Hidden Markov Model-based Methods In Condition Monitoring of Machinery Systems

Abstract

Tool condition monitoring (TCM) approaches can be dichotomized into regression and classification, based on whether the predicted output is continuous or discrete. The major focus of this thesis is on TCM approaches with continuous outputs, as crucial but rarely explored prediction methodologies. Firstly, a physically segmented hidden Markov model (HMM)-based approach is proposed to perform continuous tool wear monitoring. Then, performance of the approach is further improved by employing a hidden semi-Markov model (HSMM), with flexible duration distribution as opposed to HMM. Tunability of the HSMM-based model given a loss function is also explored. A multi-modal HMM based approach is further proposed to capture multiple regiments, compared to the previous single HMM-based approach. Furthermore as in the real-world fault detection and diagnostics applications, the true model is usually not realizable, an HMM-based semi-nonparametric approach is introduced to utilize the training data more effectively and increase accuracy of conventional HMM-based classification approach.