A rotating machinery residual service life prediction method based on time convolution network

Abstract

The application belongs to the technical field of fault prediction and health management, and discloses a rotating machinery residual service life prediction method based on a time convolution network, which comprises the following steps: (1) collecting vibration signal data of a single or multiple rotating machines in an operating stage of a full life cycle, and extracting time domain features, frequency domain features and time-frequency features in a specific direction from the vibration signal data of the full life cycle; (2) selecting effective features from the extracted features by using a comprehensive evaluation index Cri; (3) applying a t-SNE algorithm to compress the selected effective features, and then using a DBSCAN algorithm to adaptively divide the degradation stages of the rotating machines; (4) establishing a prediction model based on TCN and training the prediction model, and using the trained prediction model to predict the residual service life of the rotating machines. The application can better capture the degradation process of the rotating machines, and significantly reduce the influence of noise and fluctuations in the sensor.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of fault prediction and health management, and more particularly to a rotating machinery remaining useful life prediction method based on a time convolution network. BACKGROUND

[0002] Fault prediction and health management aims to evaluate the health state of equipment or system by applying a series of models and methods, and provides timely and effective maintenance decision-making solutions for enterprise managers based on the health state, so as to improve the safety and reliability of the system. Remaining useful life (RUL) prediction is one of the core contents of fault prediction and health management, which mainly focuses on estimating the health state before system failure by using the internal structure of the equipment and the sensor data collected during the operation process, which is of great significance to avoid disrupting production plans due to sudden failure downtime, eliminate safety hazards and major accidents, reduce equipment maintenance costs, improve enterprise production efficiency and optimize system operation and maintenance management, etc.

[0003] Rotating machinery is one of the important components of many large-scale mechanical equipment. The RUL prediction of rotating machinery is of great importance to improve the safety and reliability of large-scale mechanical equipment. Generally speaking, the RUL prediction methods mainly include the following two types: mechanism model-based method and data-driven method. The mechanism model-based method mainly establishes a physical failure model according to the equipment degradation process, which needs rich expert experience knowledge and specific analysis for specific equipment, resulting in poor generalization ability. In recent years, the rapid development of artificial intelligence and sensor technology has promoted the development of data-driven methods. The data-driven method aims to mine the potential mapping relationship between system degradation trend and state monitoring data. Since it does not need rich expert experience knowledge, only needs to process and extract effective features that can represent the system degradation from the collected sensor data, the generalization ability of this method is strong, and it has been widely used to solve the RUL prediction problem. The deep learning method has good nonlinear fitting ability and sequence data processing ability, which can significantly reduce the complexity of the RUL prediction task and improve the prediction performance, and has wide application prospect. The existing deep learning methods mainly carry out RUL prediction work from the aspects of data collection, data preprocessing, model establishment, model training and online prediction, and mainly have the following shortcomings: first, the established full life cycle unified degradation model is difficult to well fit the whole process of rotating machinery degradation. Specifically, the early rotating machinery will run smoothly for a long time, then slowly degrade at almost constant rate, and finally the rotating machinery enters the rapid degradation stage and runs to failure in a short time. Different degradation modes are significantly different, so it is necessary to establish corresponding degradation models for different degradation modes to better fit the whole process of rotating machinery degradation; second, for different degradation stages of rotating machinery, the mapping relationship between the extracted features and the health state is difficult to match; finally, some commonly used deep learning models have limitations in solving the RUL prediction problem, such as convolutional neural network has good feature extraction ability, but due to the limitation of convolution kernel size, it cannot well capture long-term sequence dependence. SUMMARY

[0004] In view of the above defects or improvement needs of the prior art, the present application provides a rotating machinery remaining useful life prediction method based on time convolution network, which solves the problems of low prediction accuracy caused by underfitting of equipment degradation process, difficulty in matching extracted features and equipment health state, and incompatibility of feature extraction ability and long sequence data processing ability in RUL prediction.

[0005] To achieve the above object, according to one aspect of the present application, a rotating machinery remaining useful life prediction method based on time convolution network is provided, which mainly includes the following steps:

[0006] (1)collecting vibration signal data of a single or multiple rotating machines in the whole life cycle in the running stage, and extracting time domain features, frequency domain features and time-frequency features in a specific direction from the vibration signal data in the whole life cycle;

[0007] (2) selecting effective features from the extracted features by using a comprehensive evaluation index Cri; wherein the comprehensive evaluation index Cri is: Cri=ω1Corr+ω2Mon-ω3Dis, wherein Dis, Corr and Mon are dynamic time warping distance index, correlation coefficient index and monotonicity index respectively, and the weight coefficients thereof are ω3, ω1 and ω2 respectively;

[0008] (3) applying t-SNE algorithm to compress the selected effective features, and then using DBSCAN algorithm to adaptively divide the degradation stages of the rotating machines;

[0009] (4) establishing a prediction model based on TCN and training the prediction model, and using the trained prediction model to predict the remaining useful life of the rotating machines.

[0010] Further, the prediction model comprises an input layer, a TCN layer and a regression layer; in the input layer, the fused effective features are input; in the TCN layer, a plurality of convolution units are added in the residual block structure, and the structure of each convolution unit is dilated causal convolution + weight normalization + ReLU activation function + Dropout; in the regression layer, a flattened layer and a fully connected layer are established to fuse all local features extracted by the TCN layer, and finally the RUL prediction value is output.

[0011] Further, the time domain, frequency domain and time-frequency features are extracted from the collected whole life cycle vibration signals along the horizontal direction and the vertical direction, first, the time domain statistical features of the vibration signals are extracted; second, the discrete time domain signals are converted into frequency domain signals by using fast Fourier transform to extract frequency domain features; finally, the time-frequency features of the vibration signals are extracted by using the Hilbert-Huang transform adaptive time-frequency analysis method.

[0012] Further, the selected features are compressed from high-dimensional space to low-dimensional space by using t-SNE algorithm, in the high-dimensional space and the low-dimensional space, the distance between state points is converted into the corresponding joint probability distribution by using Gaussian distribution and t distribution respectively, then the Kullback-Leibler divergence is used to measure the difference between the two distributions, and the gradient descent method is used to optimize the KL divergence, finally the feature state points far away from each other in the original high-dimensional space after mapping are farther away from each other, and the feature state points close to each other are closer to each other.

[0013] Furthermore, after feature compression, the distances between feature state points in the same and different degradation stages are closer and farther, respectively, which further clarifies the boundaries between different degradation stages and facilitates the subsequent adaptive division of degradation stages.

[0014] Furthermore, the DBSCAN algorithm can adaptively identify degradation patterns according to the actual health status of rotating machinery, and automatically divide the entire life cycle of rotating machinery into multiple degradation stages based on the degradation patterns.

[0015] Furthermore, the full life cycle feature sequence is X i ={x 1,2 ,…, T}, where T is the length of the feature sequence, and the cluster marker array m is introduced i :

[0016]

[0017] Where n is the number of clusters, thus the cluster label array is obtained X i It is divided into n clusters and noise point sets, that is, the entire life cycle is divided into n degradation stages, and the corresponding degradation stage labels S are generated label ; Then, the selected effective feature vector is compared with S label The features are combined to form a feature label pair, which is input into the support vector machine classification model for training. Subsequently, the valid feature vector of the test set is input into the trained SVM classification model to obtain the degradation stage label of the test set.

[0018] Furthermore, the t-SNE algorithm is first used to map the selected effective features from high-dimensional space to low-dimensional space; secondly, based on the distribution differences between different health states generated by the t-SNE algorithm, the DBSCAN algorithm is used to adaptively divide the entire life cycle of the training set into different degradation stages; then, the effective features of the training set are fused, combined with the corresponding degradation stage labels to form label pairs, which are input into the SVM classifier, and the SVM classifier is trained; finally, the effective features of the test set are fused and input into the trained SVM classifier to adaptively generate degradation stage labels for the test set.

[0019] Furthermore, the entire life cycle is adaptively divided into two stages: smooth degradation and rapid degradation. The RUL of the two stages is then predicted using a TCN-based prediction model.

[0020] Furthermore, the true value of RUL is set as the percentage of rotating machinery degradation, and the true value label of RUL at the t-th sampling point is:

[0021]

[0022] wherein RUL t and RUL0 respectively represent the order t of the sampling point and the total number of sampling points in the whole life cycle, and the input layer is normalized, and the corresponding formula is:

[0023]

[0024] wherein x i (t) and respectively represent the value of the i-th original and normalized feature sequence at time t, and are the average value, maximum value and minimum value of the i-th feature sequence value at all times.

[0025] Overall, compared with the prior art, the rotating machinery residual useful life prediction method based on the time convolution network provided by the present application mainly has the following beneficial effects:

[0026] 1. The present application selects the time domain, frequency domain and time-frequency features extracted by constructing a comprehensive evaluation index, i.e. a linear combination of dynamic time warping distance index, correlation index and monotonicity index, eliminates features with high similarity measurement, low time correlation and poor monotonicity, and retains the remaining effective features, which can better capture the degradation process of the rotating machinery and significantly reduce the influence of noise and fluctuations in the sensor.

[0027] 2. The present application applies the t-SNE algorithm to compress the retained effective features, and then uses the DBSCAN algorithm to process the compressed feature state points, which can realize adaptive identification of the degradation mode without expert prior knowledge, only according to the actual health state of the rotating machinery, and automatically divide the whole life cycle of the equipment into multiple degradation stages based on the degradation mode, so that the health state of the equipment is more matched with the divided degradation stages, thereby better fitting the whole process of the degradation of the rotating machinery and enhancing the RUL prediction accuracy.

[0028] 3. The improved TCN model designed by the present application can realize accurate prediction of RUL, and the network model has good feature extraction capability and long sequence data processing capability, which helps to accelerate the convergence speed of the network, enhance the generalization ability of the model and improve the RUL prediction accuracy.

[0029] 4. The present application first extracts time domain features from the vibration signal, then converts the time domain features into frequency domain features through Fourier transform, and then extracts frequency domain features, and finally extracts time-frequency features by using the Hilbert-Huang transform adaptive time-frequency analysis method; by extracting the above three types of features, the degradation information of the equipment can be more comprehensively reflected. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 is a flow chart of the time convolution network-based rotating machinery remaining useful life prediction method provided according to the preferred embodiment of the present application;

[0031] Figure 2 is a visualization diagram of the full life cycle vibration signal of the specific example extracted along the horizontal direction;

[0032] Figure 3 (a) and (b) in are visualization diagrams of the comprehensive evaluation index values of the time domain features, frequency domain features and time-frequency features of the specific example extracted along the horizontal direction and the vertical direction, respectively;

[0033] Figure 4 (a) and (b) in are visualization results of the application of the t-distributed neighborhood embedding algorithm to the selected effective features and the application of the density-based noisy spatial clustering algorithm to the adaptive division of the degradation stages, respectively, according to the specific example of the present application;

[0034] Figure 5 is a structural schematic diagram of the improved time convolution network model designed according to the preferred embodiment of the present application;

[0035] Figure 6 (a), (b), (c), (d) and (e) in are visualization diagrams of the RUL prediction results according to the specific example of the present application;

[0036] Figure 7 (a), (b), (c) and (d) in are visualization diagrams of the RUL prediction results according to the specific example of the present application and the comparative analysis based on the full life cycle unified degradation model. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0038] Please refer to Figure 1The application provides a rotating machinery residual useful life prediction method based on a time convolution network. The method first collects full life cycle sensor data of a single or multiple rotating machines as a training and test data set, and then applies a signal processing method to extract time domain features, frequency domain features and time-frequency features from the collected sensor data to more comprehensively reflect the degradation process of the rotating machinery. In order to better fit the degradation process of the rotating machinery, a comprehensive evaluation index is established to select the extracted features, eliminate features with less correlation with the degradation of the rotating machinery, and retain the remaining effective features. Secondly, the selected features are compressed, and the compressed feature state points are clustered to adaptively divide the degradation stages of the rotating machinery. Finally, for multiple degradation stages of the rotating machinery, a prediction model based on a time convolution network (TCN) is established, the model is trained and optimized on the training data set, and after the model training is completed, the test data set is input to obtain the RUL prediction value.

[0039] The prediction method mainly includes the following steps:

[0040] S1. Collecting full life cycle sensor data of a single or multiple rotating machines in the running stage, that is, sampling at a certain time interval, frequency, etc., so as to obtain full life cycle vibration signal data of the rotating machinery.

[0041] Collecting full life cycle vibration signals of the rotating machinery, sampling at a certain sampling time interval, sampling time, sampling frequency, etc., to obtain sampling point information.

[0042] S2. Feature extraction is performed on the collected full life cycle vibration signals of the rotating machinery to reduce the dimensionality of the vibration signal data and facilitate the subsequent establishment of an accurate degradation model. Specifically, signal processing methods are used to extract time domain features, frequency domain features and time-frequency features from the rotating machinery vibration signals in the horizontal and vertical directions to construct a feature atlas.

[0043] Signal processing methods are used to extract time domain, frequency domain and time-frequency features from the collected full life cycle vibration signals in the horizontal and vertical directions to more comprehensively capture the degradation trend of the rotating machinery. The method first extracts time domain statistical features of the vibration signal to reflect the overall trend of the degradation of the rotating machinery. Secondly, the discrete time domain signal is converted into a frequency domain signal by using fast Fourier transform to extract frequency domain features, thereby reflecting the working state of the rotating machinery from the frequency and significantly reducing the influence of noise. Finally, the Hilbert-Huang transform adaptive time-frequency analysis method is used to extract the time-frequency features of the vibration signal to obtain more hidden information in the vibration signal.

[0044] S3. Linearly combine the dynamic time warping (DTW) distance indicator, the correlation coefficient indicator and the monotonicity indicator to construct a comprehensive evaluation indicator, so as to select more effective features suitable for RUL prediction from the above extracted features, and further improve the accuracy of RUL prediction. Generally speaking, the features with smaller similarity measure, higher time correlation and better monotonicity are more effective and more suitable for RUL prediction.

[0045] The following comprehensive evaluation indicator Cri is established: Cri = ω1Corr + ω2Mon - ω3Dis, wherein Dis, Corr and Mon are the dynamic time warping distance indicator, the correlation coefficient indicator and the monotonicity indicator respectively, and the weight coefficients thereof are ω3, ω1 and ω2 respectively, and specifically:

[0046] (1) Dynamic time warping distance indicator Dis

[0047] Dis is used to measure the similarity of time series. Assuming that there are M feature sequences in the original feature space, and the length of the vibration signal sequence in the whole life cycle is T, the feature state point at time t is t = 1, …, T, the mth feature sequence is m = 1, …, M, and the center F0 of the M feature sequences is:

[0048]

[0049] For two different time series, DTW calculates the similarity distance of all sub-sequences thereof, thereby obtaining F m and the similarity distance matrix of F0. The square of the similarity distance of sequence F m [0:i] and sequence F0[0:j] is:

[0050]

[0051] Then dp[T-1][T-1] is the square of the similarity distance of F m and F0, and further The smaller the similarity distance, i.e. the smaller the Dis value, the more suitable the feature is for RUL prediction.

[0052] (2) Correlation coefficient indicator Corr

[0053] Generally speaking, rotating machinery will gradually degrade as the running time increases. Corr is a time correlation measure, which measures the linear correlation degree of the feature sequence and the running time, and the calculation formula is:

[0054]

[0055] wherein and respectively represent the feature and time value of the tth observation sample, and respectively represent the mean of the feature and time value of T observation samples, and T represents the total number of observation samples. The feature with higher correlation of time, i.e., greater Corr value, is more suitable for RUL prediction.

[0056] (3) Monotonicity index Mon

[0057] Mon is mainly used to evaluate the trend of the feature:

[0058]

[0059] wherein, dx(t) represents the difference between the feature value at time (+1) and the feature value at time t in the feature sequence, and ∑(dx(t)>0) and ∑(dx(t)<0) respectively represent the number of positive and negative values of dx(t). m >0) and ∑( m <0) are respectively dx m The greater the Mon value of the feature, the better the monotonicity of the feature, and the more suitable the feature is for RUL prediction.

[0060] Based on the above three indexes, the Cri value of the extracted feature can be calculated to determine whether to retain the feature. Specifically, the features with smaller correlation with the degradation of the rotating machinery, i.e., smaller Cri value, are removed, and the remaining features are retained, so as to better reflect the degradation trend of the rotating machinery.

[0061] S4. The t-distributed stochastic neighbor embedding (t-SNE) algorithm is used to compress the effective features selected above, so that the distance between similar state points in the same degradation stage is closer, and the distance between state points in different degradation stages is farther, which makes the boundary between different degradation stages clearer, facilitating subsequent adaptive division of degradation stages based on degradation patterns.

[0062] The selected features are compressed from high-dimensional space to low-dimensional space by t-SNE algorithm. Specifically, in high-dimensional space and low-dimensional space, Gaussian distribution and t-distribution are respectively applied to convert the distance between feature state points into corresponding joint probability distribution, then Kullback-Leibler (KL) divergence is used to measure the difference between the two distributions, and gradient descent method is used to optimize the KL divergence, finally the feature state points far away from each other in the original high-dimensional space after mapping are far away from each other, and the feature state points close to each other are close to each other. Therefore, after feature compression, the distance between feature state points in the same and different degradation stages is closer and farther respectively, which further clarifies the boundary between different degradation stages, and facilitates subsequent adaptive division of degradation stages. The specific steps of t-SNE algorithm are as follows:

[0063] (1) Let and be two feature state points in high-dimensional space, i≠j, their joint probability distribution function p ij is:

[0064]

[0065] Where σ represents the variance of the Gaussian distribution with as the center point.

[0066] (2) Let and be the feature state points mapped to low-dimensional space by and respectively, then and The joint probability distribution function q ij in low-dimensional space is:

[0067]

[0068] (3) The KL divergence is used to measure the difference between the two distributions, and the objective function is:

[0069]

[0070] (4) The KL divergence is optimized by using gradient descent method, that is, the KL divergence is minimized, and the specific method is:

[0071]

[0072] S5. Considering the potential for irregular spatial distribution of characteristic state points after implementing S4, the density-based spatial clustering of applications with noise (DBSCAN) algorithm is used to process the compressed characteristic state points and further generate degradation stage labels. This algorithm can adaptively identify degradation patterns based on the actual health status of rotating machinery and automatically divide the entire lifecycle of rotating machinery into multiple degradation stages based on these patterns, thereby better fitting the entire degradation process of rotating machinery.

[0073] Considering that the compressed feature state points may have irregular spatial distribution, the DBSCAN algorithm is used to process them and classify the state points with high enough connection density in the feature space into the same degradation stage, thereby adaptively dividing the degradation stage. Specifically, the full life cycle feature sequence is X i ={x 1,2 ,…, T}, where T is the length of the feature sequence, and the cluster marker array m is introduced i :

[0074]

[0075] Where n is the number of clusters, the cluster label array can be obtained X i It can be divided into n clusters and noise point sets, which means that the whole life cycle can be divided into n degradation stages, and the corresponding degradation stage labels S are generated. label Then, in the training data set, the selected valid feature vector is compared with S label The features are combined to form feature label pairs, which are input into the support vector machine (SVM) classification model for training. Subsequently, the valid feature vectors of the test set can be input into the trained SVM classification model to obtain the degradation stage label of the test set.

[0076] S6. Develop a corresponding prediction model based on a temporal convolution network (TCN) for each degradation stage of the rotating machinery. After the model is trained, a test dataset is used to accurately predict its RUL. Specifically, for each of the n degradation stages adaptively divided in S5, a corresponding RUL prediction model based on TCN is developed. This model is trained and optimized on the training dataset. After the model is trained, a test dataset is used to accurately predict the RUL.

[0077] The basic TCN model is improved and applied to RUL prediction. The improved TCN model mainly includes an input layer, a TCN layer and a regression layer, which are specifically described as follows:

[0078] (1) Input layer: input fused effective features. In order to accelerate the convergence speed of the network, the following normalization processing is performed before inputting the network:

[0079]

[0080] Wherein, x i (t) and x respectively represent the value of the i-th original and normalized feature sequence at time t, and x are the average, maximum and minimum values of the i-th feature sequence at all times.

[0081] (2) TCN layer: a plurality of convolution units are added in the residual block structure, and the structure of each convolution unit is dilated causal convolution + weight normalization + ReLU activation function + Dropout. Through multi-layer abstraction of the convolution unit, the TCN layer has strong feature extraction capability and can better extract local degradation features.

[0082] (3) Regression layer: a flattened layer and a fully connected layer are established to fuse all local features extracted by the TCN layer, which can more comprehensively represent the degradation process of the rotating machinery, and finally outputs the RUL prediction value. The RUL true value is set as the percentage of the degradation of the rotating machinery, and the RUL true value label of the t-th sampling point is:

[0083]

[0084] Wherein, RUL t and RUL0 represent the order t of the sampling point and the total number of sampling points in the whole life cycle.

[0085] The following takes a bearing as a specific object, and further details of the RUL prediction method based on TCN are described in combination with specific examples, and the specific steps are as follows:

[0086] (1) Extract the full life cycle vibration signal data of the bearing, and use the accelerated degradation test bearing data set of IEEE PHM 2012 challenge to train and test. The sampling frequency in the test is 25.6 kHz, the single sampling time is 0.1 s, and the sampling interval is 10 s, so 2560 sampling points can be obtained each time. Figure 2 The full life cycle vibration signal of bearing 1-1 extracted along the horizontal direction is shown. The running conditions of the bearing and the bearing test data set are shown in Table 1 and Table 2, respectively;

[0087] Table 1 Bearing operating conditions

[0088]

[0089] Table 2 Bearing test data set

[0090]

[0091]

[0092] (2) Extract time-domain features, frequency-domain features and time-frequency features from the vibration signal data of the bearing along a specific direction. First, 14 time-domain features such as mean, standard deviation, variance, peak-to-peak value, root amplitude, average amplitude, root mean square value, waveform factor, peak value, peak factor, pulse factor, margin factor, skewness and kurtosis can be extracted from the vibration signal. Then, the time-domain signal is converted to the frequency domain by using the fast Fourier transform method to extract 4 frequency-domain features such as average frequency, center of gravity frequency, root mean square frequency and frequency standard deviation. The Hilbert-Huang transform adaptive time-frequency analysis method is used to extract 4 time-frequency features such as intrinsic mode function energy, intrinsic mode function energy entropy, marginal spectrum energy and marginal spectrum energy entropy. Therefore, the vibration signal of the bearing is extracted along the horizontal and vertical directions, and finally 44 features can be extracted;

[0093] (3) According to the comprehensive evaluation index Cri constructed, effective features are selected from the extracted features, Cri = 0.5 + 0.3 - 0.2. Figure 3 The visualization result of the comprehensive evaluation index value of all the features extracted along the horizontal and vertical directions of the bearing 1_1 is shown. For each bearing, this embodiment eliminates the 14 features with the smallest comprehensive evaluation index value, and retains the remaining 30 effective features;

[0094] (4) The selected effective features are compressed by using the t-SNE algorithm, and then the DBSCAN algorithm is used to adaptively divide the degradation stages of the bearing. First, the t-SNE algorithm is used to map the selected 30 effective features from high-dimensional space to low-dimensional space, realizing feature compression, so as to better distinguish different degradation states. Figure 4 (a) in FIG. 8 is the visualization result of compressing the 30 effective feature points of the bearing 1_4 full life cycle into two dimensions by using the t-SNE algorithm, which shows the distribution difference between different health states; secondly, based on the distribution difference between different health states generated by the t-SNE algorithm, the DBSCAN algorithm is used to adaptively divide the full life cycle of the training set into different degradation stages, Figure 4(b) in FIG. 6 shows the visualization result of the DBSCAN algorithm adaptively dividing the whole life degradation process of bearing 1_4 into two degradation stages. Then, the effective features of the training set are fused, combined with the corresponding degradation stage labels to form label pairs, input into the SVM classifier, and the SVM classifier is trained; finally, the effective features of the test set are fused and input into the trained SVM classifier to adaptively generate the degradation stage labels of the test set.

[0095] (5) A prediction model based on TCN is established, and the structural diagram is as shown in FIG. 7. The specific settings of the hyperparameters are shown in Table 3. During training, the activation function of all convolutional layers is ReLU, and the activation function of the fully connected layer is Sigmoid. In the process of back propagation, the loss function is mean absolute error, and the optimization algorithm is Adam. After the model is trained, the test data set is input, and finally the RUL prediction value of the test data set is output; Figure 5

[0096] Table 3 TCN hyperparameter table

[0097]

[0098] (6) Visualize the RUL prediction result. Taking bearings 1-3, 1-5, 2-5, 2-6 and 3-3 as examples, first, the whole life of them is adaptively divided into two stages of stable degradation and rapid degradation by the degradation stage adaptive division mechanism, i.e., the segmented degradation model, proposed in the embodiment, and then the RUL of the above two degradation stages is predicted by the prediction model based on TCN proposed in the embodiment, and the prediction result is as shown in FIG. 8. Figure 6 From FIG. 8, it can be found that the embodiment performs excellently in solving the RUL prediction problem of bearings, which verifies the effectiveness of the present application. Figure 6 In order to further verify the effectiveness of the degradation stage adaptive division mechanism, i.e., the segmented degradation model, designed in the embodiment, comparative experiments are carried out on the bearing data set under working condition 1.

[0099] FIG. 9 shows the visualization result of bearings 1-3 and 1-5 applying the TCN method to predict the RUL based on the whole life unified degradation model and the segmented degradation model. From FIG. 9, it can be found that compared with the whole life unified degradation model, the degradation stage adaptive division mechanism, i.e., the segmented degradation model, proposed in the present application can achieve more and better prediction results, especially in the rapid degradation stage when the bearing is about to fail, the prediction accuracy can be improved, which verifies the effectiveness of the degradation stage adaptive division mechanism, i.e., the segmented degradation model, designed in the present application. Figure 7 Figure 7

[0100] ​​​In order to further verify the effectiveness of the improved TCN model designed in the application in solving the RUL prediction problem, it is compared with a long short-term memory (LSTM), a convolutional neural network (CNN) and a support vector regression (SVR) on the bearing data set under working condition 1. The four methods above all perform RUL prediction on the basis of the segmented degradation model, and the root mean square error (RMSE) and the mean absolute percentage error (MAPE) are used to analyze the experimental results. Their calculation methods are as follows:

[0101]

[0102]

[0103] wherein, and y i respectively represent the RUL prediction value and the true value of the i th sample. It can be found from Table 4 that the RUL prediction accuracy of TCN is higher than that of LSTM, CNN and SVR on most bearings, which verifies the effectiveness of the TCN model established in the application.

[0104] Table 4 Experimental results of TCN, LSTM, CNN and SVR on the bearing data set under working condition 1

[0105]

[0106] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the application and is not used to limit the application. Any modification, equivalent replacement and improvement made within the spirit and principle of the application shall be included in the protection scope of the application.

Claims

1. A method for predicting the remaining useful life of rotating machinery based on a time convolutional network, characterized in that: The method comprises the following steps: (1) Collect vibration signal data of a single or multiple rotating machines during the entire life cycle of the machine during operation, and extract time domain features, frequency domain features, and time-frequency features along a specific direction from the vibration signal data of the entire life cycle; (2) Using comprehensive evaluation indicators Select effective features from the extracted features; among them, the comprehensive evaluation index : ,in 、 and They are dynamic time warping distance index, correlation coefficient index and monotonicity index, and their weight coefficients are 、 and ; (3) The t-SNE algorithm is used to compress the selected effective features, and then the DBSCAN algorithm is used to adaptively divide the degradation stages of rotating machinery; (4) Establish a prediction model based on TCN and train the prediction model, and use the trained prediction model to predict the remaining service life of rotating machinery; The prediction model includes an input layer, a TCN layer, and a regression layer. In the input layer, the effective features after fusion are input. In the TCN layer, multiple convolution units are added to the residual block structure. The structure of each convolution unit is dilated causal convolution + weight normalization + ReLU activation function + Dropout. In the regression layer, a flattening layer and a fully connected layer are established to fuse all the local features extracted by the TCN layer, and finally the RUL prediction value is output.

2. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 1, wherein: Extract time domain, frequency domain and time-frequency features from the collected full life cycle vibration signals in the horizontal and vertical directions. First, extract the time domain statistical features of the vibration signal; Secondly, the fast Fourier transform is used to convert the discrete time domain signal into the frequency domain signal and extract the frequency domain features. Finally, the Hilbert-Huang transform adaptive time-frequency analysis method is used to extract the time-frequency features of the vibration signal.

3. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 1, wherein: The t-SNE algorithm is used to compress the selected features from high-dimensional space to low-dimensional space. In high-dimensional space and low-dimensional space, Gaussian distribution and t-distribution are applied respectively to convert the distance between state points into the corresponding joint probability distribution. Then, Kullback-Leibler divergence is used to measure the difference between the two distributions, and the KL divergence is optimized using the gradient descent method. Finally, after mapping, the feature state points that are far apart in the original high-dimensional space are further apart, and the feature state points that are close to each other are closer.

4. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 3, wherein: After feature compression, the distances between feature state points in the same and different degradation stages are closer and farther, respectively, which further clarifies the boundaries between different degradation stages and facilitates the subsequent adaptive division of degradation stages.

5. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 1, wherein: The DBSCAN algorithm can adaptively identify degradation patterns according to the actual health status of rotating machinery, and automatically divide the entire life cycle of rotating machinery into multiple degradation stages based on the degradation patterns.

6. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 5, wherein: Let the full life cycle feature sequence be ,in For the feature sequence length, introduce the cluster marker array : in is the number of clusters, so the cluster label array , Divided into clusters and noise point sets, that is, the entire life cycle is divided into degradation stages and generate corresponding degradation stage labels ; Then, the selected valid feature vector is compared with The features are combined to form a feature label pair, which is input into the support vector machine classification model for training. Subsequently, the valid feature vector of the test set is input into the trained SVM classification model to obtain the degradation stage label of the test set.

7. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 1, wherein: First, the t-SNE algorithm is used to map the selected effective features from high-dimensional space to low-dimensional space. Second, based on the distribution differences between different health states generated by the t-SNE algorithm, the DBSCAN algorithm is used to adaptively divide the entire life cycle of the training set into different degradation stages. Subsequently, the effective features of the training set are fused and combined with the corresponding degradation stage labels to form label pairs, which are input into the SVM classifier and trained. Finally, the effective features of the test set are fused and input into the trained SVM classifier to adaptively generate degradation stage labels for the test set.

8. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 7, wherein: The entire life cycle is adaptively divided into two stages: smooth degradation and rapid degradation. The RUL of the two stages is then predicted using a TCN-based prediction model.

9. The method for predicting the remaining useful life of rotating machinery based on a time convolutional network according to claim 1, wherein: The true value of RUL is set as the percentage of rotating machinery degradation. The true value label of the RUL of the sampling point is: in, and Represents the order of sampling points The total number of sampling points in the entire life cycle is normalized before entering the input layer. The corresponding formula is: in, and Respectively The moment The values ​​of the original and normalized feature sequences, 、 and For all moments The average, maximum, and minimum values ​​of the characteristic sequence values.

Images