Bearing residual service life prediction method based on deep mutual learning and dynamic feature construction

Abstract

The application belongs to the field of bearing life prediction, and discloses a bearing residual useful life prediction method based on deep mutual learning and dynamic feature construction. In the application, the more stable bearing feature RRMS is selected. Firstly, the improved convolutional neural network by DML is used to automatically extract features in the first stage, which is used to indicate the health condition of the bearing. When the bearing degrades to 50% (the output is less than 0.5), the last 50% is predicted through the long short-term memory network. Finally, the results of the two stages are combined to obtain the life degradation curve of the bearing. The model output obtained by the above method can be directly used for the calculation of RUL. This method avoids the selection of the bearing failure threshold. In the whole prediction process, CNN and LSTM are respectively used for different stages of bearing degradation, so that the existing full-life data and the current bearing historical data are maximally utilized to improve the precision of the model.

Description

Technical Field

[0001] This invention belongs to the field of bearing life prediction, specifically involving a method for predicting the remaining service life of bearings based on deep mutual learning and dynamic feature construction. Background Technology

[0002] Rolling bearings, as rotating components of mechanical equipment, are widely used in various industrial fields. However, the complexity of industrial environments and long-term uninterrupted operation make bearings highly susceptible to failure. As a critical component of rotating machinery, the failure of rolling bearings can cause abnormal operation of the mechanical system. Rolling bearing RUL prediction can prevent this from happening, reduce unnecessary downtime and personnel casualties, and provide a basis for developing optimal maintenance strategies.

[0003] Mechanism-based RUL prediction methods model the degradation regulation of a single fault; this requires a large amount of prior knowledge. Furthermore, the degradation process of bearings is highly complex. Therefore, establishing an accurate physical model is quite difficult. Currently, data-driven RUL prediction methods have achieved significant progress. However, some problems remain. First, constructing a feature set that can describe the degradation trend of bearings is a rather complex process. Second, the highly subjective construction of degradation labels, lacking a unified standard, introduces considerable randomness in the setting of fault thresholds. This paper addresses the problems of complex and variable feature set construction and the highly subjective selection of degradation indicators in the current field of bearing life prediction. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting the remaining service life of bearings based on deep mutual learning and dynamic feature construction, so as to solve the problems in the prior art mentioned in the background.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] The bearing remaining service prediction method based on deep mutual learning and dynamic feature construction includes the following steps:

[0007] Step 1: Construct the experimental dataset: It includes a training set and a test set. The training set consists of bearing data for bearing 1 and bearing 3, and the test set consists of bearing data for bearing 2 and bearing 4.

[0008] Step 2: Establish DML model: Preprocess the bearing data, perform FFT on the bearing vibration signal to convert the time domain signal into the frequency domain signal, and then input the obtained FFT data into the CNN model enhanced by DML;

[0009] Step 3, RUL Calculation: The first 1024 dimensions of the frequency domain signal are used as input to the DML model, and the DML model output is compared. If the model output is greater than 0.5, the RUL is calculated directly using the DML output. Otherwise, the DML model output is input into the subsequent LSTM model, and the RUL is finally calculated using the LSTM model output.

[0010] Step 4: Obtain the bearing life degradation curve: Compare the convergence speed and accuracy of the LSTM model under different time lengths, and then combine the prediction output results of the two stages to obtain the degradation curve.

[0011] Furthermore, the data acquisition for the training and test sets includes an AC motor, bearing 1, bearing 2, bearing 3, bearing 4, and a vibration sensor; the sampling frequency is 20KHz, 20480 sampling points are collected each time, and the time interval between each sampling is 10 minutes; then the degradation start point and fault point of each bearing are determined, and training labels and test labels are established, and then the original signal is processed by FFT.

[0012] Furthermore, the bearing lifecycle is divided into four stages: the healthy stage, the slow degradation stage, the rapid degradation stage, and the failure stage after the failure point. The time corresponding to the degradation initiation point is expressed as... The time corresponding to the failure point is Determining using the threshold method and The threshold values ​​are 1.1 and 5. Different bearings have different RMS values; therefore, the more stable feature RRMS is chosen to allow the same threshold to be applied to different bearings. The expression for RRMS is: ;

[0013] in, It is the average effective value of 200 to 300 data points;

[0014] According to the exponential function To fit the bearing's RRMS, thus determining the bearing's total lifespan. As shown in the following formula:

[0015] ;

[0016] Among them, parameters , , Determined by ordinary least squares, and It can be done Sure.

[0017] Furthermore, DML is applied to the field of RUL prediction. The output of the DML CNN is used to indicate the health status of the bearing. When the bearing degrades to 50%, a second-stage prediction is performed, including:

[0018] The output of the DML-based CNN is used to indicate the health status of the bearing. The DML model is achieved through mutual learning between two CNNs. The optimization objectives of the two CNNs are as follows. Applying DML to the RUL prediction domain can prevent the model from overfitting and further improve the prediction accuracy.

[0019]

[0020] .

[0021] Furthermore, a one-dimensional CNN was used to process time series data, with the last layer using sigmoid as the activation function and the learning rate set to 0.01.

[0022] When the output of the DML model is greater than 0.5, that is, when the degradation degree is less than 50%, the output of the DML can be directly used to calculate the RUL.

[0023] When the output of the DML model is less than 0.5, that is, when the degradation degree is greater than 50%, the output of the DML model is input into the LSTM model for the second stage of prediction.

[0024] Furthermore, CNN and LSTM are applied to different stages of bearing degradation, respectively, to maximize the use of existing full-life data and current bearing historical data to improve model accuracy, including:

[0025] The output of the DML model is input into the LSTM model for the second stage of prediction.

[0026] The RUL is calculated using the output of the LSTM model. In this way, the RUL of the bearing at different life stages can be predicted.

[0027] Furthermore, LSTM employs 24 steps, ensuring both prediction accuracy and model training speed.

[0028] Furthermore, the DML model also includes model training, including prediction loss and imitation loss.

[0029] Furthermore, in the prediction loss, the optimization objective of the CNN is to minimize the difference between the true value and the predicted value, given an input vector. That is, with n samples, the feature extractor will Mapping to a d-dimensional space, then the fully connected layer maps the d-dimensional vector to 1-dimensional space, based on the output. Predicting losses Using MSE, as shown in the following formula;

[0030] .

[0031] Furthermore, in the imitation loss, the distributional differences between models can be reduced by minimizing the imitation loss, and MSE can make the model perfectly fit the training data; by adding an imitation loss to measure the distributional differences between models, and then optimizing the model by minimizing this loss, training labels should be constructed before training the model. During bearing degradation, the characteristics of time t are... ,Will Input health metrics to train the model; during training, each sample... tags It is given by the following formula:

[0032] ;

[0033] In the formula, It is the bearing's RUL at time t, which can be obtained through get, It refers to the entire lifespan of the bearing, which can be determined by... get.

[0034] Technical effects and advantages of the present invention: The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction proposed in this invention has the following advantages compared with the prior art:

[0035] This invention selects the more stable bearing feature RRMS. First, in the first stage, a convolutional neural network improved by DML is used to automatically extract features to indicate the health status of the bearing. When the bearing degrades to 50% (output less than 0.5), a long short-term memory network is used to predict the remaining 50%. Finally, the results of the two stages are combined to obtain the bearing life degradation curve. The model output obtained by the above method can be directly used for RUL calculation. This method avoids the selection of bearing failure threshold. In the whole prediction process, CNN and LSTM are used for different stages of bearing degradation, respectively, to maximize the use of existing full life data and current bearing historical data to improve the accuracy of the model. Attached Figure Description

[0036] Figure 1 This is a flowchart of the prediction method of the present invention;

[0037] Figure 2This is a schematic diagram comparing the features of RMS and RRMS of the present invention;

[0038] Figure 3 This is a schematic diagram of the fitting results for the bearing of the present invention;

[0039] Figure 4 This is a schematic diagram of the convolutional neural network structure of the present invention;

[0040] Figure 5 This is a schematic diagram of the LSTM cell structure of the present invention;

[0041] Figure 6 This is a schematic diagram of the Deep Inter-Learning Model (DML) structure of the present invention;

[0042] Figure 7 This is a schematic diagram of the bearing life cycle signal of the present invention;

[0043] Figure 8 This is a schematic diagram of the loss of the LSTM model of the present invention under different step sizes;

[0044] Figure 9 This is a schematic diagram showing the results of different health indicators in this invention;

[0045] Figure 10 This is a schematic diagram of the scatter plot of the prediction results of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The specific embodiments described herein are merely used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] This invention provides, for example Figure 1 The bearing remaining service prediction method shown is based on deep mutual learning and dynamic features. It uses DML and dynamic features to construct a two-stage RUL prediction method.

[0048] Specifically, S1, constructing an experimental dataset: which includes a training set and a test set, wherein the training set consists of bearing data of bearing 1 and bearing 3, and the test set consists of bearing data of bearing 2 and bearing 4;

[0049] S2. Establish DML model: Preprocess the bearing data, perform FFT on the bearing vibration signal to convert the time domain signal into the frequency domain signal, and then input the obtained FFT data into the CNN model enhanced by DML.

[0050] S3, RUL Calculation: The first 1024 dimensions of the frequency domain signal are used as input to the DML model, and the output of the DML model is compared. If the model output is greater than 0.5, the RUL is calculated directly using the output of the DML model. Otherwise, the output of the DML model is input into the subsequent LSTM model, and the RUL is finally calculated using the output of the LSTM model.

[0051] S4. Obtain the bearing life degradation curve: Compare the convergence speed and accuracy of the LSTM model under different time lengths, and then combine the prediction output results of the two stages to obtain the degradation curve.

[0052] During the healthy phase, the RMS values ​​of different bearings vary. Therefore, the proposed method selects the more stable bearing feature RRMS, with the input signal being a frequency domain signal; training labels proportional to RUL are used. This method divides the bearing degradation process into two stages. In the first stage, the output of a CNN improved by DML is used to indicate the current RUL. When its output indicates that the RUL is less than a certain percentage, the next stage of prediction begins. In the second stage, an LSTM network is used to predict the RUL, with the output of the previous stage as its input. Finally, comparative experiments are designed to compare the proposed method with other methods, and the comparison results demonstrate the superiority of the proposed method.

[0053] The two-stage bearing remaining life prediction method is as follows: the bearing life cycle is divided into four stages, namely the health stage, the slow degradation stage, the rapid degradation stage, and the failure stage after the failure point.

[0054] The time corresponding to the degradation initiation point (DSP) is expressed as follows: The time corresponding to the failure point is This embodiment uses the threshold method to determine... and .

[0055] During the healthy phase, the RMS values ​​of different bearings vary. Therefore, this invention selects the more stable characteristic RRMS, making it easier to apply the same threshold to different bearings. The expression for RRMS is as follows:

[0056] (1);

[0057] The proposed method consists of three modules: data preprocessing, first-stage prediction, and second-stage prediction.

[0058] First, in data preprocessing, the time-domain signal is converted into a frequency-domain signal after undergoing a Fast Fourier Transform (FFT). This preprocessing is done to facilitate data input into the network.

[0059] Secondly, Direct Learning (DML) is applied to the RUL prediction domain to prevent overfitting of the model and further improve prediction accuracy. During the DML process, Network 1 and Network 2 are trained separately. Relative entropy measures the difference in predictions between the two networks. and This difference is added to the loss function of backpropagation. Here, relative entropy is measured using the Kullback Leibler (KL) divergence to measure the difference between the two distributions. The expression for the KL divergence is as follows: As shown, the entire feature extraction process is automatically completed by the CNN, and the degradation labels are directly represented by RUL. This method simplifies the feature extraction process and the construction of the degradation index.

[0060] (2);

[0061] In the first-stage prediction, the output of the DML-based CNN was used to indicate the health status of the bearing. A one-dimensional CNN was used to process the time-series data, with the last layer using a sigmoid activation function and a learning rate set to 0.01. The DML model was implemented through mutual learning between two CNNs, whose optimization objectives were respectively... Sum of formulas (4).

[0062]

[0063] (4);

[0064] Then, when the bearing degrades to 50% (output less than 0.5), LSTM is used to predict the remaining 50% based on the CNN output. Finally, the results from these two stages are combined to obtain the bearing's lifetime degradation curve.

[0065] The value corresponding to each point on the degradation curve is proportional to the RUL corresponding to that point, so there is a quantitative relationship between the constructed degradation curve and the RUL.

[0066] In summary, the bearing remaining life prediction method proposed in this invention uses CNN and LSTM to predict different stages of bearing degradation, respectively, to maximize the use of existing full-life data and current bearing historical data to improve the accuracy of the model.

[0067] Taking bearing data provided by the Intelligent Maintenance Systems (IMF) Center at the University of Cincinnati as an example, the data was classified and experiments were designed. Models were trained using the obtained data, and the output results of different models were obtained. Conclusions were drawn based on the obtained results. The specific implementation scheme is as follows:

[0068] 1. Constructing the experimental dataset

[0069] The experimental data came from the Intelligent Maintenance Systems (IMS) Center at the University of Cincinnati. The IMS bearing data acquisition system consisted of an AC motor (2000 RPM), four bearings (Rexnord ZA-2115 double-row bearings), and vibration sensors. The system's sampling frequency was 20 kHz, collecting 20,480 sampling points per sample, with a 10-minute interval between each sample. The vibration signals throughout the bearing's life cycle are shown below. Figure 7 As shown in (a). In this invention, the frequency domain signal obtained by FFT (such as...) Figure 7 (b) shows that the original information of the bearing can be preserved as much as possible. Furthermore, this invention points out that the model achieves good results when the first 1024 dimensions are selected. Therefore, this invention uses the first 1024 dimensions of the frequency domain signal as input to the model to improve the training speed.

[0070] This invention uses four bearing datasets. Data obtained in one sampling is considered as one sample. The number of samples in the four bearing datasets is shown in Table 1. Bearings 1 and 3 are selected as the training set, and bearings 2 and 4 are selected as the test set. In the experiment, the DSP and fault point of each bearing are first determined, and training and test labels are established. Then, FFT processing is performed on the original signal.

[0071] Table 1. Overview of data for the four bearings:

[0072]

[0073] The bearing lifecycle is divided into four stages: the health stage, the slow degradation stage, the rapid degradation stage, and the failure stage after the failure point. The time corresponding to the degradation initiation point (DSP) is expressed as... The time corresponding to the failure point is This invention uses a threshold method to determine... and .like Figure 2 As shown in (a), the RMS values ​​of different bearings are different during the healthy phase. Therefore, this invention selects the more stable characteristic RRMS (see Figure 3 ).in, This is the average effective value of data points 200 to 300 (stable phase). The thresholds selected in this invention are 1.1 and 5. However, the RRMS of bearings 1 and 2 did not reach 5. Therefore, this invention uses an exponential function... To fit the bearing's RRMS, thus determining the bearing's total lifespan. , as shown As shown:

[0074] ;

[0075] Among them, parameters , , Determined by ordinary least squares (OLS). The fitted graph is as follows. Figure 3 As shown. It can be done Sure.

[0076] 2. Dataset Augmentation

[0077] During data preprocessing, to facilitate data input into the network, the time-domain vibration signal is first transformed into a frequency-domain vibration signal using a Fast Fourier Transform. Feature extraction in this invention is automatically performed by a CNN, and degradation labels are directly represented by RUL. This method simplifies the feature extraction process and the construction of the degradation index. The CNN structure diagram is shown below. Figure 4 As shown. CNNs can be used as powerful feature extractors, exhibiting superior performance in feature extraction. They consist of three layers: convolutional layers, pooling layers, and fully connected layers. Convolutional and pooling layers are used for feature extraction; the former extracts high-dimensional features, while the latter reduces the feature dimensionality to that required for classification tasks. CNNs typically use the cross-entropy function as the loss function and softmax and sigmoid as commonly used functions. Softmax is generally used for multi-class classification tasks; its output is the input. The probability of belonging to each class. The sigmoid function is used for binary tasks, and its output is a value between 0 and 1. The expression for sigmoid is as follows: :

[0078] ;

[0079] While CNNs perform well in feature extraction, shallow CNN networks have limited effectiveness for complex bearing degradation features. This invention proposes Dependency Learning Model (DML) based on knowledge extraction, which can obtain more complex features. DML can improve model performance by acquiring features that a single model cannot learn from each other through mutual learning among multiple learning models. The principle of DML is as follows: Figure 6 As shown, this invention applies DML to the field of RUL prediction to prevent model overfitting and further improve prediction accuracy.

[0080] 3. Model training and parameter selection

[0081] CNN networks exhibit superior performance in feature extraction. They consist of three layers: convolutional layers, pooling layers, and fully connected layers. The convolutional and pooling layers are responsible for feature extraction. A diagram of the CNN structure is shown below. Figure 4 As shown, more detailed CNN parameter settings are shown in the table below.

[0082] Table 2 Detailed CNN parameter configuration

[0083]

[0084] The hidden layers of the LSTM are set to 12 nodes. The stride size of the LSTM affects the mean squared error (MSE) loss of the model. Therefore, this invention discusses the loss at different steps with 300 representations. Figure 8 This paper details how the model can converge at different step sizes, but its convergence speed and accuracy are affected by the step size. As the step size increases, the loss of the LSTM decreases; however, when the step size exceeds a certain value, the model becomes less likely to converge with further increases in step size. In this invention, the LSTM uses 24 steps, ensuring both prediction accuracy and model training speed.

[0085] Before training the model, a custom loss function needs to be defined.

[0086] For DML networks, the loss consists of two parts: prediction loss and imitation loss. For LSTM networks, the loss is prediction loss.

[0087] 1) Prediction Loss. The optimization goal of CNNs is to minimize the difference between the true and predicted values. Given an input vector... (n samples), the feature extractor will The vector is mapped to a d-dimensional space, and then the fully connected layer maps the d-dimensional vector to 1-dimensional space. Based on the output... Predicting losses Using MSE, as shown in the formula As shown:

[0088] .

[0089] 2) Mimicry Loss. Distributional discrepancies between models can be reduced by minimizing the mimicry loss. MSE can make a model perfectly fit the training data; however, this leads to overfitting because the distributions of the test and training data are not entirely identical. Therefore, it is necessary to add a mimicry loss to measure the distributional discrepancies between models, and then optimize the model by minimizing this loss. The mimicry loss is expressed by equation... The KL divergence representation in [the text].

[0090] Training labels should be constructed before training the model. During bearing degradation, the characteristics of time t are... .Will Input health metrics to train the model. During training, each sample... tags , by formula Given:

[0091] .

[0092] In the formula middle, It is the bearing's RUL at time t, which can be obtained through get. It refers to the entire lifespan of the bearing, which can be determined by... get.

[0093] The training process is divided into two parts: DML model training and LSTM model training. The output of the former model is the input of the latter. Using this strategy to train the model can make full use of existing bearing life data and bearing historical data.

[0094] During model training, the vibration signal (training dataset) is first subjected to FFT, and then the FFT data is input into the DML model. The output of the DML model can be used to describe the degree of bearing degradation. Finally, when the prediction result of the DML model is less than 0.5, the next stage of prediction model training begins. Specifically, the output value of the DML model (>0.5) is input into the LSTM model. The DML model is achieved through mutual learning between two CNNs. The optimization objectives of the two CNNs are (7) and (8), respectively. The optimization objective of the LSTM model is (6).

[0095] In the model testing (predicting the bearing's relative degradation limit (RUL)), the degree of bearing degradation at a certain moment is obtained through the Dynamic Modeling (DML) model. When the degradation degree is less than 50% (i.e., the DML model output is greater than 0.5), the DML output can be directly used to calculate the RUL. When the DML determines that the bearing degradation degree is 50% at this moment (i.e., the DML model output is less than 0.5), the FFT signal of the bearing at the next moment is first input into the DML model, and then the output of the DML model is input into the LSTM model. At this time, the output of the LSTM model can be used to calculate the RUL. In this way, the RUL of the bearing at different stages of its lifespan can be predicted. The reason for adopting this two-stage method is that in the early stage of the bearing's lifespan, the DML model can fit its degradation curve well, while in the later stage of degradation, the bearing's vibration signal fluctuates greatly. At this time, the bearing's degradation characteristics are complex and variable. Although DML can extract complex features well, its role is still limited in the later stage of bearing degradation. Considering that LSTM can well describe sequence data with spatiotemporal correlation, this invention adds LSTM to the DML model.

[0096] Based on the above parameters, the present invention conducts experiments to obtain a trained model, and selects a pre-divided test set for model testing.

[0097] 4. Analysis of Experimental Results

[0098] In this invention, the outputs of different models are used as indicators of bearing health. These models include CNN models, LSTM models, CNN-LSTM (a combination of CNN and LSTM) models, and DML models (composed of two CNNs). Furthermore, the method proposed in this invention is compared with DML. Figure 9 The results of different health indicators for bearings 2 and 4 are shown. Figure 9 The results cannot be directly used to compare the performance of health indicators. However, as can be seen from the figure, in the later stages of the bearing life cycle, the outputs of the four health indicators do not show a significant degradation trend and are far from the true RUL. The method adopted in this invention is based on this situation. Experimental results show that the proposed method can improve the prediction accuracy in the later stages. In addition, DML can learn more bearing degradation features, which can be confirmed by comparing DML with CNN, LSTM, and CNN-LSTM.

[0099] Figure 10The scatter plot, composed of label values ​​and predicted values, visually demonstrates the predictive performance of the proposed method. It clearly shows that the prediction accuracy for orientation 2 is higher than that for orientation 4. To quantitatively compare the performance of the above methods, this invention uses more detailed indicators, including correlation, root mean square error, and mean absolute error, to measure the predictive effectiveness of each method. Correlation can be used to measure the linear correlation between health indicators and time. The closer the value is to 1, the higher the linear correlation between the two.

[0100] RMSE and MAE can be used to describe the error between predicted and observed values. RMSE is sensitive to outliers, while MAE is sensitive to the mean. The evaluation expressions for RMSE and MAE are shown in equations (9) and (10), respectively.

[0101] (9)

[0102] (10);

[0103] Table 3 shows the values ​​of the three metrics mentioned above in the test set. From the table, it can be concluded that CNN, LSTM, and CNN-LSTM each have their own advantages in the three different metrics. DML is an improvement on CNN, and its performance is superior to CNN. Compared with the four different health metrics mentioned above, the method proposed in this invention performs the best. The reason is that a single method is difficult to track complex degradation processes, especially in the later stages of bearing degradation (such as...). Figure 9 As shown in the figure, the proposed method makes full use of the historical data of the bearing and combines the memory function of LSTM, thereby effectively predicting the degradation trend of the bearing in the later stage.

[0104] Table 3 Prediction results of various methods

[0105]

[0106] The value of the health index can be used to calculate the RUL of the bearing, which can be calculated by equation (11) at a certain time t.

[0107] (11);

[0108] In equation (11), This is the output of the health indicator at time t. The prediction results are shown in Table 4. Here, points 120 and 180 (calculated by DSP) are selected as prediction points for bearings 2 and 4, respectively. In this invention, a percentage error is introduced. To evaluate the performance of the prediction results. Predicting point i. The calculation formula is as follows.

[0109] (12);

[0110] As shown in Table 4, CNN-LSTM... The accuracy is lower than that of CNN and LSTM, while DML further improves the prediction accuracy of CNN. Furthermore, it can be concluded that the health metric constructed by the proposed method can predict the bearing's RUL with the highest accuracy.

[0111] Table 4 Comparison of RUL prediction results from different methods

[0112]

[0113] This invention selects the more stable bearing feature RRMS. First, in the first stage, a convolutional neural network improved by DML is used to automatically extract features to indicate the health status of the bearing. When the bearing degrades to 50% (output less than 0.5), a long short-term memory network is used to predict the remaining 50%. Finally, the results of the two stages are combined to obtain the bearing life degradation curve. The model output obtained by the above method can be directly used for RUL calculation. This method avoids the selection of bearing failure threshold. In the whole prediction process, CNN and LSTM are used for different stages of bearing degradation, respectively, to maximize the use of existing full life data and current bearing historical data to improve the accuracy of the model.

[0114] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of bearings based on deep mutual learning and dynamic feature construction, characterized in that... It includes the following steps: Step 1: Construct the experimental dataset: It includes a training set and a test set. The training set consists of bearing data for bearing 1 and bearing 3, and the test set consists of bearing data for bearing 2 and bearing 4. Step 2: Establish DML model: Preprocess the bearing data, perform FFT on the bearing vibration signal to convert the time domain signal into the frequency domain signal, and then input the obtained FFT data into the CNN model enhanced by DML; Step 3, RUL Calculation: The first 1024 dimensions of the frequency domain signal are used as input to the DML model, and the DML model output is compared. If the model output is greater than 0.5, the RUL is calculated directly using the DML output. Otherwise, the DML model output is input into the subsequent LSTM model, and the RUL is finally calculated using the LSTM model output. Step 4: Obtain the bearing life degradation curve: Compare the convergence speed and accuracy of the LSTM model under different time lengths, and then combine the prediction output results of the two stages to obtain the degradation curve. Applying DML to the RUL prediction domain, the output of the DML CNN is used to indicate the health status of the bearing. When the bearing degrades to 50%, a second-stage prediction is performed, including: The output of the DML-based CNN is used to indicate the health status of the bearing. The DML model is achieved through mutual learning between two CNNs. The optimization objectives of the two CNNs are as follows. Applying DML to the RUL prediction domain can prevent the model from overfitting and further improve the prediction accuracy. ； A one-dimensional CNN was used to process time series data. The last layer used sigmoid as the activation function, and the learning rate was set to 0.

01. When the output of the DML model is greater than 0.5, that is, when the degradation degree is less than 50%, the output of the DML can be directly used to calculate the RUL. When the output of the DML model is less than 0.5, that is, when the degradation degree is greater than 50%, the output of the DML model is input into the LSTM model for the second stage of prediction. CNN and LSTM are applied to different stages of bearing degradation, respectively, to maximize the use of existing full-life data and current bearing historical data to improve model accuracy, including: The output of the DML model is input into the LSTM model for the second stage of prediction. The RUL is calculated using the output of the LSTM model. In this way, the RUL of the bearing at different life stages can be predicted.

2. The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction according to claim 1, characterized in that: The data acquisition for the training and test sets included an AC motor, bearing 1, bearing 2, bearing 3, bearing 4, and a vibration sensor; the sampling frequency was 20kHz, with 20480 sampling points collected each time, and the time interval between each sampling was 10 minutes; then the degradation start point and fault point of each bearing were determined, and training labels and test labels were established, and the original signals were then processed by FFT.

3. The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction according to claim 2, characterized in that: The life cycle of a bearing is divided into four stages: the healthy stage, the slow degradation stage, the rapid degradation stage, and the failure stage after the failure point. The time corresponding to the degradation initiation point is expressed as... The time corresponding to the failure point is Determining using the threshold method and The threshold values ​​are 1.1 and 5. Since the RMS values ​​differ for different bearings, the more stable feature RRMS is chosen to allow for the application of the same threshold to different bearings. The expression for RRMS is as follows: ； in, It is the average effective value of 200 to 300 data points; According to the exponential function To fit the bearing's RRMS, thus determining the bearing's total lifespan. As shown in the following formula: ； Among them, parameters , , Determined by ordinary least squares, and It can be done Sure.

4. The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction according to claim 3, characterized in that: LSTM uses 24 steps, which ensures both prediction accuracy and training speed.

5. The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction according to claim 1, characterized in that: The DML model also includes model training, which includes prediction loss and imitation loss.

6. The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction according to claim 5, characterized in that: In the prediction loss, the optimization objective of the CNN is to minimize the difference between the true value and the predicted value, given an input vector. That is, with n samples, the feature extractor will Mapping to a d-dimensional space, then the fully connected layer maps the d-dimensional vector to 1-dimensional space, based on the output. Predicting losses Using MSE, as shown in the following formula; 。 7. The bearing remaining service life prediction method based on deep mutual learning and dynamic feature construction according to claim 6, characterized in that: In the aforementioned imitation loss, the distributional differences between models can be reduced by minimizing the imitation loss, and MSE can enable the model to perfectly fit the training data. An imitation loss measuring the distributional differences between models is added, and the model is then optimized by minimizing this loss. Before training the model, training labels should be constructed. During bearing degradation, the characteristics of time t are... ,Will Input health metrics to train the model; during training, each sample... tags It is given by the following formula: ； In the formula, It is the bearing's RUL at time t, which can be obtained through get, It refers to the entire lifespan of the bearing, which can be determined by... get.

Images