Rolling bearing performance degradation prediction method and device based on SAE and TCN-Attention model
By combining SAE with TCN-Attention model, using the Sec activation function and Attention mechanism, the problem of gradient vanishing and high-dimensional data processing difficulties in rolling bearing performance decay prediction is solved, and rolling bearing performance decay prediction with high precision and high generalization capabilities is achieved.
Patent Information
- Application Number
- CN202210525015.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-14
- Publication Date
- 2024-11-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The prior art has gradient vanishing problems and difficulty in processing high-dimensional data in rolling bearing performance decay prediction, resulting in low prediction accuracy and long processing time, difficulty in memory time series for a long time, and inability to accurately predict the performance decay trend of mechanical equipment.
Using a combination of SAE and TCN-Attention model, the time-domain and frequency-domain feature indexes are extracted through CEEMDAN and wavelet semi-soft threshold noise reduction technology, the SAE model of the Sec activation function is used for feature fusion, and the rolling bearings are weighted by the Attention mechanism. Performance decay prediction.
The prediction accuracy and the sensitivity of the model to fault information are improved, the gradient disappearance problem is overcome, the influence of external factors is reduced, and a network model with high generalization ability is established, which can accurately predict the performance decay trend of rolling bearings.
Smart Images

Figure CN114841208B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of mechanical performance degradation prediction, and in particular to a rolling bearing performance degradation prediction method and device based on SAE and TCN-Attention models. Background Art
[0002] Now, with the increasing maturity of science and technology, the competitiveness of science and technology is increasing, and the development of industrial technology is inseparable from the country's science and technology. While science and technology are developing, mechanical equipment is developing towards intelligence and precision. As the running time of rotating machinery increases, its performance degradation will accelerate, and the probability of damage will be greatly increased. Therefore, it is very necessary to predict the performance degradation of mechanical equipment. Performance degradation prediction is one of the indispensable and important components of predictive maintenance, which has extraordinary significance for maintaining the life, health status and safe operation of the entire production line of mechanical equipment.
[0003] The internal structure of mechanical equipment is complex and diverse, and the working conditions are unknown. Under high load conditions, vulnerable parts such as bearings will be impacted to varying degrees, which will cause failures of rolling bearings and other parts over time. Generally speaking, performance degradation prediction is very important for detecting the performance status of equipment, which can detect and isolate early faults and determine the performance status of equipment.
[0004] At present, the performance degradation prediction methods for mechanical equipment are divided into two categories: model-driven and data-driven methods. The driving method mainly judges the performance degradation of mechanical equipment based on the changes in the physical information of the equipment. Its methods are mainly based on mechanics, stress, strain and other methods. Lei predicted the service life of the machine using a random process model method. Cubillo et al. proposed a physical model to predict the performance degradation of rotating machinery. Although the model-based method can analyze its essence and accurately predict the performance degradation trend of rotating machinery, it is difficult to establish some prescribed physical models in many engineering applications, which increases the difficulty of prediction. The data-driven method can monitor the mechanical equipment of interest in real time, and it is easy to update and adjust the model parameters and the performance degradation trend of the equipment. Liu et al. used a model-based particle filter method to monitor the performance degradation of rolling bearings for a series of problems caused by the running state of rolling bearings. It can be seen that although the model-driven prediction method can effectively predict the performance degradation trend of rotating machinery, it is difficult to process high-dimensional data and requires certain manual experience, which will cause problems such as low prediction accuracy and long processing time. Artificial intelligence methods can adaptively establish models in the case of unknown mechanical equipment status, and improve the prediction accuracy of the model by repeatedly training the model. Among them, recurrent neural networks and long short-term memory (LSTM) are often used to predict the performance degradation of mechanical equipment. Feng Li et al. proposed a quantum recursive encoder-decoder neural network to predict the performance degradation trend of mechanical equipment when traditional neural networks have low accuracy in predicting the performance degradation of rotating machinery. Qiaoping Tian et al. proposed a LSTM prediction model to monitor the performance degradation state of bearings and predict their remaining service life based on the impact of the operating state of bearings on the operation and maintenance costs of mechanical equipment. Experiments have shown that the prediction accuracy of the model is significantly improved. Mengfu He et al. proposed an LSTM model to predict the performance degradation of bearings in response to issues such as the impact of the operating state of bearings on the normal operation of mechanical systems. Deep learning algorithms have powerful time series processing capabilities and prediction models can adaptively adjust their own parameters and achieve optimization. They are widely used in the prediction of performance degradation of rotating mechanical equipment. However, these models have the problem of gradient disappearance and cannot memorize longer time series for a long time due to the large amount of data and limited model memory capacity when processing time series, thereby reducing the prediction accuracy of mechanical equipment performance degradation. Therefore, in order to meet the demand for the performance change trend of the monitored rolling bearings, a prediction method with high prediction accuracy and strong generalization ability is needed to predict them. Summary of the invention
[0005] In order to solve the above problems, the present invention provides a rolling bearing performance degradation prediction method and device based on SAE and TCN-Attention models to achieve accurate prediction of rolling bearing performance degradation.
[0006] On the one hand, the present invention provides a rolling bearing performance degradation prediction method based on SAE and TCN-Attention model, comprising the following steps:
[0007] S1. Obtaining the original vibration signal data of the rolling bearing, performing feature extraction after noise reduction, and obtaining a time domain feature decay index and a frequency domain feature decay index;
[0008] S2, inputting the time domain feature decay index and the frequency domain feature decay index into the SAE model to obtain a fused feature index;
[0009] S3. Input the fused feature indicators into the TCN-Attention model to obtain the rolling bearing performance degradation trend.
[0010] Furthermore, the method for performing noise reduction and feature extraction on the original signal vibration signal data is as follows:
[0011] S11, using CEEMDAN method to decompose the original vibration signal data into several IMF components;
[0012] S12. Setting the correlation coefficient threshold r s The wavelet semi-soft threshold denoising method is used to reduce the correlation coefficient value greater than the correlation coefficient threshold r s The IMF component at this time is used for noise reduction;
[0013] S13, the IMF component after denoising in step S12 and the correlation coefficient value are less than or equal to the correlation coefficient threshold r s The IMF components are reconstructed to obtain the time domain signal data and the frequency domain signal data of the reconstructed signal;
[0014] S14. Obtain a time domain characteristic decay index and a frequency domain characteristic decay index according to the time domain signal data and the frequency domain signal data.
[0015] Furthermore, the time domain feature decay index types include dimensional indicators and dimensionless indicators. The dimensional indicators include mean, variance, root mean square value, root amplitude, kurtosis, peak-to-peak value, skewness and peak value. The dimensionless indicators include waveform index, kurtosis index, peak index, margin index, skewness index and pulse index.
[0016] Furthermore, the SAE model includes an input layer, a first hidden layer, a second hidden layer and an output layer.
[0017] Furthermore, the activation function of the SAE model is the Sec function, specifically:
[0018]
[0019]
[0020] Among them, f1(x) is the Sec function, and s1(x) is the derivative of the Sec function.
[0021] Furthermore, the TCN-Attention model includes an input layer, a TCN layer, an Attention layer, a fully connected layer and an output layer.
[0022] Furthermore, the Attention layer calculates the weight value output by the TCN layer, and the weight calculation method includes:
[0023] Through the weight scoring function e ki Determine the ratio of the total hidden layer weight of the sequence output by the TCN layer to the weight of each hidden layer in the TCN layer output sequence. The weight scoring function is:
[0024] e ki =vtanh(Wh k +Uh i + b);
[0025] Where: h k is the total hidden layer weight of the sequence output by the TCN layer, h i is the weight of the i-th hidden layer in the output sequence, W, U, and b represent the total input weight coefficient in the Attention layer, the weight coefficient corresponding to the i-th hidden layer state, and the bias value respectively;
[0026] Calculate the total proportion of the hidden layer at the current moment
[0027] Calculate the feature vector at the next moment by weighting
[0028] The second aspect of the present invention provides a rolling bearing performance degradation prediction device based on SAE and TCN-Attention models, comprising:
[0029] An index acquisition unit acquires the original vibration signal data of the rolling bearing, performs feature extraction after noise reduction, and obtains a time domain feature decay index and a frequency domain feature decay index;
[0030] A fusion unit, inputting the time domain feature decay index and the frequency domain feature decay index into the SAE model to obtain a fused feature index;
[0031] The prediction unit selects excellent feature indicators from the fused feature indicators, inputs the excellent feature indicators into the TCN-Attention model, and obtains the rolling bearing performance degradation trend.
[0032] As described above, compared with the prior art, the present invention has the following effects:
[0033] 1. This application adopts the SAE fusion method to realize the fusion of various characteristic indicators of rolling bearings, and overcomes the gradient vanishing problem existing in traditional neural networks through an improved activation function.
[0034] 2. This application uses the TCN-Attention prediction model to predict the decline of rolling bearing performance. The Attention model can give corresponding weights to important information, improve the sensitivity of the model to fault information, and select different model parameters to reduce the impact of external factors, and establish a network model. Through two public data sets, it is verified and compared with multi-layer perceptron, LSTM, TCN, and Attention-LSTM. The TCN-Attention model has high prediction accuracy and is more sensitive to fault information.
[0035] 3. The Attention layer of the present invention calculates the intermediate output state weight of the TCN model through a weight scoring function, increases the weighting of important information, and improves prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 It is a flow chart of a rolling bearing performance degradation prediction method according to a specific embodiment of the present invention;
[0037] Figure 2 A flow chart of a method for performing noise reduction and feature extraction on original signal vibration signal data according to a specific embodiment of the present invention;
[0038] Figure 3 is the original vibration signal of a specific embodiment of the present invention, Figure 3 a is the time domain diagram, Figure 3 b is the frequency domain diagram;
[0039] Figure 4 The original vibration signal of the specific embodiment of the present invention is decomposed by CEEMDAN to obtain 10 IMF components;
[0040] Figure 5 is a signal after CEEMDAN-WSST denoising in a specific embodiment of the present invention, Figure 5 a Time domain diagram, Figure 5 b is the frequency domain diagram;
[0041] Figure 6are the dimensioned index and the dimensionless index obtained by feature extraction of the noise-reduced signal in a specific embodiment of the present invention, Figure 6 a is the mean indicator, Figure 6 b is the variance index, Figure 6 c is the absolute average index, Figure 6 d is the skewness index, Figure 6 e is the minimum value index, Figure 6 f is the kurtosis index, Figure 6 g is the root mean square index, Figure 6 h is the peak-to-peak value indicator, Figure 6 i is the margin index, Figure 6 j is the waveform indicator, Figure 6 k is the kurtosis index, Figure 6 l is the peak index, Figure 6 m is the skewness index, Figure 6 n is the pulse index, Figure 6 o is the average frequency, Figure 6 p is the standard deviation frequency, Figure 6 q is the root mean square frequency, Figure 6 r is the center frequency;
[0042] Figure 7 It is a schematic diagram of the principle structure of the SAE network of a specific embodiment of the present invention;
[0043] Figure 8 It is the characteristic index after SAE model fusion in the specific embodiment of the present invention;
[0044] Fig. 9 This is a rolling bearing performance stage division diagram for a specific embodiment of the present invention, Fig. 9 a is the trend curve, Fig. 9 b is the residual curve;
[0045] Fig.10 It is a principle structure diagram of the TCN-Attention prediction model of a specific embodiment of the present invention;
[0046] Fig.11 It is a prediction effect diagram of the TCN-Attention model, multi-layer perceptron, LSTM neural network, TCN neural network, and Attention-LSTM model of a specific embodiment of the present invention. Fig.11 a is the prediction effect diagram of the TCN-Attention model. Fig.11 b is the prediction effect diagram of the multi-layer perceptron. Fig.11 c is the prediction effect diagram of LSTM neural network. Fig.11 d is the prediction effect diagram of TCN neural network. Fig.11 e is the prediction effect diagram of the Attention-LSTM model. DETAILED DESCRIPTION
[0047] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.
[0048] It should be noted that the illustrations provided in the following embodiments are only used to illustrate the basic concept of the present invention in a schematic manner, and thus the illustrations only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0049] In a specific embodiment, a rolling bearing performance degradation prediction method based on SAE and TCN-Attention model is provided. Figure 1 As shown, the following steps are included:
[0050] S1. Obtaining the original vibration signal data of the rolling bearing, performing feature extraction after noise reduction, and obtaining a time domain feature decay index and a frequency domain feature decay index;
[0051] In a specific embodiment, the time domain diagram and frequency domain diagram of the original vibration signal are as follows: Figure 3 As shown, the original vibration signal is collected by an acceleration sensor, and will be interfered by different working conditions during the collection process, so that the collected signal contains interference signals. The existing signal noise reduction method will have problems of modal aliasing and loss of extracted valuable information during the noise reduction process. In order to solve the above problems, in a specific embodiment, a method for noise reduction and feature extraction using the original signal vibration signal data is as follows Figure 2 As shown, the effective information in the signal can be fully preserved to a great extent, and has a high signal-to-noise ratio and a small root mean square error. The above noise reduction and feature extraction methods include:
[0052] S11, using the effect and other issues, it is proposed to decompose the original vibration signal data into several IMF components based on the complete set empirical mode decomposition (CEEMDAN) method. The error caused by signal reconstruction can be avoided by decomposing it through CEEMDAN.
[0053] In this embodiment, the original vibration signal is decomposed into 10 IMF components, such as Figure 4 shown.
[0054] S12. Setting the correlation coefficient threshold rs The wavelet semi-soft threshold (WSST) denoising method is used to reduce the correlation coefficient value greater than the correlation coefficient threshold r s The IMF component at this time is used for noise reduction;
[0055] The correlation coefficient r is the Pearson correlation coefficient. The size of the Pearson correlation coefficient r can be used to determine the proportion of noise components in each IMF. The larger the correlation coefficient r is, the more noise signals it contains. The correlation coefficient r is used to filter out IMF components with higher noise components for further noise reduction, so that effective information is fully retained. In this embodiment, the correlation coefficient threshold r is set s The Pearson correlation analysis method is used to perform correlation analysis on the IMF decomposed by CEEMDAN and the original data and calculate the corresponding correlation coefficient value, as shown in Table 1:
[0056] Table 1 Correlation coefficient values of various IMFs
[0057]
[0058] S13, the IMF component after denoising in step S12 and the correlation coefficient value are less than or equal to the correlation coefficient threshold r s The IMF components are reconstructed to obtain the reconstructed signal time domain signal data and frequency domain signal data.
[0059] As can be seen from Table 1, the correlation coefficient between IMF1 and IMF2 is higher than 0.6. The IMF with a correlation coefficient greater than 0.6 is further denoised using the wavelet semi-soft threshold. Then the denoised IMF is reconstructed with other IMFs to complete the signal denoising process. The denoised signal is obtained as follows: Figure 5 As shown by Figure 5 It can be seen that the noise reduction effect of the high-frequency part of the signal after denoising is very obvious. A large amount of noise is removed, while the effective value information is retained, achieving the ideal noise removal effect.
[0060] S14. Obtain a time domain characteristic decay index and a frequency domain characteristic decay index according to the time domain signal data and the frequency domain signal data.
[0061] The original vibration signal includes a time domain feature decay index and a frequency domain feature decay index. The time domain feature decay index types include dimensional indexes and dimensionless indexes. The dimensional indexes include mean, variance, root mean square value, root square amplitude, kurtosis, peak-to-peak value, skewness and peak value. The dimensionless indexes include waveform index, kurtosis index, peak index, margin index, skewness index and pulse index. The time domain feature decay index and the frequency domain feature decay index are as follows: Figure 6 As shown, Figure 6It can be seen that the performance status of the rolling bearing reflected by the above indicators is not very obvious. Among them, the variance, peak-to-peak value and root amplitude have good monotonicity and clearly represent the late failure of the rolling bearing, but fail to express the early failure of the bearing well; the margin index, skewness index, standard deviation frequency and center frequency and other indicators have a lot of redundant information and the information is messy; and the remaining performance degradation indicators cannot well represent the operating status of the rolling bearing throughout its life cycle and contain a lot of redundant information. It can be seen that a single indicator cannot well represent the performance change status of the bearing, and the above indicators need to be integrated through the SAE model.
[0062] S2, inputting the time domain feature decay index and the frequency domain feature decay index into the SAE model to obtain a fused feature index;
[0063] The signals collected by rolling bearing monitoring are rich in content, interconnected and mixed with the influence of uncertain factors. The fault form of the bearing can be characterized by each characteristic indicator, but the fault forms are diverse, and different characteristic indicators characterize the motion state of the bearing differently. Therefore, how to use a characteristic indicator with good comprehensive performance to characterize the performance change trend of the bearing is a key technical issue of feature fusion. At present, common feature fusion methods include wavelet transform fusion, evidence theory, Kalman filter, principal component analysis, kernel principal component analysis (KPCA), local linear embedding (LLE) and other methods. However, the existing fusion algorithm is not effective for nonlinear data fusion and is subject to data distribution, parameter selection and sensitivity to noise. Stacked Auto-encoder (SAE), as a model of deep learning, belongs to a type of unsupervised learning, and can achieve effective expression of a large number of unlabeled data features. When SAE processes high-dimensional data, it can extract high-value information without losing important information and thus achieve information fusion.
[0064] In a specific embodiment, the structure of the SAE model is as follows Figure 7 As shown, it includes an input layer, a first hidden layer, a second hidden layer and an output layer.
[0065] In the prior art, the SAE model usually uses the sigmoid function as the activation function, which is very likely to cause the gradient vanishing problem when processing long time series, affecting the effect of feature extraction, and thus failing to well characterize the performance change trend of the rolling bearing. In order to solve this problem, in a specific embodiment, the activation function is the Sec function, specifically:
[0066]
[0067]
[0068] Among them, f1(x) is the Sec function, and s1(x) is the derivative of the Sec function.
[0069] The minimum value of the derivative of the Sec function is 0.5, so the gradient vanishing problem will not occur in the SAE training process due to the appearance of zero; and the Sec function has no power operation, which will greatly shorten the calculation time and accelerate its convergence; the output average value of the Sec function is zero, which will reduce the impact on the subsequent neural network output.
[0070] The method for performing feature fusion of the SAE model includes:
[0071] S21, build Figure 2 SAE model shown;
[0072] S22, normalizing the time domain feature index and the frequency domain feature index, and the normalized data is the input layer data;
[0073] S23, the data of the input layer is input into the first hidden layer for parameter training and calculation of the output of this layer;
[0074] S24, using the output of the first hidden layer as the input of the second hidden layer for training parameters, thereby obtaining the output of the layer;
[0075] S25, using the back propagation algorithm to repeat S22-S23 so that the parameters in all layers are optimized and the results are optimal, completing T iterations of SAE;
[0076] S26, by retaining the training results of the weights of all layers and setting a weight threshold τ, connecting those above the threshold τ and discarding other connections;
[0077] S27, by analyzing all output results of the neural network, obtaining the final neuron weight matrix;
[0078] S28. Obtain corresponding features by analyzing the retained neuronal connections.
[0079] The characteristic indicators after SAE model fusion are as follows: Figure 8 As shown, it can be seen that the fused characteristic curve has a certain monotonicity, which can better reflect the performance degradation trend of the rolling bearing, especially the early abnormal situation can be clearly seen.
[0080] The performance change trend of rolling bearings is a random process. Under normal conditions, the bearing motion performance is stable and the vibration signal amplitude fluctuates slightly. Once the bearing is damaged, its motion state is unstable at this stage, and it is impossible to grasp how the bearing's operating trend changes at each stage. A characteristic indicator with good comprehensive performance is needed to characterize the changes in the bearing at each stage. Characteristic indicators with good comprehensive performance should have the same uniformity with the performance degradation of mechanical equipment, universality, and generalization ability to resist interference. This embodiment evaluates the comprehensive performance through linear weighting of three evaluation indicators: correlation, monotonicity, and robustness. Characteristic indicators with good comprehensive performance are excellent characteristic indicators. The linear weighted algorithm is:
[0081] maxE=e1Corr(X)+e2Mon(X)+e3Rob(X)
[0082]
[0083] Among them: E represents the linear weighted value of the three evaluation indicators, Corr(X) represents correlation, Mon(X) represents monotonicity, Rob(X) represents robustness, e i The weighted value E represents the attribute weight of the indicator. The weighted value E is positively correlated with the three evaluation indicators. The larger the E is, the better the comprehensive performance of the state detection indicator is, and it can better reflect the changing trend of the rolling bearing performance degradation.
[0084] The correlation refers to the degree of correlation between the health monitoring sequence of the mechanical equipment and the performance change trend of the equipment, and the formula is:
[0085]
[0086] The monotonicity refers to the consistency of the performance degradation trend of mechanical equipment. The performance degradation process of any mechanical equipment during use is inevitable, so the acquired characteristic index should have a monotonic change trend. The formula is:
[0087]
[0088] The robustness refers to the tolerance of characteristic indicators under abnormal conditions, and its formula is:
[0089]
[0090] Where: K represents the total number of sampling points, x i represents the trend sequence of the acquired condition monitoring indicators, δ(·) represents the step function, X=(x1,x2,...,x k ) represents the acquired characteristic sequence of status monitoring indicators, T = (t1, t2, ..., t k) represents the corresponding time sequence.
[0091] The essence of each characteristic indicator is different, and it also has different attribute weights. In order to avoid external interference, the weighting formula method is selected to calculate the weight. The calculation method is as follows:
[0092]
[0093] Where: w i1 =1, n is the number of attributes, i represents the i-th attribute, and j is the queuing level.
[0094] The time domain evaluation index, frequency domain evaluation index and fused feature index of the above-mentioned embodiment are respectively subjected to comprehensive performance calculation. According to the final weighted calculation result, it can be known that the feature index fused by the SAE model has the best comprehensive performance, that is, the feature index obtained after fusion contains the most valuable information.
[0095] The process of rolling bearing performance degradation is a random process, and the performance degradation index sequence is also random. In order to avoid the occurrence of these randomness, it is necessary to perform smooth trend analysis on the fused state detection characteristic index, decompose the fused state detection characteristic index curve into the trend sequence and residual sequence of the corresponding characteristic index sequence, and better analyze the change of rolling bearing performance state.
[0096] The fused feature index sequence is composed of a trend sequence and a residual sequence. The specific formula is as follows:
[0097] X=X+X re
[0098] Where: X=(x1,x2,...,x n )、X=(x1,x2,...,x n ),X re =(x1-x1,x2-x2,...,x n -x n ) are the characteristic indicator sequence, trend sequence, and residual sequence after fusion.
[0099] The fixed window averaging method is used to process the fused feature sequence, and the following is obtained: Fig. 9 The trend curve and residual curve shown in the figure are as follows. In this embodiment, the size of the window ΔT is used to obtain the value of r unit time periods before and after the moment, which is the trend value at the moment.
[0100] ΔT=2r
[0101]
[0102] Where: x Tt is the trend value of the rolling bearing at time t, xt is the eigenvalue of the fused state monitoring at time t, r is the radius of the window ΔT;
[0103] According to the trend curve and residual curve shown, the whole life cycle of rolling bearing can be accurately divided into four stages. The first stage is the normal operation stage, in which the bearing performance is good; in the second stage, the bearing is slightly damaged, the trend value and residual value increase, and the residual value shows that the system fluctuates and the bearing operation is unstable; the third stage is the sub-healthy stable operation stage, the system has self-regulation ability, and the trend value and residual value show that the bearing operation tends to be stable; in the fourth stage, the bearing is seriously damaged and cannot reach a stable operation state, and the bearing fails.
[0104] S3. Input the fused indicators into the TCN-Attention model to obtain the rolling bearing performance degradation trend.
[0105] Temporal convolutional network (TCN) is a sequence prediction model proposed on the basis of convolutional neural network, which aims to learn the time dependency in the input sequence and regards the convolutional network as the natural starting point for sequence modeling tasks. The TCN model has a powerful receptive field and can effectively deal with the above-mentioned problems, but giving the same weight to important information in the output layer can easily cause important information to be forgotten, which leads to low model prediction accuracy. The attention mechanism first appeared in the field of vision. In a specific embodiment, the Attention mechanism is combined with the TCN neural network to predict the decline trend of bearing performance, which can increase the ability to handle complex problems of long time series and the generalization ability.
[0106] like Fig.10 As shown, the TCN-Attention model of this embodiment includes an input layer, a TCN layer, an Attention layer, a fully connected layer and an output layer.
[0107] The TCN layer in this embodiment includes a WeightNorm module, an ELU module and a Dropout module. The WeightNorm module is used to increase the computational efficiency in the calculation. The Dropout module can regularize the network and prevent the network from overfitting during operation. In addition, a 1×1 convolution is set in the residual module to prevent changes in the sequence dimension during the input and output processing.
[0108] The Attention layer calculates the weight value of the TCN layer output to solve the corresponding weight distribution problem when the TCN layer outputs. The input of the Attention layer corresponds to the output of the TCN prediction model. The Attention layer includes the Permute layer, the Dense layer and the Multiply layer. The purpose of the Permute layer is to perform the corresponding dimensional transformation. The Multiply layer multiplies the weight calculated by the Attention layer with the output of the TCN layer, thereby realizing the distribution and weighting process of the corresponding weights.
[0109] The method for the Attention layer to calculate the weight value output by the TCN layer includes:
[0110] Through the weight scoring function e ki Determine the ratio of the total hidden layer weight of the sequence output by the TCN layer to the weight of each hidden layer in the TCN layer output sequence. The weight scoring function is:
[0111] e ki =vtanh(Wh k +Uh i +b)
[0112] Where: h k is the total hidden layer weight of the sequence output by the TCN layer, h i is the weight of the i-th hidden layer in the output sequence, v is the parameter of the Attention layer, W, U, and b represent the total input weight coefficient in the Attention layer, the weight coefficient corresponding to the i-th hidden layer state, and the bias value respectively;
[0113] Calculate the total proportion of the hidden layer at the current moment
[0114] Calculate the feature vector at the next moment by weighting
[0115] The fully connected layer is used to process the output of Attention to achieve performance degradation prediction of rolling bearings, and the output layer is used to output the prediction results.
[0116] The method for constructing the TCN-Attention prediction model includes:
[0117] S31, build Figure 4 The TCN-Attention model shown;
[0118] S32. Initialize the weights of the TCN-Attention model and set parameters such as the number of iterations, expansion coefficient, and convolution kernel size as shown in Table 2.
[0119] Table 2 TCN-Attention prediction model parameters
[0120]
[0121] S33, input the training set into the TCN-Attention model, obtain the output error, and back-propagate the error to update the network parameters;
[0122] S34. Repeat step 3 until the network parameter update reaches the optimization requirement and the training is completed.
[0123] S35. Input the test set into the TCN-Attention model to verify the prediction results of the rolling bearing performance degradation trend.
[0124] In this embodiment, the fused data obtained in step S2 is used as a prediction and to verify the prediction effect of this embodiment. The first 67% of the data length of the fused data sequence is used as a training set, and the last 33% of the data length is used as a test set. They are respectively input into the TCN-Attention model, the multi-layer perceptron, the LSTM model, the TCN model and the Attention-LSTM model, and the effect of the prediction curve is obtained as shown in FIG. Fig.11 shown. Fig.11 In each figure, the dark lines are the actual trend curves of rolling bearings, and the light lines are the trend curves of the training set and the test set. The left side of the vertical dividing line is the training set curve, and the right side is the test set trend curve. Fig.11 It can be seen that the prediction curves of the multilayer perceptron, LSTM and TCN models all have a certain degree of deviation. In particular, the prediction result of the multilayer perceptron has a large error. It can be clearly seen from the figure that the prediction curve deviates a lot from the true curve. The Attention-LSTM prediction model has better prediction results than the above three models, but in the later stage, its prediction curve and the true curve have obvious lack of fit. TCN-Attention has a good prediction effect in each stage of the bearing, and the prediction curve has a good fit with the true curve, especially the trend prediction in the later stage of the whole life cycle is the best. The model has a certain prediction stability and robustness.
[0125] By calculating the error MAE and RMSE values of each prediction model, as shown in Table 3, it can be seen that the prediction error of the multilayer perceptron is the largest and the model performance is unstable, while the prediction error based on the TCN-Attention model is the smallest and the prediction accuracy is high, which can accurately predict the performance degradation trend of the rolling bearing.
[0126] Table 3 Model prediction results
[0127]
[0128] As can be seen from the table above, the TCN-Attention prediction model of this embodiment performs best in terms of prediction effect and has strong generalization ability. RMSE and e MAE All of them are smaller than the other four prediction models. Among them, the e of the TCN-Attention model RMSE and e MAE Compared with multilayer perceptron, LSTM, Attention-LSTM and TCN, the accuracy is reduced by 54.99%, 28.49%, 17.19%, 22.16% and 69.34%, 51.60%, 27.97% and 44.48% respectively. Therefore, after experimental comparison and verification, the TCN-Attention prediction model greatly improves the prediction accuracy of bearing performance degradation.
[0129] In a specific embodiment, a rolling bearing performance degradation prediction device based on SAE and TCN-Attention model is provided, comprising:
[0130] An index acquisition unit acquires the original vibration signal data of the rolling bearing, performs feature extraction after noise reduction, and obtains a time domain feature decay index and a frequency domain feature decay index;
[0131] A fusion unit, inputting the time domain feature decay index and the frequency domain feature decay index into the SAE model to obtain a fused feature index;
[0132] The prediction unit inputs the fused feature indicators into the TCN-Attention model to obtain the rolling bearing performance degradation trend.
[0133] Each unit in the prediction device can be implemented by a hardware module, a software program, or a combination of hardware and software. The specific working process of the instruction acquisition unit, the fusion unit and the prediction unit is as described in the above embodiment and will not be repeated here.
[0134] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.
Claims
1. A rolling bearing performance degradation prediction method based on SAE and TCN-Attention model, characterized in that: The steps include: S1. Obtaining the original vibration signal data of the rolling bearing, performing feature extraction after noise reduction, and obtaining a time domain feature decay index and a frequency domain feature decay index; S2. Input the time domain feature decay index and the frequency domain feature decay index into the SAE model to obtain a fused feature index; the activation function of the SAE model is the Sec function, specifically: ; ; in, is the Sec function, is the derivative of the Sec function; S3, input the fused feature index into the TCN-Attention model to obtain the rolling bearing performance degradation trend; The TCN-Attention model includes an input layer, a TCN layer, an Attention layer, a fully connected layer, and an output layer; The Attention layer calculates the weight value output by the TCN layer. The weight calculation method includes: Through the weighted scoring function Determine the ratio of the total hidden layer weight of the sequence output by the TCN layer to the weight of each hidden layer in the TCN layer output sequence. The weight scoring function is: ; in: is the total hidden layer weight of the sequence output by the TCN layer, The output sequence is i hidden layer weights, , , Represents the total input weight coefficient in the Attention layer, corresponding to the i The weight coefficient and bias value corresponding to the hidden layer state; Calculate the total proportion of the hidden layer at the current moment ; Calculate the feature vector at the next moment by weighting .
2. The rolling bearing performance degradation prediction method based on SAE and TCN-Attention model according to claim 1 is characterized in that: Method for noise reduction and feature extraction of the original signal vibration signal data: S11, using CEEMDAN method to decompose the original vibration signal data into several IMF components; S12. Setting the correlation coefficient threshold r s , the wavelet semi-soft threshold denoising method is used to reduce the correlation coefficient value greater than the correlation coefficient threshold r s The IMF component at this time is used for noise reduction; S13, the IMF component after denoising in step S12 and the correlation coefficient value are less than or equal to the correlation coefficient threshold r s The IMF components are reconstructed to obtain the time domain signal data and the frequency domain signal data of the reconstructed signal; S14. Obtain a time domain characteristic decay index and a frequency domain characteristic decay index according to the time domain signal data and the frequency domain signal data.
3. The rolling bearing performance degradation prediction method based on SAE and TCN-Attention model according to claim 2 is characterized in that: The time domain feature decay index types include dimensional indexes and dimensionless indexes. The dimensional indexes include mean, variance, root mean square value, root square amplitude, kurtosis, peak-to-peak value, skewness and peak value. The dimensionless indexes include waveform index, kurtosis index, peak index, margin index, skewness index and pulse index.
4. The rolling bearing performance degradation prediction method based on SAE and TCN-Attention model according to claim 1 is characterized in that: The SAE model includes an input layer, a first hidden layer, a second hidden layer and an output layer.
5. A rolling bearing performance degradation prediction device based on SAE and TCN-Attention models, characterized in that: include: An index acquisition unit acquires the original vibration signal data of the rolling bearing, performs feature extraction after noise reduction, and obtains a time domain feature decay index and a frequency domain feature decay index; The fusion unit inputs the time domain feature decay index and the frequency domain feature decay index into the SAE model to obtain a fused feature index; the activation function of the SAE model is the Sec function, specifically: ; ; in, is the Sec function, is the derivative of the Sec function; A prediction unit selects excellent feature indicators from the fused feature indicators, inputs the excellent feature indicators into the TCN-Attention model, and obtains the rolling bearing performance degradation trend; The TCN-Attention model includes an input layer, a TCN layer, an Attention layer, a fully connected layer, and an output layer; The Attention layer calculates the weight value output by the TCN layer. The weight calculation method includes: Through the weighted scoring function Determine the ratio of the total hidden layer weight of the sequence output by the TCN layer to the weight of each hidden layer in the TCN layer output sequence. The weight scoring function is: ; in: is the total hidden layer weight of the sequence output by the TCN layer, The output sequence is i hidden layer weights, , , Represents the total input weight coefficient in the Attention layer, corresponding to the i The weight coefficient and bias value corresponding to the hidden layer state; Calculate the total weight of the hidden layer at the current moment ; Calculate the feature vector at the next moment by weighting .
