CN115660198B - Method for predicting residual service life of rolling bearing - Google Patents
Method for predicting residual service life of rolling bearing Download PDFInfo
- Publication number
- CN115660198B CN115660198B CN202211386138.0A CN202211386138A CN115660198B CN 115660198 B CN115660198 B CN 115660198B CN 202211386138 A CN202211386138 A CN 202211386138A CN 115660198 B CN115660198 B CN 115660198B
- Authority
- CN
- China
- Prior art keywords
- sample
- module
- layer
- bearing
- life
- 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.)
- Active
Links
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 provides a method for predicting the residual service life of a rolling bearing, which comprises the steps of constructing a training sample set; training the constructed bearing service life prediction model by adopting a training sample set to obtain a trained bearing service life prediction model; in actual use, for a certain appointed bearing B, the residual service life of the appointed bearing B at the current moment is predicted based on a trained bearing service life prediction model. The method for predicting the residual service life of the rolling bearing can effectively improve the accuracy and the efficiency of predicting the residual service life of the rolling bearing.
Description
Technical Field
The invention belongs to the technical field of prediction of residual service life of rotary machinery, and particularly relates to a method for predicting residual service life of a rolling bearing.
Background
With the proposal of industry 4.0, engineering equipment is continuously developed towards intellectualization and complicating. Therefore, in practical application, extremely strict requirements are put on the reliability and safety of engineering equipment. Bearings are widely used in rotating machinery as one of the core components of modern engineering equipment. Because the bearing is always operated under certain load conditions, some degradation tendency is unavoidable and normal use of the engineering equipment is ultimately affected. In the modern industry context, unexpected failure of bearings can not only result in significant property damage, but even catastrophic consequences. Related studies have shown that about 30% of rotary machine failures are caused by bearing failures. Therefore, in order to effectively avoid engineering accidents caused by bearing faults, predicting the residual service life (RUL) of the bearing has great research value.
The existing researches show that the RUL prediction method is mainly divided into three types: a physical model-based method, a data driven method, and a hybrid method. As engineering equipment becomes increasingly complex, it becomes extremely difficult to obtain a physical model of the equipment failure mechanism. Meanwhile, with the development of artificial intelligence technology, various data-driven algorithms have been developed. The data driving method can extract the characteristic information related to the performance degradation trend from the service life data of engineering equipment, and can effectively save cost and time. Therefore, the data driving method is a mainstream method in the field of RUL prediction.
With the advent of the age of 'big data and cloud computing', most of data which can be acquired in an industrial system has the characteristics of mass, high and nonlinearity. The traditional data driving method is difficult to automatically process a large amount of high-dimensional nonlinear data, and deep learning is a new technology developed by a neural network, so that a new solution idea is provided for training a large amount of high-dimensional nonlinear data by using a powerful feature extraction function. Currently, the residual life prediction method based on deep learning mainly includes a CNN-based method and an RNN-based method.
However, both CNN-based and RNN-based RUL prediction methods have limitations. For the former, CNN is limited and has a self structure, the capability of capturing long-term dependency is very limited, and more data space features are extracted. In the latter case, the RNN architecture is designed to process data in a serial fashion and cannot be calculated in parallel, thus requiring significant time and cost in training and prediction.
Therefore, there is a need to develop a RUL prediction method that can compensate for the defects of CNN and RNN in processing long sequence data.
Disclosure of Invention
Aiming at the defects existing in the prior art, the invention provides a method for predicting the residual service life of a rolling bearing, which can effectively solve the problems.
The technical scheme adopted by the invention is as follows:
the invention provides a method for predicting the residual service life of a rolling bearing, which comprises the following steps:
step 1, constructing a training sample set V= { V 1 ,V 2 ,...,V n-C+1 }；
Step 1.1, selecting a plurality of rolling bearings with the same type, and presetting a plurality of working conditions; setting the corresponding relation between each rolling bearing and the working condition;
for each rolling bearing, degrading the rolling bearing under the corresponding working condition, collecting vibration signals of the rolling bearing in the whole process from starting to complete degradation through a vibration acceleration sensor, and arranging the rolling bearings according to the collection time sequence to form an original sample set S= { S 1 ,S 2 ,...,S n }；
Step 1.2, for the original sample set s= { S 1 ,S 2 ,...,S n Each original sample S in } i Where i=1, 2, n, standardized treatment, obtaining a normalized sample S' i Thereby forming a normalized sample set S '= { S' 1 ,S' 2 ,...,S' n }；
Step 1.3, for the normalized sample S' i Inputting into a feature extractor for multi-scale feature extraction to obtain a sample S' after feature extraction " i Thereby forming a feature-extracted sample set S "= { S". 1 ,S" 2 ,...,S" n }；
Step 1.4, for the sample S' after feature extraction " i Calculating the ratio of the sampling time to the total service life to obtain a corresponding service life ratio P i As its tag data, the sample S' after feature extraction " i Ratio to lifetime P i Splicing to obtain a spliced sample with a label, denoted as (S' i ,P i )；
Thus, n time-ordered labeled samples were obtained, each: (S) " 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" n ,P n )；
Step 1.5, for labeled sample set (S', 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" n ,P n ) C continuous samples with labels form a training sample, and a training sample set V= { V is constructed in a mode that the moving step length is 1 1 ,V 2 ,...,V n-C+1 }；
The specific method is as follows:
from sample 1 labeled (S', 1 ,P 1 ) Beginning, to sample C labeled (S', C ,P C ) Until (S) " 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" C ,P C ) Formation of training sample V1 1 ＝(S" 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" C ,P C )；
Moving 1 step from sample 2 with label (S', 2 ,P 2 ) Beginning, sample (S' to the C+1th labeled sample " C+1 ,P C+1 ) Until (S) " 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" C+1 ,P C+1 ) Form training sample No. 2V 2 ＝(S" 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" C+1 ,P C+1 )；
And so on
Up to the nth-C+1 labeled samples (S', n-C+1 ,P n-C+1 ) Beginning, to the nth labeled sample (S'; n ,P n ) Until (S) " n-C+1 ,P n-C+1 ),(S" n-C+2 ,P n-C+2 ),...,(S" n ,P n ) Form the n-C+1 training sample V n-C+1 ＝(S" n-C+1 ,P n-C+1 ),(S" n-C+2 ,P n-C+2 ),...,(S" n ,P n )；
Thereby forming a training sample set v= { V 1 ,V 2 ,...,V n-C+1 }；
Step 2, training sample set v= { V 1 ,V 2 ,...,V n-C+1 Performing a prediction model of the service life of the constructed bearingTraining to obtain a trained bearing service life prediction model;
the specific training mode is as follows:
for each training sample V j ＝(S" j ,P j ),(S" j+1 ,P j+1 ),...,(S" j+C-1 ,P j+C-1 ) Wherein j=n-C+1, and inputting the j=n-C+1 into a constructed bearing service life prediction model to obtain a life ratio predicted value P 'of the bearing in the next step length' j Predicted value of life ratio P' j And the next labeled sample (S', j+C ,P j+C ) Tag P of (2) j+C Comparing to obtain a life ratio prediction difference value, adjusting model parameters of the bearing service life prediction model according to the life ratio prediction difference value, further training the bearing service life prediction model with the adjusted parameters by adopting a next training sample, and iterating continuously until a trained bearing service life prediction model is obtained;
And 3, in actual use, predicting the residual service life of a certain appointed bearing B at the current moment by adopting the following modes:
step 3.1, designating bearing B from the beginning of operation time t start To the current time t now Continuously collecting vibration signals to form a sample set S B ＝{S B 1 ,S B 2 ,...,S B u -a }; u is the number of samples acquired from the bearing B;
step 3.2, for sample set S B ＝{S B 1 ,S B 2 ,...,S B u Each sample in the sequence is subjected to standardization treatment and feature extraction to obtain a sample set S' after feature extraction " B ＝{S" B 1 ,S" B 2 ,...,S" B u }；
Step 3.3, for the feature extracted sample set S' B ＝{S" B 1 ,S" B 2 ,...,S" B u Constructing a test sample by C continuous samples with a moving step length of 1Sample set V B ＝{V B 1 ,V B 2 ,...,V B u-C+1 }；
Test sample set V B ＝{V B 1 ,V B 2 ,...,V B u-C+1 Each test sample in the sequence is sequentially input into a trained bearing service life prediction model to sequentially obtain life ratio predicted values at corresponding sampling moments, so that u-C+1 life ratio predicted values which are sequentially arranged are obtained, and the predicted values are respectively expressed as: p'. 1 B ,P' B 2 ,...,P' B u-C+1 ；
Step 3.4, predicting the value P 'of the u-C+1 life ratio' 1 B ,P' B 2 ,...,P' B u-C+1 Smoothing and curve fitting are carried out to obtain the designated bearing B from the starting working time t start To the current time t now A life prediction curve Z of (2);
analyzing the extending trend of the life prediction curve Z to obtain a life prediction curve Z 'after the extending trend, wherein the intersection point of the life prediction curve Z' after the extending trend and the straight line y=1 is a coordinate point at the end of the life of the appointed bearing B, and the corresponding abscissa time is the total life prediction value of the appointed bearing B;
Thus, the bearing B is specified at the current time t now Residual life prediction value = total life prediction value-used life value, which is the current time t now Thereby completing the prediction of the remaining service life of the specified bearing B.
Preferably, in step 1.1, the original sample set s= { S 1 ,S 2 ,...,S n The } is obtained by:
when the working time of the rolling bearing is t=0, after 1 Δt time length, collecting vibration signals in the T time according to the sampling frequency f, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:form 1 st originalSample S 1 Thus, it is->Wherein the two-dimensional vibration signal includes a horizontal direction vibration signal and a vertical direction vibration signal;
starting from the time t=0, collecting vibration signals in the time T according to the sampling frequency f after 2 Δt time lengths, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:form the 2 nd original sample S 2 Thus, it is->
And so on
When the device is completely degraded, assuming that n delta T time lengths pass, collecting vibration signals in T time according to sampling frequency f, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as: Form the nth original sample S n Thus, it is->
Therefore, n original samples are acquired in a total time sequence in the whole life process of the rolling bearing to form an original sample set S= { S 1 ,S 2 ,...,S n }。
Preferably, in step 1.2, for each raw sample S i Where i=1, 2, n, standardized treatment, obtaining a normalized sample S' i The method specifically comprises the following steps:
original sampleIs>Wherein k=1, 2,..m, performing normalization processing using the following formula to obtain a normalized two-dimensional vibration signal +.>
Wherein: mean is the original sample set s= { S 1 ,S 2 ,...,S n The mean value of all two-dimensional vibration signals in }, σ is the original sample set s= { S 1 ,S 2 ,...,S n Variance of all two-dimensional vibration signals in };
thereby obtaining a normalized sample
Preferably, the step 1.3 specifically comprises:
step 1.3.1: the model structure of the preset feature extractor comprises a CNNbLock1 module, a CNNbLock2 module, a CNNbLock3 module, a CNNbLock4 module, a CNNbLock5 module, a residual network structure, a global tie pooling layer, a GRU module and a feature splicing module;
the CNNblock1 module and the GRU module are connected in series to form a GRU path;
the CNNbLock1 module, the CNNbLock2 module, the CNNbLock3 module, the CNNbLock4 module, the CNNblock5 module, the residual network structure and the global tie pooling layer form a depth convolution path with residual;
Step 1.3.2: normalized sample S' i Inputting the shallow features into a CNNblock1 module to extract shallow features B' i ；
Step 1.3.3: shallow features B' i Deep feature extraction is carried out through a CNNblock2 module, a CNNblock3 module, a CNNblock4 module and a CNNblock5 module in sequence to obtain deep features D' i ；
Make shallow layer feature B' i And deep features D' i Commonly input to a residual network structure to enable shallow layer characteristics B' i And deep features D' i Adding, inputting the added features into a global tie pooling layer for pooling to obtain spatial features space' i ；
Step 1.3.4: will be shallow layer feature B' i Input to GRU module for time feature extraction to obtain time feature time' i ；
Step 1.3.5: space the space features' i And time feature time' i The three samples are input into a characteristic splicing module together for characteristic splicing to obtain multi-scale space-time characteristics, namely, a sample S' after characteristic extraction " i 。
Preferably, the step 2 specifically comprises:
the constructed bearing service life prediction model comprises four layers of encoders, a full-connection layer and a Sigmoid layer; each layer of encoder comprises a multi-head self-attention mechanism layer, a feedforward full-connection layer, a first normalization layer, a second normalization layer, a first residual error layer and a second residual error layer;
Each training sample V j ＝(S" j ,P j ),(S" j+1 ,P j+1 ),...,(S" j+C-1 ,P j+C-1 ) Inputting the first intermediate feature Mid to a self-attention mechanism layer;
training sample V j Adding the first residual layer and the intermediate feature Mid, and inputting the added first residual layer and the intermediate feature Mid into a first normalization layer to obtain a second intermediate feature Mid';
the second intermediate feature Mid 'passes through the feedforward full connecting layer to obtain a third intermediate feature Mid';
adding the second intermediate feature Mid ' with the third intermediate feature Mid ' through a second residual layer, and inputting the added second intermediate feature Mid ' into a second normalization layer to obtain an advanced feature High;
the advanced feature High sequentially passes through the full connection layer and the Sigmoid layer to obtain the life ratio predicted value P 'of the advanced feature High in the next step length' j 。
Preferably, in step 3.4, the predicted values P for the u-C+1 lifetime ratios' 1 B ,P' B 2 ,...,P' B u-C+1 The smoothing process is carried out, specifically:
presetting a moving average term number g; g is an odd number; the result of dividing g by 2 is taken down as an integer, g 1
Predicted value P 'for each life ratio' B z Wherein z=1, 2,..u-c+1, centered around it, is taken forward adjacent g 1 Taking g backward from life ratio predictive value 1 Predicted values of the life ratio, and then P' B z 2g is obtained 1 Taking average value of life ratio predicted values as life ratio predicted value P' B z A lifetime ratio predicted value after smoothing treatment;
If a certain life ratio prediction value P' B z With no adjacent g at the front 1 Predicted values of the ratio of the service lives, or that the rear surface does not have adjacent g 1 Predicted values of the life ratio are then calculated as the predicted values of the life ratio P' B z And discarding.
The method for predicting the residual service life of the rolling bearing has the following advantages:
the method for predicting the residual service life of the rolling bearing can effectively improve the accuracy and the efficiency of predicting the residual service life of the rolling bearing.
Drawings
FIG. 1 is a flow chart of an implementation of the rolling bearing RUL prediction method based on MSGCNN-TR.
FIG. 2 is a diagram of the MSGCNN-TR network model in accordance with the present invention.
Fig. 3 is a time domain diagram of a rolling bearing 1_1 in the embodiment of the invention.
Fig. 4 is a time domain diagram of the rolling bearing 1_3 in the embodiment of the invention.
Fig. 5 is a root mean square of the rolling bearing 1_1 in the embodiment of the present invention.
Fig. 6 is a root mean square of the rolling bearing 1_3 in the embodiment of the present invention.
Fig. 7 is a feature extraction result of the rolling bearing 1_1 in the embodiment of the invention.
Fig. 8 is a MA processing result of the rolling bearing 1_3 in the embodiment of the invention.
Fig. 9 is a MA processing result of the rolling bearing 1_7 in the embodiment of the invention.
Fig. 10 is a polynomial fit result of the rolling bearing 1_3 in the embodiment of the invention.
Fig. 11 is a polynomial fit result of the rolling bearing 1_7 in the embodiment of the invention.
Detailed Description
In order to make the technical problems, technical schemes and beneficial effects solved by the invention more clear, the 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 for purposes of illustration only and are not intended to limit the scope of the invention.
The invention aims to overcome the defects of the prior art and provides a rolling bearing residual service life prediction method based on MSGCNN-TR. The method predicts the RUL of the rolling bearing through a transducer prediction model (MSGCNN-TR) with a multiscale gated convolutional network. The model mainly comprises two parts: feature extraction combining CNN and GRU and transform-based RUL prediction. The part 1 is actually a feature extractor, which consists of a depth convolution path with residual errors and a GRU path, and can effectively extract spatial features and temporal features which can reflect the degradation trend of the bearing in the life data of the rolling bearing. The 2 nd part is an RUL prediction model formed by four layers of transformers, and a multi-head self-attention mechanism in the transformers can effectively capture long-period and short-period dependency relations in time sequence data, so that RUL prediction precision and efficiency of the rolling bearing can be effectively improved.
FIG. 1 is a flow chart of an implementation of a method for predicting the residual service life of a rolling bearing based on MSGCNN-TR. As shown in fig. 1, the specific steps of the present invention include:
step 1, constructing a training sample set V= { V 1 ,V 2 ,...,V n-C+1 }；
Step 2 of the method, in which the step 2,employing training sample set v= { V 1 ,V 2 ,...,V n-C+1 Training the constructed bearing service life prediction model to obtain a trained bearing service life prediction model;
and 3, in actual use, predicting the residual service life of a certain designated bearing B at the current moment by adopting the following way.
The specific implementation steps of the invention are shown in fig. 2, and the method comprises the following steps:
step 1, constructing a training sample set V= { V 1 ,V 2 ,...,V n-C+1 }；
Step 1.1, selecting a plurality of rolling bearings with the same type, and presetting a plurality of working conditions; setting the corresponding relation between each rolling bearing and the working condition;
for each rolling bearing, arranging a vibration acceleration sensor on the rolling bearing to enable the vibration acceleration sensor to degenerate under corresponding working conditions, collecting vibration signals of the rolling bearing in the whole process from starting to complete degradation through the vibration acceleration sensor, and arranging the vibration signals according to the collection time sequence to form an original sample set S= { S 1 ,S 2 ,...,S n }；
In this step, the original sample set s= { S 1 ,S 2 ,...,S n The } is obtained by:
When the working time of the rolling bearing is t=0, after 1 Δt time length, collecting vibration signals in the T time according to the sampling frequency f, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:form 1 st original sample S 1 Thus, it is->Wherein the two-dimensional vibration signal includes a horizontal direction vibration signal and a vertical direction vibration signal;
starting from the time t=0, collecting vibration signals in the time T according to the sampling frequency f after 2 deltat time lengths, and sharingAcquiring m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:form the 2 nd original sample S 2 Thus, it is->
And so on
When the device is completely degraded, assuming that n delta T time lengths pass, collecting vibration signals in T time according to sampling frequency f, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:form the nth original sample S n Thus, it is->
Therefore, n original samples are acquired in a total time sequence in the whole life process of the rolling bearing to form an original sample set S= { S 1 ,S 2 ,...,S n }。
Step 1.2, for the original sample set s= { S 1 ,S 2 ,...,S n Each original sample S in } i Where i=1, 2, n, standardized treatment, obtaining a normalized sample S' i Thereby forming a normalized sample set S '= { S' 1 ,S' 2 ,...,S' n }；
In this step, for each original sample S i Where i=1, 2, n, standardized treatment, obtaining a normalized sample S' i The method specifically comprises the following steps:
original sampleIs>Wherein k=1, 2,..m, performing normalization processing using the following formula to obtain a normalized two-dimensional vibration signal +.>
Wherein: mean is the original sample set s= { S 1 ,S 2 ,...,S n The mean value of all two-dimensional vibration signals in }, σ is the original sample set s= { S 1 ,S 2 ,...,S n Variance of all two-dimensional vibration signals in };
thereby obtaining a normalized sample
Step 1.3, for the normalized sample S' i Inputting into a feature extractor for multi-scale feature extraction to obtain a sample S' after feature extraction " i Thereby forming a feature-extracted sample set S "= { S". 1 ,S" 2 ,...,S" n }；
The method specifically comprises the following steps:
step 1.3.1: the model structure of the preset feature extractor comprises a CNNbLock1 module, a CNNbLock2 module, a CNNbLock3 module, a CNNbLock4 module, a CNNbLock5 module, a residual network structure, a global tie pooling layer, a GRU module and a feature splicing module;
the CNNblock1 module and the GRU module are connected in series to form a GRU path;
The CNNbLock1 module, the CNNbLock2 module, the CNNbLock3 module, the CNNbLock4 module, the CNNblock5 module, the residual network structure and the global tie pooling layer form a depth convolution path with residual;
step 1.3.2: normalized sample S' i Inputting the shallow features into a CNNblock1 module to extract shallow features to obtain a shallow layerFeature B' i The method comprises the steps of carrying out a first treatment on the surface of the The CNNblock1 module is a large convolution kernel, the size of the convolution kernel is 48 x 1, and the purpose is to extract shallow features and delete feature information which is useless for a target task.
Specifically, the CNN block1 module comprises a convolution layer, a ReLU activation function layer, a batch normalization layer and a MAXPooling layer. Wherein the size of the convolution kernel is 48 x 1. Thus, the normalized sample S' i Feature extraction is sequentially carried out through a convolution layer, a ReLU activation function layer is activated, a batch normalization layer is normalized, and a MAXPooling layer is subjected to pooling treatment to obtain a sample S i 'shallow features B' i 。
Step 1.3.3: shallow features B' i Deep feature extraction is carried out through a CNNblock2 module, a CNNblock3 module, a CNNblock4 module and a CNNblock5 module in sequence to obtain deep features D' i ；
The CNNblock2 module, the CNNblock3 module, the CNNblock4 module and the CNNblock5 module have the same structure as the CNN block1 module, but the size of a convolution kernel is 3*1, and the convolution kernel is a small convolution kernel and is used for performing dimension reduction processing on data, so that on one hand, the interference of redundant information can be greatly reduced, and the deep space features in the data can be extracted; on the other hand, the method is helpful for deepening the network structure and inhibiting the occurrence of the over-fitting condition. In order to reduce the feature dimension, reduce the complexity of the model and avoid gradient disappearance and gradient explosion, a ReLU activation function is used in each convolution layer, a batch normalization layer and a MAXPooling layer are used after each convolution layer, and a dropout is used for randomly inactivating neurons, so that the phenomenon of overfitting is prevented.
Make shallow layer feature B' i And deep features D' i Commonly input to a residual network structure to enable shallow layer characteristics B' i And deep features D' i Adding, inputting the added features into a global tie pooling layer for pooling to obtain spatial features space' i ；
The global tie pooling layer is adopted, so that parameters of a network can be further reduced, and the phenomenon of over-fitting is prevented.
Step 1.3.4: will be shallow layer feature B' i Input to GRU moduleExtracting time features to obtain time features time' i ；
Specifically, the GRU module is formed by a gate cycle unit. The core structure of the gate cycle unit mainly comprises two parts, namely an update gate and a reset gate. Shallow features B' i Sequentially passing through an update gate and a reset gate of the gate circulation unit to obtain time characteristic time' i ；
Step 1.3.5: space the space features' i And time feature time' i The three samples are input into a characteristic splicing module together for characteristic splicing to obtain multi-scale space-time characteristics, namely, a sample S' after characteristic extraction " i 。
The feature extractor provided by the invention has the following characteristics:
the model structure of the feature extractor is constructed from a depth convolution path with residuals and a GRU path. In order to make up for the limitation that a deep convolution path with residual connection is not good at extracting time sequence characteristics, a GRU path is introduced to extract time sequence characteristics of data.
Furthermore, the GRU path may help to improve the ability of the model to fuse local contexts at each time, as the attention mechanism in the transducer module used in the residual life prediction process makes the output features insensitive to local contexts. In order not to destroy the extracted features of the two paths, the features of the two paths are fused by adopting a splicing method, long-distance dependency relationship in time sequence data is captured, noise in an input signal is restrained, and the extracted features are richer. And after the feature extraction is completed, splicing the extracted multi-scale feature information with target information representing the residual life in the time sequence data to obtain a sample with a label.
Step 1.4, for the sample S' after feature extraction " i Calculating the ratio of the sampling time to the total service life to obtain a corresponding service life ratio P i As its tag data, the sample S' after feature extraction " i Ratio to lifetime P i Splicing to obtain a spliced sample with a label, denoted as (S' i ,P i )；
Thus, n time-ordered labeled samples were obtained, each: (S) " 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" n ,P n )；
Step 1.5, for labeled sample set (S', 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" n ,P n ) C continuous samples with labels form a training sample, and a training sample set V= { V is constructed in a mode that the moving step length is 1 1 ,V 2 ,...,V n-C+1 }；
The specific method is as follows:
from sample 1 labeled (S', 1 ,P 1 ) Beginning, to sample C labeled (S', C ,P C ) Until (S) " 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" C ,P C ) Formation of training sample V1 1 ＝(S" 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" C ,P C )；
Moving 1 step from sample 2 with label (S', 2 ,P 2 ) Beginning, sample (S' to the C+1th labeled sample " C+1 ,P C+1 ) Until (S) " 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" C+1 ,P C+1 ) Form training sample No. 2V 2 ＝(S" 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" C+1 ,P C+1 )；
And so on
Up to the nth-C+1 labeled samples (S', n-C+1 ,P n-C+1 ) Beginning, to the nth labeled sample (S'; n ,P n ) Until (S) " n-C+1 ,P n-C+1 ),(S" n-C+2 ,P n-C+2 ),...,(S" n ,P n ) Form the n-C+1 training sample V n-C+1 ＝(S" n-C+1 ,P n-C+1 ),(S" n-C+2 ,P n-C+2 ),...,(S" n ,P n )；
Thereby forming a training sample set v= { V 1 ,V 2 ,...,V n-C+1 }；
Step 2, training sample set v= { V 1 ,V 2 ,...,V n-C+1 Training the constructed bearing service life prediction model to obtain a trained bearing service life prediction model;
the specific training mode is as follows:
for each training sample V j ＝(S" j ,P j ),(S" j+1 ,P j+1 ),...,(S" j+C-1 ,P j+C-1 ) Wherein j=n-C+1, and inputting the j=n-C+1 into a constructed bearing service life prediction model to obtain a life ratio predicted value P 'of the bearing in the next step length' j Predicted value of life ratio P' j And the next labeled sample (S', j+C ,P j+C ) Tag P of (2) j+C Comparing to obtain a life ratio prediction difference value, adjusting model parameters of the bearing service life prediction model according to the life ratio prediction difference value, further training the bearing service life prediction model with the adjusted parameters by adopting a next training sample, and iterating continuously until a trained bearing service life prediction model is obtained;
As a preferred mode, the constructed bearing service life prediction model comprises four layers of encoders, a full-connection layer and a Sigmoid layer; each layer of encoder comprises a multi-head self-attention mechanism layer, a feedforward full-connection layer, a first normalization layer, a second normalization layer, a first residual error layer and a second residual error layer;
each training sample V j ＝(S" j ,P j ),(S" j+1 ,P j+1 ),...,(S" j+C-1 ,P j+C-1 ) Inputting the first intermediate feature Mid to a self-attention mechanism layer;
training sample V j Adding the first residual layer and the intermediate feature Mid, and inputting the added first residual layer and the intermediate feature Mid into a first normalization layer to obtain a second intermediate feature Mid';
the second intermediate feature Mid 'passes through the feedforward full connecting layer to obtain a third intermediate feature Mid';
adding the second intermediate feature Mid ' with the third intermediate feature Mid ' through a second residual layer, and inputting the added second intermediate feature Mid ' into a second normalization layer to obtain an advanced feature High;
the advanced feature High sequentially passes through the full connection layer and the Sigmoid layer to obtain the life ratio predicted value P 'of the advanced feature High in the next step length' j 。
And 3, in actual use, predicting the residual service life of a certain appointed bearing B at the current moment by adopting the following modes:
step 3.1, designating bearing B from the beginning of operation time t start To the current time t now Continuously collecting vibration signals to form a sample set S B ＝{S B 1 ,S B 2 ,...,S B u -a }; u is the number of samples acquired from the bearing B;
step 3.2, for sample set S B ＝{S B 1 ,S B 2 ,...,S B u Each sample in the sequence is subjected to standardization treatment and feature extraction to obtain a sample set S' after feature extraction " B ＝{S" B 1 ,S" B 2 ,...,S" B u }；
Step 3.3, for the feature extracted sample set S' B ＝{S" B 1 ,S" B 2 ,...,S" B u Constructing a test sample set V by using C continuous samples to form a test sample and moving the sample set V by a step length of 1 B ＝{V B 1 ,V B 2 ,...,V B u-C+1 }；
Test sample set V B ＝{V B 1 ,V B 2 ,...,V B u-C+1 Each test sample in the sequence is sequentially input into a trained bearing service life prediction model to sequentially obtain life ratio predicted values at corresponding sampling moments, so that u-C+1 life ratio predicted values which are sequentially arranged are obtained, and the predicted values are respectively expressed as: p'. 1 B ,P' B 2 ,...,P' B u-C+1 ；
Step 3.4, predicting values for the u-C+1 lifetime ratiosP' 1 B ,P' B 2 ,...,P' B u-C+1 Because of the large fluctuation, the smoothing process is performed first, and then curve fitting is performed, for example, a polynomial fitting method is adopted to obtain the designated bearing B from the starting operation time t start To the current time t now A life prediction curve Z of (2);
analyzing the extending trend of the life prediction curve Z to obtain a life prediction curve Z 'after the extending trend, wherein the intersection point of the life prediction curve Z' after the extending trend and the straight line y=1 is a coordinate point at the end of the life of the appointed bearing B, and the corresponding abscissa time is the total life prediction value of the appointed bearing B;
Thus, the bearing B is specified at the current time t now Residual life prediction value = total life prediction value-used life value, which is the current time t now Thereby completing the prediction of the remaining service life of the specified bearing B.
In this step, predicted values of the ratio of u-C+1 life times P' 1 B ,P' B 2 ,...,P' B u-C+1 The smoothing process is carried out, specifically:
presetting a moving average term number g; g is an odd number; the result of dividing g by 2 is taken down as an integer, g 1
Predicted value P 'for each life ratio' B z Wherein z=1, 2,..u-c+1, centered around it, is taken forward adjacent g 1 Taking g backward from life ratio predictive value 1 Predicted values of the life ratio, and then P' B z 2g is obtained 1 Taking average value of life ratio predicted values as life ratio predicted value P' B z A lifetime ratio predicted value after smoothing treatment;
if a certain life ratio prediction value P' B z With no adjacent g at the front 1 Predicted values of the ratio of the service lives, or that the rear surface does not have adjacent g 1 Predicted values of the life ratio are then calculated as the predicted values of the life ratio P' B z And discarding.
The present invention and its effects will be specifically described below by way of one example.
The experimental data of this example was derived from the rolling bearing dataset provided by the IEEE PHM challenge in 2012, which was acquired by the proctisia experimental platform. The experimental device mainly comprises 3 modules: the degradation generation module, the rotation module and the measurement module realize the accelerated degradation of 17 rolling bearings under 3 different working conditions. Wherein operating mode 1 and operating mode 2 contain the life data of 7 bearings respectively, and operating mode 3 contains the life data of 3 bearings. The detailed information is shown in table 1. The data acquisition process is to acquire vibration signals within 0.1s every 10s by using an acceleration sensor, wherein the sampling frequency is 25.6kHz, namely 2560 two-dimensional vibration signals are included in each sample. The two-dimensional vibration signals comprise a horizontal vibration signal and a vertical vibration signal, and the horizontal vibration signal and the vertical vibration signal comprise useful information of bearing degradation, so that the invention adopts the vibration signals in the horizontal direction and the vertical direction for experiments.
Table 1I working condition information of EEE PHM2012 bearing dataset
According to the information provided by the data set, the first two bearings under each working condition are selected, the total life data (all two-dimensional vibration signals from the beginning of running to complete degradation of the bearings) of the total 6 bearings are used as basic data for constructing a training sample set, and the non-life data (two-dimensional vibration signals from the beginning of running to a certain moment) of the remaining 11 bearings are used as basic data for constructing a test sample set, so that an RUL prediction experiment is performed, as shown in table 2. The number of samples used for constructing the training sample set is 7534, and the number of samples used for constructing the test sample set is 13959.
In order to understand the performance degradation trend of the rolling bearing used in the experimental process, the life data of each bearing in the horizontal direction and the vertical direction are visually displayed. It is found that the bearings contained in the data set have two failure modes in total, one is fatigue failure belonging to normal degradation, and the other is transient failure belonging to abnormal degradation. As shown in fig. 3 and 4, time domain diagrams of the bearing 1_1 and the bearing 1_2, respectively, represent life data in the horizontal direction of the bearing. The amplitude of the life data of the bearing 1_1 in the horizontal direction gradually increases with the increase of time, and the degradation trend of the bearing shows gradual characteristics, and the bearing belongs to fatigue failure. The amplitude of the life data of the bearing 1_2 in the horizontal direction always tends to be stable before the bearing fails, and suddenly increases at the end of the life, and belongs to transient failure. To further analyze the degradation trend of the bearing, the change in Root Mean Square (RMS) of the life data of the bearing was also analyzed. As shown in fig. 5 and 6, the root mean square of the rolling bearing 1_1 and the root mean square of the rolling bearing 1_3 show life data of the bearing 1_1 and the bearing 1_2 in the horizontal direction. The root mean square of the life data in the horizontal direction of the bearing 1_1 gradually increases with time, while the root mean square of the life data in the horizontal direction of the bearing 1_2 tends to stabilize and does not suddenly increase until the end of the life. Thus, the bearing is more difficult to predict when it encounters a transient failure.
Table 2 data information of IEEE phm2012 bearing dataset
In order to implement the proposed method, appropriate experimental parameters need to be set to train the network.
Firstly, normalizing an original sample, then inputting the normalized sample into a constructed multi-scale feature extractor for feature extraction, and reducing 2560 dimensions of each sample into 32 dimensions to obtain a sample after feature extraction. Taking the bearing 1_1 as an example, a feature map after dimension reduction is shown in fig. 7. As can be seen from fig. 7, most of the extracted features monotonically increase, have a significant degradation tendency, and although some feature values do not show a monotonically increasing tendency, they approach zero, so that the predicted result is almost unaffected, and in general, the feature extractor is excellent.
Then, for the samples after feature extraction, if the samples are samples for constructing a training sample set, adopting a mode of step 1.4, and taking the life ratio of each sample after feature extraction as the label data of each sample to obtain a sample with a label;
setting C to 10 in step 1.5, and starting from sample S' with the 1 st label " 1 Beginning, sample S' to 10 th labeled " 10 Until the 1 st training sample V is formed 1 ＝(S" 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" 10 ,P 10 ) Inputting the predicted value into a constructed bearing service life prediction model, and predicting to obtain a predicted value P 'of the service life ratio of the bearing in the next step length' 11 Then the life ratio predictive value P 'is calculated' 11 Sample S' with 11 th label " 11 Is a tag value P of (2) 11 Comparing, and adjusting model parameters of a bearing service life prediction model according to the difference value;
from sample S' with label 2, using a movement step of 1 " 2 Beginning, sample S' to 11 th labeled " 11 Until the 2 nd training sample V is formed 2 ＝(S" 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" 11 ,P 11 ) And the like, a plurality of training samples are obtained.
That is, for samples used to construct the training sample set, one training sample is constructed with 10 time-step samples, and the life ratio prediction for the next time step is performed. And the step size of each window movement is 1.
Taking the Beraing1_1 as an example, 2803 samples with labels are totally included, and 2792 training samples are obtained through the processing in the above manner, so as to form a training sample set V.
And inputting the constructed training sample set V into the constructed transducer prediction model for training, and carrying out 6-fold cross validation according to the training sample set consisting of 6 bearings. In the model training process, the optimizer uses a random gradient descent method, the learning rate is set to be 0.001, and the loss function uses MSE loss.
The test sample set is not provided with a label, and is constructed by a two-dimensional vibration signal from the beginning of running to a certain moment. The construction method is the same as that of the training sample set, and will not be described here again.
In order to evaluate the effect of a trained transducer prediction model on a test sample set, objective evaluation indexes are required to be adopted for measurement, and three evaluation indexes are mainly adopted in the invention: scoring functions, mean Absolute Error (MAE) and Normalized Root Mean Square Error (NRMSE).
The scoring function is a scoring standard specified by the IEEE PHM2012 challenge game, and the calculation method is as follows:
wherein: actrll i Andthe actual residual life and the predicted residual life of the ith test sample are respectively A i The score for ith bearing, r for the number of bearings tested,% Er for the error magnitude.
MAE and NRMSE are commonly used as predictive problem performance evaluation indicators, calculated as follows:
wherein: actrll i Andthe actual remaining life and the predicted remaining life of the ith test sample are respectively, and r represents the number of bearings tested.
After the transducer predictive model is pre-trained, the test sample set is input into the trained transducer predictive model, and the degradation trend of the bearing in the test sample set can be obtained. Taking the bearings 1_3 and 1_7 as an example, as shown in fig. 8 and 9, MA processing results of the rolling bearing 1_3 and MA processing result diagrams of the rolling bearing 1_7 are respectively.
As can be seen from fig. 8 and 9, the prediction result directly output from the MSGCNN-TR has a serious oscillation phenomenon. To alleviate this, the prediction results were smoothed using the method of step 3.4. Meanwhile, as can be seen from fig. 8 and 9, the obtained curve after smoothing can effectively alleviate the random oscillation phenomenon in prediction. And then fitting the obtained smoothed curve by using a polynomial fitting model.
As shown in fig. 10 and 11, the polynomial fitting results of the rolling bearing 1_3 and the rolling bearing 1_7 are respectively obtained, and after the fitting is completed, the overall life degradation trend of the bearing can be obtained, and the predicted value of the remaining service life of the bearing can be obtained. As can be seen from fig. 10 and 11, the fitted curve can well complete the prediction task.
To further demonstrate the advantages of the present invention, three comparative experiments were set up. The three groups of comparison experiments are based on the invention, only the transducer prediction model in the model is respectively changed into RNN, LSTM and GRU modules, and other parts are kept unchanged. RNN networks are classical methods in time series data processing, LSTM and GRU are variants of RNNs that are better at handling long-term dependencies in time series than RNNs. Table 3 shows the experimental results of the life prediction mode (called MSGCNN-TR) based on the transducer prediction model and the other three comparison methods, and it can be seen from Table 3 that the RUL prediction method of the MSGCNN-TR of the invention has better effect on the three evaluation indexes of MAE and NRMSE than the other three methods and the average error is minimum. It was demonstrated that the transducer predictive model of the invention performed better than RNN and its variants in this experiment.
To demonstrate the effectiveness of each part of the overall structure in the present invention, 3 ablation experiments were set up, see table 4 for details. The results of the ablation experiments are shown in Table 5, and it can be seen from Table 5 that in the MSGCNN-TR architecture of the present invention, each module has an important effect on the RUL prediction results. The GRU unit adopted in the characteristic extraction process can capture long-term dependency relationship in an input sample, so that a network learns more time information, and similarly, a deep convolution path with residual errors can be adopted to extract more spatial characteristics, thereby improving RUL prediction performance of the whole learning architecture. If the RUL output result does not adopt smoothing processing or polynomial fitting in the prediction process, the RUL prediction result is seriously affected.
Table 3 results of comparative experiments
Table 4 ablation protocol
Table 5 ablation experimental results
In summary, according to the method for predicting the residual service life of the rolling bearing based on the MSGCNN-TR, the space features and the time features of the rolling bearing time sequence are respectively captured by adopting a deep convolution path with residual errors and a GRU path, and then the captured multi-scale features are input into a model constructed by a multi-layer transform encoder for training and prediction. Experiments prove that the method can effectively improve the prediction precision and efficiency of RUL.
The invention has the following beneficial effects:
1) The invention provides a rolling bearing RUL prediction model based on MSGCNN-TR, which utilizes CNN and GRU to construct a feature extractor and utilizes a transducer to conduct RUL prediction and has the advantages of high prediction life precision and accuracy.
2) The invention constructs a feature extractor based on a multi-scale gating convolution network for the first time, and the feature extractor can capture the space features and the time features implied by the life data of the rolling bearing by using a depth convolution path with residual errors and a GRU path.
3) The invention sets a series of comparison experiments and a plurality of ablation experiments on the IEEE PHM2012 data set, the comparison experiments prove that the performance of the model provided by the invention is obviously superior to that of other comparison methods, and meanwhile, the ablation experiments prove the necessity and the effectiveness of each module in the model provided by the invention.
While the foregoing describes illustrative embodiments of the present invention to facilitate an understanding of the present invention by those skilled in the art, it should be understood that the present invention is not limited to the scope of the embodiments, but is to be construed as protected by the accompanying claims insofar as various changes are within the spirit and scope of the present invention as defined and defined by the appended claims.
Claims (3)
1. The method for predicting the residual service life of the rolling bearing is characterized by comprising the following steps of:
step 1, constructing a training sample set V= { V 1 ,V 2 ,...,V n-C+1 }；
Step 1.1, selecting a plurality of rolling bearings with the same type, and presetting a plurality of working conditions; setting the corresponding relation between each rolling bearing and the working condition;
for each rolling bearing, degrading it under the corresponding working condition and adding by vibrationThe speed sensor collects vibration signals of the rolling bearing in the whole process from the beginning of working to complete degradation, and the vibration signals are arranged according to the collection time sequence to form an original sample set S= { S 1 ,S 2 ,...,S n }；
Step 1.2, for the original sample set s= { S 1 ,S 2 ,...,S n Each original sample S in } i Where i=1, 2, n, standardized treatment, obtaining a normalized sample S' i Thereby forming a normalized sample set S '= { S' 1 ,S' 2 ,...,S' n }；
Step 1.3, for the normalized sample S' i Inputting into a feature extractor for multi-scale feature extraction to obtain a sample S' after feature extraction " i Thereby forming a feature-extracted sample set S "= { S". 1 ,S" 2 ,...,S" n }；
Step 1.4, for the sample S' after feature extraction " i Calculating the ratio of the sampling time to the total service life to obtain a corresponding service life ratio P i As its tag data, the sample S' after feature extraction " i Ratio to lifetime P i Splicing to obtain a spliced sample with a label, denoted as (S' i ,P i )；
Thus, n time-ordered labeled samples were obtained, each: (S) " 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" n ,P n )；
Step 1.5, for labeled sample set (S', 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" n ,P n ) C continuous samples with labels form a training sample, and a training sample set V= { V is constructed in a mode that the moving step length is 1 1 ,V 2 ,...,V n-C+1 }；
The specific method is as follows:
from sample 1 labeled (S', 1 ,P 1 ) Beginning, to sample C labeled (S', C ,P C ) Until (S) " 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" C ,P C ) Formation of training sample V1 1 ＝(S" 1 ,P 1 ),(S" 2 ,P 2 ),...,(S" C ,P C )；
Moving 1 step from sample 2 with label (S', 2 ,P 2 ) Beginning, sample (S' to the C+1th labeled sample " C+1 ,P C+1 ) Until (S) " 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" C+1 ,P C+1 ) Form training sample No. 2V 2 ＝(S" 2 ,P 2 ),(S" 3 ,P 3 ),...,(S" C+1 ,P C+1 )；
And so on;
up to the nth-C+1 labeled samples (S', n-C+1 ,P n-C+1 ) Beginning, to the nth labeled sample (S'; n ,P n ) Until (S) " n-C+1 ,P n-C+1 ),(S" n-C+2 ,P n-C+2 ),...,(S" n ,P n ) Form the n-C+1 training sample V n-C+1 ＝(S" n-C+1 ,P n-C+1 ),(S" n-C+2 ,P n-C+2 ),...,(S" n ,P n )；
Thereby forming a training sample set v= { V 1 ,V 2 ,...,V n-C+1 }；
Step 2, training sample set v= { V 1 ,V 2 ,...,V n-C+1 Training the constructed bearing service life prediction model to obtain a trained bearing service life prediction model;
the specific training mode is as follows:
for each training sample V j ＝(S" j ,P j ),(S" j+1 ,P j+1 ),...,(S" j+C-1 ,P j+C-1 ) Wherein j=n-C+1, and inputting the j=n-C+1 into a constructed bearing service life prediction model to obtain a life ratio predicted value P 'of the bearing in the next step length' j Predicted value of life ratio P' j And the next labeled sample (S', j+C ,P j+C ) Tag P of (2) j+C Comparing to obtain life ratio predictionThe difference value is predicted according to the life ratio, the model parameters of the bearing service life prediction model are adjusted, the next training sample is adopted to further train the bearing service life prediction model with the adjusted parameters, and iteration is continued until a trained bearing service life prediction model is obtained;
and 3, in actual use, predicting the residual service life of a certain appointed bearing B at the current moment by adopting the following modes:
step 3.1, designating bearing B from the beginning of operation time t start To the current time t now Continuously collecting vibration signals to form a sample set S B ＝{S B 1 ,S B 2 ,...,S B u -a }; u is the number of samples acquired from the bearing B;
step 3.2, for sample set S B ＝{S B 1 ,S B 2 ,...,S B u Each sample in the sequence is subjected to standardization treatment and feature extraction to obtain a sample set S' after feature extraction " B ＝{S" B 1 ,S" B 2 ,...,S" B u }；
Step 3.3, for the feature extracted sample set S' B ＝{S" B 1 ,S" B 2 ,...,S" B u Constructing a test sample set V by using C continuous samples to form a test sample and moving the sample set V by a step length of 1 B ＝{V B 1 ,V B 2 ,...,V B u-C+1 }；
Test sample set V B ＝{V B 1 ,V B 2 ,...,V B u-C+1 Each test sample in the sequence is sequentially input into a trained bearing service life prediction model to sequentially obtain life ratio predicted values at corresponding sampling moments, so that u-C+1 life ratio predicted values which are sequentially arranged are obtained, and the predicted values are respectively expressed as: p'. 1 B ,P' B 2 ,...,P' B u-C+1 ；
Step 3.4Predicted values of the ratio of u-C+1 life times P' 1 B ,P' B 2 ,...,P' B u-C+1 Smoothing and curve fitting are carried out to obtain the designated bearing B from the starting working time t start To the current time t now A life prediction curve Z of (2);
analyzing the extending trend of the life prediction curve Z to obtain a life prediction curve Z 'after the extending trend, wherein the intersection point of the life prediction curve Z' after the extending trend and the straight line y=1 is a coordinate point at the end of the life of the appointed bearing B, and the corresponding abscissa time is the total life prediction value of the appointed bearing B;
thus, the bearing B is specified at the current time t now Residual life prediction value = total life prediction value-used life value, which is the current time t now Thereby completing the prediction of the remaining service life of the designated bearing B;
in step 1.1, the original sample set s= { S 1 ,S 2 ,...,S n The } is obtained by:
when the working time of the rolling bearing is t=0, after 1 Δt time length, collecting vibration signals in the T time according to the sampling frequency f, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as: Form 1 st original sample S 1 Thus, it is->Wherein the two-dimensional vibration signal includes a horizontal direction vibration signal and a vertical direction vibration signal;
starting from the time t=0, collecting vibration signals in the time T according to the sampling frequency f after 2 Δt time lengths, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:formation ofSample No. 2 2 Thus, it is->
And so on;
when the device is completely degraded, assuming that n delta T time lengths pass, collecting vibration signals in T time according to sampling frequency f, and collecting m two-dimensional vibration signals to form a two-dimensional vibration signal sequence, wherein the two-dimensional vibration signal sequence is expressed as:form the nth original sample S n Thus, it is->
Therefore, n original samples are acquired in a total time sequence in the whole life process of the rolling bearing to form an original sample set S= { S 1 ,S 2 ,...,S n }；
The step 1.3 specifically comprises the following steps:
step 1.3.1: the model structure of the preset feature extractor comprises a CNNbLock1 module, a CNNbLock2 module, a CNNbLock3 module, a CNNbLock4 module, a CNNbLock5 module, a residual network structure, a global tie pooling layer, a GRU module and a feature splicing module;
the CNNblock1 module is a large convolution kernel, the size of the convolution kernel is 48 x 1, and the CNNblock1 module is used for extracting shallow features and deleting feature information useless for a target task; the CNN block1 module comprises a convolution layer, a ReLU activation function layer, a batch normalization layer and a MAXPooling layer;
The CNNblock2 module, the CNNblock3 module, the CNNblock4 module and the CNNblock5 module have the same structure as the CNN block1 module, but the size of the convolution kernel is 3*1, and the convolution kernel is a small convolution kernel for performing dimension reduction processing on data; using a ReLU activation function in each convolution layer, using a batch normalization layer and a MAXPooling layer after each convolution layer, and using dropout to randomly inactivate neurons;
the CNNblock1 module and the GRU module are connected in series to form a GRU path;
the CNNbLock1 module, the CNNbLock2 module, the CNNbLock3 module, the CNNbLock4 module, the CNNblock5 module, the residual network structure and the global tie pooling layer form a depth convolution path with residual;
step 1.3.2: normalized sample S' i Inputting the shallow features into a CNNblock1 module to extract shallow features B' i ；
Step 1.3.3: shallow features B' i Deep feature extraction is carried out through a CNNblock2 module, a CNNblock3 module, a CNNblock4 module and a CNNblock5 module in sequence to obtain deep features D' i ；
Make shallow layer feature B' i And deep features D' i Commonly input to a residual network structure to enable shallow layer characteristics B' i And deep features D' i Adding, inputting the added features into a global tie pooling layer for pooling to obtain spatial features space' i ；
Step 1.3.4: will be shallow layer feature B' i Input to GRU module for time feature extraction to obtain time feature time' i ；
Step 1.3.5: space the space features' i And time feature time' i The three samples are input into a characteristic splicing module together for characteristic splicing to obtain multi-scale space-time characteristics, namely, a sample S' after characteristic extraction " i ；
The step 2 is specifically as follows:
the constructed bearing service life prediction model comprises four layers of encoders, a full-connection layer and a Sigmoid layer; each layer of encoder comprises a multi-head self-attention mechanism layer, a feedforward full-connection layer, a first normalization layer, a second normalization layer, a first residual error layer and a second residual error layer;
each training sample V j ＝(S" j ,P j ),(S" j+1 ,P j+1 ),...,(S" j+C-1 ,P j+C-1 ) Inputting the first intermediate feature Mid to a self-attention mechanism layer;
training sample V j Added to the intermediate feature Mid by a first residual layer, and thenInputting the second intermediate feature Mid' to a first normalization layer;
the second intermediate feature Mid 'passes through the feedforward full connecting layer to obtain a third intermediate feature Mid';
adding the second intermediate feature Mid ' with the third intermediate feature Mid ' through a second residual layer, and inputting the added second intermediate feature Mid ' into a second normalization layer to obtain an advanced feature High;
the advanced feature High sequentially passes through the full connection layer and the Sigmoid layer to obtain the life ratio predicted value P 'of the advanced feature High in the next step length' j 。
2. The method for predicting the remaining life of a rolling bearing according to claim 1, wherein in step 1.2, for each raw sample S i Where i=1, 2, n, standardized treatment, obtaining a normalized sample S' i The method specifically comprises the following steps:
original sampleIs>Wherein k=1, 2,..m, performing normalization processing using the following formula to obtain a normalized two-dimensional vibration signal +.>
Wherein: mean is the original sample set s= { S 1 ,S 2 ,...,S n The mean value of all two-dimensional vibration signals in }, σ is the original sample set s= { S 1 ,S 2 ,...,S n Variance of all two-dimensional vibration signals in };
thereby obtaining a normalized sample
3. The method for predicting remaining life of rolling bearing according to claim 1, wherein in step 3.4, the predicted value P 'of the ratio of u-c+1 life is calculated for' 1 B ,P' B 2 ,...,P' B u-C+1 The smoothing process is carried out, specifically:
presetting a moving average term number g; g is an odd number; the result of dividing g by 2 is taken down as an integer, g 1 ；
Predicted value P 'for each life ratio' B z Wherein z=1, 2,..u-c+1, centered around it, is taken forward adjacent g 1 Taking g backward from life ratio predictive value 1 Predicted values of the life ratio, and then P' B z 2g is obtained 1 Taking average value of life ratio predicted values as life ratio predicted value P' B z A lifetime ratio predicted value after smoothing treatment;
if a certain life ratio prediction value P' B z With no adjacent g at the front 1 Predicted values of the ratio of the service lives, or that the rear surface does not have adjacent g 1 Predicted values of the life ratio are then calculated as the predicted values of the life ratio P' B z And discarding.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211386138.0A CN115660198B (en) | 2022-11-07 | 2022-11-07 | Method for predicting residual service life of rolling bearing |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211386138.0A CN115660198B (en) | 2022-11-07 | 2022-11-07 | Method for predicting residual service life of rolling bearing |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115660198A CN115660198A (en) | 2023-01-31 |
| CN115660198B true CN115660198B (en) | 2023-08-01 |
Family
ID=85017025
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211386138.0A Active CN115660198B (en) | 2022-11-07 | 2022-11-07 | Method for predicting residual service life of rolling bearing |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115660198B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117191394B (en) * | 2023-11-06 | 2024-01-19 | 南京凯奥思数据技术有限公司 | Fault detection method and device for air compressor equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112257333A (en) * | 2020-09-24 | 2021-01-22 | 浙江工业大学 | Mechanical equipment internal component service life prediction method based on deep learning |
| CN113947017A (en) * | 2021-10-18 | 2022-01-18 | 北京化工大学 | Method for predicting residual service life of rolling bearing |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111832216B (en) * | 2020-04-14 | 2022-07-12 | 新疆大学 | Rolling bearing residual service life prediction method based on EEMD-MCNN-GRU |
| CN112036547B (en) * | 2020-08-28 | 2023-09-22 | 徐工汉云技术股份有限公司 | Rolling bearing residual life prediction method combining automatic feature extraction with LSTM |
| US20220187819A1 (en) * | 2020-12-10 | 2022-06-16 | Hitachi, Ltd. | Method for event-based failure prediction and remaining useful life estimation |
| CN114626625A (en) * | 2022-03-25 | 2022-06-14 | 华南理工大学 | Bearing life prediction method based on multi-module U-BiLSTM network |
| CN114925723B (en) * | 2022-05-06 | 2023-04-07 | 盐城工学院 | Method for predicting residual service life of rolling bearing by adopting encoder and decoder |
-
2022
- 2022-11-07 CN CN202211386138.0A patent/CN115660198B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112257333A (en) * | 2020-09-24 | 2021-01-22 | 浙江工业大学 | Mechanical equipment internal component service life prediction method based on deep learning |
| CN113947017A (en) * | 2021-10-18 | 2022-01-18 | 北京化工大学 | Method for predicting residual service life of rolling bearing |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115660198A (en) | 2023-01-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Xiang et al. | LSTM networks based on attention ordered neurons for gear remaining life prediction | |
| CN111274737A (en) | Method and system for predicting remaining service life of mechanical equipment | |
| CN113642754B (en) | Complex industrial process fault prediction method based on RF noise reduction self-coding information reconstruction and time convolution network | |
| Fu et al. | A re-optimized deep auto-encoder for gas turbine unsupervised anomaly detection | |
| CN111459144A (en) | Airplane flight control system fault prediction method based on deep cycle neural network | |
| Jiang et al. | Dual residual attention network for remaining useful life prediction of bearings | |
| CN115660198B (en) | Method for predicting residual service life of rolling bearing | |
| CN114880925A (en) | Equipment life prediction method based on time convolution network and multi-layer self-attention | |
| CN112434390A (en) | PCA-LSTM bearing residual life prediction method based on multi-layer grid search | |
| CN115791174B (en) | Rolling bearing abnormality diagnosis method, system, electronic equipment and storage medium | |
| CN114297918A (en) | Aero-engine residual life prediction method based on full-attention depth network and dynamic ensemble learning | |
| Chen et al. | Real-time bearing remaining useful life estimation based on the frozen convolutional and activated memory neural network | |
| CN114841072A (en) | Differential fusion Transformer-based time sequence prediction method | |
| Kong et al. | A contrastive learning framework enhanced by unlabeled samples for remaining useful life prediction | |
| Tang et al. | Prediction of bearing performance degradation with bottleneck feature based on LSTM network | |
| Du et al. | A remaining useful life prediction method based on time–frequency images of the mechanical vibration signals | |
| CN114169091A (en) | Method for establishing prediction model of residual life of engineering mechanical part and prediction method | |
| Shi et al. | A dual attention LSTM lightweight model based on exponential smoothing for remaining useful life prediction | |
| Yang et al. | Physics-informed multi-state temporal frequency network for RUL prediction of rolling bearings | |
| CN116465623A (en) | Gearbox service life prediction method based on sparse converter | |
| Guo et al. | Operation stage division and RUL prediction of bearings based on 1DCNN-ON-LSTM | |
| Wang et al. | Few‐shot multiscene fault diagnosis of rolling bearing under compound variable working conditions | |
| Zhang et al. | A Flexible Monitoring Framework via Dynamic-Multilayer Graph Convolution Network | |
| CN114841063A (en) | Aero-engine residual life prediction method based on deep learning | |
| Lin | Intelligent Fault Diagnosis of Consumer Electronics Sensor in IoE via Transformer |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |