CN115982621A - Rotary machine residual service life prediction method based on time convolution network - Google Patents
Rotary machine residual service life prediction method based on time convolution network Download PDFInfo
- Publication number
- CN115982621A CN115982621A CN202211699953.2A CN202211699953A CN115982621A CN 115982621 A CN115982621 A CN 115982621A CN 202211699953 A CN202211699953 A CN 202211699953A CN 115982621 A CN115982621 A CN 115982621A
- Authority
- CN
- China
- Prior art keywords
- degradation
- time
- service life
- rotating machine
- predicting
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000006731 degradation reaction Methods 0.000 claims abstract description 123
- 230000015556 catabolic process Effects 0.000 claims abstract description 111
- 238000012549 training Methods 0.000 claims abstract description 27
- 230000036541 health Effects 0.000 claims abstract description 16
- 238000011156 evaluation Methods 0.000 claims abstract description 13
- 238000012360 testing method Methods 0.000 claims description 23
- 238000009826 distribution Methods 0.000 claims description 21
- 238000005070 sampling Methods 0.000 claims description 18
- 238000012706 support-vector machine Methods 0.000 claims description 12
- 230000003044 adaptive effect Effects 0.000 claims description 11
- 238000013507 mapping Methods 0.000 claims description 8
- 238000013145 classification model Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 5
- 230000006835 compression Effects 0.000 claims description 5
- 238000007906 compression Methods 0.000 claims description 5
- 238000011478 gradient descent method Methods 0.000 claims description 4
- 208000037170 Delayed Emergence from Anesthesia Diseases 0.000 claims description 3
- 239000003550 marker Substances 0.000 claims description 3
- 230000001364 causal effect Effects 0.000 claims description 2
- 230000003862 health status Effects 0.000 claims description 2
- 238000012163 sequencing technique Methods 0.000 claims description 2
- 230000006870 function Effects 0.000 description 8
- 230000007246 mechanism Effects 0.000 description 7
- 238000012800 visualization Methods 0.000 description 7
- 238000000605 extraction Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 238000013527 convolutional neural network Methods 0.000 description 5
- 230000008569 process Effects 0.000 description 5
- 238000000638 solvent extraction Methods 0.000 description 5
- 238000010586 diagram Methods 0.000 description 3
- 239000000284 extract Substances 0.000 description 3
- 230000007774 longterm Effects 0.000 description 3
- 238000012423 maintenance Methods 0.000 description 3
- 238000003672 processing method Methods 0.000 description 3
- 230000000007 visual effect Effects 0.000 description 3
- 238000004364 calculation method Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 238000005315 distribution function Methods 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 230000001788 irregular Effects 0.000 description 2
- 238000004519 manufacturing process Methods 0.000 description 2
- 238000005259 measurement Methods 0.000 description 2
- 238000005192 partition Methods 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 108010076504 Protein Sorting Signals Proteins 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010835 comparative analysis Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 239000011159 matrix material Substances 0.000 description 1
- 238000005065 mining Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 238000001228 spectrum Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Testing Of Devices, Machine Parts, Or Other Structures Thereof (AREA)
Abstract
The invention belongs to the technical field related to fault prediction and health management, and discloses a method for predicting the residual service life of a rotary machine based on a time convolution network, which comprises the following steps: (1) Acquiring vibration signal data of a single or multiple rotary machines in a full life cycle of an operation stage, and extracting time domain characteristics, frequency domain characteristics and time-frequency characteristics from the vibration signal data of the full life cycle along a specific direction; (2) Selecting effective features from the extracted features by adopting a comprehensive evaluation index Cri; (3) Compressing the selected effective characteristics by using a t-SNE algorithm, and then adaptively dividing the degradation stage of the rotary machine by using a DBSCAN algorithm; (4) And establishing a prediction model based on the TCN, training the prediction model, and predicting the residual service life of the rotary machine by adopting the trained prediction model. The invention can better capture the degradation process of the rotating machinery and obviously reduce the influence of noise and fluctuation in the sensor.
Description
Technical Field
The invention belongs to the technical field related to fault prediction and health management, and particularly relates to a method for predicting the residual service life of a rotary machine based on a time convolution network.
Background
The fault prediction and health management aim to apply a series of models and methods to evaluate the health state of equipment or a system and provide a timely and effective maintenance decision scheme for an enterprise manager based on the health state, so that the safety and the reliability of the system are improved. Residual service life (RUL) prediction is one of core contents of fault prediction and health management, and the method mainly focuses on predicting the health state before system fault by using the internal structure of equipment, sensor data acquired in the operation process and the like, and has great significance in avoiding production plan disturbance due to sudden fault shutdown, avoiding potential safety hazards and major accidents, reducing the maintenance cost of equipment, improving the production efficiency of enterprises, optimizing the operation and maintenance management of the system and the like.
The rotating machine is one of important parts of a large-scale mechanical device, and prediction of the RUL of the rotating machine is important for improving the safety and reliability of the large-scale mechanical device. Generally, the RUL prediction methods mainly have the following two types: a mechanism-based model and a data-driven based method. The method based on the mechanism model is mainly used for establishing a physical failure model according to the equipment degradation process, and needs rich expert experience knowledge, and specific equipment is specifically analyzed, so that the generalization capability of the method is poor. In recent years, the development of data-driven based methods has been driven by the rapid development of artificial intelligence and sensor technologies. The method based on data driving aims at mining potential mapping relation between system degradation trend and state monitoring data, and only needs to process collected sensor data without abundant expert experience knowledge to extract effective characteristics capable of representing system degradation, so that the method has strong generalization capability and is widely applied to solving the RUL prediction problem. The deep learning method has good nonlinear fitting capability and sequence data processing capability, can improve the prediction performance while remarkably reducing the complexity of an RUL prediction task, and has wide application prospect. The conventional deep learning method mainly develops RUL prediction work from the aspects of data acquisition, data preprocessing, model establishment, model training, online prediction and the like, and mainly has the following defects: first, it is difficult for the established full-life-cycle unified degradation model to fit well the entire process of the degradation of the rotating machine. Specifically, the rotating machine operates smoothly for a long time in an early stage, then slowly degrades at an almost constant rate, and finally enters a rapid degradation stage and operates to a failure in a short time. Different degradation modes have obvious difference, so that corresponding degradation models need to be established aiming at the different degradation modes to better fit the whole degradation process of the rotary machine; secondly, for different degradation stages of the rotary machine, the mapping relation between the extracted features and the health state is difficult to match; finally, some common deep learning models have limitations in solving the RUL prediction problem, for example, a convolutional neural network has good feature extraction capability, but cannot capture the long-term sequence dependency well due to the limitation of the size of a convolutional kernel.
Disclosure of Invention
Aiming at the defects or the improvement requirements of the prior art, the invention provides a method for predicting the residual service life of a rotary machine based on a time convolution network, which solves the problems of low prediction precision caused by under-fitting of an equipment degradation process, difficult matching of extracted features and equipment health states, incompatibility of feature extraction capability and long-term sequence data processing capability and the like in RUL prediction.
To achieve the above object, according to an aspect of the present invention, there is provided a method for predicting remaining service life of a rotating machine based on a time convolution network, the method mainly includes the following steps:
(1) Acquiring vibration signal data of a single or multiple rotary machines in a full life cycle of an operation stage, and extracting time domain characteristics, frequency domain characteristics and time-frequency characteristics from the vibration signal data of the full life cycle along a specific direction;
(2) Selecting effective features from the extracted features by adopting a comprehensive evaluation index Cri; wherein the comprehensive evaluation index Cri is as follows: cri = ω 1 Corr+ω 2 Mon-ω 3 Dis, where Dis, corr and Mon are respectively dynamic time warping distance index, correlation coefficient index and monotonicity index, and their weighting coefficients are respectively omega 3 、ω 1 And omega 2 ；
(3) Compressing the selected effective characteristics by using a t-SNE algorithm, and then adaptively dividing the degradation stage of the rotary machine by using a DBSCAN algorithm;
(4) And establishing a prediction model based on the TCN, training the prediction model, and predicting the residual service life of the rotary machine by adopting the trained prediction model.
Further, the prediction model comprises an input layer, a TCN layer and a regression layer; inputting the fused effective characteristics in the input layer; in the TCN layer, a plurality of convolution units are added in a residual block structure, and the structure of each convolution unit is expansion causal convolution + weight normalization + ReLU activation function + Dropout; in the regression layer, a flattening layer and a full connection layer are established to fuse all local features extracted by the TCN layer, and finally the RUL predicted value is output.
Further, extracting time domain, frequency domain and time-frequency characteristics from the acquired full-life-cycle vibration signals along the horizontal direction and the vertical direction, and firstly extracting time domain statistical characteristics of the vibration signals; secondly, converting the discrete time domain signals into frequency domain signals by adopting fast Fourier transform, and extracting frequency domain characteristics; and finally, extracting the time-frequency characteristics of the vibration signals by using a Hilbert-Huang transform self-adaptive time-frequency analysis method.
Further, compressing the selected features from a high-dimensional space to a low-dimensional space by adopting a t-SNE algorithm, converting the distance between the state points into corresponding joint probability distribution in the high-dimensional space and the low-dimensional space by respectively applying Gaussian distribution and t distribution, measuring the difference between the two distributions by adopting Kullback-Leibler divergence, and optimizing the KL divergence by adopting a gradient descent method, thereby finally realizing that the feature state points which are far away from each other in the original high-dimensional space are farther away from each other and the feature state points which are near to each other are closer to each other after mapping.
Further, after feature compression, the distances between feature state points in the same degradation stage and the distances between feature state points in different degradation stages are respectively closer and farther, so that the boundaries between different degradation stages are further clarified, and the subsequent self-adaptive degradation stage division is facilitated.
Further, the DBSCAN algorithm enables adaptive identification of degradation modes according to the actual health status of the rotating machine and automatic division of the full life cycle of the rotating machine into multiple degradation stages based on the degradation modes.
Further, the full lifecycle feature sequence is X i ＝{x 1,2 ,…, T Where T is the characteristic sequence length, introduce a cluster marker array m i ：
Where n is the number of clusters, thereby obtaining an array of cluster markers
X i Is divided into n clusters and noise point sets, i.e. the full life cycle is divided into n degradation phases, and corresponding degradation phase labels S are generated label (ii) a Subsequently, the selected valid feature vector is summed with S label Combining to form feature label pair, inputting into classification model of support vector machine for training, and trainingAnd inputting the effective characteristic vector of the test set in the finished SVM classification model so as to obtain the degradation stage label of the test set.
Further, firstly, mapping the selected effective characteristics from a high-dimensional space to a low-dimensional space by adopting a t-SNE algorithm; secondly, based on the distribution difference between different health states generated by the t-SNE algorithm, the DBSCAN algorithm is used for adaptively dividing the full life cycle of the training set into different degradation stages; then, fusing the effective characteristics of the training set, combining the effective characteristics with corresponding degradation stage labels to form label pairs, inputting the label pairs into an SVM classifier, and training the SVM classifier; and finally, fusing the effective characteristics of the test set, inputting the effective characteristics into the trained SVM classifier, and adaptively generating a degradation stage label of the test set.
Further, the full life cycle is adaptively divided into two stages of smooth degradation and rapid degradation, and then the prediction is carried out by applying a prediction model based on TCN to the RUL of the two stages of smooth degradation and rapid degradation.
Further, the real RUL value is set as the percentage of the degradation of the rotating machine, and the real RUL value of the t-th sampling point is labeled as:
wherein, RUL t And RUL 0 Respectively representing the sequencing t of the sampling points and the total number of the sampling points in the whole life cycle; before input into the input layer, normalization is carried out, and the corresponding formula is as follows:
wherein x is i (t) and
respectively represent the value of the ith original and normalized characteristic sequence at time t>
And/or>
Respectively, the average value, the maximum value and the minimum value of the ith characteristic sequence value at all the time instants.
In general, compared with the prior art, the method for predicting the remaining service life of the rotary machine based on the time convolution network provided by the invention has the following beneficial effects:
1. the method selects the extracted time domain, frequency domain and time frequency characteristics by constructing a linear combination of a comprehensive evaluation index, namely a dynamic time warping distance index, a correlation index and a monotonicity index, eliminates the characteristics of large similarity measurement, low time correlation and poor monotonicity, retains the residual effective characteristics, can better capture the degradation process of the rotary machine, and remarkably reduces the influence of noise and fluctuation in the sensor.
2. The invention compresses the reserved effective characteristics by applying the t-SNE algorithm, then processes the compressed characteristic state points by applying the DBSCAN algorithm, can realize that the degradation mode can be adaptively identified only according to the actual health state of the rotating machine without expert priori knowledge, and automatically divides the full life cycle of the equipment into a plurality of 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 degradation process of the rotating machine and enhancing the prediction precision of the RUL.
3. The improved TCN model designed by the invention can realize the accurate prediction of the RUL, and the network model has good feature extraction capability and long-term sequence data processing capability, so that the convergence speed of the network is increased, the generalization capability of the model is enhanced, and the RUL prediction accuracy is improved.
4. Firstly, extracting time domain features from a vibration signal, secondly, converting the time domain features into frequency domain features through Fourier transform, further extracting the frequency domain features, and finally, extracting time frequency features by adopting a Hilbert-Huang transform adaptive time frequency analysis method; by extracting the three types of characteristics, the degradation information of the equipment can be reflected more comprehensively.
Drawings
FIG. 1 is a flow chart of a method for predicting remaining useful life of a rotating machine based on a time convolution network according to a preferred embodiment of the present invention;
FIG. 2 is a visualization of the extraction of a full-life cycle vibration signal of a specific example in a horizontal direction;
fig. 3 (a) and (b) are respectively a visualization diagram of extracting the comprehensive evaluation index values of the time domain feature, the frequency domain feature and the time-frequency feature of the specific example along the horizontal direction and the vertical direction;
FIG. 4 (a) and (b) are the visualization results of the adaptive partition degradation stage of the noise spatial clustering algorithm based on density and the compression of the selected effective features by the t-distribution neighborhood embedding algorithm and the application of the noise spatial clustering algorithm applied to the specific example of the present invention, respectively;
FIG. 5 is a schematic diagram of the structure of an improved time convolution network model as contemplated by the preferred embodiment of the present invention;
FIGS. 6 (a), (b), (c), (d), (e) are respectively the visual representation of the RUL prediction results applied to the embodiment of the present invention;
FIGS. 7 (a), (b), (c), and (d) are respectively the visual images of the RUL prediction results of the comparative analysis of the present invention applied to the specific example and the one based on the full life cycle unified degradation model.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. In addition, the technical features involved in the embodiments of the present invention described below may be combined with each other as long as they do not conflict with each other.
Referring to fig. 1, the present invention provides a method for predicting remaining useful life of a rotary machine based on a time convolution network, which first collects full-life-cycle sensor data of one or more rotary machines as a training and testing data set, and then extracts time-domain features, frequency-domain features and time-frequency features from the collected sensor data by using a signal processing method to more fully reflect the degradation process of the rotary machine. In order to better fit the degradation process of the rotary machine, a comprehensive evaluation index is established to select the extracted features, the features with small degradation correlation with the rotary machine are removed, and the remaining effective features are reserved. And secondly, compressing the selected features, and clustering the compressed feature state points, thereby adaptively dividing the degradation stage of the rotary machine. And finally, establishing a corresponding prediction model based on a Time Convolution Network (TCN) for a plurality of degradation stages of the rotating machine, training and optimizing the model on a training data set, and inputting a test data set after the model training is finished to obtain an RUL prediction value.
The prediction method mainly comprises the following steps:
s1, collecting the data of a full life cycle sensor of one or more rotating machines in an operation stage, namely sampling according to a certain time interval, frequency and the like, thereby obtaining the full life cycle vibration signal data of the rotating machines.
The method comprises the steps of collecting full-life-cycle vibration signals of the rotary machine, sampling according to certain sampling time intervals, sampling time, sampling frequency and the like, and obtaining sampling point information.
And S2, performing feature extraction on the collected rotating machinery full-life-cycle vibration signal to reduce the dimensionality of the vibration signal data and facilitate the subsequent establishment of an accurate degradation model. Specifically, a signal processing method is used for extracting time domain features, frequency domain features and time-frequency features from the rotary mechanical vibration signals along the horizontal direction and the vertical direction so as to construct a feature map.
And extracting time domain, frequency domain and time-frequency characteristics from the acquired full-life-cycle vibration signals along the horizontal direction and the vertical direction by using a signal processing method so as to more comprehensively capture the degradation trend of the rotary machine. The method first extracts the time domain statistical features of the vibration signal to reflect the general overview of the degradation trend of the rotating machine. Secondly, a discrete time domain signal is converted into a frequency domain signal by adopting fast Fourier transform, and frequency domain characteristics are extracted, so that the working state of the rotary machine can be reflected on the frequency, and the influence of noise can be obviously reduced. And finally, extracting the time-frequency characteristics of the vibration signals by using a Hilbert-Huang transform self-adaptive time-frequency analysis method, and acquiring more hidden information in the vibration signals.
And S3, carrying out linear combination on a Dynamic Time Warping (DTW) distance index, a correlation coefficient index and a monotonicity index to construct a comprehensive evaluation index, so that effective characteristics more suitable for RUL prediction are selected from the extracted characteristics, and the accuracy of the RUL prediction is further improved. Generally, features with smaller similarity measures, higher temporal correlation and better monotonicity are more efficient and more suitable for RUL prediction.
Establishing the following comprehensive evaluation index Cri: cri = ω 1 Corr+ω 2 Mon-ω 3 Dis, where Dis, corr and Mon are respectively dynamic time warping distance index, correlation coefficient index and monotonicity index, and their weighting coefficients are respectively omega 3 、ω 1 And omega 2 The method specifically comprises the following steps:
(1) Dynamic time warping distance index Dis
Dis is used to measure similarity in time series. Assuming that M characteristic sequences exist in the original characteristic space and the length of the vibration signal sequence of the full life cycle is T, the characteristic state point at the time T is
T =1, \\ 8230, T, the mth characteristic sequence is->
M =1, \ 8230, M, the center F of M characteristic sequences 0 Comprises the following steps:
for two different time sequences, DTW calculates the similarity distance of all their corresponding subsequences, thus obtaining F m And F 0 The similarity distance matrix of (2). Sequence F m [0:i]And sequence F 0 [0:j]The square of the similarity distance of (a) is:
then dp [ T-1][T-1]Is F m And F 0 Is further obtained by the square of the similarity distance of
Features with smaller similarity distance, i.e. smaller Dis value, are more suitable for RUL prediction.
(2) Correlation coefficient index Corr
In general, rotary machines are subject to degradation as operating time increases. Corr is a time correlation measurement which measures the linear correlation degree of the characteristic sequence and the running time, and the calculation formula is as follows:
wherein
And &>
Respectively, a characteristic and a time value representing the tth observation sample, respectively>
And &>
The mean values of the characteristic and time values of the T observation samples are respectively, and T is the total number of the observation samples. Features with higher temporal correlation, i.e. larger Corr values, are more suitable for RUL prediction.
(3) Monotonicity index Mon
Mon is mainly used to evaluate the trend of features:
wherein,
represents the difference value of the characteristic value between the (+ 1) time and the t time in the characteristic sequence, sigma ( m >0) And ∑ ( m <0) Then respectively dx in the signature sequence m A number of positive and negative values. The degradation process of a rotating machine from start-up operation to eventual failure is assumed to be monotonically linear. Therefore, a larger Mon value of a feature indicates a better monotonicity of the feature, and is more suitable for RUL prediction.
Based on the three indexes, the Cri value of the extracted feature can be calculated, and whether the feature is reserved or not can be selected. Specifically, the features with small correlation with the degradation of the rotary machine, i.e., the feature with small Cri value, are removed, and the remaining features are retained, thereby better reflecting the degradation trend of the rotary machine.
S4, compressing the selected effective features by using a T-distributed stored neighboring region embedding algorithm (T-SNE), so that the distances between similar state points in the same degradation stage are closer, and the distances between state points in different degradation stages are farther, so that the boundaries between different degradation stages are clearer, and the subsequent degradation mode-based adaptive degradation division stages are facilitated.
And compressing the selected features from a high-dimensional space to a low-dimensional space by adopting a t-SNE algorithm. Specifically, in a high-dimensional space and a low-dimensional space, gaussian distribution and t distribution are respectively applied to convert the distance between characteristic state points into corresponding joint probability distribution, then Kullback-Leibler (KL) divergence is adopted to measure the difference between the two distributions, a gradient descent method is used to optimize the KL divergence, and finally the characteristic state points which are far away from each other in the original high-dimensional space are farther away from each other and the characteristic state points which are near to each other are closer to each other after mapping. Therefore, after feature compression, the distances between feature state points in the same degradation stage and the distances between feature state points in different degradation stages are respectively closer and farther, so that the boundaries between different degradation stages are further clarified, and the subsequent self-adaptive degradation stage division is facilitated. the specific steps of the t-SNE algorithm are as follows:
(1) Order to
And &>
For two characteristic state points in the high dimensional space, i ≠ j, their joint probability distribution function p ij Comprises the following steps:
wherein σ represents
Is the variance of the gaussian distribution of the center point.
(2) Order to
And &>
Are respectively based on>
And/or>
A characteristic status point mapped to the low dimension space, then +>
And &>
Joint probability distribution function q in low dimensional space ij Comprises the following steps:
(3) The difference between the two distributions is measured by KL divergence with the objective function:
(4) And optimizing the KL divergence by using a gradient descent method, namely minimizing the KL divergence, wherein the specific method comprises the following steps of:
and S5, considering the problem that the space distribution shape of the characteristic state points is irregular after S4 is implemented, processing the compressed characteristic state points by using a Density-based noisy spatial clustering of applications with noise (DBSCAN) algorithm, and further generating a label in a degradation stage. Therefore, the DBSCAN algorithm can realize self-adaptive identification of the degradation mode according to the actual health state of the rotary machine, and automatically divide the full life cycle of the rotary machine into a plurality of degradation stages based on the degradation mode, so that the whole degradation process of the rotary machine can be better fitted.
Considering that the compressed characteristic state points may have irregular spatial distribution, the DBSCAN algorithm is used for processing the compressed characteristic state points, and the state points with high enough connection density in the characteristic space are divided into the same degradation stage, so that the degradation stage is adaptively divided. Specifically, the full lifecycle signature sequence is X i ＝{x 1,2 ,…, T Where T is the characteristic sequence length, introduce a cluster marker array m i ：
Where n is the number of clusters, from which an array of cluster markers can be derived
X i Can be divided into n clusters and noise point sets, i.e. the whole life cycle can be divided into n degradation stages and corresponding degradation stage labels S can be generated label . Subsequently, the selected significant feature vector is summed with S in the training dataset label And combining to form a feature label pair, inputting the feature label pair into a Support Vector Machine (SVM) classification model for training, and subsequently inputting the effective feature vector of the test set into the trained SVM classification model so as to obtain the degradation stage label of the test set.
S6, establishing a corresponding prediction model based on a Time Convolution Network (TCN) aiming at a plurality of degradation stages of the rotating machine, inputting a test data set after model training is finished, and accurately predicting the RUL. Specifically, aiming at n degradation stages divided in a self-adaptive mode in S5, a corresponding RUL prediction model is established based on TCN, the model is trained and optimized on a training data set, and a test data set is input after model training is finished, so that the accurate prediction of the RUL is realized.
The basic TCN model is improved and then applied to RUL prediction. The improved TCN model provided in this embodiment mainly includes an input layer, a TCN layer, and a regression layer, which is specifically described as follows:
(1) An input layer: and inputting the fused effective features. In order to accelerate the convergence speed of the network, the following normalization processing is carried out before the network is input:
wherein x is i (t) and
respectively represent the value of the ith original and normalized characteristic sequence at time t>
And/or>
Respectively, the average value, the maximum value and the minimum value of the ith characteristic sequence value at all the time instants.
(2) A TCN layer: and adding a plurality of convolution units in the residual block structure, wherein the structure of each convolution unit is dilation-causal convolution + weight normalization + ReLU activation function + Dropout. Through the multilayer abstraction of the convolution unit, the TCN layer has strong feature extraction capability and can better extract local degradation features.
(3) Regression layer: the degradation process of the rotary machine can be more comprehensively characterized by establishing a flattening layer and a full connection layer to fuse all local characteristics extracted by the TCN layer, and finally, an RUL predicted value is output. The real RUL value is set as the percentage of the degradation of the rotating machine, and the real RUL value label of the t-th sampling point is as follows:
wherein, RUL t And RUL 0 Respectively representing the ordering t of the samples and the total number of samples over the life cycle.
The method for predicting RUL based on TCN of the present invention will be further described in detail with reference to specific examples, with a rotating machine, i.e. a bearing, as a specific object, and the specific steps are as follows:
(1) Extracting the vibration signal data of the bearing in the full life cycle, training and testing by adopting an accelerated degradation test bearing data set of IEEE PHM 2012 challenge race, wherein the sampling frequency in the test is 25.6kHz, the single sampling time is 0.1s, and the sampling interval is 10s, so that 2560 sampling points can be obtained in each sampling. Fig. 2 shows the full life cycle vibration signal of the bearing 1-1 taken in the horizontal direction. The operating conditions and bearing test data sets for the bearings are shown in tables 1 and 2, respectively;
TABLE 1 bearing operating conditions
TABLE 2 bearing test data set
(2) And extracting time domain characteristics, frequency domain characteristics and time-frequency characteristics from the vibration signal data of the bearing in the full life cycle along a specific direction. Firstly, 14 time domain characteristics of a vibration signal, such as a mean value, a standard deviation, a variance, a peak-to-peak value, a root amplitude, an average amplitude, a root-mean-square value, a wave form factor, a peak value factor, a pulse factor, a margin factor, skewness, kurtosis and the like, can be extracted, then the time domain signal is converted into a frequency domain by using a fast Fourier transform method so as to extract 4 frequency domain characteristics of average frequency, center-of-gravity frequency, root-mean-square frequency, frequency standard deviation and the like, and 4 time frequency characteristics of natural modal function energy, natural modal function energy entropy, marginal spectrum energy entropy and the like are extracted by using a Hilbert-yellow transform self-adaptive time frequency analysis method. Therefore, the vibration signal of the bearing is characterized along the horizontal direction and the vertical direction, and finally 44 characteristics can be extracted;
(3) And selecting effective characteristics from the extracted characteristics according to the constructed comprehensive evaluation index Cri, wherein Cri =0.5+0.3-0.2. Fig. 3 shows the visualization result of the integrated evaluation index values of all the features extracted in the horizontal and vertical directions by the bearing 1_1. For each bearing, 14 features with the minimum comprehensive evaluation index value are removed, and the rest 30 effective features are reserved;
(4) And compressing the selected effective characteristics by applying a t-SNE algorithm, and then adaptively dividing the degradation stage of the bearing by applying a DBSCAN algorithm. Firstly, mapping 30 selected effective characteristics from a high-dimensional space to a low-dimensional space by adopting a t-SNE algorithm to realize characteristic compression, thereby better distinguishing different degradation states. Fig. 4 (a) is a visualization result of compressing 30 valid feature points of the bearing 1 _4full life cycle into two dimensions by using t-SNE algorithm, showing the distribution difference between different health states; secondly, based on the distribution difference between different health states generated by the t-SNE algorithm, the full life cycle of the training set is adaptively divided into different degradation stages by using the dbss algorithm, and (b) in fig. 4 shows the visualization result of the dbss algorithm adaptively dividing the full life cycle degradation process of the bearing 1 xu 4 into two degradation stages. Then fusing the effective characteristics of the training set, combining the effective characteristics with corresponding degradation stage labels to form label pairs, inputting the label pairs into an SVM classifier, and training the SVM classifier; and finally, fusing the effective characteristics of the test set, inputting the effective characteristics into the trained SVM classifier, and adaptively generating a degradation stage label of the test set.
(5) The TCN-based prediction model is established, the structural schematic diagram is shown in FIG. 5, and the specific setting of the hyper-parameters is 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 the mean absolute error, and the optimization algorithm is Adam. After the model training is finished, inputting a test data set, and finally outputting an RUL predicted value of the test data set;
TABLE 3TCN hyperparameter Table
(6) Visualization of the RUL prediction results. Taking the bearings 1-3, the bearings 1-5, the bearings 2-6 and the bearings 3-3 as examples, firstly, the degradation stage adaptive partitioning mechanism, i.e., the segmented degradation model, provided by the embodiment is applied to partition the whole life cycle of the bearings into two stages of stable degradation and rapid degradation in an adaptive manner, and then, the prediction model based on TCN provided by the embodiment is applied to the RUL in the two degradation stages for prediction, and the prediction result is shown in fig. 6. As can be seen from fig. 6, the present embodiment is excellent in solving the problem of RUL prediction of bearings, which verifies the effectiveness of the present invention.
In order to further verify the effectiveness of the degradation stage adaptive partitioning mechanism, namely the segmented degradation model, designed by the embodiment, a comparison experiment is performed on the degradation stage adaptive partitioning mechanism and a bearing data set of a full-life-cycle unified degradation model under a working condition 1. FIG. 7 shows the visual results of the bearings 1-3 and 1-5 applying the TCN method to predict RUL based on the complete life cycle unified degradation model and the segmented degradation model. From fig. 7, it can be found that, compared with the establishment of a full-life-cycle unified degradation model, the degradation stage adaptive partitioning mechanism, namely the segmented degradation model, provided by the invention can obtain more and better prediction effects, and particularly, the prediction accuracy can be improved in the rapid degradation stage of the impending failure of a bearing, which verifies the effectiveness of the degradation stage adaptive partitioning mechanism, namely the segmented degradation model, designed by the invention.
In order to further verify the effectiveness of the improved TCN model designed by the present invention in solving the RUL prediction problem, the improved TCN model is compared with a Long-short-term memory network (LSTM), a Convolutional Neural Network (CNN), and a Support Vector Regression (SVR) on a bearing data set under a working condition 1, and all the four methods perform RUL prediction on the basis of a piecewise degradation model, and analyze the experimental results by using Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE), and the calculation method specifically includes:
wherein,
and y i Respectively representing the predicted value and the real value of the RUL of the ith sample. As can be seen from Table 4, TCN is large relative to LSTM, CNN and SVRThe RUL on most bearings has higher prediction precision, which verifies the effectiveness of the TCN model established by the invention.
TABLE 4 experimental results of TCN with LSTM, CNN and SVR on bearing data set for Condition 1
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.
Claims (10)
1. A method for predicting the residual service life of a rotary machine based on a time convolution network is characterized by comprising the following steps:
(1) Acquiring vibration signal data of a single or multiple rotary machines in a full life cycle of an operation stage, and extracting time domain characteristics, frequency domain characteristics and time-frequency characteristics from the vibration signal data of the full life cycle along a specific direction;
(2) Selecting effective features from the extracted features by adopting a comprehensive evaluation index Cri; wherein the comprehensive evaluation index Cri is as follows: cri = ω 1 Corr+ω 2 Mon-ω 3 Dis, where Dis, corr and Mon are respectively dynamic time warping distance index, correlation coefficient index and monotonicity index, and their weighting coefficients are respectively omega 3 、ω 1 And omega 2 ；
(3) Compressing the selected effective characteristics by using a t-SNE algorithm, and then adaptively dividing the degradation stage of the rotary machine by using a DBSCAN algorithm;
(4) And establishing a prediction model based on the TCN, training the prediction model, and predicting the residual service life of the rotary machine by adopting the trained prediction model.
2. The method for predicting the remaining service life of a rotating machine based on a time convolution network as claimed in claim 1, wherein: the prediction model comprises an input layer, a TCN layer and a regression layer; inputting the fused effective characteristics in the input layer; in the TCN layer, a plurality of convolution units are added in a residual block structure, and the structure of each convolution unit is expansion causal convolution + weight normalization + ReLU activation function + Dropout; in the regression layer, all local features extracted by the TCN layer are fused by establishing a flattening layer and a full connection layer, and finally the RUL predicted value is output.
3. The method for predicting the remaining service life of a rotating machine based on a time convolution network as claimed in claim 1, wherein: extracting time domain, frequency domain and time-frequency characteristics from the acquired full-life-cycle vibration signals along the horizontal direction and the vertical direction, and firstly extracting time domain statistical characteristics of the vibration signals; secondly, converting the discrete time domain signals into frequency domain signals by adopting fast Fourier transform, and extracting frequency domain characteristics; and finally, extracting the time-frequency characteristics of the vibration signals by using a Hilbert-Huang transform self-adaptive time-frequency analysis method.
4. The method for predicting the remaining service life of a rotating machine based on a time convolution network according to claim 1, wherein: compressing the selected features from a high-dimensional space to a low-dimensional space by adopting a t-SNE algorithm, respectively applying Gaussian distribution and t distribution to convert the distance between state points into corresponding joint probability distribution in the high-dimensional space and the low-dimensional space, then measuring the difference between the two distributions by adopting Kullback-Leibler divergence, and optimizing the KL divergence by adopting a gradient descent method, thereby finally realizing that the feature state points which are far away from each other in the original high-dimensional space are farther apart and the feature state points which are near to each other are closer to each other after mapping.
5. The method of claim 4, wherein the time convolutional network-based rotating machine remaining service life prediction method comprises: after the characteristic compression, the distances between the characteristic state points in the same degradation stage and the different degradation stages are respectively closer and farther, so that the boundaries between the different degradation stages are further clarified, and the subsequent self-adaptive degradation stage division is facilitated.
6. The method for predicting the remaining service life of a rotating machine based on a time convolution network as claimed in claim 1, wherein: the DBSCAN algorithm enables adaptive identification of degradation modes according to the actual health status of the rotating machine and automatically divides the full life cycle of the rotating machine into multiple degradation stages based on the degradation modes.
7. The method of claim 6, wherein the method comprises: let the full lifecycle signature sequence be X i ＝{x 1 ,x 2 ,…,x T Where T is the characteristic sequence length, introduce a cluster marker array m i ：
Where n is the number of clusters, thereby obtaining an array of cluster markers
X i Is divided into n clusters and noise point sets, i.e. the full life cycle is divided into n degradation stages, and corresponding degradation stage labels S are generated label (ii) a Subsequently, the selected valid feature vector is summed with S label And combining to form a feature label pair, inputting the feature label pair into a support vector machine classification model for training, and subsequently inputting the effective feature vector of the test set into the trained SVM classification model so as to obtain the degradation stage label of the test set.
8. The method for predicting the remaining service life of a rotating machine based on a time convolution network as claimed in claim 1, wherein: firstly, mapping the selected effective characteristics from a high-dimensional space to a low-dimensional space by adopting a t-SNE algorithm; secondly, based on the distribution difference between different health states generated by the t-SNE algorithm, the DBSCAN algorithm is used for adaptively dividing the full life cycle of the training set into different degradation stages; then, fusing the effective characteristics of the training set, combining the effective characteristics with corresponding degradation stage labels to form label pairs, inputting the label pairs into an SVM classifier, and training the SVM classifier; and finally, fusing the effective characteristics of the test set, inputting the effective characteristics into the trained SVM classifier, and adaptively generating a degradation stage label of the test set.
9. The method of claim 8, wherein the time convolutional network-based rotating machine remaining service life prediction method comprises: the full life cycle is adaptively divided into two stages of stable degradation and rapid degradation, and then the prediction is carried out on the RUL of the two stages of stable degradation and rapid degradation by applying a prediction model based on TCN.
10. The method for predicting the remaining service life of a rotating machine based on a time convolution network as claimed in claim 2, wherein: the real RUL value is set as the percentage of the degradation of the rotating machine, and the real RUL value label of the t-th sampling point is as follows:
wherein, RUL t And RUL 0 Respectively representing the sequencing t of the sampling points and the total number of the sampling points in the whole life cycle; before input into the input layer, normalization is carried out, and the corresponding formula is as follows:
wherein x is i (t) and
respectively representing the values of the ith original and normalized feature sequences at time t,
and/or>
Respectively, the average value, the maximum value and the minimum value of the ith characteristic sequence value at all the time instants. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211699953.2A CN115982621A (en) | 2022-12-28 | 2022-12-28 | Rotary machine residual service life prediction method based on time convolution network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211699953.2A CN115982621A (en) | 2022-12-28 | 2022-12-28 | Rotary machine residual service life prediction method based on time convolution network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115982621A true CN115982621A (en) | 2023-04-18 |
Family
ID=85967690
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211699953.2A Pending CN115982621A (en) | 2022-12-28 | 2022-12-28 | Rotary machine residual service life prediction method based on time convolution network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115982621A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116226646A (en) * | 2023-05-05 | 2023-06-06 | 国家石油天然气管网集团有限公司 | Method, system, equipment and medium for predicting health state and residual life of bearing |
-
2022
- 2022-12-28 CN CN202211699953.2A patent/CN115982621A/en active Pending
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116226646A (en) * | 2023-05-05 | 2023-06-06 | 国家石油天然气管网集团有限公司 | Method, system, equipment and medium for predicting health state and residual life of bearing |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109102005B (en) | Small sample deep learning method based on shallow model knowledge migration | |
| US8630962B2 (en) | Error detection method and its system for early detection of errors in a planar or facilities | |
| US11840998B2 (en) | Hydraulic turbine cavitation acoustic signal identification method based on big data machine learning | |
| CN110703057A (en) | Power equipment partial discharge diagnosis method based on data enhancement and neural network | |
| CN110647830B (en) | Bearing fault diagnosis method based on convolutional neural network and Gaussian mixture model | |
| JP6783443B2 (en) | Information processing equipment, information processing systems, information processing methods, programs, and recording media | |
| Tran et al. | Single and combined fault diagnosis of reciprocating compressor valves using a hybrid deep belief network | |
| Qiu et al. | A piecewise method for bearing remaining useful life estimation using temporal convolutional networks | |
| CN117611015B (en) | Real-time monitoring system for quality of building engineering | |
| CN109034076A (en) | A kind of automatic clustering method and automatic cluster system of mechanical fault signals | |
| CN112183643A (en) | Hard rock tension-shear fracture identification method and device based on acoustic emission | |
| CN115982621A (en) | Rotary machine residual service life prediction method based on time convolution network | |
| Hong et al. | Supervised-learning-based intelligent fault diagnosis for mechanical equipment | |
| Kaloni et al. | Nonparametric damage detection and localization model of framed civil structure based on local gravitation clustering analysis | |
| CN112200238A (en) | Hard rock tension-shear fracture identification method and device based on sound characteristics | |
| Pang et al. | Discrete Cosine Transformation and Temporal Adjacent Convolutional Neural Network‐Based Remaining Useful Life Estimation of Bearings | |
| Mobtahej et al. | An LSTM‐Autoencoder Architecture for Anomaly Detection Applied on Compressors Audio Data | |
| CN115063664A (en) | Model learning method, training method and system for industrial vision detection | |
| CN117782198B (en) | Highway electromechanical equipment operation monitoring method and system based on cloud edge architecture | |
| Kong et al. | Vibration fault analysis of hydropower units based on extreme learning machine optimized by improved sparrow search algorithm | |
| CN115293030A (en) | Bearing residual service life prediction method based on deep mutual learning and dynamic feature construction | |
| CN117689321A (en) | Service data processing method and device | |
| Hao et al. | New fusion features convolutional neural network with high generalization ability on rolling bearing fault diagnosis | |
| CN112348700B (en) | Line capacity prediction method combining SOM clustering and IFOU equation | |
| JP2022036054A (en) | Inspection device, inspection method, and inspection program for strung wire |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |