CN115575127A - Bearing residual life prediction method based on whale optimization LSSVM - Google Patents
Bearing residual life prediction method based on whale optimization LSSVM Download PDFInfo
- Publication number
- CN115575127A CN115575127A CN202211228857.XA CN202211228857A CN115575127A CN 115575127 A CN115575127 A CN 115575127A CN 202211228857 A CN202211228857 A CN 202211228857A CN 115575127 A CN115575127 A CN 115575127A
- Authority
- CN
- China
- Prior art keywords
- data
- bearing
- residual life
- prediction
- model
- 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
Images
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
- G01M13/045—Acoustic or vibration analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
- G06N20/10—Machine learning using kernel methods, e.g. support vector machines [SVM]
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Acoustics & Sound (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a whale optimized LSSVM-based bearing residual life prediction method, which comprises the steps of data feature extraction, feature screening, feature fusion, degradation index construction and the like. The experimental prediction data of the method is closer to the real residual life time, and the method shows good prediction effect, the mean error of the model prediction result constructed by the method is respectively improved by 45.13% and 38.95% compared with the CNN-LSTM model and CNN-GRU model method, the scores are respectively improved by 0.16 and 0.12, the method provides a new thought for the prediction of the residual service life of the bearing, and the method has important guiding significance.
Description
Technical Field
The invention belongs to the technical field of bearing residual life prediction, relates to a machine learning prediction analysis method, and particularly relates to a bearing residual life prediction method based on whale optimization LSSVM.
Background
With the rapid progress of the manufacturing technology, the mechanical equipment is continuously developed in the direction of large-scale, complex, precise, intelligent and integrated, and the reliability problem becomes increasingly prominent. The bearing is used as one of the most widely applied and important parts in mechanical equipment, and plays a vital role in the health state of the operation of the whole mechanical equipment. At present, the service time of equipment in some domestic mechanical equipment plants is long, some bearing parts are close to the service life of the bearing parts and even work beyond the service life, once the bearing exceeds the service life of the bearing, the running precision of the bearing is rapidly reduced, which is undoubtedly a great potential safety hazard for mechanical equipment, once a fault occurs, economic loss is inevitably caused, and serious casualties are possibly caused. Therefore, it is very important to predict the residual life of the mechanical equipment accurately and then perform predictive maintenance.
Since the 21 st century, significant products and technologies closely related to fault diagnosis and health guarantee technologies have been set as key research directions in development and planning of China. In recent years, prediction of the residual life of a bearing has been highly valued by a large number of researchers. A plurality of representative methods for predicting the residual life of the bearing are sequentially proposed, such as a model method and a data driving method. The model method describes the degradation process of the bearing by establishing a physical model based on a failure mechanism, but the model method is difficult to actually establish an accurate degradation model of the bearing. The data driving method learns the degradation process of the bearing through a statistical model or an artificial intelligence technology to realize the residual life prediction of the bearing, however, the data driving method depends on historical data, and the accuracy of prediction is not ideal actually.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides a bearing residual life prediction method based on whale optimization LSSVM, and a set of feature sets with high signal-to-noise ratio is obtained by a method of decomposing original data by EMD and then extracting features at the front end; the middle end screens the feature set by using a method of combining original forest and monotonicity to obtain an optimal feature subset, and the feature screened by the method has obvious degradation trend and contains main information of original data degradation; the LSSVM is optimized and constructed by the whale optimizing algorithm at the rear end, performance degradation indexes are established, optimizing time is greatly shortened, and convergence rate of the algorithm is improved.
The invention is realized by the following technical scheme:
a bearing residual life prediction method based on whale optimized LSSVM comprises data feature extraction, feature screening, feature fusion (dimensionality reduction fusion) and degradation index construction.
Specifically, the data feature extraction step includes:
(1) Data acquisition: and adopting a PHM2012 challenge open competition data set, and carrying out an accelerated degradation experiment on the bearing by the bearing data under different operating conditions, thereby obtaining measured data in the whole life cycle of the bearing to carry out fault detection, fault diagnosis and residual life prediction related algorithm verification. The measured data comprises three working conditions, namely load 4000N and rotating speed 1800r/min (working condition 1); load is 4200N, and rotating speed is 1650r/min (working condition 2); load 5000N, and rotating speed 1500r/min (working condition 3). The measured data under each working condition comprises a vibration signal and a temperature signal, and partial data of the temperature signal is missing. The vibration signal includes vibration information in horizontal and vertical directions, the sampling frequency is 25.6kHz, the recording is carried out once every 10s, the recording time is 0.1s, and 2560 points are included. The invention uses the bearing numbers of 1_1, 1_2, 2_1, 2_2, 3_1 and 3_2 as training sets, the bearing numbers of 1_3, 1_4, 1_5, 1_6, 1_7, 2_3, 2_4, 2_5, 2_6, 2_7 and 3_3, and 11 groups of bearing non-full-life data are used as experimental data to carry out the residual life prediction experiment of the bearing.
(2) EMD decomposition: the noise in the original data is too high, the signal-to-noise ratio can be improved after EMD decomposition, and two component IMF1 and IMF2 are obtained through EMD decomposition.
(3) Feature extraction: in the invention, 50 features are extracted. Extracting 16 different time domain characteristics including a mean value, a root mean square amplitude, an absolute mean value, a skewness, a kurtosis, a variance, a maximum value, a minimum value, a peak-to-peak value, a waveform index, a peak index, a margin index, a pulse index, a skewness index and a kurtosis index; dimensionless features mainly include skewness index, kurtosis index, peak index, pulse index, margin index and waveform index; the frequency domain features are typically used to find periodic information in the signal, extracting 9 different frequency domain features, 4 varying frequency domain features reflecting the location of the dominant frequency band and 5 frequency domain features representing the degree of dispersion or concentration of the spectrum. The characteristics extracted in the experiment are too many, the variation trend is too many, the performance degradation index cannot be directly constructed, the characteristics need to be further screened, meanwhile, the bearing degradation information cannot be effectively extracted due to too large characteristic noise directly extracted from an original signal, and therefore the original vibration information needs to be processed firstly.
Specifically, the feature screening and feature fusing step comprises:
(1) Random forest screening (random forest sorting): the random forest can not only process the classification and regression problems, but also be used for reducing the characteristic dimensionality and has good applicability. And calculating the importance of all extracted features by using an importance analysis algorithm in the random forest, wherein the value range of the normalized indexes of the random forest is [0,1], the more 1 index of the indexes represents that the importance of the features is higher, and the random forest ranks the importance of all the extracted features by using the importance algorithm.
The importance evaluation index of the random forest is measured by a Gini Index (GI), the importance score of the random forest is represented by IM, the number of features is represented by m, and the GI score of each feature is calculated:
in formula (1), K represents the group of data-a common K class, P mk Representing the proportion of k in the node. The GI variation formula before and after the node m branches is:
in formula (2), two new nodes of GI use GI l And GIr. The formula of the importance index of a single feature is:
(2) Monotonicity screening (monotonicity ranking): the characteristics with consistent degradation trend can be selected by carrying out monotonicity screening on the characteristics, the value range of the monotonicity index is [0,1], the characteristics with the index close to 1 represent that the monotonicity trend is better, and the training error of the training model can be reduced.
In the formula (4), n is the number of measuring points, m is the number of target products, and m =1 in the example; x is the number of j Represents the jth eigenvalue, pdiff represents the positive difference, and Ndiff represents the negative difference.
(3) Composite screening (mixed index screening): on the basis of the importance of the random forest index and the importance of the monotonicity index, the two importance are mixed and superposed to obtain a new importance index.
(4) Dimension reduction and feature fusion: principal Component Analysis (PCA) is the most commonly used linear dimension reduction method, and its objective is to map high-dimensional data into low-dimensional space by some kind of linear projection, and expect the maximum information amount (maximum variance) of the data in the projected dimension, so as to use less data dimension, and retain the characteristics of more raw data points. And performing dimensionality reduction on the filtered data after the previous filtering operation, obtaining several groups of features with strong monotonicity through the extracted features by dimensionality reduction, and then fusing the preferred features to obtain principal component data, which is called as a health index and used for representing the degradation features of the object.
Specifically, the step of constructing the degradation indicator includes:
(1) Model optimization (calculating optimal parameters of training models) of the same group of data sets under the same condition: and (3) searching an optimal solution for parameters of a Least Square Support Vector Machine (LSSVM) by utilizing a whale algorithm (WOA). The WOA algorithm achieves the purpose of optimization by simulating the predation behavior of the wolf colony and based on a wolf colony cooperation mechanism, has the characteristics of less parameters, easiness in realization, strong convergence performance and the like, and the training of the WOA-LSSVM model can be divided into 4 steps:
step1, carrying out normalization operation on the training set data, wherein the normalization formula is as follows:
step2, transposing the normalized data to adapt to the model;
step3, initializing model parameters of a Least Square Support Vector Machine (LSSVM);
and Step4, carrying out model training and searching for an optimal solution.
In Step1, X min Indicating that the minimum number, X, in the normalized data is required max Representing the maximum number in the data normalized data.
(2) Data filtering (moving average filtering): the degree of the fitting of the later-period data is influenced by the overlarge up-and-down fluctuation of the data, so the filtering processing is carried out by adopting the pre-estimated data of the moving average filter, and the smoothing operation is carried out on the data by using a local weighted regression (lowess) method.
(3) And (3) fitting data: because the test data is not the complete service life data of the bearing, the final result of the residual service life prediction of the bearing cannot be obtained, a data fitting method is introduced to fit the incomplete service life prediction data, a complete degradation curve can be obtained through fitting, and the final predicted residual service life of the bearing can be obtained according to the intersection point of the failure threshold and the fitted curve. Setting the full life cycle distribution of the bearing as a [0,1] interval, wherein "0" represents no damage state of the bearing, and "1" represents that the bearing is worn to reach a failure threshold value and can not continue to work, and the adopted polynomial expression is as follows:
y=a 1 x k +a 2 x k-1 +…+a k x 1 +c, (6)
in the formula (6), a 1 、a 2 、…、a k Representing the polynomial coefficients, c representing a constant term, these parameters being calculated from the fitted curve.
(4) Comparing the life prediction results (establishing prediction results by a prediction model): and importing the test set data into the obtained optimal parameter Least Square Support Vector Machine (LSSVM) model to obtain a final prediction result. And performing reverse normalization operation on the obtained result to obtain a curve close to the actual life curve. And calculating an error value between the predicted data and the real data of the test set through a root mean square error, wherein the root mean square error formula is as follows:
in the formula (7), X obs,i Denotes the i-th set of test values, X model,i Representing the ith set of true values, X obs,i -X model,i Representing the deviation of a set of test values from the true values.
The real residual service life time of all test sets participating in the experiment and the predicted residual service life time are respectively recorded, and the condition of a prediction result is verified through an experimental error formula which is as follows:
ActRIL in formula (8) i Representing the actual residual life value, RUL, of the bearing i And (4) representing the predicted residual life value of the bearing obtained by the test.
The values of the predicted results are compared with the true values, underestimation and overestimation cannot be considered in the same way, and the scores of each set of experiments are calculated by different formulas as follows:
in the formula (9), er i The experimental error for each set of experiments is shown.
The final score for the predicted results for all experiments is formulated as:
in the formula (10), A i Score values for each set of experiments are shown.
The invention has the beneficial effects that: the experimental prediction data of the method is closer to the real residual life time, and the method shows good prediction effect, the mean error of the model prediction result constructed by the method is respectively improved by 45.13% and 38.95% compared with the CNN-LSTM model and CNN-GRU model method, the scores are respectively improved by 0.16 and 0.12, the method provides a new thought for the prediction of the residual service life of the bearing, and the method has important guiding significance.
Drawings
FIG. 1 is a general flow-sheet framework diagram of the whale optimized LSSVM based bearing remaining life prediction method of the present invention;
FIG. 2 is a graph of 5 sets of test set horizontal direction onset vibration data under 2012PHM data set conditions;
FIG. 3 is a diagram of a new feature subset from a hybrid screen;
FIG. 4 is a graph of the model partial prediction results;
FIG. 5 is a graph of data fit;
FIG. 6 is a graph comparing each experimental group with CNN-LSTM and CNN-GRU.
Detailed Description
The following detailed description of the preferred embodiments of the present invention, taken in conjunction with the accompanying drawings, will make the advantages and features of the invention more readily understood by those skilled in the art, and thus will more clearly and distinctly define the scope of the invention.
As shown in FIG. 1, the method for predicting the residual life of the bearing based on whale optimization LSSVM of the invention comprises the following steps: data feature extraction, feature screening and feature fusion (dimension reduction fusion), and construction of a degradation index.
Step one, extracting vibration data characteristics.
And respectively extracting features from the training set and the test set in the 2012PHM public challenge data set to obtain two feature sets.
And step two, feature screening and dimensionality reduction fusion.
And respectively carrying out importance ordering on random forests and monotonicity on the feature sets, then superposing the importance of the random forests and the monotonicity to obtain a new importance ordering, and dividing the top 10% of features to form a new feature set. And finally, carrying out PCA dimension reduction fusion on the new feature set to obtain a first principal component as model input data.
And step three, constructing a degradation index.
Initializing parameters of a wolf algorithm, firstly carrying out normalization operation on the fused and screened main component data, updating the current position of the wolf, transposing the normalized data to adapt to the model, then carrying out optimization training on a least square support vector machine model through a whale optimization algorithm to obtain optimal parameters c and g, training a support vector regression model, and importing the fused data of the test set into the optimal parameter model to obtain a degradation index. And performing data filtering operation on the degradation indexes, and performing data fitting on the filtered data by using a fitting function to obtain a final prediction result.
The selected comparison methods are life prediction combining the CNN and LSTM, CNN and GRU methods respectively. As shown in FIG. 6, the bearing residual life prediction method based on whale optimization LSSVM of the invention is compared with other two methods, and the results show that the mean error is respectively reduced by 45.13% and 38.95%, and the final scores are respectively improved by 0.16 and 0.12. 2-6 are absolute error direct views of predicted values and true values of 11 test experimental groups based on the LSSVM model, and are compared with absolute errors of predicted results of two selected documents. From the error result, the prediction errors of the test groups 1-4, 1-5, 2-3, 2-4, 2-5 and 2-7 are obviously better than those of other methods, the absolute error of a single test group is larger than that of a reference, but the average error and the final score are higher than those of the reference in the whole, and the method has better prediction results. According to the method, the two methods for preprocessing data, which are provided on the basis of the WOA-LSSVM model, are tested and verified in the PHM2012 data set, the WOA optimization algorithm can reduce the model training time, the experimental result shows that the accuracy of prediction can be effectively improved by processing the data by adopting the method, and the method has great advantages in bearing life prediction.
The above-mentioned embodiments only express several embodiments of the present invention, and the description thereof is more specific and detailed, but not construed as limiting the scope of the invention. It should be noted that, for a person skilled in the art, several variations and modifications can be made without departing from the inventive concept, which falls within the scope of the present invention.
Claims (5)
1. A bearing residual life prediction method based on whale optimization LSSVM is characterized by comprising the following steps:
s1, extracting data characteristics;
s2, screening and fusing the features extracted in the step S1;
and S3, establishing a prediction model for the data fused in the S2, establishing a degradation index, and predicting the residual life time of the bearing.
2. The method for predicting the residual life of a bearing based on whale optimization LSSVM of claim 1, wherein the data feature extraction in the step S1 comprises the following specific steps:
(1) EMD decomposition: performing EMD on the divided training set and test set respectively to obtain IMF1 and IMF2 component vectors;
(2) Feature extraction: and respectively extracting time domain and frequency domain characteristics and dimensionless characteristics from the component vectors IMF1 and IMF2 obtained by decomposition to form a characteristic set.
3. The method for predicting the residual life of the bearing based on the whale optimized LSSVM as claimed in claim 2, wherein the specific steps of feature screening and feature fusion in the step S2 are as follows:
(1) And (3) feature screening: carrying out feature selection by a method of random forest and monotonicity composite screening on features respectively extracted from the divided training set and test set, wherein the screened features have obvious degradation trend and contain main degradation information of original data;
(2) Feature fusion: and fusing the characteristic values subjected to the dimensionality reduction into principal components by using a principal component analysis algorithm.
4. The whale optimized LSSVM based bearing residual life prediction method as claimed in claim 3, wherein the concrete steps of data set life prediction under different conditions in the S3 step are as follows:
(1) Model optimization of the same group of datasets under the same condition: putting the principal component data extracted from the early training set after feature fusion into a trainer, optimizing two parameters of the trainer by using a whale algorithm, evaluating the optimal trainer by all results of the two parameters obtained by optimizing the algorithm result, and obtaining a performance degradation index by using an optimal model;
(2) Data filtering: carrying out filtering operation on the degradation index;
(3) And (3) fitting data: performing data fitting on the filtered data through a polynomial function to obtain the final residual life of the bearing;
(4) Comparing the life prediction results: and evaluating the final residual life of the bearing by adopting an evaluation index to evaluate the prediction error of each group of experiments.
5. The method for predicting the residual life of a bearing based on whale optimization LSSVM of claim 4, wherein the method for optimizing two parameters of a trainer by using a whale algorithm comprises the following specific steps:
step1, carrying out normalization operation on the training set data, wherein the normalization formula is as follows:
step2, transposing the normalized data to adapt to the model;
step3, initializing the parameters of the least square support vector machine model;
and Step4, carrying out model training and searching for an optimal solution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211228857.XA CN115575127A (en) | 2022-10-08 | 2022-10-08 | Bearing residual life prediction method based on whale optimization LSSVM |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211228857.XA CN115575127A (en) | 2022-10-08 | 2022-10-08 | Bearing residual life prediction method based on whale optimization LSSVM |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115575127A true CN115575127A (en) | 2023-01-06 |
Family
ID=84584167
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211228857.XA Pending CN115575127A (en) | 2022-10-08 | 2022-10-08 | Bearing residual life prediction method based on whale optimization LSSVM |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115575127A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115879248A (en) * | 2023-03-03 | 2023-03-31 | 山东亿宁环保科技有限公司 | Full life cycle management method and system suitable for vacuum pump |
-
2022
- 2022-10-08 CN CN202211228857.XA patent/CN115575127A/en active Pending
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN115879248A (en) * | 2023-03-03 | 2023-03-31 | 山东亿宁环保科技有限公司 | Full life cycle management method and system suitable for vacuum pump |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110929918A (en) | 10kV feeder line fault prediction method based on CNN and LightGBM | |
CN116226646B (en) | Method, system, equipment and medium for predicting health state and residual life of bearing | |
CN112434390B (en) | PCA-LSTM bearing residual life prediction method based on multi-layer grid search | |
CN109902632A (en) | A kind of video analysis device and video analysis method towards old man's exception | |
CN115575127A (en) | Bearing residual life prediction method based on whale optimization LSSVM | |
CN112906764A (en) | Communication safety equipment intelligent diagnosis method and system based on improved BP neural network | |
CN114266297A (en) | Semantic knowledge base of thermal power equipment, construction method and zero sample fault diagnosis method | |
CN114004059B (en) | Health portrait method for hydroelectric generating set | |
CN114548494B (en) | Visual cost data prediction intelligent analysis system | |
CN115496108A (en) | Fault monitoring method and system based on manifold learning and big data analysis | |
US20240184678A1 (en) | Deep Learning Method Integrating Prior Knowledge for Fault Diagnosis | |
CN114091334A (en) | Partial discharge fault diagnosis method based on improved bat algorithm and support vector machine | |
CN116502155A (en) | Safety supervision system for numerical control electric screw press | |
CN116842433A (en) | Power distribution network fault classification method based on ensemble learning and longhorn beetle whisker algorithm | |
CN114547796B (en) | Ball mill feature fusion fault diagnosis method based on optimized BN network | |
CN115407753A (en) | Industrial fault diagnosis method for multivariate weighted ensemble learning | |
CN115153549A (en) | BP neural network-based man-machine interaction interface cognitive load prediction method | |
CN115359857A (en) | Method for predicting PM2.5 concentration based on CART model and optimized RVFLNN model | |
CN116224950A (en) | Intelligent fault diagnosis method and system for self-organizing reconstruction of unmanned production line | |
CN115293189A (en) | Rotating machinery state monitoring method based on stack self-coding dimension reduction | |
Zheng et al. | Research on predicting remaining useful life of equipment based on health index | |
CN112001530A (en) | Predictive maintenance method and system for transformer oil chromatography online monitoring device | |
CN111062430A (en) | Pedestrian re-identification evaluation method based on probability density function | |
CN114024820B (en) | Equipment health portrait method | |
Zou et al. | Maintenance Evaluation of Power Grid Fault Equipment Based on Random Forest |
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 |