CN117252083A - Bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption - Google Patents
Bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption Download PDFInfo
- Publication number
- CN117252083A CN117252083A CN202310849683.7A CN202310849683A CN117252083A CN 117252083 A CN117252083 A CN 117252083A CN 202310849683 A CN202310849683 A CN 202310849683A CN 117252083 A CN117252083 A CN 117252083A
- Authority
- CN
- China
- Prior art keywords
- residual life
- fuzzy
- data
- domain data
- life prediction
- 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 70
- 238000006731 degradation reaction Methods 0.000 title claims abstract description 29
- 230000015556 catabolic process Effects 0.000 title claims abstract description 28
- 230000036541 health Effects 0.000 claims abstract description 77
- 238000012549 training Methods 0.000 claims abstract description 57
- 230000001373 regressive effect Effects 0.000 claims description 27
- 230000006978 adaptation Effects 0.000 claims description 19
- 239000011159 matrix material Substances 0.000 claims description 17
- 238000012360 testing method Methods 0.000 claims description 17
- 230000008569 process Effects 0.000 claims description 13
- 238000009499 grossing Methods 0.000 claims description 12
- 238000010219 correlation analysis Methods 0.000 claims description 10
- 238000005457 optimization Methods 0.000 claims description 10
- 238000012544 monitoring process Methods 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 9
- 238000010276 construction Methods 0.000 claims description 7
- 230000006835 compression Effects 0.000 claims description 6
- 238000007906 compression Methods 0.000 claims description 6
- 238000013433 optimization analysis Methods 0.000 claims description 6
- 238000013528 artificial neural network Methods 0.000 claims description 4
- 235000014653 Carica parviflora Nutrition 0.000 claims description 2
- 241000243321 Cnidaria Species 0.000 claims description 2
- 238000013508 migration Methods 0.000 abstract description 5
- 230000005012 migration Effects 0.000 abstract description 5
- 230000008901 benefit Effects 0.000 abstract description 2
- 238000005259 measurement Methods 0.000 abstract description 2
- 238000009826 distribution Methods 0.000 description 15
- 230000006870 function Effects 0.000 description 15
- 230000008859 change Effects 0.000 description 8
- 238000004364 calculation method Methods 0.000 description 6
- 238000000605 extraction Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 239000013598 vector Substances 0.000 description 5
- 238000000638 solvent extraction Methods 0.000 description 4
- 230000003044 adaptive effect Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 3
- 230000003862 health status Effects 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 2
- 230000007850 degeneration Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 239000000284 extract Substances 0.000 description 2
- 238000012804 iterative process Methods 0.000 description 2
- 238000013507 mapping Methods 0.000 description 2
- 238000005070 sampling Methods 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000035772 mutation Effects 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 230000003449 preventive effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 108020001568 subdomains Proteins 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
- 230000001052 transient effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
- G01M13/00—Testing of machine parts
- G01M13/04—Bearings
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/04—Ageing analysis or optimisation against ageing
-
- 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
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Geometry (AREA)
- Computer Hardware Design (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a bearing residual life prediction method and a system combining degradation phase division and subdomain self-adaption, wherein the method comprises the following steps of S1, constructing a health index curve; s2, dividing health stages; s3, constructing and training a residual life prediction model; s4, predicting the residual life. The advantages are that: different health phases present in the life cycle can be subtly mapped to different subfields, dispersing global aligned global differences to locally aligned phase differences. And in the migration prediction stage, utilizing a multi-order measurement based on the data label and the membership degree to pull up each degradation stage, and completing the alignment of the fuzzy substructure and the cross-domain regression. The invention can realize the characteristic alignment of finer granularity, construct a model with better generalization and realize more accurate residual life prediction.
Description
Technical Field
The invention relates to the technical field of bearing equipment state monitoring and health management, in particular to a bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption.
Background
Bearings are important components of mechanical equipment, and as the running time increases, the performance of the bearings gradually decreases, faults occur and even the usability of the equipment is affected, and how to guarantee the safety and reliability of the bearings is an important problem at present. Therefore, the prediction of faults and the health management (Prognostics and Health Management, PHM) of the bearings are an important issue, and the PHM of the bearings aims to predict possible problems by monitoring the health condition of the equipment, so as to avoid serious faults or accidents. The residual life prediction (Remaining useful life prediction, RUL prediction) is a core part of the PHM task, and its main task is to predict the length of time from the current moment to the time before the running device or component loses its operational capability, and then to formulate a preventive maintenance strategy according to the prediction result.
Currently, the residual life prediction methods of bearings can be roughly classified into two types according to the technology applied in the prediction process: a model-based residual life prediction method and a data-driven residual life prediction method. (1) a model-based residual life prediction method: the model-based approach utilizes a priori knowledge to build a model that characterizes bearing degradation, i.e., an empirical model is constructed from various mathematical physical models, to predict the remaining life of the bearing. The method has good interpretability, and can obtain good prediction effect under the conditions of completely understanding failure mechanism and constructing a proper model. However, the model-based residual life prediction method requires a great deal of expert knowledge and is poor in generalization, and physical characteristics of the bearing are difficult to comprehensively understand, so that the application of the method is limited. (2) a data-driven-based residual life prediction method: the data-driven residual life prediction method maps the degradation process to a relational function between the health status and the monitored data, and extracts and learns patterns from the available data to characterize the degradation behavior. The traditional data driving method adopts a simple machine learning algorithm to predict, but the network structure adopted by the algorithm is simple, and potential information among data cannot be better mined. With the development of artificial intelligence technology, deep learning has excellent performance in the task of predicting the residual life of a bearing due to the strong feature extraction and data processing capability.
However, deep learning models require a large amount of high quality training data to understand the underlying pattern of data and require the data to be tested to meet the same data distribution as the training data, which assumes the following challenges in actual remaining life prediction: (1) Firstly, each device is often under a complex and time-varying working condition, the training data and the test data are subjected to covariant offset, the acquisition of high-quality and uniformly distributed training data cannot be ensured, and when one RUL prediction model predicts other devices, the performance is drastically reduced, and the phenomenon is called as a cross-domain problem. (2) Meanwhile, the reliability and the safety of each device are continuously enhanced, the whole life cycle is more difficult to acquire, and the incomplete data to be tested and the training data have larger distribution deviation.
In view of the above challenges, deep learning model-based transfer learning methods have been developed and are now the focus of data-driven method research.
In the problem of predicting the residual life of the bearing under the cross-domain condition, the data distribution of the source domain and the target domain is different, and the difference is reduced by extracting the domain invariant common characteristics of the two domains. However, in practical problems, the overall distribution of the target domain data of the missing part period will change, and a larger difference is generated between the target domain data and the source domain data of the complete period, so that global domain self-adaption is difficult to directly perform, different health stages exist in the whole life period of the bearing, structural features of each stage are obvious, the distribution difference is large, the difference of the internal structures of the life period is ignored due to global alignment, the false matching of different substructures can be caused, and common potential local features of the data cannot be mined. Therefore, how to accurately predict the residual life of the bearing under the conditions of incomplete data and cross-domain conditions is a problem to be solved at present.
Disclosure of Invention
The invention aims to provide a bearing residual life prediction method and a system combining degradation phase division and sub-domain self-adaption, so as to solve the problems in the prior art.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
a method for predicting the residual life of bearing by combining degradation phase division and sub-domain self-adaption comprises the following steps,
s1, constructing a health index curve:
acquiring an original data set, extracting time domain and frequency domain characteristics of source domain data and target domain data in the original data set, and performing secondary index optimization and correlation analysis on the obtained multidimensional characteristics to acquire a health index curve;
s2, health stage division:
processing the health index curve by adopting a time sequence window weighted clustering algorithm, obtaining the health stage labels and the fuzzy membership degree of each source domain data and each target domain data, and dividing the health stage labels and the fuzzy membership degree into a training set and a testing set according to a proportion;
s3, constructing and training a residual life prediction model:
the residual life prediction model comprises a fuzzy subdomain feature extractor and a RUL regressive; inputting the training set into a residual life prediction model to train the residual life prediction model; in the training process, based on the alignment loss of the sub-structure fuzzy alignment module in the fuzzy subdomain feature extractor and the regression loss of the RUL regressive, acquiring total loss, optimizing parameters of the fuzzy subdomain feature extractor and the RUL regressive by minimizing the total loss, and acquiring and storing a trained residual life prediction model;
S4, residual life prediction:
and inputting the test set into a trained residual life prediction model to perform residual life prediction, and obtaining a residual life prediction result.
Preferably, the secondary index optimization specifically includes performing inter-feature difference alignment operation on the multi-dimensional features to obtain a relative index, and performing feature compression smoothing operation on the relative index to obtain a denoising index.
Preferably, the correlation analysis specifically includes measuring the correlation degree between the data at each moment and the data at the initial moment by using pearson correlation coefficient on the index after the secondary optimization, and constructing the health index.
Preferably, the source domain data is electromechanical equipment monitoring data collected when electromechanical equipment with a bearing runs to failure; the target domain data is monitoring data of electromechanical equipment from the operation of the electromechanical equipment without the bearing to the fault;
the source domain data are supervised data with the residual life of the bearing as a label; the target domain data is unsupervised data without a bearing remaining life tag.
Preferably, step S2 specifically includes,
s21, dividing the health index curve into windows, and replacing the current single-point numerical value with the statistical value of the windows to realize fuzzy smoothing of the health index curve;
S22, clustering the health index curves subjected to fuzzy smoothing, and obtaining health stage labels and fuzzy membership degrees of the health stage labels of the source domain data and the target domain data through iterative optimization of fuzzy membership degrees of the data points and clustering centers of the categories;
s23, dividing the source domain data, the target domain data, the health-stage labels to which the source domain data and the target domain data belong and the fuzzy membership degree into a training set and a testing set in proportion.
Preferably, step S3 comprises in particular,
s31, inputting a training set into a fuzzy subdomain feature extractor, and respectively acquiring feature potential representations of source domain data and target domain data in the training set by using a deep neural network of the fuzzy subdomain feature extractor to acquire high-dimensional feature matrixes corresponding to the source domain data and the target domain data;
s32, inputting high-dimensional feature matrixes corresponding to the source domain data and the target domain data into a substructure fuzzy alignment module of a fuzzy subdomain feature extractor, and acquiring alignment loss by calculating FLMMD between the feature matrixes of the source domain data and the target domain data and FLCORAL between the feature matrixes of the source domain data and the target domain data and the feature matrix time of the source domain data and the target domain data;
s33, inputting the time sequence feature matrix of the aligned source domain data and the target domain data in the training set, which are output by the fuzzy subdomain feature extractor, into an RUL regressive device, outputting the residual life predicted value of the source domain data and the target domain data in the training set, and calculating the mean square error of the residual life predicted value and the true value of the source domain as regression loss;
S34, calculating total loss according to the alignment loss and the regression loss;
and S35, minimizing total loss, and feeding back and adjusting network parameters of the fuzzy subdomain feature extractor and the RUL regressive to realize network training of the fuzzy subdomain feature extractor and the RUL regressive until training is completed, so as to obtain a trained residual life prediction model.
Preferably, step S32 specifically includes,
s321, alignment of the first portion: fuzzy local maximum mean value difference based on the maximum mean value difference, and meanwhile, the probability that each sample belongs to all categories is considered, so that the alignment of fuzzy subdomains with finer granularity is realized; calculating FLMMD between the source domain data and the target domain data feature matrix as a first partial loss;
s322, aligning the second part: fine granularity alignment is performed on the second order statistics based on Fuzzy Local CORAL of the second order statistics Correlation Alignment while considering the probability that each sample belongs to all categories; calculating FLCORAL of the source domain data and the target domain data feature matrix time as a second partial loss;
s323, integrating the first partial loss and the second partial loss to obtain the alignment loss.
Preferably, the fuzzy subdomain feature extractor is a ResNet50 feature extractor.
Preferably, the RUL regressor is a fully connected network-based regression predictor.
It is also an object of the present invention to provide a bearing residual life prediction system combining degradation phase partitioning and sub-domain adaptation, for implementing any of the above methods, the system comprising,
and a curve construction module: the method is used for constructing a health index curve;
acquiring an original data set, extracting time domain and frequency domain characteristics of source domain data and target domain data in the original data set, and performing secondary index optimization and correlation analysis on the obtained multidimensional characteristics to acquire a health index curve;
the stage division module: for dividing the health phase;
processing the health index curve by adopting a time sequence window weighted clustering algorithm, obtaining the health stage labels and the fuzzy membership degree of each source domain data and each target domain data, and dividing the health stage labels and the fuzzy membership degree into a training set and a testing set according to a proportion;
the prediction model building module: the method is used for constructing and training a residual life prediction model;
the residual life prediction model comprises a fuzzy subdomain feature extractor and a RUL regressive; inputting the training set into a residual life prediction model to train the residual life prediction model; in the training process, based on the alignment loss of the sub-structure fuzzy alignment module in the fuzzy subdomain feature extractor and the regression loss of the RUL regressive, acquiring total loss, optimizing parameters of the fuzzy subdomain feature extractor and the RUL regressive by minimizing the total loss, and acquiring and storing a trained residual life prediction model;
Life prediction module: for predicting remaining life;
and inputting the test set into a trained residual life prediction model to perform residual life prediction, and obtaining a residual life prediction result.
The beneficial effects of the invention are as follows: 1. the phase difference of global alignment to local alignment can be dispersed by fully and skillfully utilizing the phase existing in the life cycle corresponding to different subfields. 2. In the health stage division stage, a time sequence fuzzy clustering algorithm suitable for time sequence window data is provided, and unified and standardized health index construction and stage division processes are realized. In the migration stage, the degradation stages are pulled up by using a multi-order metric based on the data labels and membership degrees, so that the fuzzy substructure alignment and cross-domain regression are completed. 3. The method can realize the characteristic alignment of finer granularity, construct a model with better generalization, and simultaneously realize more accurate residual life prediction.
Drawings
FIG. 1 is a schematic flow chart of a method for predicting the residual life of a bearing in an embodiment of the invention;
FIG. 2 is a schematic diagram of the basic principle of transfer learning in an embodiment of the present invention;
FIG. 3 is a schematic diagram of intra-feature compression smoothing operations in an embodiment of the invention;
FIG. 4 is a schematic diagram of a TWW-FCM embodiment of the present invention;
fig. 5 is a schematic diagram of a framework of a Fuzzy Subdomain Adaptive Regression Network (FSARN) in an embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings, in order to make the objects, technical solutions and advantages of the present invention more apparent. It should be understood that the detailed description is presented by way of example only and is not intended to limit the invention.
Example 1
As shown in fig. 1 to 5, in the present embodiment, there is provided a bearing remaining life prediction method combining degradation phase division and sub-field adaptation, comprising the steps of,
1. health index curve construction:
the method comprises the steps of obtaining an original data set, extracting time domain and frequency domain characteristics of source domain data and target domain data in the original data set, and carrying out secondary index optimization and correlation analysis on the obtained multidimensional characteristics to obtain a health index curve.
The secondary index optimization comprises a relative index and a denoising index, wherein the relative index is obtained by performing inter-feature difference alignment operation on the multidimensional features, and the denoising index is obtained by performing intra-feature compression smoothing operation (shown in fig. 3) on the relative index.
The correlation analysis specifically comprises the step of measuring the correlation degree of data at each moment and initial moment data by using a pearson correlation coefficient on the index after secondary optimization, and constructing a health index HI.
The source domain data are electromechanical equipment monitoring data acquired when electromechanical equipment with a bearing runs to failure; the target domain data is the monitoring data from the operation of the bearingless electromechanical equipment to the failure electromechanical equipment, and only a part of the monitoring data is collected in the early stage.
The source domain data are supervised data with the residual life of the bearing as a label; the target domain data is unsupervised data without a bearing remaining life tag.
2. Health stage division:
and processing the health index curve by adopting a time sequence window weighted clustering algorithm (TWW-FCM algorithm), acquiring health stage labels and fuzzy membership degrees of the health stage labels to which each source domain data and each target domain data belong, and dividing the health stage labels and the fuzzy membership degrees into a training set and a testing set according to a proportion.
As shown in fig. 4, the TWW-FCM algorithm is a clustering algorithm suitable for phase-dividing time series data, and after the HI curve is subjected to fuzzy smoothing, the HI curve is clustered to obtain a health-stage label HS and a fuzzy membership degree. The method specifically comprises the following steps:
1. Dividing the health index curve into windows, and replacing the current single-point numerical value with the statistical value of the windows to realize fuzzy smoothing of the health index curve.
2. Clustering the fuzzy smoothed health index curves, and obtaining the health stage labels and the fuzzy membership degree of each source domain data and the target domain data by iteratively optimizing the fuzzy membership degree of each data point and the clustering center of each class.
3. The source domain data, the target domain data, the health stage labels to which the source domain data belong and the fuzzy membership degree are proportionally divided into a training set and a testing set.
3. Building and training a residual life prediction model:
the residual life prediction model comprises a fuzzy subdomain feature extractor and a RUL regressive; inputting the training set into a residual life prediction model to train the residual life prediction model; in the training process, based on the alignment loss of the sub-structure fuzzy alignment module in the fuzzy subdomain feature extractor and the regression loss of the RUL regressive, the total loss is obtained, and the parameters of the fuzzy subdomain feature extractor and the RUL regressive are optimized by minimizing the total loss, so that a trained residual life prediction model is obtained and stored.
Specifically comprises the following steps of,
1. Inputting the training set into a fuzzy subdomain feature extractor, and respectively acquiring feature potential representations of source domain data and target domain data in the training set by using a deep neural network of the fuzzy subdomain feature extractor to acquire high-dimensional feature matrixes corresponding to the source domain data and the target domain data;
2. inputting the high-dimensional feature matrixes corresponding to the source domain data and the target domain data into a substructure fuzzy alignment module of the fuzzy subdomain feature extractor, and obtaining the alignment loss by calculating FLMMD between the feature matrixes of the source domain data and the target domain data and FLCORAL between the feature matrixes of the source domain data and the target domain data.
The method comprises the steps of inputting a high-dimensional feature matrix of source domain data and target domain data into a sub-structure fuzzy alignment module, and learning domain invariant representation of a sub-domain hierarchy by minimizing feature distribution differences among sub-structures, wherein the sub-structure fuzzy alignment module comprises two parts of alignment, and specifically comprises the following steps:
2.1, first part alignment: based on the fuzzy local maximum mean difference (Fuzzy Local Maximum Mean Discrepancy, FLMMD) of the maximum mean difference (Maximum Mean Discrepancy, MMD), the FLMMD simultaneously considers the probability that each sample belongs to all categories, and realizes the fuzzy subdomain alignment of finer granularity, so that the source domain data and the target domain data are closer in the mapped feature space; the substructure blurring alignment module takes FLMMD between the source domain data and the target domain data feature matrix as a first partial loss
2.2, second part alignment: based on Fuzzy Local CORAL (FLCORAL) of the second order statistic Correlation Alignment (CORAL), FLCORAL also considers the probability that each sample belongs to all classes at the same time, performs fine-granularity alignment on the second order statistic, thereby capturing the complex difference between the two domains; the substructure blurring alignment module uses FLCORAL between the source domain data and the target domain data feature matrix as a second partial loss
2.3, integrating the first partial loss and the second partial loss to obtain the alignment loss. The two partial losses of the substructure fuzzy alignment module are combined, and the alignment loss is calculated by adopting the following formula
Wherein: gamma isWeight coefficient of (c) in the above-mentioned formula (c).
3. Inputting the time sequence feature matrix of the aligned source domain data and the target domain data in the training set, which are output by the fuzzy subdomain feature extractor, into an RUL regressive device, outputting the residual life predicted value of the source domain data and the target domain data in the training set, and calculating the mean-square error (MSE) of the residual life predicted value of the source domain and the true value as regression loss
4. Calculating model total loss from alignment loss and regression lossThe following calculation is used and,
wherein: beta isPenalty coefficients of (a).
5. And minimizing total loss, and feeding back and adjusting network parameters of the fuzzy subdomain feature extractor and the RUL regressive to realize network training of the fuzzy subdomain feature extractor and the RUL regressive until training is completed, so as to obtain a trained residual life prediction model.
By minimizing alignment lossDomain invariant representation can be learned to obtain domain invariant feature matrices between source and target domains, specifically, < >>For minimizing the first moment of the source domain and the target domain,/for the first moment of the source domain and the target domain>For minimizing the second moment between the source and target domains, the difference between the two domains is measured by calculating the distance between the covariance of the source and target samples. The two function together to enable the data distribution of the source domain and the target domain to be more approximate.
For the fuzzy subdomain feature extractor and the RUL regressor connected in series, a model total loss can be obtainedWill minimize the model total loss->To optimize the target, the network parameters of the fuzzy subdomain feature extractor and the RUL regressor are adjusted in a feedback manner to realize the fuzzy subdomainAnd (3) carrying out network training on the feature extractor and the RUL regressor until a training termination condition is met, and obtaining a residual life model after training.
In this embodiment, the fuzzy subfield feature extractor is a ResNet50 feature extractor. The RUL regressor is a regression predictor based on a fully connected network.
4. Residual life prediction:
and inputting the test set into a trained residual life prediction model to perform residual life prediction, and obtaining a residual life prediction result.
It is also an object of the present invention to provide a bearing residual life prediction system combining degradation phase partitioning and sub-domain adaptation, the system for implementing the method, the system comprising,
and a curve construction module: the method is used for constructing a health index curve;
acquiring an original data set, extracting time domain and frequency domain characteristics of source domain data and target domain data in the original data set, and performing secondary index optimization and correlation analysis on the obtained multidimensional characteristics to acquire a health index curve;
the stage division module: for dividing the health phase;
processing the health index curve by adopting a time sequence window weighted clustering algorithm, obtaining the health stage labels and the fuzzy membership degree of each source domain data and each target domain data, and dividing the health stage labels and the fuzzy membership degree into a training set and a testing set according to a proportion;
the prediction model building module: the method is used for constructing and training a residual life prediction model;
the residual life prediction model comprises a fuzzy subdomain feature extractor and a RUL regressive; inputting the training set into a residual life prediction model to train the residual life prediction model; in the training process, based on the alignment loss of the sub-structure fuzzy alignment module in the fuzzy subdomain feature extractor and the regression loss of the RUL regressive, acquiring total loss, optimizing parameters of the fuzzy subdomain feature extractor and the RUL regressive by minimizing the total loss, and acquiring and storing a trained residual life prediction model;
Life prediction module: for predicting remaining life;
and inputting the test set into a trained residual life prediction model to perform residual life prediction, and obtaining a residual life prediction result.
Example two
In this embodiment, by way of a specific example, the implementation process of the method of the present invention will be described in detail:
1. experimental data set acquisition and Health Index (HI) construction
The data set used in this example was the bearing data set of IEEE PHM Challenge 2012, which was obtained on a proctisia test platform consisting of three parts: a rotating part, a loading part and a data acquisition part.
The electric power of the rotating part was 250W, and the power was transmitted to the bearing through the rotating shaft. The load part is a pneumatic jack, and provides 4000N dynamic load for the bearing, so that the bearing is rapidly degraded. The data acquisition part comprises vibration data and temperature data, the vibration sensor consists of two micro accelerometers which are mutually positioned at 90 degrees, the micro accelerometers are respectively arranged on a horizontal axis and a vertical axis, the sampling frequency is 25.6kHz, sampling is carried out once every 10s for 0.1s, and the vibration sensor stops when the vibration amplitude reaches 20 g.
The data of the data set includes three conditions, the first is 4000N load, 1800rpm, the second is 4200N load, 1650rpm, and the last is 5000N load, 1500rpm. The HI creation includes the following steps:
1. And (3) obtaining time-frequency statistical characteristics:
as shown in fig. 2, 17 time domain and 8 frequency domain features of the original signal are extracted in order to fully extract potential features of the time sequence data, so that the variation trend of the bearing vibration signal is studied in multiple angles. The time domain signal refers to the change characteristics of the signal on a time axis, so that the operation trend of the equipment can be well described, the frequency domain signal can well represent the operation state of the equipment, and meanwhile, the noise information can be conveniently separated and removed. After a series of time-domain frequency domain feature extraction, the time-frequency feature is expressed as p= { p 1 ,p 2 ,…,p i ,…,p I I=25.
2. Secondary characteristic index extraction:
the features extracted in step 1 have individual differences, so there is no identical starting height between features, and there are spurious fluctuations and noise inside. To eliminate these effects, the present invention proposes two quadratic optimization feature indices, a relative index and a denoising index. The relative index has the functions of eliminating individual difference, pulling in distribution gaps and aligning various feature vectors, and the absolute change range is influenced due to the fact that the relative gaps exist in early-stage gentle data parts due to the fact that the initial positions of all statistical features are different in height, and the absolute change range is calculated by adopting a formula shown in the following formula (1):
Wherein p is i Representing a feature, i representing the serial number of the feature, p norm Representing the average number of features in steady operation. K represents the duration of the steady phase. After the inter-individual difference alignment operation, the time-frequency characteristic becomes the relative index p r ＝{p r1 ,p r2 ,…,p ri ,…,p rI }。
The denoising index is used for reducing abnormal data points, accurately controlling the fluctuation compression range of the local sequence and ensuring the monotonicity of the data trend. Ith denoising index p di The calculation formula in the slidable window is:
wherein p is rj Is the j-th data of the relative index obtained by the formula (1), wi is the length of the sliding window, x i Is p r The ith data of the corresponding time series. k (k) i And b i The parameters of the linear regression model fitted on the ith sliding window can be obtained by the least square method. Formulas (3) and (4) represent k, respectively i And b i Is expressed in closed form. H i Is a sliding window sample after compressing local sequence fluctuation, the expression of which is shown in formula (5), d i Is H i Starting point of (2):
wherein y is p r Is a value of (2). min (H) i ) And max (H) i ) Representing the upper and lower bounds of the index, which function to reduce the impact of outlier data on the trend. They can be represented by the following formula:
min(H i )＝μ-3σ
max(H i )＝μ+3σ
in the formula (6), μ and σ are each W i Mean and standard deviation of (a). After the relative index is subjected to intra-individual compression smoothing operation, a denoising index p is obtained d ＝{p d1 ,p d2 ,…,p ri ,…,p dI }。
The obtained denoising index p d I.e. the final data p' obtained after the secondary feature index extraction operation.
3. Constructing HI curves:
HI is an indicator of the change in health status throughout the life cycle and is typically represented as a trend curve. Since the bearing is degenerated into a gradual change process, the initial stage is considered as a normal state, and the degeneration stage data deviate from the normal state data along with the occurrence of operation faults, and the correlation between the degeneration stage data and the initial state data also changes, the correlation between the later operation stage data and the initial stage data can be reflected into a group of data points of different health stage states. Therefore, the invention carries out correlation analysis on the 25-dimensional characteristics after transformation, effectively measures the correlation degree of the data at two time points by using Pearson correlation coefficients, obtains the variation trend on a time axis, and constructs the universal HI. The specific method comprises the following steps: the initial part of the optimized 25-dimensional secondary characteristic index p' is selected as a basic value of a normal state, and then the Pearson correlation coefficient between the initial time and other times is calculated to be used as HI for measuring the state of the equipment, and the HI is expressed as follows:
taking the source domain dataset as an example, where x and y represent multidimensional feature vectors at different points in time after processing, x, y=p' j ,p′ k ,j,k∈{1,2,…,N s }. By sequentially calculating Pearson coefficients, we can obtain a correlation coefficient vector r= { R 11 ,r 12 ,…,r 1N I.e. HI curve. The HI curve will be used as a basis for the subsequent HS partitioning and as one of the features to be entered into the deep migration network.
2. Health Stage (HS) partitioning
Considering that the fault threshold indexes of various bearings are usually different, an adaptive HS dividing method needs to be established, and the invention provides a Time window weighted clustering method (Time-series Window Weighted Fuzzy C-Means Algorithm, TWW-FCM), and because the HI curve obtained before has local noise and fluctuation, the traditional FCM Algorithm usually carries out error classification on the judgment of noise points, so that the problem of discontinuous health stages such as transient abrupt change or false stage occurs. The current single point prediction is not reliable, and the stage is determined by other points in a certain range, so TWW-FCM suitable for stage division time sequence data is proposed, and the schematic diagram is shown in fig. 4. The specific flow is as follows:
firstly, the HI curve is required to be divided into windows, and the statistics value of the windows is used for replacing the current single-point value, so that the data can embody the state within a period of time, is smoother and more continuous, and reduces the mutation and oscillation of the data, thereby being more in line with the degradation rule of equipment and being beneficial to data analysis and modeling. The invention selects a Gaussian function as a window to process, which is defined as:
Wherein r is i For the i-th value on the HI curve, μ represents the mean of the samples, σ represents the standard deviation of the samples. The data is then windowed, for each data point r i Taking the value of the gaussian function at this point as a weight, weighted average is performed on all data points in the window, and the processed HI, i.e. the sequence G, is represented by formula (9):
G＝(g 1 ,g 2 ,…,g N ) (9)
after the HI curve is subjected to fuzzy smoothing treatment, the HI curve is clustered to obtain the health-stage index. First, determining the number N of clusters, and randomly initializing the membership value u of each data point to each cluster ij . Membership u for each data point i ij Representing data pointsi belongs to the degree of the j-th cluster, which is a value between 0 and 1, the fuzzy membership degree u ij The calculation of (2) is shown in formula (10):
wherein II is the characteristic sample x j To class center c i And the euclidean distance norm index between the two, and q represents a blurring factor. Thereafter, based on the fuzzy membership degree of each data point, the center point c= (C) of each cluster is calculated 1 ,c 2 ,…,c i ,…c N ) For the ith cluster, the coordinates c common to its center points i Calculation is performed by the formula (11):
TWW-FCM optimizes the membership value and the cluster center for each class through an iterative process, minimizing the objective function J until the stop condition is met. The objective function J and corresponding constraints are shown in equation (12):
0≤u ij ≤1,1≤i≤C,1≤j≤N (12)
Since each class center is initially initialized randomly at the beginning of the iterative process, iteration is stopped when a local minimum or saddle point of the objective function is obtained. The stop condition is defined as the change of the membership matrix between two successive iteration steps meeting a termination threshold epsilon. The convergence condition is expressed as follows:
||U k+1 -U k ||<ε (13)
finally, the health stage c to which each data point belongs is obtained through TWW-FCM algorithm processing i E (1, 2, …, N) and corresponding fuzzy membership u i ＝(u i1 ,u i2 ,…,u iN ) The health-stage labels and the fuzzy membership matrix are used as input data of the deep migration network.
3. The fuzzy subdomain feature extractor performs depth feature extraction and feature alignment:
and the fuzzy subdomain feature extractor combines the health stage category to which each data point belongs with the corresponding fuzzy membership degree to perform depth feature extraction and feature alignment on the acquired multidimensional time-frequency features. Unlike the general sub-domain adaptation using a determined domain label, the present invention introduces phase uncertainty into the sub-domain adaptation, and uses Fuzzy membership (Fuzzy membership) generated in the TWW-FCM algorithm to represent the probability that a sample belongs to each sub-domain, thereby achieving better sub-domain adaptation. Firstly, a deep neural network is used for acquiring the characteristic potential representation of the source domain and the target domain data, and then the characteristic representation of the source domain and the target domain is input into a Sub-structure fuzzy alignment module (Sub-structure Fuzzy alignment module, SFAM) to learn the domain invariant representation of the Sub-domain hierarchy by minimizing the domain difference between the Sub-structures. In the present invention, two multi-order sub-domain metric distances are included in the SFAM: fuzzy Local Maximum Mean Discrepancy (FLMMD) and Fuzzy Local CORAL (FLCORAL).
First, FLMMD, which is a non-parametric distance estimation for measuring the difference between two distributions, is widely used in field adaptation. The original MMD focuses on the distribution differences of the whole metrics, and ignores the category information of each sub-domain in one domain. The FLMMD provided by the invention simultaneously considers the probability that each sample belongs to all categories, realizes the alignment of fuzzy subdomains with finer granularity, and ensures that the source domain and the target domain data are closer in the mapping feature space.
Assuming source domainAnd the target Domain->Wherein x is S ,x T The source domain and the target domain are respectively samples, u S ,u T Corresponding fuzzy membership degree, x, of source domain and target domain samples respectively T Is a source domain label. D (D) S And D T Is divided into N health phases, i.e. sub-domain +.>And->Where N ε 1,2, …, N is the class label. FLMMD between the two distributions p and q is defined as the desire for Regenerated Kernel Hilbert Space (RKHS) distance between mean embedding within each subdomain. Thus, the square form of MMD is shown in formula (14):
wherein x is S And x T Respectively areAnd->Samples of (b), p (n) And q (n) Are respectively->And->Is a distribution of (a). Phi (·) represents some function mapping the input to RKHS with characteristic kernel k, k (x) S ,x T )＝<φ(x S ),φ(x T )>. Each hypothesis is givenThe samples are according to the weight w n Belonging to each category, its unbiased estimate is expressed as:
wherein,and->Respectively represent +.>And->Is a weight of (2).And->Are all equal to 1, and->Is a weighted sum over class n. For sample x i Weight +.>The calculation of (2) is as shown in formula (16):
wherein u is in Is the vector u i Membership of the nth element of (c). Because the stage fuzzy membership can well reflect one sampleProbability distributions belonging to different health phases, therefore, the present invention uses fuzzy membership +.>And->Calculate sample weight +.>And->
While MMD is one of the most common metrics in domain adaptation, MMD only computes the first moment and may not fully capture the complex differences between the two domains. The invention thus introduces in turn a second order statistic CORAL loss between the source and target features, in particular a CORAL loss that measures the difference between the two domains by calculating the distance between the covariance of the source and target samples. The invention provides that FLCORAL loss can make the model more comprehensively consider fuzzy differences among subdomains, thereby further improving the generalization performance of the model. The calculation formula of the CORAL loss is as follows:
wherein,representing the square matrix Frobenius norm. V (V) S And V T Is the covariance matrix of the source data and the target data with dimension d, which can be calculated by equation (18): />
Where 1 is a column vector with all elements equal to 1, n S And n T Representing the number of source data and target data, respectively. By giving weight to the sample data and introducing sub-domain information with finer granularity, the second-order statistical distance pull-up of the sub-domain level is realized. Its unbiased estimate is expressed as:
wherein,representing input data +.>And->Is a covariance matrix of (a). Consistent with FLMMD, source and target Domain sample weights +.>And->Can be determined by fuzzy membership u i And (3) calculating:
to align the hidden features of the source domain and the target domain, we need to activate the hidden state h T And h S . In one batch, giveN is fixed b From source domainsLabeled sample of (b) n From the target domain->Is a label-free sample of (1), where n b Representing the batch size. As shown in fig. 5, the feature extractor in the FSARN network will generate an active hidden stateAnd +.>Since we cannot calculate directly φ (·), the two distances FLMMD and FLCORAL are calculated by equation (21):
wherein k is mmd (. Cndot. Cndot.) and k coral (. Cndot.). Cndot.represents the kernel functions of FLMMD and FLCORAL, respectively. After obtaining an estimate of the weighted distribution difference between the source domain and the target domain, each corresponding sub-domain may be aligned by minimizing the difference between them.
4. The RUL regressor performs RUL prediction:
the RUL regressor is a prediction network with three fully connected layers, and the result obtained by receiving the characteristics obtained by the characteristic extractor to perform RUL prediction is expressed as:
y t ＝σ(w o g t +b o ) (22)
wherein the method comprises the steps ofRepresenting the output sequence of the fuzzy subdomain feature extractor at time t, y t Is the result of the RUL prediction at time t, σ (&) is the sigmoid function, & lt + & gt>Representing trainable parameters, b o Is a scalar. The regressor uses mean-square error (MSE) as a loss function of the regression:
wherein the method comprises the steps ofAnd->Sample->Target regression values and estimation values of (a).
In order to shorten the distance between the source domain and the target domain, the prediction result of the model needs to be optimized to solve the problem shown in the following formula (24):
wherein J (·, ·) represents the real value y of the target domain RUL i And the predicted value f (x i ) The loss function of the sum of the two values,for measuring all subdomain differences between the source domain and the target domain.
In summary, the loss of FSARN consists of two parts: fuzzy subdomain alignment lossRUL predictive lossObjective function->Expressed as:
where β is the weight coefficient of the subfield alignment loss and γ is the weight coefficient of the CORAL loss. Thus, the subdomain adaptation and regression fit are performed simultaneously, and the solution equation is as shown in (26):
In the proposed depth network, by minimizing the subfield alignment loss, the network can bring the source and target domain data closer by minimizing the differences between subfields; RUL prediction knowledge in the source domain is learned by minimizing RUL prediction loss. In the training process of the neural network model, a back propagation method is adopted for parameter updating, and a random gradient descent (SGD) algorithm is used as weight optimization to update network parameters. At each step of training, the model parameters are updated by equation (27):
where η represents the learning rate, which represents the learning steps taken by the SGD algorithm as training proceeds. θ E And theta R Parameters of the feature extractor and the RUL predictor, respectively. By minimizing the gradient using the loss function, the data distribution of the source and target domains is pulled as close as possible, and advanced features in the hidden layer can be automatically learned, thereby performing RUL prediction on unlabeled target domain data.
In the domain-adaptive-based residual life prediction method, the existing method focuses on global alignment of a source domain and a target domain, and local information contained in a health stage is ignored. By locally based alignment, finer granularity feature alignment can be achieved, enhancing the applicability and applicability of the model. But how to mine the common potential features of data in cases where the target domain data is incomplete and the health status varies greatly from one health stage to another is a difficult problem. According to the bearing residual life prediction method combining degradation phase division and subdomain self-adaption, disclosed by the invention, the fuzzy subdomain adaptive regression network (Fuzzy Subdomain Adaptation Regression Network, FSARN) is utilized to pull up each degradation phase by utilizing multi-order measurement based on the data label and membership, so that the fuzzy substructure alignment and cross-domain regression are completed, and the prediction accuracy and the generalization of a model are improved.
By adopting the technical scheme disclosed by the invention, the following beneficial effects are obtained:
the invention provides a bearing residual life prediction method and a system combining degradation phase division and sub-domain self-adaption, which can fully and skillfully utilize the phase difference existing in the service life period corresponding to different sub-domains and disperse the overall difference of global alignment to the phase difference of local alignment. In the health stage division stage, a time sequence fuzzy clustering algorithm suitable for time sequence window data is provided, and unified and standardized health index construction and stage division processes are realized. In the migration stage, the degradation stages are pulled up by using a multi-order metric based on the data labels and membership degrees, so that the fuzzy substructure alignment and cross-domain regression are completed. The method can realize the characteristic alignment of finer granularity, construct a model with better generalization, and simultaneously realize more accurate residual life prediction.
The foregoing is merely a preferred embodiment of the present invention and it should be noted that modifications and adaptations to those skilled in the art may be made without departing from the principles of the present invention, which is also intended to be covered by the present invention.
Claims (10)
1. A bearing residual life prediction method combining degradation phase division and sub-domain self-adaption is characterized in that: comprises the following steps of the method,
s1, constructing a health index curve:
acquiring an original data set, extracting time domain and frequency domain characteristics of source domain data and target domain data in the original data set, and performing secondary index optimization and correlation analysis on the obtained multidimensional characteristics to acquire a health index curve;
s2, health stage division:
processing the health index curve by adopting a time sequence window weighted clustering algorithm, obtaining the health stage labels and the fuzzy membership degree of each source domain data and each target domain data, and dividing the health stage labels and the fuzzy membership degree into a training set and a testing set according to a proportion;
s3, constructing and training a residual life prediction model:
the residual life prediction model comprises a fuzzy subdomain feature extractor and a RUL regressive; inputting the training set into a residual life prediction model to train the residual life prediction model; in the training process, based on the alignment loss of a sub-structure fuzzy alignment module in a fuzzy subdomain feature extractor and the regression loss of an RUL regressive, obtaining a model total loss, optimizing parameters of the fuzzy subdomain feature extractor and the RUL regressive by minimizing the model total loss, and obtaining and storing a trained residual life prediction model;
S4, residual life prediction:
and inputting the test set into a trained residual life prediction model to perform residual life prediction, and obtaining a residual life prediction result.
2. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: the secondary index optimization is specifically that the inter-characteristic difference alignment operation is carried out on the multidimensional characteristics to obtain relative indexes, and then the intra-characteristic compression smoothing operation is carried out on the relative indexes to obtain denoising indexes.
3. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: the correlation analysis is specifically to use the pearson correlation coefficient to measure the correlation degree of the data at each moment and the data at the initial moment on the index after the secondary optimization, and construct the health index.
4. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: the source domain data are electromechanical equipment monitoring data acquired when electromechanical equipment with a bearing runs to failure; the target domain data is monitoring data of electromechanical equipment from the operation of the electromechanical equipment without the bearing to the fault;
The source domain data are supervised data with the residual life of the bearing as a label; the target domain data is unsupervised data without a bearing remaining life tag.
5. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: step S2 specifically includes the following,
s21, dividing the health index curve into windows, and replacing the current single-point numerical value with the statistical value of the windows to realize fuzzy smoothing of the health index curve;
s22, clustering the health index curves subjected to fuzzy smoothing, and obtaining health stage labels and fuzzy membership degrees of the health stage labels of the source domain data and the target domain data through iterative optimization of fuzzy membership degrees of the data points and clustering centers of the categories;
s23, dividing the source domain data, the target domain data, the health-stage labels to which the source domain data and the target domain data belong and the fuzzy membership degree into a training set and a testing set in proportion.
6. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: step S3 specifically includes the following,
s31, inputting a training set into a fuzzy subdomain feature extractor, and respectively acquiring feature potential representations of source domain data and target domain data in the training set by using a deep neural network of the fuzzy subdomain feature extractor to acquire high-dimensional feature matrixes corresponding to the source domain data and the target domain data;
S32, inputting high-dimensional feature matrixes corresponding to the source domain data and the target domain data into a substructure fuzzy alignment module of a fuzzy subdomain feature extractor, and acquiring alignment loss by calculating FLMMD between the feature matrixes of the source domain data and the target domain data and FLCORAL between the feature matrixes of the source domain data and the target domain data and the feature matrix time of the source domain data and the target domain data;
s33, inputting the time sequence feature matrix of the aligned source domain data and the target domain data in the training set, which are output by the fuzzy subdomain feature extractor, into an RUL regressive device, outputting the residual life predicted value of the source domain data and the target domain data in the training set, and calculating the mean square error of the residual life predicted value and the true value of the source domain as regression loss;
s34, calculating the total loss of the model according to the alignment loss and the regression loss;
and S35, minimizing the total loss of the model, and feeding back and adjusting network parameters of the fuzzy subdomain feature extractor and the RUL regressive to realize network training of the fuzzy subdomain feature extractor and the RUL regressive until the training is completed, so as to obtain a trained residual life prediction model.
7. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 6, wherein: step S32 specifically includes the following,
S321, alignment of the first portion: fuzzy local maximum mean value difference based on the maximum mean value difference, and meanwhile, the probability that each sample belongs to all categories is considered, so that the alignment of fuzzy subdomains with finer granularity is realized; calculating FLMMD between the source domain data and the target domain data feature matrix as a first partial loss;
s322, aligning the second part: fine granularity alignment is performed on the second order statistics based on Fuzzy Local CORAL of the second order statistics Correlation Alignment while considering the probability that each sample belongs to all categories; calculating FLCORAL between the source domain data and the target domain data feature matrix as a second partial loss;
s323, integrating the first partial loss and the second partial loss to obtain the alignment loss.
8. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: the fuzzy subdomain feature extractor is a ResNet50 feature extractor.
9. The method for predicting bearing residual life by combining degradation phase division and sub-domain adaptation according to claim 1, wherein: the RUL regressor is a regression predictor based on a fully connected network.
10. A bearing residual life prediction system combining degradation phase division and sub-domain self-adaption is characterized in that: a system for implementing the method of any one of the preceding claims 1 to 9, said system comprising,
and a curve construction module: the method is used for constructing a health index curve;
acquiring an original data set, extracting time domain and frequency domain characteristics of source domain data and target domain data in the original data set, and performing secondary index optimization and correlation analysis on the obtained multidimensional characteristics to acquire a health index curve;
the stage division module: for dividing the health phase;
processing the health index curve by adopting a time sequence window weighted clustering algorithm, obtaining the health stage labels and the fuzzy membership degree of each source domain data and each target domain data, and dividing the health stage labels and the fuzzy membership degree into a training set and a testing set according to a proportion;
the prediction model building module: the method is used for constructing and training a residual life prediction model;
the residual life prediction model comprises a fuzzy subdomain feature extractor and a RUL regressive; inputting the training set into a residual life prediction model to train the residual life prediction model; in the training process, based on the alignment loss of the sub-structure fuzzy alignment module in the fuzzy subdomain feature extractor and the regression loss of the RUL regressive, acquiring total loss, optimizing parameters of the fuzzy subdomain feature extractor and the RUL regressive by minimizing the total loss, and acquiring and storing a trained residual life prediction model;
Life prediction module: for predicting remaining life;
and inputting the test set into a trained residual life prediction model to perform residual life prediction, and obtaining a residual life prediction result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310849683.7A CN117252083A (en) | 2023-07-12 | 2023-07-12 | Bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310849683.7A CN117252083A (en) | 2023-07-12 | 2023-07-12 | Bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117252083A true CN117252083A (en) | 2023-12-19 |
Family
ID=89133924
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310849683.7A Pending CN117252083A (en) | 2023-07-12 | 2023-07-12 | Bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117252083A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118364296A (en) * | 2024-06-19 | 2024-07-19 | 中国特种设备检测研究院 | Rolling bearing residual life prediction method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113609770A (en) * | 2021-08-06 | 2021-11-05 | 电子科技大学 | Rolling bearing RUL prediction method based on piecewise linear fitting HI and LSTM |
| CN114548152A (en) * | 2022-01-17 | 2022-05-27 | 上海交通大学 | Method for predicting residual life of marine sliding bearing based on transfer learning |
| CN116147917A (en) * | 2023-01-17 | 2023-05-23 | 西安交通大学 | Construction method, device and equipment for predicting residual service life label of bearing |
| CN116306289A (en) * | 2023-03-21 | 2023-06-23 | 中国科学院空间应用工程与技术中心 | Multi-source domain self-adaption-based electromechanical device cross-domain residual life prediction method |
| CN116415485A (en) * | 2022-12-29 | 2023-07-11 | 电子科技大学中山学院 | Multi-source domain migration learning residual service life prediction method based on dynamic distribution self-adaption |
-
2023
- 2023-07-12 CN CN202310849683.7A patent/CN117252083A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113609770A (en) * | 2021-08-06 | 2021-11-05 | 电子科技大学 | Rolling bearing RUL prediction method based on piecewise linear fitting HI and LSTM |
| CN114548152A (en) * | 2022-01-17 | 2022-05-27 | 上海交通大学 | Method for predicting residual life of marine sliding bearing based on transfer learning |
| CN116415485A (en) * | 2022-12-29 | 2023-07-11 | 电子科技大学中山学院 | Multi-source domain migration learning residual service life prediction method based on dynamic distribution self-adaption |
| CN116147917A (en) * | 2023-01-17 | 2023-05-23 | 西安交通大学 | Construction method, device and equipment for predicting residual service life label of bearing |
| CN116306289A (en) * | 2023-03-21 | 2023-06-23 | 中国科学院空间应用工程与技术中心 | Multi-source domain self-adaption-based electromechanical device cross-domain residual life prediction method |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN118364296A (en) * | 2024-06-19 | 2024-07-19 | 中国特种设备检测研究院 | Rolling bearing residual life prediction method and system |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108900346B (en) | Wireless network flow prediction method based on LSTM network | |
| CN106951695B (en) | Method and system for calculating residual service life of mechanical equipment under multiple working conditions | |
| CN112328588B (en) | Industrial fault diagnosis unbalanced time sequence data expansion method | |
| CN114218872B (en) | DBN-LSTM semi-supervised joint model-based residual service life prediction method | |
| CN111768000A (en) | Industrial process data modeling method for online adaptive fine-tuning deep learning | |
| CN113468720B (en) | Service life prediction method for digital-analog linked random degradation equipment | |
| CN114500325B (en) | SDN controller fault self-adaptive intelligent detection method based on unsupervised transfer learning | |
| Mo et al. | Few-shot RUL estimation based on model-agnostic meta-learning | |
| CN112784920A (en) | Cloud-side-end-coordinated dual-anti-domain self-adaptive fault diagnosis method for rotating part | |
| CN117252083A (en) | Bearing residual life prediction method and system combining degradation phase division and sub-domain self-adaption | |
| CN111638707A (en) | Intermittent process fault monitoring method based on SOM clustering and MPCA | |
| CN114580545A (en) | Wind turbine generator gearbox fault early warning method based on fusion model | |
| CN116306289B (en) | Multi-source domain self-adaption-based electromechanical device cross-domain residual life prediction method | |
| CN112507479A (en) | Oil drilling machine health state assessment method based on manifold learning and softmax | |
| CN113469013B (en) | Motor fault prediction method and system based on transfer learning and time sequence | |
| CN117909881A (en) | Fault diagnosis method and device for multi-source data fusion pumping unit | |
| CN117972585A (en) | Fault enhancement diagnosis method based on PCA-DDPM and CNN under small sample condition | |
| CN113420815A (en) | Semi-supervised RSDAE nonlinear PLS intermittent process monitoring method | |
| CN117520809A (en) | Transformer fault diagnosis method based on EEMD-KPCA-CNN-BiLSTM | |
| CN117973511A (en) | Elevator fault diagnosis method integrating knowledge graph and neural network | |
| CN114036947B (en) | Small sample text classification method and system for semi-supervised learning | |
| Luo et al. | A novel method for remaining useful life prediction of roller bearings involving the discrepancy and similarity of degradation trajectories | |
| CN115630582A (en) | Multi-sliding-window model fused soft rock tunnel surrounding rock deformation prediction method and equipment | |
| CN115794805A (en) | Medium-low voltage distribution network measurement data supplementing method | |
| CN115129029A (en) | Industrial system fault diagnosis method and system based on sub-field adaptive dictionary learning |
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 |