Font Size: a A A

Application Research For Fault Diagnosis Based On Statistical Clustering And Particle Filter

Posted on:2015-10-07Degree:DoctorType:Dissertation
Country:ChinaCandidate:L M LiFull Text:PDF
GTID:1222330452465474Subject:Aeronautical and Astronautical Science and Technology
Abstract/Summary:PDF Full Text Request
Several problems existed in fault features extraction, multiple faults diagnosis, lifeprognosis, abrupt failure diagnosis and prognosis in fault diagnosis of rotating machineryhave been solved by statistical clustering and particle filter, this problems include:1) how toeffectively fuse multiple features according to their respective contributions for faultdiagnosis is a problem needed to be solved;2) In the process of multiple faults diagnosis, thesame kind of fault often corresponds to various kinds of data, including data under differentenvironment and different operating conditions, which increase the time complexity intraining classifier, how to minimize the time complexity of the trained classifier is a newproblem;3) In life prognosis of rotating machinery, particle filter has been extensively used,and unscented particle filter is the improvement of particle filter technology, which can makefull use of prior information and observations to correct prognosis model, but the technologyis still not strong robustness, therefore it is essential to improve it in this aspect;4) Featuresof abrupt failure are hard to captured due to uncertain happened time and short duration,which increased the difficulty of diagnosis and prognosis for abrupt failure, therefore, it isurgent to explore the effective and rapid fault diagnosis and prognosis algorithm. Accordingto those problems, the main research contents and contributions are as follows:Aiming at the first problem, two methods based on statistical clustering analysis wereproposed: a feature selection method base on affinity propagation (AP) clustering methodwas proposed in this paper, which using the advantage of it that don’t need to initialize theclustering center, by automatic clustering to find all centers of the test data, and these centerscan represent all the features, experiment verify the method we proposed is more accuratethan other feature selection methods, and at the same time it took less time; In the foundationof this theory, another method named linear feature fusion method based on AP algorithmand affinity aggregation for spectrum clustering (AASC) is proposed, after obtained weightsof all features, we combined it linearly into one feature. Which makes it easier to diagnoseand prognose mechanical fault using machine learning and data-driven methods, at the sametime the diagnosis accuracy is improved, experiment verify the effective of this method.Aiming at the second problem, we proposed using hierarchical clustering to manage thehistorical data according to speed and the degree of damage, after hierarchical clustering, weonly need to train the small data set, when there are new data need to judge which kind of fault it belongs, we can firstly judge it which layer it belongs, and then using thecorresponding classifier to diagnose it, which reduces the training time when put all historicaldata together. We use rolling element bearing multiple faults data from Case Western ReserveUniversity as test data, and the experimental result shows that this method can diagnosemultiple fault effectively.Aiming at the third problem, we proposed an improved unscented particle filter (IUPF)method for improving the robustness of unscented particle filter (UPF) algorithm while it wasused for prognosis, in which the sigma samples of unscented transformation (UT) intraditional UPF are generated by singular value decomposition (SVD), and then those sigmapoints are propagated by the standard unscented Kalman filter (UKF) to generate asophisticated proposal distribution. In order to verify this improved method, two data setwere used, they respectively are Lithium-ion battery capacity data and rolling elementbearing fault data from NASA, and after experiment, it shows that our method can improvethe robustness of prognosis while using UPF on remaining useful life prognosis.Aiming at abrupt failure diagnosis and prognosis in the fourth problem, we adopt themethod of cluster analysis of the data before the abrupt failure, on behalf of its trend ofclustering center, by real-time monitoring the online data, when there is data matched withthose centers, we can immediately take emergency measures to prevent catastrophicaccidents. Aiming at the problem of abrupt failure prognosis, we improved the mixture ofGaussian Hidden Markov Model (MOG-HMM) method of multivariate statistics, multiplestates MOG-HMM was adopted to adapt to the abrupt failure characteristics of short durationof each state. At last, by collecting data from abrupt failure simulation equipment and using iton testing the diagnosis and prognosis, it shows the feasibility and effectiveness.
Keywords/Search Tags:Fault diagnosis and prognosis, Hybrid clustering, Hierarchical clustering, Improved unscented particle filter, Multiple states mixture of Gaussian HiddenMarkov Model
PDF Full Text Request
Related items