《FAULT DIAGNOSIS APPROACH BASED ON HIDDEN MARKOV MODEL AND SUPPORT VECTOR MACHINE .doc》由会员分享，可在线阅读，更多相关《FAULT DIAGNOSIS APPROACH BASED ON HIDDEN MARKOV MODEL AND SUPPORT VECTOR MACHINE .doc（5页珍藏版）》请在三一办公上搜索。
1、CHINESE JOURNAL OF MECHANICAL ENGINEERING92- Vol.20, No. 5, 2007LIU GuanjunLIU XinminQlUJingHU NiaoqingCollege of Mechatronics Engineeringand Automation,National University of DefenseTechnology,Changsha 410073, ChinaFAULT DIAGNOSIS APPROACH BASED ON HIDDEN MARKOV MODEL AND SUPPORT VECTOR MACHINE*Abs
2、tract: Aiming at solving the problems of machine-learning in fault diagnosis, a diagnosis approach is proposed based on hidden Markov model (HMM) and support vector machine (SVM). HMM usually describes intra-class measure well and is good at dealing with continuous dynamic signals. SVM expresses int
3、er-class difference effectively and has perfect classify ability. This approach is built on the merit of HMM and SVM. Then, the experiment is mad; in the transmission system of a helicopter. With the features extracted from vibration signals in gea.box, this HMM-SVM based diagnostic approach is trai
4、ned and used to monitor and diagnose th; ge-vaox s faults. The result shows that this method is better than !iMM-based and SVM-basod duvgnosiig methods in higher diagnostic accuracy with small training samples Key words: Hidden Markov mode. Support vector imchijie Fzl: diagnosis0 INTRODUCTIONGearbox
5、es are very imporiaul to the transmission system of a helicopter and they affect trie helicopters reliability and safety directly. It is significant to diagnose the faults of the gearboxes rapidly and efficiently. But the present machine-learning approaches (neural network, for example) that be used
6、 widely in condition monitoring and fault diagnosing have some shortcomings, such as: Diagnosis is the matching result of the present information and the templates. The relation between former information and latter information is ignored. Experiential risk minimization (ERM) principle is adopted. S
7、o lots of samples are essential to train the practicable model. But these samples are very difficult to be acquired.Hidden Markov model (HMM) is a statistical model extended from Markov model. HMM is capable of characterizing a doubly embedded stochastic process with an underlying stochastic process
8、 that, also unobservable (hidden), can be observed through another set of stochastic processes. HMM is a parametric model characterized by the state transition probabilities, the instantaneous probabilities of test outcomes given the system state and the initial state distribution. These parameters
9、can be adaptively estimated by the well-known Baum-Welch algorithm. HMM, which is the statistical model of continuous dynamic series, has the precise data structures and reliable computation. Now it becomes the dominating approach for speech recognitionI Many researchers have applied HMM to conditio
10、n monitoring and fault diagnosing2 3, and the favorable results that is better than that of the neural network is obtained4.Support vector machine (SVM) is a novel powerful machine-learning method with small-samples based on VC dimension theory and structural risk minimization (SRM) principle. SVM i
11、mplements well trade-off between the quality of the approximation of the given data and the complexity of the approximating function by SRM, and owns high generalization performance. The SVM has more excellent feature than ANN. And these features are as follows: The SVM could take the optimal soluti
12、on in the condition of a small number of samples. The optimal solution for the SVM is transformed to solve a quadratic programming problem. In this method, the global optimum solution could be taken, but only the local optimum solution could be gained for the ANN algorithm. (D Algorithm for the SVM
13、transforms the sample space (SS) into the high dimensional feature space (HDFS) by the nonlinear transformation. In the HDFS,it structures linear classification function to achieve the nonlinear classification in the SS. It indicates that the machine learning has good generalization performance, and
14、 solves the dimensionality problem51. The SVM has been successfully applied to fault diagnose because of their excellent classification ability6 71.HMM is good at dealing with sequential inputs, while SVM shows superior performance in classification. Furthermore, HMM usually provides an intra-class
15、measure while SVM proposes inter-class difference. Since these two classifiers use different criteria, they can be combined to yield an ideal one. So a hybrid HMM and SVM fault diagnostic approach is presented to solve the unstable fault diagnosing problems.1 HMM-SVM BASED DIAGNOSTIC MODEL1.1 Hidden
16、 Markov modelsHMM are extensions of Markov models to include the case where the observations are probabilistic functions of the states rather than the states themselves. An HMM is characterized by several parameters. The first parameter is the transition probability distribution A=ag, where ay is th
17、e probability of being in state Sj at time M-l provided that the state at time / is Sh i.e.(1)Kq.SqS,) i,jNwhere q, denotes the state at time / and N is the number of states. The second parameter of an HMM is the observation probability distribution, B= bjk)(2)b)(k) = P(okq,=SJ) Kj(=5,) 1/JVwhich is
18、 the probability of 5, being the initial state.A compact notation X=(AJI,k) is used to define an HMM. The probability of a given observation sequence, 0=o, o2,-,ot, can be calculated as(4)P(0A) = Z Xs0fls,s,AJosJ* This project is supported by National Natural Science Foundation of China(No. 50375153
19、). Received January 9, 2007; received in revised form May 24, 2007; accepted June 14,2007The maximum likelihood (ML) method can be used to reestimate the model parameters, X=(AJ, ic), as followsCHINESE JOURNAL OF MECHANICAL ENGINEERING93 (7)(8)Emj ok)(5)bl(k) =(,)where ntj is the expected number of
20、transitions from S, to Sj, nt is the expected number of transitions from Sj, rrtj is the expected number of times in Sj.Training of an HMM for a given observation sequence can be realized by the so-called Baum-Welch algorithm. Starting with initial or pre-estimated HMM parameters, the algorithm upda
21、tes the parameters, by calculating the ML estimates, step by step increasing the probability of the observation sequence in each step. The training procedure along with the other features of hidden Markov models is explained in detail in Ref. 1.1.2 SVM algorithmsThis section briefly introduces the t
22、heory of SVM. A more detailed description of SVM can be founr! in Ref. 5.Statistical learning theory(SlT), which is a small-sample sia tistics introduced by VA JN1 iC, si al in 1970s, provides us an uniform framework for i-iiral learning problem. And a novel powerful learning meihod ca led SVM is de
23、veloped based on it. The SVM, which can solve small-sample learning problem, has been successfully applied in pattern recognition and function approximation.Based on the structural risk minimization principle from the computational learning theory, SVM seeks a decision surface to separate the traini
24、ng data points into two classes and makes decisions based on the support vectors that are selected as the only effective element from the training set.As for the binary classification, assume that the training set is(Xl,yMXI,yI),-,(X.,y.) fle(l,-l) = l,2,-, (6)A separating hyper-plane divides it int
25、o two sides, each side containing points with the same class label only. The goal of the SVM learning is to find the optimal separating hyper-plane (OSH) that has the maximal margin to both sides. This can be formula-rized asj min (W) = -wf = -(WW) s.t. yi(W.X,) + b-QThe dual problem isi nitmin Q(a)
26、 = -alaJylyJ(XlXJ)-Yiats.t. a, 0 i=-,2,n X.y,a, =0 The decision function ismin Q(a) = -aiaJylyJK(XiXJ)-1al(10)2 ij-i*-ins.t. 0z, C ( = l,2,-,n Xjyz, =0rV,The weights can be calculated by minimizing the mean square value of the residual error over an analysis window31.The observational data of vibrat
27、ion signal of mechanical system behaves just like the linear auto-regressive model in that the next sample of the signal is relate to the p previous samples. So in this paper, the reflection coefficients of the polynomial transfer function of the linear auto-regressive model was chosen as condition
28、features. The p was suitably chosen by the preexistent knowledge and system experiments.If there is periodicity in the fault signal, the preprocessing of getting rid of the periodicity must be taken. Then, the linear auto-regressive model can be built.After preprocessing, every vibration signal is d
29、ivided into T windows of equal length. The windows overlap each other with some points. Here every window is 128 points long and overlaps 64 points each other. A set of features was extracted from each window. The features for a single window were selected to be the reflection coefficients of the po
30、lynomial transfer function of the linear auto-regressive model for that window. The order of the model here is 6. These features were taken as the state observation of the helicopters gearbox.2.2 Training of the HMM-SVM diagnostic model2.2.1 Training of HMMFig. 2 Initial Markov chainBecause the glob
31、al optimum result depends on the initial values of X, segmental *-means segmentation with clustering isBecause continuous Gaussian mixed HMM is better than discrete HMM in smaller distortion and clearer classification, continuous Gaussian mixed HMM with rapid left-to-right Markov chain was adopted here. The number of states and mixed Gaussian distribution should be chosen properly to arrive practicable HMM according to actual condition, the smaller the samples is, the fewer the number of states and mixed Gaussian distribution are. In this paper, 4 states Markov chain (F
