average value of the RMSE error, is shown. The method of the present invention also obtains the lowest value. Compared with some existing prediction methods, the bearing RUL prediction model constructed by the present invention has better feature extraction ability than some existing deep learning RUL prediction models and achieves good results in terms of the two evaluation indexes of MAE and RMAE, which proves that the model proposed by the present invention can have good performance.

[0133] The present invention provides a method for predicting the remaining life of a bearing based on a parallel TCN network and an LSTM network with CBAM attention, including: preprocessing the collected bearing vibration data, fusing and normalizing the data, and the operation state data includes data in the normal operation stage and the fault stage; dividing the operation state data into a training data set and a test data set, where the training data set is used for model training and the test data set is used for life prediction; constructing a model of a parallel TCN network and an LSTM network with CBAM attention, training the model with the training set data, saving the parameters of the model when the training effect is good, and inputting the test data set into the trained model for bearing life prediction. Compared with ordinary deep learning prediction methods, the technical solution of the present invention can extract richer features and can predict the bearing life more accurately and reliably.

[0134] The embodiments described above are only descriptions of the preferred embodiments of the present invention and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solution of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. A bearing life prediction method based on deep learning of CBAM attention, characterized in that: include: Collect bearing vibration signals to be predicted; The bearing vibration signal to be predicted is input into a preset bearing life prediction model to predict the bearing life; wherein the bearing life prediction model is a parallel time convolutional network and long short-term memory network model based on the CBAM attention mechanism; and is obtained through training with a constructed training set, wherein the training set includes the bearing vibration signals of the entire life cycle of the bearing in the normal operation stage and the failure stage.

2. The bearing life prediction method based on deep learning of CBAM attention according to claim 1 is characterized in that: Constructing the training set includes: Simultaneously collect bearing vibration signals in the horizontal and vertical directions throughout the entire life cycle of the bearing; The horizontal vibration signal and the vertical vibration signal of the bearing are fused and normalized to construct the training set.

3. The bearing life prediction method based on deep learning of CBAM attention according to claim 1 is characterized in that: The bearing life prediction model is constructed using a long short-term memory network module, a temporal convolutional network module and a CBAM attention mechanism.

4. The bearing life prediction method based on deep learning of CBAM attention according to claim 3 is characterized in that: The bearing life prediction model predicts the bearing life including: Use one-dimensional convolution to perform preliminary feature extraction on the input bearing vibration signal; The long short-term memory network module and the temporal convolutional network module are used to extract deep features from the initially extracted features, and the extracted deep features are fused; The CBAM attention mechanism is used to perform weighted processing on the fused features in different dimensions, enhance the weight distribution of preset important features, and enter the fully connected layer to make result predictions.

5. The bearing life prediction method based on deep learning of CBAM attention according to claim 4 is characterized in that: The temporal convolutional network module is constructed by stacking a 5-layer temporal convolutional network and two one-dimensional convolutions; The method for the temporal convolutional network module to perform deep feature extraction on the initially extracted features is: in, For input The corresponding network output; For the Input filtering operations; is the filter window size; ; is the expansion coefficient.

6. The bearing life prediction method based on deep learning of CBAM attention according to claim 4 is characterized in that: The long short-term memory network module is constructed by stacking two layers of long short-term memory networks and two one-dimensional convolutions; The method for the long short-term memory network module to perform deep feature extraction on the initially extracted features is: in, represents the forget gate, represents the input gate, represents the output gate, represents the sigmoid activation function, Represents the weights of the three gates, represents the hidden layer information at time t-1, represents the information input at time t, Represents the offset of the three gates.

7. The bearing life prediction method based on deep learning of CBAM attention according to claim 4 is characterized in that: The CBAM attention mechanism includes: a channel attention part and a spatial attention part; The operation of the channel attention part is expressed as: in, Represents the channel attention feature map; Represents the Sigmoid function; represents a multi-layer perceptron; represents average pooling; represents maximum pooling; Represents the input features of the channel attention part; feature; feature; Represents element-wise multiplication operation; Represents the channel attention output feature; The operation of the spatial attention part is expressed as: in, Represents the spatial attention feature map; Represents the Sigmoid function; represents a multi-layer perceptron; represents average pooling; represents maximum pooling; Represents the channel attention output feature; feature; feature; Represents element-wise multiplication operation; Represents the channel attention output feature; Represents the convolution kernel as The convolution operation.

Images