CN115293030A - Bearing residual service life prediction method based on deep mutual learning and dynamic feature construction - Google Patents

Abstract

The invention belongs to the field of bearing service life prediction, and discloses a method for predicting the residual service life of a bearing based on deep mutual learning and dynamic feature construction, wherein a more stable bearing feature RRMS is selected, firstly, a convolutional neural network improved by DML is used for automatically extracting features in a first stage to indicate the health condition of the bearing, when the bearing degrades to 50 percent (the output is less than 0.5), the last 50 percent of the bearing is predicted through a length memory network, and finally, the results of the two stages are combined to obtain a service life degradation curve of the bearing, the model output obtained by the method can be directly used for calculating RUL, the method avoids the selection of a bearing failure threshold, in the whole prediction process, CNN and LSTM are respectively used in different stages of the bearing degradation, and the existing full-life data and the current bearing historical data are utilized to the maximum extent to improve the accuracy of the model.

Description

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

Technical Field

The invention belongs to the field of bearing service life prediction, and particularly relates to a method for predicting the residual service life of a bearing based on deep mutual learning and dynamic feature construction.

Background

Rolling bearings are widely used in various industrial fields as rotating parts of mechanical equipment. However, the complexity of the industrial environment and the long-term uninterrupted operation make the bearings highly susceptible to failure, which causes abnormal operation of the mechanical system, as a critical component of the rotating machinery. The RUL prediction of the rolling bearing can avoid the situation, reduce unnecessary shutdown and casualties and provide basis for making an optimal maintenance strategy.

The mechanism-based RUL prediction method is modeled based on the degradation modulation of a single fault; it requires a large amount of a priori knowledge. Furthermore, the process of degradation of the bearing is very complex. Therefore, it is difficult to establish an accurate physical model. Currently, the data-driven RUL prediction method has achieved a great deal of success. Still from the problems, first, it is a rather complicated process to construct a feature set that can describe the bearing degradation trend. Second, the highly subjective construction of degenerate labels is not a uniform standard, but rather creates a great deal of randomness in the setting of failure thresholds. The method aims at the problems of complexity, changeability, high subjectivity in degradation index selection and the like in feature set construction in the field of bearing service life prediction at present.

Disclosure of Invention

The invention aims to provide a method for predicting the residual service life of a bearing based on deep mutual learning and dynamic characteristic construction, so as to solve the problems in the prior art in the background technology.

In order to achieve the purpose, the invention adopts the following technical scheme:

the method for predicting the residual service life of the bearing based on deep mutual learning and dynamic characteristic construction comprises the following steps:

step one, establishing an experimental data set: it comprises a training set consisting of bearing data for bearing 1 and bearing 3 and a test set consisting of bearing data for bearing 2 and bearing 4;

step two, building a DML model: preprocessing the bearing data, performing FFT (fast Fourier transform) on a vibration signal of the bearing, converting a time domain signal into a frequency domain signal, and inputting the obtained FFT data into a CNN (computer network node) model enhanced by DML (digital multiplex) technology;

step three, RUL calculation: taking the first 1024-dimensional data in the frequency domain signals as the input of a DML model, and comparing the output of the DML model; if the output of the model is more than 0.5, directly using the output result of the DML to calculate the RUL; otherwise, the output result of the DML model is input into a subsequent LSTM model, and the output of the LSTM model is finally used for calculating the RUL;

step four, obtaining a service life degradation curve of the bearing: and comparing the convergence speed and the accuracy of the LSTM model under different step lengths, and then combining the predicted output results of the two stages to obtain a degradation curve.

Furthermore, the device comprises an alternating current motor, a bearing 1, a bearing 2, a bearing 3, a bearing 4 and a vibration sensor during data acquisition of the training set and the test set; the sampling frequency is 20KHz, 20480 sampling points are collected in each sampling, and the time interval of each sampling is 10 minutes; then, determining a degradation starting point and a failure point of each bearing, establishing a training label and a testing label, and then performing FFT processing on the original signal.

Further, the life cycle of the bearing is divided into four phases, namely a healthy phase, a slow degradation phase, a fast degradation phase and a failure phase after a failure point. The time corresponding to the start of degradation is denoted t _d The time corresponding to the point of failure is t _f Determining t by thresholding _d And t _f The threshold values are 1.1 and 5;the RMS values of different bearings are different and therefore a more stable characteristic RRMS is chosen, so that the same threshold can be applied to different bearings, the RRMS being expressed as:

wherein RMS _norm Is the average effective value of the data points 200 to 300;

fitting the RRMS of the bearing according to an exponential function λ (t) to determine the full life λ of the bearing _t As shown in formula:

λ(t)＝y ₀ +αt ^β ；

wherein the parameter y ₀ α, β are determined by the ordinary least squares method, and t _f Can be determined by λ (t) = 5.

Further, applying DML to the field of RUL prediction, the CNN output of DML being used to indicate the health of the bearing, and when the bearing degrades to 50%, performing a second stage of prediction comprising:

the output of CNN based on DML is used for indicating the health condition of the bearing, the DML model is realized by mutual learning of two CNNs, the optimization targets of the two CNNs are respectively formula (1) and formula (2), and the DML is applied to the RUL prediction field to prevent overfitting of the model and further improve the prediction precision;

Loss _net1 ＝L _y +D _KL (p ₂ ||p ₁ ) (1)

Loss _net2 ＝L _y +D _KL (p ₁ ||p ₂ ) (2)。

further, one-dimensional CNN is used to process time series data, the last layer uses sigmoid as an activation function, and the learning rate is set to 0.01;

when the output of the DML model is greater than 0.5, i.e., when the degree of degradation 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, namely when the degradation degree is more than 50%, the output result of the DML model is input into the LSTM model to carry out the prediction of the second stage.

Further, the CNN and the LSTM are respectively used in different stages of bearing degradation, and the existing full-life data and the current bearing historical data are utilized to the maximum extent to improve the accuracy of the model, and the method comprises the following steps:

inputting the output result of the DML model into the LSTM model for second-stage prediction;

the RUL is calculated using the output of the LSTM model, and in this way the RUL for the bearing at different stages of life can be predicted.

Furthermore, the LSTM adopts 24 steps, so that the training speed of the model is ensured while the prediction precision is ensured.

Further, the DML model also comprises model training, including loss prediction and loss simulation.

Further, in the prediction loss, the optimization goal of CNN is to minimize the difference between the true value and the predicted value, given an input vector

I.e., n samples, the feature extractor will

Mapping to d-dimensional space, then mapping d-dimensional vector to 1-dimensional by full-connected layer, and outputting H _t Predicting the loss L _y MSE is used, as shown in formula;

further, in the loss of imitability, the distribution difference between models can be reduced by minimizing the loss of imitations, and the MSE can enable the models to be completely fitted to the training data; increasing the simulation loss of the distribution difference between the measurement models, and then optimizing the model by minimizing the loss, before training the model, a training label should be constructed, and during the bearing degradation, the time t is characterized by

Will be provided with

Inputting health index to train model, and during training, each sample X _t Label y of _t Given by the formula:

in the formula, actRIL _t Is the RUL of the bearing at time t, can pass through _f T is obtained, RUL ₀ Is the total life of the bearing, can pass t _f -t _d Thus obtaining the product.

The invention has the technical effects and advantages that: compared with the prior art, the method for predicting the residual service life of the bearing based on the deep mutual learning and dynamic characteristic construction has the following advantages:

according to the method, a more stable bearing characteristic RRMS is selected, firstly, a convolution neural network improved by DML is used for automatically extracting characteristics in a first stage to indicate the health condition of a bearing, when the bearing degrades to 50% (the output is less than 0.5), the later 50% is predicted through a long-time and short-time memory network, finally, the results of the two stages are combined to obtain a life degradation curve of the bearing, the model output obtained through the method can be directly used for calculating the RUL, the method avoids the selection of a bearing failure threshold value, in the whole prediction process, a CNN and an LSTM are respectively used in different stages of the bearing degradation, and the accuracy of the model is improved by utilizing the existing full-life data and the current bearing historical data to the maximum extent.

Drawings

FIG. 1 is a flow chart of a prediction method of the present invention;

FIG. 2 is a schematic diagram comparing the characteristics of RMS and RRMS in accordance with the present invention;

FIG. 3 is a schematic representation of the fitting results for a bearing of the present invention;

FIG. 4 is a schematic diagram of a convolutional neural network structure of the present invention;

FIG. 5 is a schematic diagram of the LSTM cell structure of the present invention;

FIG. 6 is a schematic diagram of the structure of the depth reciprocity model (DML) of the present invention;

FIG. 7 is a schematic representation of a full life cycle signal for a bearing of the present invention;

FIG. 8 is a graph showing the loss of the LSTM model of the present invention at different step sizes;

FIG. 9 is a graph illustrating the results of different health indicators according to the present invention;

FIG. 10 is a schematic view of a prediction result scatter plot according to the present invention.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. The specific embodiments described herein are merely illustrative of the invention and do not delimit the invention. All other embodiments, which can be obtained by a person skilled in the art without making any creative effort based on the embodiments in the present invention, belong to the protection scope of the present invention.

The invention provides a bearing residual service life prediction method based on deep mutual learning and dynamic feature construction as shown in figure 1, and a two-stage RUL prediction method is constructed by adopting DML and dynamic features.

Specifically, S1, establishing an experimental data set: it comprises a training set consisting of bearing data for bearing 1 and bearing 3 and a test set comprising bearing data for bearing 2 and bearing 4;

s2, establishing a DML model: preprocessing the bearing data, performing FFT (fast Fourier transform) on a vibration signal of the bearing, converting a time domain signal into a frequency domain signal, and inputting the obtained FFT data into a CNN (CNN) model enhanced by DML (digital signal library);

s3, calculating RUL: taking the first 1024-dimensional data in the frequency domain signal as the input of a DML model, and comparing the output of the DML model; if the output of the model is more than 0.5, directly using the output result of the DML to calculate the RUL; otherwise, inputting the output result of the DML model into a subsequent LSTM model, and finally calculating the RUL by using the output of the LSTM model;

s4, obtaining a service life degradation curve of the bearing: and comparing the convergence speed and the accuracy of the LSTM model under different step lengths, and then combining the predicted output results of the two stages to obtain a degradation curve.

During the healthy phase, the RMS values of the different bearings are different. Therefore, in the proposed method, a more stable bearing characteristic RRMS is selected, and the input signal is a frequency domain signal; training labels proportional to RUL were used. The method divides the degradation process of the bearing into two stages. The previous stage uses the output of the CNN modified by the DML to indicate the current RUL. When its output indicates a RUL of less than a certain percentage, the next stage of prediction is started. In the latter stage, the LSTM network is used to predict RUL, the input of which is the output of the previous stage. Finally, a comparison experiment is designed, the method is compared with other methods, a conclusion is drawn according to a comparison result, and the superiority of the method is proved.

The two-stage bearing residual life prediction method specifically comprises the following step that the life cycle of the bearing is divided into four stages, namely a healthy stage, a slow degradation stage, a rapid degradation stage and a failure stage after a failure point.

The time corresponding to the start of Degradation (DSP) is denoted t _d The time corresponding to the point of failure is t _f . The embodiment adopts a threshold value method to determine t _d And t _f 。

During the healthy phase, the RMS values of the different bearings are different. Therefore, the present invention selects a more stable characteristic RRMS, allowing different bearings to apply the same threshold. RRMS is represented by the formula

The proposed method comprises three modules: data preprocessing, first-stage prediction and second-stage prediction.

First, in data preprocessing, a time domain signal is converted into a frequency domain signal after being subjected to Fast Fourier Transform (FFT). In data preprocessing. The purpose of this processing is to facilitate the input of data into the network.

And secondly, applying the DML to the RUL prediction field to prevent overfitting of the model and further improve the prediction accuracy. During the DML process, network 1 and network 2 are trained forward, respectively. Relative entropy measures the difference p between two network predictions ₁ And p ₂ (ii) a This difference is added as a loss function of the back propagation. Here, the relative entropy uses Kullback Leibler (KL) divergence to measure the difference between the two distributions. The expression of KL divergence is shown in the formula. The whole feature extraction process is automatically completed by CNN, and the degraded label is directly represented by RUL. The method simplifies the process of feature extraction and the construction of the degradation index.

In the first stage prediction, the output of the CNN based on the DML is used to indicate the health of the bearing. Where one-dimensional CNN is used to process time series data, the last layer uses sigmoid as an activation function, and the learning rate is set to 0.01. The DML model is realized by mutual learning of two CNNs, and the optimization targets of the two CNNs are respectively formula (3) and formula (4).

Loss _net1 ＝L _y +D _KL (p ₂ ||p ₁ ) (3)

Loss _net2 ＝L _y +D _KL (p ₁ ||p ₂ ) (4)；

Then, when the bearing degrades to 50% (output less than 0.5), LSTM is used to predict the last 50% based on CNN output. Finally, the results of the two stages are combined to obtain a life degradation curve of the bearing.

The value corresponding to each point on the degradation curve is in direct proportion to the RUL corresponding to the point, so that a quantitative relation exists between the constructed degradation curve and the RUL.

In conclusion, according to the method for predicting the residual life of the bearing, provided by the invention, in the whole prediction process, the CNN and the LSTM are respectively used in different stages of bearing degradation, and the precision of a model is improved by utilizing the existing full-life data and the current historical data of the bearing to the maximum extent.

Taking bearing data provided by an intelligent maintenance system (IMF) center of the university of Xinxina as an example, classifying the data and designing an experiment, and training a model by using the obtained data to obtain output results of different models. And analyzing and drawing a conclusion based on the obtained result. The specific embodiment is implemented as follows:

1. construction of an Experimental dataset

The experimental data come from the center of the Intelligent Maintenance System (IMS) of the Xinxinacati university, and a bearing data acquisition system of the IMS comprises an alternating current motor (the speed is 2000 RPM), four bearings (Rexnord ZA-2115 double-row bearings) and a vibration sensor, wherein the sampling frequency of the system is 20KHz, 20480 sampling points are collected in each sampling process, and the time interval of each sampling process is 10 minutes. The vibration signal of the bearing at the full life cycle is shown in fig. 7 (a). In the present invention, the frequency domain signal obtained by FFT (as shown in fig. 7 (b)) can retain the original information of the bearing as much as possible. Furthermore, the present invention indicates that the model works well when the first 1024 dimensions are selected. Therefore, the invention takes the first 1024 dimensions of the frequency domain signal as the input of the model to improve the training speed of the model.

The present invention uses four bearing datasets. The data obtained in one sample is treated as one sample. The number of samples for the four bearing data sets is shown in table 1. Bearing 1 and bearing 3 were selected as the training set and bearing 2 and bearing 4 were selected as the test set. In the experiment, the DSP and the fault point of each bearing are firstly determined, and a training label and a testing label are established. The original signal is then FFT processed.

Table 1 summary of four bearing data:

the life cycle of the bearing is divided into four stages, namely a healthy stage, a slow degradation stage, a fast degradation stage and a failure stage after a failure point. The time corresponding to the start of Degradation (DSP) is denoted t _d The time corresponding to the point of failure is t _f . The invention adopts a threshold value method to determine t _d And t _f . As shown in fig. 2 (a), during the healthy phase, the RMS values of different bearings are different. Thus, the present invention selects a more stable characteristic RRMS (see fig. 3). Wherein RMS _norm Is the average valid value of the data points 200 to 300 (stable phase). The thresholds chosen in the present invention are 1.1 and 5. However, the RRMS of bearing 1 and bearing 2 did not reach 5. To this end, the invention fits the RRMS of the bearing with an exponential function λ (t) to determine the full life λ of the bearing _t Is shown as the formula:

λ(t)＝y ₀ +αt ^β (5)；

wherein the parameter y ₀ α, β are determined by the ordinary least squares method (OLS). The fitted graph is shown in fig. 3. And t is _f It can be determined by λ (t) = 5.

2. Data set enhancement

In data preprocessing, in order to input data into a network conveniently, a time domain vibration signal is converted into a frequency domain vibration signal through fast Fourier transform. The feature extraction of the invention is automatically completed by CNN, and the degraded label is directly represented by RUL. The method simplifies the process of feature extraction and the construction of the degradation index. The structure of the CNN is shown in FIG. 4. The CNN can be used as a powerful feature extractor and has super-strong performance in feature extraction. It consists of three layers: convolutional layers, pooling layers, and full-link layers. The convolution layer and the pool layer are used for feature extraction; the former extracts high-dimensional features, and the latter reduces the dimensionality of the features to a classification task. CNN usually uses cross entropy function as loss function, softmax and sigmoid as common function, softmax is generally used for multi-classification task; its output is input x _i Probability belonging to each class. The sigmoid function is used for binary tasks, the output of which is a value between 0 and 1. The expression of sigmoid is as follows:

although CNN performs well in feature extraction, shallow CNN networks have limited effectiveness for complex bearing degradation features. The DML provided by the invention on the basis of knowledge extraction can obtain more complex characteristics. The DML can obtain features that a single model cannot learn each other through mutual learning among a plurality of learning models, thereby improving the effect of the models. The principle of DML is shown in fig. 6. The invention applies DML to the field of RUL prediction to prevent overfitting of the model and further improve the prediction precision.

3. Model training and parameter selection

CNN networks have excellent performance in feature extraction. It consists of three layers: convolutional layers, pooling layers, and full-link layers. The function of convolutional and pooling layers is feature extraction. The structure of CNN is shown in fig. 4, and the more detailed CNN parameter settings are shown in the following table.

Table 2 CNN detailed parameter configuration

The hidden layer of LSTM is set to 12 nodes. The step size of the LSTM may affect the Mean Square Error (MSE) loss of the model. Thus, the present invention discusses the loss of different steps under 300 characterizations. Fig. 8 details that the model can converge at different step sizes, but the convergence speed and accuracy are affected by the step size. When the step length becomes longer, the loss of the LSTM is reduced; however, when the step length is larger than a certain value, the model is not easy to converge with the increase of the step length. In the invention, LSTM adopts 24 steps, thereby ensuring the training speed of the model while ensuring the prediction precision.

Before model training, custom loss functions are required.

For DML networks, the loss consists of two parts: predicted loss and modeled loss. For LSTM networks, the loss is a predicted loss.

1) And (6) predicting loss. The optimization goal of CNN is to minimize the difference between the true and predicted values. Given an input vector

(n samples), the feature extractor will

To a d-dimensional space and then the fully connected layer maps the d-dimensional vector to 1-dimensional. According to the output H _t Predicting the loss L _y MSE is used, as shown in the formula:

2) Loss of imitability. The distribution differences between models can be reduced by minimizing the simulation loss. MSE may fit the model to the training data completely; but this will result in an overfitting because the distribution of the test data and the training data is not exactly the same. Therefore, it is necessary to increase the simulation loss that measures the distribution difference between models and then optimize the models by minimizing this loss. The mimicry loss is represented by the KL divergence in the formula.

Before training the model, training labels should be constructed. During the bearing degradation, time t is characterized by

Will be provided with

Health indicators are input to train the model. During the training process, each sample X _t Label y of _t Given by the formula:

in the formula, actRIL _t Is the RUL of the bearing at time t, can pass through _f -t is obtained. RUL ₀ Is the total life of the bearing, can pass t _f -t _d Thus obtaining the compound.

The training process is divided into two parts. DML model training and LSTM model training. The output of the previous model is the input of the next model. The strategy training model can make full use of the existing bearing life data and bearing historical data.

In the model training process, the vibration signal (training data set) 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 degradation of the bearing. And finally, when the prediction result of the DML model is less than 0.5, training the prediction model at the later stage. The specific method is to input the output value (> 0.5) of the DML model into the LSTM model. The DML model is implemented by mutual learning of two CNNs. The optimization goals for the two CNNs are (7) and (8), respectively. The optimization goal of the LSTM model is (6)

During model testing (predicting the RUL of a bearing), the degree of degradation of the bearing at a certain time is obtained by the DML model. When the degree of degradation is less than 50% (i.e., the output of the DML model is greater than 0.5), the output of the DML can be used directly to calculate the RUL. When the DML judges that the degradation degree of the bearing at the moment is 50 percent (namely the output of the DML model is less than 0.5), firstly, the FFT signal of the bearing at the next moment is input into the DML model, and then the output result of the DML model is input into the LSTM model. In this case, the output of the LSTM model can be used to calculate RUL. In this way, the RUL of the bearing at different stages of life can be predicted. The reason for this two-stage approach is that the DML model fits its degradation curve well at early stages of the bearing life, while the vibration signal of the bearing fluctuates greatly at later stages of degradation. In this case, the degradation characteristics of the bearing are complex and variable. Although DML can extract complex features well, its role is still limited at later stages of bearing degradation. The present invention adds LSTM based DML, considering that LSTM can well describe sequence data with spatio-temporal correlation.

The invention carries out experiments according to the parameters to obtain a trained model, and selects a test set which is divided in advance to carry out model tests.

4. Analysis of Experimental results

In the present invention, the outputs of the different models are used as health indicators for the bearings. These models include the CNN model, the LSTM model, the CNN-LSTM (combination of CNN and LSTM) model, and the DML model (consisting of two CNNs). Furthermore, the method proposed by the present invention was compared with DML. Fig. 9 shows the results of different health indicators for the bearing 2 and the bearing 4. The results in fig. 9 cannot be used directly to compare the performance of the health indicators. However, it can be seen that the outputs of the four health indicators do not have a significant tendency to degrade later in the life cycle of the bearing and are far from the true RUL. The method employed by the present invention is based on this situation. Experimental results show that the method can improve the later prediction precision. In addition, DMLs may learn more bearing degradation characteristics, which may be confirmed by comparison of DMLs to CNN, LSTM and CNN-LSTM.

Fig. 10 is a scatter plot consisting of label values and predicted values, which can visually demonstrate the prediction effect of the proposed method, and it can be clearly seen that the prediction accuracy of the azimuth 2 is higher than that of the azimuth 4. In order to quantitatively compare the performances of the methods, the invention uses more detailed indexes comprising correlation, root mean square error and average absolute error so as to measure the prediction effect of the methods. The correlation may be used to measure a linear correlation between the health indicator and time. The closer the value is to 1, the higher the linear correlation between the two.

RMSE and MAE can be used to describe the error between predicted and observed values; while MAE is sensitive to mean values. The evaluation expressions of RMSE and MAE are shown in the formula (9) respectively.

Table 3 shows the values of the three above-mentioned criteria in the test set. From the table, it can be concluded that CNN, LSTM and CNN-LSTM have respective advantages in three different criteria. DML is an improvement over CNN, with better performance. Compared with the four different health indexes, the method provided by the invention has the best effect. The reason is that it is difficult to track the complicated degradation process with a single method. Particularly, in the later stage of the bearing degradation (as shown in fig. 9), the proposed method makes full use of the historical data of the bearing and combines the memory function of the LSTM, thereby effectively predicting the degradation trend of the bearing in the later stage.

TABLE 3 prediction results of various methods

The value of the health indicator can be used to calculate the RUL of the bearing, which at a certain time t can be calculated by the equation.

In the formula, H _t Is the output of the health indicator at time t. The predicted results are shown in table 4. Here, the 120 th and 180 th points (calculated by the DSP) are selected as predicted points for the bearings 2 and 4, respectively. In the present invention, a percentage error, er, was introduced to evaluate the performance of the predicted results. Er of the predicted point i, the calculation formula is as follows.

As can be seen from Table 4, er of CNN-LSTM is smaller than those of CNN and LSTM, and DML further improves the prediction accuracy of CNN. Furthermore, it can be concluded that the health indicators constructed by the proposed method can predict the RUL of the bearing with the highest accuracy.

TABLE 4 comparison of RUL prediction results for different methods

According to the method, a more stable bearing characteristic RRMS is selected, firstly, a convolution neural network improved by DML is used for automatically extracting characteristics in a first stage to indicate the health condition of a bearing, when the bearing degrades to 50% (the output is less than 0.5), the later 50% is predicted through a long-time and short-time memory network, finally, the results of the two stages are combined to obtain a life degradation curve of the bearing, the model output obtained through the method can be directly used for calculating the RUL, the method avoids the selection of a bearing failure threshold value, in the whole prediction process, a CNN and an LSTM are respectively used in different stages of the bearing degradation, and the accuracy of the model is improved by utilizing the existing full-life data and the current bearing historical data to the maximum extent.

Finally, it should be noted that: although the present invention has been described in detail with reference to the foregoing embodiments, it will be apparent to those skilled in the art that modifications may be made to the embodiments or portions thereof without departing from the spirit and scope of the invention.

Claims (10)

1. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction is characterized by comprising the following steps of:

step one, establishing an experimental data set: it comprises a training set consisting of bearing data for bearing 1 and bearing 3 and a test set consisting of bearing data for bearing 2 and bearing 4;

step two, building a DML model: preprocessing the bearing data, performing FFT (fast Fourier transform) on a vibration signal of the bearing, converting a time domain signal into a frequency domain signal, and inputting the obtained FFT data into a CNN (computer network node) model enhanced by DML (digital multiplex) technology;

step three, RUL calculation: taking the first 1024-dimensional data in the frequency domain signal as the input of a DML model, and comparing the output of the DML model; if the output of the model is more than 0.5, directly using the output result of the DML to calculate the RUL; otherwise, the output result of the DML model is input into a subsequent LSTM model, and the output of the LSTM model is finally used for calculating the RUL;

step four, obtaining a service life degradation curve of the bearing: and comparing the convergence speed and the accuracy of the LSTM model under different step lengths, and then combining the predicted output results of the two stages to obtain a degradation curve.

2. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 1, wherein: the device comprises an alternating current motor, a bearing 1, a bearing 2, a bearing 3, a bearing 4 and a vibration sensor when data of a training set and a test set are acquired; the sampling frequency is 20KHz, 20480 sampling points are collected in each sampling, and the time interval of each sampling is 10 minutes; then, determining a degradation starting point and a failure point of each bearing, establishing a training label and a testing label, and then performing FFT processing on the original signal.

3. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 2, wherein: the life cycle of a bearing is divided into four phases, namely a healthy phase, a slow degradation phase, a fast degradation phase and a failure phase after a failure point. The time corresponding to the start of degradation is denoted t _d The time corresponding to the fail point is t _f Determining t by thresholding _d And t _f The threshold values are 1.1 and 5; the RMS values of the different bearings are different and therefore a more stable characteristic RRMS is chosen, so that the same threshold can be applied for the different bearings, the expression for RRMS being as follows:

wherein RMS _norm Is the average effective value of the data points 200 to 300;

fitting the RRMS of the bearing according to an exponential function λ (t) to determine the full life λ of the bearing _t As shown in formula:

λ(t)＝y ₀ +αt ^β ；

wherein the parameter y ₀ α, β are determined by the ordinary least squares method, and t _f Can be determined by λ (t) = 5.

4. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 3, wherein: applying DML to the field of RUL prediction, the output of CNN of DML being used to indicate the health of the bearing, when the bearing degrades to 50%, making a second stage of prediction comprising:

the output of CNN based on DML is used for indicating the health condition of the bearing, the DML model is realized by mutual learning of two CNNs, the optimization targets of the two CNNs are respectively formula (1) and formula (2), and the DML is applied to the RUL prediction field to prevent overfitting of the model and further improve the prediction precision;

Loss _net1 ＝L _y +D _KL (p ₂ ||p ₁ ) (1)

Loss _net2 ＝L _y +D _KL (p ₁ ||p ₂ ) (2)。

5. the method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 4, wherein: one-dimensional CNN is used for processing time series data, the last layer uses sigmoid as an activation function, and the learning rate is set to 0.01;

when the output of the DML model is greater than 0.5, i.e., when the degree of degradation 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, namely when the degradation degree is more than 50%, the output result of the DML model is input into the LSTM model to carry out the second stage of prediction.

6. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 5, wherein: the CNN and the LSTM are respectively used for different stages of bearing degradation, existing full-life data and current bearing historical data are utilized to the maximum extent to improve the accuracy of the model, and the method comprises the following steps:

inputting the output result of the DML model into the LSTM model for second-stage prediction;

the RUL is calculated using the output of the LSTM model, and in this way the RUL for the bearing at different stages of life can be predicted.

7. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 6, wherein: the LSTM adopts 24 steps, so that the training speed of the model is ensured while the prediction precision is ensured.

8. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 1, wherein: model training, including predicting loss and simulating loss, is also included in the DML model.

9. The method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 8, wherein: in the prediction loss, the optimization goal of CNN is to minimize the difference between the true value and the predicted value, and to give an input vector

I.e. n samples, the feature extractor will

Mapping to d-dimensional space, then mapping d-dimensional vector to 1-dimensional by full-connected layer, and outputting H _t Predicting the loss L _y MSE is used, as shown in formula;

10. the method for predicting the residual service life of the bearing based on deep mutual learning and dynamic feature construction as claimed in claim 9, wherein: in the imitative loss, the distribution difference between models can be reduced by minimizing the imitative loss, and the MSE can enable the models to be completely fitted to training data; increasing the simulation loss of the distribution difference between the measurement models, and then optimizing the model by minimizing the loss, before training the model, a training label should be constructed, and during the bearing degradation, the time t is characterized by

Will be provided with

Inputting health index to train model, and during training, each sample X _t Label y of _t Given by the formula:

in the formula, actRIL _t Is the RUL of the bearing at time t, can pass through _f T is obtained, RUL ₀ Is the total life of the bearing, can pass t _f -t _d Thus obtaining the product.

Images