Mechine Fault Diagnosis Method Based On The Infinite Hidden Markov Model

This thesis was supported by the National Natural Science Foundation of China(No. 51261024, 51075372, 51265039, 50775208), Science and Technology Projects of Education Department of Jiangxi Province, China(No. GJJ12405), the Open Fund of The State Key Laboratory of Mechanical Transmissions(No. SKLMT-KFKT-201514), the Guangdong Province Figures Key Laboratory of Signal Processing Research Topics(No. 2014GDDSIPL-01). For the serious shortage in the definition and estimation process of fault diagnosis methods model based on the traditional hidden Markov model(HMM), this thesis applied a new machine learning method--the infinite hidden Markov model(iHMM) to the mechanical fault diagnosis, put forward some new fault diagnosis method based on iHMM, and achieved some innovative achievements. The main contents of this thesis include the following sections:The first chapter discusses the significance for the proposed and expansion of the research. It also discusses the research status of the traditional hidden Markov model(HMM), especially in the field of mechanical fault diagnosis. It discusses the research status of iHMM in detail. On this basis, it puts forward the main research contents of this thesis, the structural arrangements and innovations.The second chapter discusses the theory and the shortage of traditional hidden Markov models, and on this basis discusses the theory and algorithms of the infinite Hidden Markov Model in detail. In iHMM, first, Calculating the inference of state transition probability in Dirichlet process. Second, it uses a hierarchical Dirichlet process to infer hidden state mechanism and the emission mechanism. Lastly, it discusses the model hyperparameter optimization and likelihood estimation. The good performance of the inference algorithm of iHMM is tested and verified through the simulation examples, and the results show that iHMM is equipped with a good ability to explore the number of states, and reflect the state of the sequence of structural features accurately. In the last of this chapter, the feasibility of the iHMM for machinery fault diagnosis is demonstrated from the perspective of probability and statistics.In the third chapter, a new rotating machinery fault diagnosis method is proposed, according to the unique characteristics iHMM. The propsoed method extracts four different fault feature by calculating the spectral kurtosis, and puts it into iHMM to train different diagnostic model, then the trained model is used to recognize faults type of the collected data. The performance of the proposed method and the traditional fault recognition method based on HMM are compared, and the results show that the proposed method is effective and can recognize different faults types effectively. The proposed method outperforms the traditional fault recognition methods based on HMM, it overcomes the shortage of traditional fault recognition methods based on HMM.The fourth chapter discusses another iHMM theory and algorithm based on Beam sampling and proposes a new method of degradation trend prediction of rolling bearings based on iHMM in which the Beaming sampling is used. The proposed method extracts fault feature by calculating the wavelet entropy, and puts it into iHMM to complete the degradation trend prediction. The performance of the proposed method and the traditional method based on HMM in degradation trend prediction of rolling bearings is compared, and the results show that the proposed method is more effective than the methods based on HMM in the filed of degradation trend prediction. In order to test and verify the advantages of feature extraction by calculating wavelet entropy, the root mean square value, kurtosis and information entropy have been put into the iHMM to complete the degradation trend prediction, and compared the method using wavelet entropy, the results show that using wavelet entropy as fault feature is better then other feature in the degradation trend prediction of rolling bearings.The fifth chapter discusses the possibility of degradation trend prediction of rolling bearings in the case of missing data. By simulating the case of missing data proposes a new method of degradation trend prediction of rolling bearings based on i HMM under missing data. It also discusses prediction model and prediction methods with iHMM in the case of missing data. The proposed method can preprocess the missing data effectively. In the proposed method, it extracts fault feature from the preprocessed missing data, and and puts it into iHMM to train the prediction model and complete the degradation trend prediction. The proposed method and the methods in chapter four are compared, and the results show that the proposed method can also be effective under missing data in the filed of degradation trend prediction of rolling bearings.The sixth chapter summarizes the main research content of this thesis, and points directions for further research in the future.