Uncertainty quantification and its properties for hidden Markov models with application to condition based maintenance
MetadataShow full item record
Condition-based maintenance (CBM) can be viewed as a transformation of data gathered from a piece of equipment into information about its condition, and further into decisions on what to do with the equipment. Hidden Markov model (HMM) is a useful framework to probabilistically model the condition of complex engineering systems with partial observability of the underlying states. Condition monitoring and prediction of such type of system requires accurate knowledge of HMM that describes the degradation of such a system with data collected from the sensors mounted on it, as well as understanding of the uncertainty of the HMMs identified from the available data. To that end, this thesis proposes a novel HMM estimation scheme based on the principles of Bayes theorem. The newly proposed Bayesian estimation approach for estimating HMM parameters naturally yields information about model parametric uncertainties via posterior distributions of HMM parameters emanating from the estimation process. In addition, a novel condition monitoring scheme based on uncertain HMMs of the degradation process is proposed and demonstrated on a large dataset obtained from a semiconductor manufacturing facility. Portion of the data was used to build operating mode specific HMMs of machine degradation via the newly proposed Bayesian estimation process, while the remainder of the data was used for monitoring of machine condition using the uncertain degradation HMMs yielded by Bayesian estimation. Comparison with a traditional signature-based statistical monitoring method showed that the newly proposed approach effectively utilizes the fact that its parameters are uncertain themselves, leading to orders of magnitude fewer false alarms. This methodology is further extended to address the practical issue that maintenance interventions are usually imperfect. We propose both a novel non-ergodic and non-homogeneous HMM that assumes imperfect maintenances and a novel process monitoring method capable of monitoring the hidden states considering model uncertainty. Significant improvement in both the log-likelihood of estimated HMM parameters and monitoring performance were observed, compared to those obtained using degradation HMMs that always assumed perfect maintenance. Finally, behavior of the posterior distribution of parameters of unidirectional non- ergodic HMMs modeling in this thesis for degradation was theoretically analyzed in terms of their evolution as more data become available in the estimation process. The convergence problem is formulated as a Bernstein-von Mises theorem (BvMT), and under certain regularity conditions, the sequence of posterior distributions is proven to converge to a Gaussian distribution with variance matrix being the inverse of the Fisher information matrix. An example of a unidirectional HMM is presented for which the regularity conditions are verified, and illustrations of expected theoretical results are given using simulation. The understanding of such convergence of posterior distributions enables one to determine when Bayesian estimation of degradation HMMs is justified and converges toward true model parameters, as well as how much data one then needs to achieve desired accuracy of the resulting model. Understanding of these issues is of utmost important if HMMs are to be used for degradation modeling and monitoring.