Nonlinear System Remaining Life Prediction Based On Spectral Clustering And CHSMM

With the rapid development of science and technology, the size of the control logic and the complexity is increasing in the large complicated system, so the factors that affect the system's stability is also increasing, corresponding also lead to the system failure and the possibility of fault more and more. Once the system failure will result in substantial human, material and financial losses, or even caused catastrophe, so the maintenance of repair of the system is particularly important. But most of the traditional maintenance strategy also remain in the regular maintenance and subsequent maintenance, lacking accurate state judgment and health analysis on the running system. CBM as an advanced maintenance concept, can monitoring and recognizes the state of the system, and track the degradation process, and predict the futures state of the system, thus achieving the health management for system.HMM as a pattern recognition technology has used widely in the areas of speech recognition and Face recognition. As the extension of HMM, Hidden Semi-Markov Model(HSMM), because of its stronger ability to model and analysis of time-series, it more suitable for system degradation state recognition and prediction of time-related RUL(Remaining Useful Life). At present there are some problem in the researching of failure prediction based on the HSMM, such as lacking of category labels and discrete training, according to these problem, in this paper, carried out the prediction of RUL method based on spectral clustering and CHSMM.According to the problem of lacking of samples'label and it is difficult to models, establish the RUL prediction framework based on spectral clustering and CHSMM, and describes the overall solution of nonlinear system degradation state recognition and RUL predictions.According to the high complexity of feature extraction for nonlinear system, propose a joint feature extraction and spectral clustering framework. The method directly uses PCA (Principal Component Analysis) in the kernel space of established by spectral clustering, this is equivalent to KPCA (Kernel Principal Component Analysis) in the original feature space, testing proved that this method can significantly improve computational efficiency.For the problem of initialization parameters and loss of information in the process of training the HSMM, study of the improved algorithm of HSMM and evolved it into CHSMM. And on this basis, a RUL prediction method based on the state proposed by using spectral clustering and CHSMM is proposed.An experimental plan of C-MAPSS Aircraft Engine Simulation platform is design, to time-series as samples, through the experiments to validate the method of degradation state recognition and RUL prediction in this paper. And the result shows that this method can accurately positioning the system degradation state, and predict the RUL of the system, and has good feasibility and effectiveness.