Bayesian optimization-based DCAE-RNN rolling bearing early fault prediction method

Abstract

The invention discloses a Bayesian optimization-based DCAE-RNN rolling bearing early fault prediction method, and the method comprises the steps: constructing a DCAE-RNN model, the DCAE-RNN model comprises a DCAE model and an RNN model, the DCAE model is composed of an SADE and a CAE, and carrying out the hyper-parameter optimization of the SDAE and the CAE through a Bayesian optimization function; the training set is adopted to train the DCAE-RNN model, and a trained DCAE-RNN model is obtained; and inputting a real-time vibration signal of rolling bearing operation into the trained DCAE-RNN model, performing residual calculation by using a predicted value output by the model and a true value, and detecting an early fault of the rolling bearing. According to the method, the running state of the rolling bearing can be well described, the early failure of the rolling bearing can be predicted earlier and more stably, the method has high accuracy and robustness, the model is convenient to deploy, and the method is very important for degradation monitoring of the running state of the rolling bearing.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing fault detection, and in particular relates to a rolling bearing early fault prediction method based on Bayesian optimized DCAE-RNN. Background Art

[0002] As an important component of rotating machinery, rolling bearings often fail during operation due to factors such as harsh working environment, high load, high speed operation and manufacturing quality, which seriously affects the operation of mechanical equipment and industrial production. Accurate diagnosis of the initial failure of rolling bearings can effectively prevent mechanical damage and casualties. The information of the initial failure characteristics of rolling bearings is very weak and is often accompanied by background noise, making it difficult to effectively extract early failure characteristics. How to effectively diagnose the initial failure of rolling bearings in actual operating environments has always been a difficult problem in the field of mechanical equipment maintenance management.

[0003] The information of early fault characteristics of rolling bearings is difficult to extract effectively. At present, the main methods for early fault diagnosis are mathematical models and data-driven methods. The mathematical modeling method relies on the signals collected by sensors during the operation of rotating machinery for fault diagnosis, and analyzes them in the time domain, frequency domain and time-frequency domain respectively, and extracts a large number of characteristic parameters from the signals as the criteria for traditional fault diagnosis. However, with the complexity and diversification of industrial materials, different rolling bearing materials are not the same, and rolling bearings are coupled with related components during operation, which leads to the establishment of physical failure models. In the end, it may only be possible to obtain an approximate solution, making the prediction accuracy of RUL insufficient. Time domain features (root mean square error, kurtosis) are widely used in bearing performance evaluation. Chen ZQ et al. proposed a time domain statistical index based on the root mean square error (RMS) for rolling bearing state identification in "Shock and Vibration". RMS reflects signal energy, and the slower the amplitude value changes over time, the more obvious its effect is. Obviously, it is ineffective compared with the weak early fault amplitude. Yuan YL et al. in New Technology & New Process used the time domain statistical index of kurtosis to identify the state of rolling bearings. Since kurtosis has a greater response to impact signals, it has a greater response to early abnormal fault signals, and thus has a higher sensitivity to early abnormal fault signals. However, for small fluctuations, it may be too sensitive to cause false alarms, which affects the reliability of rolling bearing fault assessment. Moreover, the fluctuations caused by the uncertainty of vibration during the operation of rolling bearings have a great impact on the prediction of the bearing operation state. Therefore, time domain feature analysis is not conducive to the early fault prediction of rolling bearings.

[0004] Data-driven deep learning methods can automatically learn to enrich different features and have strong feature extraction capabilities, making it possible to apply them to the initial fault research of rolling bearings. Moreover, the RUL prediction method based on deep learning can make full use of the data processing capabilities of the deep learning network model, has a weak dependence on physical knowledge, and does not require too much expert experience and signal processing knowledge. Li Heng et al. proposed in the "Vibration and Shock" document to perform short-time Fourier transform on the vibration signal of the rolling bearing, and use the obtained time-frequency spectrum samples to train the Convolutional Neural Networks (CNN), which has a high recognition accuracy for different types of faults. Although the single-class support vector machine (SVM) method proposed by Zhang XY et al. in the "Measurement" document can be mapped to a high-dimensional space, it is difficult for large-scale training data, especially when performing intelligent tuning of hyperparameters, which will invest a lot of time and effort. The samples trained by the above neural network models are independent of each other, and the temporal correlation between the data is ignored. The Long Short-term Memory Network (LSTM) network is a special type of recurrent neural network (RNN) that is effective in machine learning problems involving a wide range of sequence data. Compared with RNN, it effectively alleviates the gradient vanishing problem in RNN, enabling it to capture and retain long-term dependency information when processing long sequences. However, under the same hardware conditions, the training and inference speed of LSTM is usually slower than that of RNN, and it requires more computing resources. Du Xiaolei et al. proposed a bearing fault diagnosis model based on deep wavelet convolutional autoencoder and LSTM, using wavelet autoencoder to extract fault features for LSTM network training and fault diagnosis. Zhao Jianpeng et al. used empirical mode decomposition data to extract feature parameters and input them into the LSTM network to predict the working state of the rotating machinery at the next moment. However, these methods have problems such as simple use of network models and uncertain mechanisms of the temporal correlation characteristics of extracted data. In the "Measurement" paper, WANG Q et al. established multiple deep autoencoder models to extract features, and used clustering methods to select the original feature set to improve the prediction accuracy of bearings. Xu F et al. proposed a feature extraction method of exponential function stacking autoencoder based on moving window in the paper "Measurement", combining the local minimum point of slope to extract degradation trend and improve its monotonicity. However, when using autoencoder, the internal distribution of the hidden layer data of the autoencoder is often ignored, which limits the effectiveness of feature extraction.Li Junhuai et al. introduced convolutional neural network into denoising autoencoder (DAE) in the patent "A method for short-term daily load curve prediction based on DCAE-LSTM", and used convolutional layer to replace the fully connected layer of denoising autoencoder to obtain convolutional denoising autoencoder, in which each neuron is composed of convolutional neurons; denoised data is obtained. This patent uses SDAE model based on Bayesian optimization and contraction autoencoder (CAE) to form DCAE model, which enhances the robustness of feature learning while achieving data denoising.

[0005] The deeper the network and the more nodes, the better the learning and fitting effects will be, but the number of hyperparameters will also increase. Hyperparameters are a series of parameters set before model training begins, such as the number of hidden layers, the number of nodes in each layer, the number of iterations, and the learning rate. The hyperparameters of the model are difficult to select and have no regularity. In addition, there are unpredictable effects between different hyperparameters, and the evaluation of each hyperparameter combination requires a large number of iterative calculations. Different hyperparameter combinations will affect the final accuracy and stability of the model. Usually, a large number of hyperparameter experiments are required to determine the final hyperparameter combination of the model, which consumes a lot of time and effort. Summary of the invention

[0006] In view of the above problems existing in the prior art, the present invention proposes a rolling bearing early fault prediction method based on DCAE-RNN with Bayesian optimization, which has a reasonable design, solves the shortcomings of the prior art and has good effects.

[0007] A rolling bearing early fault prediction method based on Bayesian optimization DCAE-RNN, comprising the following steps:

[0008] S1. Obtain historical vibration signals of rolling bearings in normal state and establish a training set;

[0009] S2. Construct a DCAE-RNN model, wherein the DCAE-RNN model includes a DCAE model and an RNN model, wherein the DCAE includes an input layer, three hidden layers and an output layer, wherein the first hidden layer and the second hidden layer are both composed of a stacked denoising autoencoder SADE, and the third hidden layer is composed of a contraction autoencoder CAE;

[0010] The DCAE model is used to extract the depth features of the rolling bearing vibration signal, and the RNN model is used to extract the temporal correlation characteristics between the depth features;

[0011] S3, obtain the optimal hyperparameter combination of the DCAE model through the Bayesian optimization algorithm;

[0012] S4, using the training set to train the DCAE-RNN model to obtain a trained DCAE-RNN model;

[0013] S5. Input the real-time vibration signal of the rolling bearing into the trained DCAE-RNN model, use the predicted value and the true value output by the model to calculate the residual and detect the early failure of the rolling bearing.

[0014] Furthermore, in S1, a rolling bearing test bench is constructed, including an AC motor, a motor speed controller, a support shaft, two support bearings, a hydraulic loading system, and an acceleration sensor. The AC motor is used to provide a rotational speed, and the hydraulic loading system is used to provide a radial force. The rolling bearing test bench is used to perform accelerated degradation tests on the rolling bearings, and the vibration signals of the rolling bearings from operation to failure are collected using an acceleration sensor. The historical normal signals are used as training and test sets for the SDAE model, CAE model, and RNN model, and the full life cycle data is used as the training and test sets for the DCAE-RNN model.

[0015] Furthermore, the encoding and decoding equation of the SADE is as follows:

[0016]

[0017] in, is the encoded feature, f e (·) is the encoding function from the input layer to the output layer, a h is the result of linear transformation of input data, σ(·) is the Sigmoid activation function, and W xh Its weight value, is the input vector, b xh is the bias value, is the reconstructed input data, f d (·) is the decoding function from hidden layer to output function, W hr is the weight value, b hr is its bias value;

[0018] Therefore, the reconstruction error of SDAE is:

[0019]

[0020] The reconstruction error of CAE is:

[0021]

[0022]

[0023] Among them, J CAE is the total loss function of the contraction autoencoder, θ represents the weight, L(x,r) is the reconstruction error of the contraction autoencoder, λ is the contraction factor, is the contraction penalty term, J h(x) is the Jacobian matrix, h i is the output of the i-th encoder, i.e., the hidden layer feature, i=1,...,M, x j is the j-th input data, j=1,...,N.

[0024] Furthermore, the S3 includes the following steps:

[0025] S3.1, set the value range of each hyperparameter, the hyperparameters include the number of hidden layers, the number of neurons in the hidden layer, the optimization function and the activation function of the SDAE and CAE models, and then perform hyperparameter optimization for SDAE and CAE respectively through the following steps;

[0026] S3.2, firstly, based on the value range of the set hyperparameters, a sampling algorithm is used to obtain candidate combinations of hyperparameters as a hyperparameter search space;

[0027] S3.3, input candidate combinations, train the SDAE model and CAE model, and output the prediction accuracy of the candidate combinations on the validation set;

[0028] S3.4, using the Gaussian process regression method, calculate the regression function between the prediction accuracy of different candidate combinations on the validation set;

[0029] S3.5, taking the regression function as the prior distribution in the Gaussian process, traversing the hyperparameter search space, and obtaining the regression function with the highest accuracy as the objective function;

[0030] S3.6. Based on Gaussian process regression, the posterior probability distribution P(E|D) is constructed according to formula (7):

[0031]

[0032] Where P(E|D) is the probability of the training data D appearing given E; P(E) is the prior probability of E; P(D) is the probability of the data D appearing;

[0033] S3.7. Input the objective function, solve it iteratively through the Bayesian algorithm, and output the corresponding hyperparameter combination with the highest prediction accuracy as the optimal hyperparameter combination X n =[X1,X2,…,X n ];

[0034] S3.8. Set up the SDAE and CAE models according to the optimal hyperparameter combination obtained; test the prediction accuracy of the entire DCAE-RNN model to determine whether the model meets the accuracy requirements. If it does, the model hyperparameters are X n,If the accuracy requirement is not met, the Bayesian optimization range is expanded and the above steps are repeated to re-perform Bayesian optimization on the hyperparameters of SDAE and CAE.

[0035] Furthermore, in S4, the data in the training set is rearranged into a data set Y with the same neuron dimension as the input layer of the DCAE model. According to formulas (1) to (3), the data set Y is used to train the SDAE after hyperparameter optimization. The output layer of SDAE is used as the input layer of CAE. According to formulas (4) to (6), the CAE after hyperparameter optimization is trained to obtain the trained DCAE model. The data reconstructed by the DCAE model is used as the input of RNN to construct a DCAE-RNN model.

[0036] Furthermore, in S5, the residual is calculated between the predicted value and the true value of the real-time data to obtain a reconstruction error, and the support vector data description SVDD is used to determine whether the reconstruction error exceeds a set threshold. If it exceeds the threshold, the rolling bearing has an early failure and an alarm is issued.

[0037] Beneficial technical effects brought by the present invention:

[0038] The method of the present invention can well describe the operating state of rolling bearings, predict early failures of rolling bearings earlier and more stably with high accuracy and robustness, and the model is lightweight and easy to deploy, which is very important for degradation monitoring of the operating state of rolling bearings. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a structural schematic diagram of the DCAE model in the present invention;

[0040] Figure 2 Schematic diagram of the DCAE-RNN model structure in the present invention;

[0041] Figure 3 This is a flow chart for predicting early failure of rolling bearings in the present invention;

[0042] Figure 4 It is an original time domain vibration signal diagram of the entire life cycle of the rolling bearing in an embodiment of the present invention;

[0043] Among them, (a) is the horizontal original time domain vibration signal diagram of the data set bearing2_3; (b) is the horizontal original time domain vibration signal diagram of the data set bearing2_5; (c) is the horizontal original time domain vibration signal diagram of the data set bearing3_4; (d) is the vertical original time domain vibration signal diagram of the data set bearing3_4;

[0044] Figure 5It is a full life cycle operation state diagram of a rolling bearing described based on the DCAE-RNN model in an embodiment of the present invention;

[0045] Among them, (a) is the full life cycle operation status of the dataset bearing2_3; (b) is the full life cycle operation status of the dataset bearing2_5;

[0046] Figure 6 The full life cycle operation status diagram described for the RNN model;

[0047] Figure 7 This is a comparison chart of the full life cycle operation status of two time domain indicator recognition datasets bearing2_3 for DCAE-RNN and rolling bearings;

[0048] Figure 8 A comparison chart of the full life cycle operation status of the data set bearing2_3 described by the DCAE-LSTM model, DCAE+RMS method, and DCAE+Kurtosis method in an embodiment of the present invention;

[0049] Fig. 9 This is a comparison diagram of the rolling bearing operating status of the DCAE-RNN and time domain statistical index identification data set bearing2_5 in an embodiment of the present invention;

[0050] Fig.10 A comparison chart of the full life cycle operation status of the data set bearing2_5 is described for the DCAE-LSTM model, DCAE+RMS method, and DCAE+Kurtosis method in an embodiment of the present invention;

[0051] Fig.11 This is a comparison chart of the full life cycle operation status of the DCAE-RNN model and the DCAE+LSTM model describing the data set bearing3_4 in an embodiment of the present invention;

[0052] Among them, (a) is a comparison diagram of the full life cycle operation status described by the DCAE-RNN model; (b) is a comparison diagram of the full life cycle operation status described by the DCAE+LSTM model;

[0053] Fig.12 The number of samples when different models in the embodiments of the present invention predict early failures;

[0054] Fig.13 This is a comparison chart of the prediction time of the overall operating status of the DCAE+RNN and DCAE+LSTM models in an embodiment of the present invention; DETAILED DESCRIPTION

[0055] The specific implementation of the present invention is further described below in conjunction with specific embodiments:

[0056] A rolling bearing early fault prediction method based on Bayesian optimization DCAE-RNN, comprising the following steps:

[0057] S1. Obtain historical vibration signals of rolling bearings in normal state and establish a training set;

[0058] S2. Construct a DCAE-RNN model. The DCAE-RNN model includes a DCAE model and an RNN model. The DCAE includes an input layer, three hidden layers and an output layer. The first hidden layer and the second hidden layer are both composed of a stacked denoising autoencoder SDAE, and the third hidden layer is composed of a contraction autoencoder CAE.

[0059] The DCAE model is used to extract the deep features of the rolling bearing vibration signal, and the RNN model is used to extract the temporal correlation characteristics between the deep features;

[0060] S3, obtain the optimal hyperparameter combination of the DCAE model through the Bayesian optimization algorithm;

[0061] S4, using the training set to train the DCAE-RNN model to obtain a trained DCAE-RNN model;

[0062] S5. During actual monitoring, the real-time vibration signal of the rolling bearing is input into the DCAE-RNN model, and the predicted value of the vibration signal at the next moment is output. The residual is calculated between the predicted value and the true value of the real-time data to obtain the reconstruction error. The support vector data description SVDD is used to determine whether the reconstruction error exceeds the set threshold. If it exceeds the threshold, the rolling bearing has an early fault and an alarm is issued.

[0063] Specifically, in S1, a rolling bearing test bench is constructed, including an AC motor, a motor speed controller, a support shaft, two support bearings, a hydraulic loading system, and an acceleration sensor. The AC motor is used to provide the rotational speed, and the hydraulic loading system is used to provide the radial force. The rolling bearing test bench is used to perform accelerated degradation tests on rolling bearings at different rotational speeds, different sizes, and different states. The acceleration sensor is used to collect the vibration signals of the rolling bearings from operation to failure, and the historical normal signals are used as training sets.

[0064] Specifically, Auto Encoder (AE) is an unsupervised feature expression network, and consists of an input layer (x), a hidden layer (h), and an output layer (r). The purpose of AE is to reconstruct data through the process of encoding and decoding so that the error between the output data and the input data is minimized. Stacked Denoising Auto Encoder (SDAE) is an improved method of AE. SDAE reconstructs data from damaged or noisy data samples. CAE enhances the robustness of learning features by adding a penalty term (Jacobi) to the loss function. In other words, CAE uses the internal information of the data, mainly for the output of the data to enhance the robustness of feature learning; SDAE uses the external information of the data, mainly for the input of the data to enhance the robustness of feature learning. Therefore, CAE and SDAE are combined to construct DCAE to enhance the ability of data feature learning.

[0065] Based on their respective advantages, SDAE and CAE are used to learn latent features from rolling bearing vibration data and learn deeper features on top of latent features, respectively. Finally, the learned deep features are robust to small perturbations of the input.

[0066] like Figure 1 As shown, DCAE consists of three hidden layers, the first and second hidden layers consist of SDAE, and the third hidden layer consists of CAE. Feature learning is performed based on input and output data to improve the robustness of rolling bearing initial fault feature extraction.

[0067] The unprocessed vibration data contains a lot of noise, so the noisy input data is used Instead of the noise-free input data x.

[0068] The encoding and decoding equations of the stacked denoising autoencoder (SDAE) are as follows:

[0069]

[0070] in, is the encoded feature, f e (·) is the encoding function from the input layer to the output layer, a h is the result of linear transformation of input data, σ(·) is the Sigmoid activation function, and W xh Its weight value, is the input vector, b xh is the bias value, is the reconstructed input data, f d (·) is the decoding function from hidden layer to output function, W hr is the weight value, b hr is its bias value;

[0071] Therefore, the reconstruction error of SDAE is:

[0072]

[0073] The reconstruction error of CAE is:

[0074]

[0075] Among them, J CAE is the total loss function of the contraction autoencoder, θ represents the weight, L(x,r) is the reconstruction error of the contraction autoencoder, λ is the contraction factor, is the contraction penalty term, J h (x) is the Jacobian matrix, h i is the output of the i-th encoder, i.e., the hidden layer feature, i=1,...,M, x j is the j-th input data, j=1,...,N.

[0076] Recurrent Neural Network (RNN) is a type of neural network model that can process and model sequence data. Unlike traditional feedforward neural networks, RNN introduces recurrent connections in the hidden layers of the network, allowing the network to share information between time steps and capture the temporal dependency of the input data. This design enables RNN to perform well in processing time series data, natural language processing, speech recognition, and other tasks involving order or sequence.

[0077] The core feature of RNN is its hidden state, which is updated at each time step and carries the information inherited from the previous time step. This information transmission mechanism allows RNN to retain and utilize previous information when there are long-distance dependencies in the sequence, enhancing the ability to predict subsequent elements in the sequence.

[0078] In summary, the DCAE model and the RNN model are combined to construct a neural network model, such as Figure 2 As shown, this is the DCAE-RNN model framework.

[0079] The model mainly consists of an input layer, a feature extraction layer, a prediction layer, and an output layer. The processed data set is input into the DCAE network, and the DCAE network is trained according to equations (1) to (6). The DCAE model is used to capture more useful fault features of the data, and obtain stronger robustness from the original data, as well as the temporal dependency characteristics between RNN data. The output of the DCAE model is used as the input of the RNN model to predict the data.

[0080] Specifically, in S3, the Bayesian optimization algorithm is used to set and optimize the optimal hyperparameter combination for SDAE and CAE respectively. Bayesian optimization is aimed at complex optimization problems where the objective function expression of the hyperparameters of deep learning models is unknown and the search cost is high. It uses a continuously updated probability model to update the posterior probability of the optimization function through a small number of objective function evaluations to obtain the optimal model hyperparameter combination.

[0081] The following steps are involved:

[0082] S3.1, set the value range of each hyperparameter, the hyperparameters include the number of hidden layers, the number of neurons in the hidden layer, the optimization function and the activation function of the SDAE and CAE models, and then perform hyperparameter optimization for SDAE and CAE respectively through the following steps;

[0083] S3.2, firstly, based on the value range of the set hyperparameters, a sampling algorithm is used to obtain candidate combinations of hyperparameters as a hyperparameter search space;

[0084] S3.3, input candidate combinations, train the SDAE model and CAE model, and output the prediction accuracy of the candidate combinations on the validation set;

[0085] S3.4, using the Gaussian process regression method, calculate the regression function between the prediction accuracy of different candidate combinations on the validation set;

[0086] S3.5, taking the regression function as the prior distribution in the Gaussian process, traversing the hyperparameter search space, and obtaining the regression function with the highest accuracy as the objective function;

[0087] S3.6. Based on Gaussian process regression, the posterior probability distribution P(E|D) is constructed according to formula (7):

[0088]

[0089] Where P(E|D) is the probability of the training data D appearing given E; P(E) is the prior probability of E; P(D) is the probability of the data D appearing;

[0090] S3.7. Input the objective function, solve it iteratively through the Bayesian algorithm, and output the corresponding hyperparameter combination with the highest prediction accuracy as the optimal hyperparameter combination X n =[X1,X2,…,X n ];

[0091] S3.8. Set up the SDAE and CAE models according to the optimal hyperparameter combination obtained; test the prediction accuracy of the entire DCAE-RNN model to determine whether the model meets the accuracy requirements. If it does, the model hyperparameters are X n,If the accuracy requirement is not met, the Bayesian optimization range is expanded and the above steps are repeated to re-perform Bayesian optimization on the hyperparameters of SDAE and CAE.

[0092] Specifically, in S4, the data in the training set are rearranged into a data set Y with the same neuron dimension as the input layer of the DCAE model. According to equations (1) to (3), the data set Y is used to train the SDAE after hyperparameter optimization. The output layer of SDAE is used as the input layer of CAE. According to equations (4) to (6), the CAE after hyperparameter optimization is trained to obtain the trained DCAE model. The data reconstructed by the DCAE model is used as the input of the RNN to train the RNN network and construct a trained DCAE-RNN model.

[0093] In S5, the residual of the predicted value and the real value of the real-time data is calculated to generate a reconstruction error. The support vector data description SVDD is used to determine whether the reconstruction error exceeds the set threshold. If it exceeds the threshold, the rolling bearing has an early failure and an alarm is issued, such as Figure 3 shown.

[0094] In order to verify the effectiveness of this method in the prediction of initial failure of rolling bearings, the full life cycle data of rolling bearings obtained by the rolling bearing test platform (XJTU-SY) was used for experimental analysis. As shown in Table 1, five rolling bearing experiments with different fault sizes were set at different speeds, sampled at 25.6kHz, and each sample had 32768 points;

[0095] Table 1 Rolling bearing full life cycle data set

[0096]

[0097] Select the bearing2_3, bearing2_5, and bearing3_4 data sets in Experiment B and Experiment C, select the historical normal data in the data set to divide the training set and the test set to train the SDAE, CAE, and RNN single models respectively, and then use the entire data set to divide the training set and the test set to train the entire model. The actual life of the three data sets is 8h53min, 5h39min, and 25h15min. The fault locations include the outer ring, inner ring, and rolling element of the rolling bearing.

[0098] Bayesian optimization model verification,In order to ensure that the model can more accurately identify early failures of rolling bearings, the present invention uses a Bayesian optimization algorithm for intelligent parameter adjustment.

[0099] According to the size of the training set, the number of hidden layers is set to an integer range of [2,5], and the number of neurons in the hidden layer is set to [100,1000], and the hyperparameters of SDAE and CAE are optimized respectively. At present, the main optimization functions of neural networks are Adam, RMSProp, and Nadam. The BO algorithm is used to optimize the optimization function that best suits the model. The activation function suitable for the model is iteratively searched in the activation functions sigmoid, tanh, and relu. Because the model proposed in the present invention uses the Bayesian optimization method to determine the model hyperparameters.

[0100] The formula for the expected value of the square of the difference between the parameter estimate and the true parameter value is:

[0101]

[0102] The above formula is used to evaluate the degree of data change. When the MSE value is smaller, the hyperparameter prediction value has a higher accuracy. Bayes uses Gaussian function instead of black box function in the process of optimizing parameters, and obtains the optimal solution through the consistent and continuous weak assumption of local smoothness. The weak assumption of smoothness enables the BO algorithm to effectively use the interval proximity information for more accurate inference, thereby selecting more "potential" points. That is, the most "potential" point in the search interval is determined by actively selecting a strategy, which is the optimal solution for the hyperparameter combination. It can be seen that each hyperparameter to be optimized will affect the final accuracy of the model.

[0103] In order to verify whether the accuracy of the optimal hyperparameter combination selected by the BO algorithm meets the requirements, the optimized hyperparameter combination is brought into the model proposed by the present invention, and the model is iteratively trained. The hyperparameter combination selected by Bayesian optimization has a faster convergence speed and less fluctuation during the model training process.

[0104] After verifying that the model meets the accuracy requirements, a display program is added during the model training hyperparameter process. At the end of the training, the optimized hyperparameter combination is displayed. The hyperparameter combination of the model of the present invention is shown in Table 2.

[0105] Table 2 Optimized model hyperparameters

[0106]

[0107] The original time domain vibration signals of the rolling bearings in the data sets bearing2_3, bearing2_5, and bearing3_4 over the entire life cycle are shown in the figure below: Figure 4 As shown. Figure 4(a) It can be seen that before about 4100000 (bearing operation time is 2.1 hours) data, the time domain response of bearing 2_3 is relatively stable, and then the time domain signal shows very obvious abnormal vibration, that is, bearing 2_3 has suffered relatively obvious damage, and then the time domain signal shows a trough phenomenon, which is a typical "healing phenomenon" in rolling bearing outer ring failure. After the damage occurs, under the action of circulating force, the damaged position is briefly run-in to smoothness, and then severe damage begins to appear at 11000000 (bearing operation time is 5.6 hours) data, and the rolling bearing is no longer "healed" until it fails. Figure 4 In (b), before about 4,000,000 data points (bearing operation time is 2 hours), the time domain response of bearing 2_5 is relatively stable, and then the time domain signal shows very obvious abnormal vibration, that is, bearing 2_5 has suffered relatively obvious damage, and then it appears to be "healed", and then it begins to show severe damage at 8,000,000 data points (bearing operation time is 4.07 hours). Figure 4 In (c) and (d), before about 34,000,000 data points (the bearing has been running for about 16 hours), the time domain response of bearing 3_4 is relatively stable. After that, the time domain signal shows very obvious abnormal vibration, that is, bearing 3_4 has suffered relatively obvious damage, and then it appears to be "healed", and then it begins to show severe damage at 47,000,000 data points (the bearing has been running for about 23 hours).

[0108] like Figure 5 As shown in the figure, the DCAE-RNN model based on Bayesian optimization proposed in the present invention describes the operating state of the rolling bearing, predicts the initial failure time of the rolling bearing, and classifies the operating state of the rolling bearing. The DCAE-RNN model can better reflect the operating state of the rolling bearing, predict the time when the initial failure of the rolling bearing occurs, and can also judge the damage degree of the rolling bearing according to the acceleration. Figure 5 (a) Bearing2_3: Within about 5400 sampling cycles, the first abnormal test data has a larger amplitude than the normal signal. The rolling bearing runs for about 1.9 hours, and then the data tends to be normal. Figure 4 Compared with (a), the abnormality was detected 12 minutes earlier. Severe damage began to appear at about 20,000 sampling cycles (bearing operation time was about 7.1 hours), which was different from Figure 4 Similarly, bearing2_5 first detected abnormal data in the 5100th sampling cycle (bearing operation time is about 1.8h), which is about 1.5 hours earlier than (a). Figure 4 (b) The abnormality was detected about 18 minutes in advance. Severe damage began to appear at about 9000 sampling cycles (bearing operation time was about 7.1 hours). Figure 4(b) is about 0.8 hours earlier. In summary, the proposed model can improve the detection of initial faults by an average of 15 minutes and the detection of severe damage by an average of 1.2 hours compared with the original signal. Therefore, the proposed DCAE-RNN model can reflect the initial fault and degradation stage of rolling bearings more accurately and earlier.

[0109] In order to further verify the effectiveness of the method proposed in this paper, the proposed DCAE+RNN algorithm is compared with Kurtosis, RMS, RNN, and DCAE+LSTM. Taking the bearing2_3 dataset as an example, the comparison results are as follows: Figure 6 and 7 As shown;

[0110] from Figure 6 It can be seen that although the RNN model can better reflect the degradation trend of rolling bearings, serious impact signals appear at the beginning of 12,000 samples, which is different from Figure 4 (a) The original data of the whole life is seriously distorted. Time-domain statistical indicators are often used in rolling bearing state identification. Commonly used time-domain statistical indicators include mean value, root mean square error, and kurtosis.

[0111] from Figure 7 The green curve shows that kurtosis is very sensitive to abnormal changes in the signal. When the rolling bearing operates abnormally, transient spikes will appear in the vibration signal, so the kurtosis will change significantly, but it will lead to unstable diagnostic results, often resulting in false alarms, and increasing the chance of misdiagnosis. Figure 7 The red curve shows that the vibration signal did not produce abnormalities at the initial failure of the rolling bearing, but only fluctuated in the later stage of the failure. Therefore, it cannot be used as an evaluation indicator to well represent the operating condition of the rolling bearing. Although statistical indicators based on the time domain can reflect the degradation trend of the rolling bearing to a certain extent, they cannot well reflect the changes in the early failure of the rolling bearing and the various stages of the later degradation. Substitute the optimal hyperparameter combination selected by the BO algorithm into the model for training, such as Figure 7 The purple curve shows the operating curve of the DCAE-RNN model based on Bayesian optimization proposed in the present invention for the operating state of the rolling bearing.

[0112] In order to further verify the accuracy of the training model prediction, the present invention proposes three methods: DCAE+LSTM, DCAE+RMS and DCAE+Kurtosis. Figure 8 shown.

[0113] from Figure 8 It can be seen that Figure 7Compared with the DCAE+RNN model in the previous example, the DCAE+LSTM running curve has no obvious difference. The first abnormal vibration occurs at around 5400 samples, and the start of severe bearing damage is around 20000 samples. However, the DCAE+RNN model is easier to deploy. The other two methods (DCAE+RMS and DCAE+Kurtosis) are Figure 8 It can be seen that although stable characteristics are learned through DCAE training, it is not sensitive to state changes. Fig. 9 and 10 As shown in Figure 2, similar results also appear in the bearing dataset bearing2_5 by analogy.

[0114] Considering that the LSTM model may not be able to fully explore the intrinsic connection between the current moment data and the past moment data due to insufficient training data, the horizontal data and vertical data of bearing bearing3_4 are used as data sets for model training. The training results are shown in Figure 2. Fig.11 shown.

[0115] Depend on Fig.11 (a) and (b) show that when the data doubles, both the DCAE+RNN model and the DCAE+LSTM model begin to experience early failures around 70,000 data samples (bearing operation time is about 13 hours).

[0116] In order to further quantitatively verify the effectiveness of the DCAE-RNN model based on Bayesian optimization proposed in this paper, this paper takes the bearing2_3 dataset as an example and uses SVDD to represent the signal position when the fault occurs. Fig.12 The effectiveness of the different methods is evaluated by comparing the running time when the fault alarm occurs by showing the period in which each method predicts the early fault.

[0117] Depend on Fig.12 It can be seen that the DCAE-RNN model after BO can perform fault alarm within 5369 cycles. Compared with DCAE+LSTM, DCAE+Kurtosis, DCAE+RMS, and RNN methods, the DCAE-RNN network optimized by BO algorithm can detect early faults earlier, and the effect is better than other methods. Compared with the DCAE+LSTM model, although it is only 121 cycles (2.6 minutes) earlier, it is Fig.13It can be seen that the DCAE+RNN model predicts the overall operating status in 325 seconds, which is 83 seconds faster than the DCAE+LSTM. Under the same hardware conditions, the calculation speed is faster, which is convenient for the later deployment of the model to achieve online monitoring of the machinery. In summary, the DCAE-RNN algorithm after BO proposed in the present invention is not only very sensitive to the early abnormalities of rolling bearings, but also can reflect the degradation trend of rolling bearings in the later stage. Moreover, the model calculation speed is faster, which is convenient for deployment. The signal of rolling bearings is stable during the diagnosis process, so it can predict the early failure of rolling bearings, proving the effectiveness and reliability of the DCAE-RNN model optimized by the BO algorithm.

[0118] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A rolling bearing early fault prediction method based on DCAE-RNN with Bayesian optimization, characterized in that: The following steps are involved: S1. Obtain historical vibration signals of rolling bearings in normal state and establish a training set; S2. Construct a DCAE-RNN model, wherein the DCAE-RNN model includes a DCAE model and an RNN model, wherein the DCAE includes an input layer, three hidden layers and an output layer, wherein the first hidden layer and the second hidden layer are both composed of a stacked denoising autoencoder SDAE, and the third hidden layer is composed of a contraction autoencoder CAE; The DCAE model is used to extract the depth features of the rolling bearing vibration signal, and the RNN model is used to extract the temporal correlation characteristics between the depth features; S3. Obtain the optimal hyperparameter combination of the DCAE model through the Bayesian optimization algorithm; S4, using the training set to train the DCAE-RNN model to obtain a trained DCAE-RNN model; S5. Input the real-time vibration signal of the rolling bearing into the trained DCAE-RNN model, use the predicted value and the true value output by the model to calculate the residual and detect the early failure of the rolling bearing.

2. The method for predicting early failure of rolling bearings based on DCAE-RNN with Bayesian optimization according to claim 1 is characterized in that: In S1, a rolling bearing test bench is constructed, including an AC motor, a motor speed controller, a support shaft, two support bearings, a hydraulic loading system, and an acceleration sensor. The AC motor is used to provide a rotational speed, and the hydraulic loading system is used to provide a radial force. The rolling bearing test bench is used to perform accelerated degradation tests on the rolling bearings, and the vibration signals of the rolling bearings from operation to failure are collected using an acceleration sensor. The historical normal signals are used as training and test sets for the SDAE model, the CAE model, and the RNN model, and the overall data is used as the training and test sets for the DCAE-RNN model.

3. The method for predicting early failure of rolling bearings based on DCAE-RNN with Bayesian optimization according to claim 2 is characterized in that: The encoding and decoding equations of the SADE are as follows: in, is the encoded feature, f e (·) is the encoding function from the input layer to the output layer, a h is the result of linear transformation of input data, σ(·) is the Sigmoid activation function, and W xh Its weight value, is the input vector, b xh is the bias value, is the reconstructed input data, f d (·) is the decoding function from hidden layer to output function, W hr is the weight value, b hr is its bias value; Therefore, the reconstruction error of SDAE is: The reconstruction error of CAE is: Among them, J CAE is the total loss function of the contraction autoencoder, θ represents the weight, L(x,r) is the reconstruction error of the contraction autoencoder, λ is the contraction factor, is the contraction penalty term, J h (x) is the Jacobian matrix, h i is the output of the i-th encoder, i.e., the hidden layer feature, i=1,...,M, x j is the j-th input data, j=1,...,N.

4. The method for predicting early failure of rolling bearings based on DCAE-RNN with Bayesian optimization according to claim 3 is characterized in that: The S3 includes the following steps: S3.1, set the value range of each hyperparameter, the hyperparameters include the number of hidden layers, the number of neurons in the hidden layer, the optimization function and the activation function of the SDAE and CAE models, and then perform hyperparameter optimization for SDAE and CAE respectively through the following steps; S3.2, firstly, based on the value range of the set hyperparameters, a sampling algorithm is used to obtain candidate combinations of hyperparameters as a hyperparameter search space; S3.3, input candidate combinations, train the SDAE model and CAE model, and output the prediction accuracy of the candidate combinations on the validation set; S3.4, using the Gaussian process regression method, calculate the regression function between the prediction accuracy of different candidate combinations on the validation set; S3.5, taking the regression function as the prior distribution in the Gaussian process, traversing the hyperparameter search space, and obtaining the regression function with the highest accuracy as the objective function; S3.

6. Based on Gaussian process regression, the posterior probability distribution P(E|D) is constructed according to formula (7): Where P(E|D) is the probability of the training data D appearing given E; P(E) is the prior probability of E; P(D) is the probability of the data D appearing; S3.

7. Input the objective function, solve it iteratively through the Bayesian algorithm, and output the corresponding hyperparameter combination with the highest prediction accuracy as the optimal hyperparameter combination X n =[X1,X2,…,X n ]; S3.

8. Set up the SDAE and CAE models according to the optimal hyperparameter combination obtained; test the prediction accuracy of the entire DCAE-RNN model to determine whether the model meets the accuracy requirements. If it does, the model hyperparameters are X n ,If the accuracy requirements are not met, the Bayesian optimization range is expanded and the ,above steps are repeated to re-perform Bayesian optimization on the ,hyperparameters of SDAE and CAE.

5. The method for predicting early failure of rolling bearings based on DCAE-RNN with Bayesian optimization according to claim 4 is characterized in that: In S4, the data in the training set are rearranged into a data set Y with the same neuron dimension as the input layer of the DCAE model. According to formulas (1) to (3), the data set Y is used to train the SDAE after hyperparameter optimization. The output layer of SDAE is used as the input layer of CAE. According to formulas (4) to (6), the CAE after hyperparameter optimization is trained to obtain the trained DCAE model. The data reconstructed by the DCAE model is used as the input of the RNN to construct the DCAE-RNN model.

6. The method for predicting early failure of rolling bearings based on DCAE-RNN with Bayesian optimization according to claim 5 is characterized in that: In S5, residual calculation is performed between the predicted value and the true value of the real-time data to obtain a reconstruction error. The support vector data description SVDD is used to determine whether the reconstruction error exceeds a set threshold. If the threshold is exceeded, the rolling bearing has an early failure and an alarm is issued.

Images