An unsupervised degradation estimation framework for diagnostics and prognostics in cyber-physical system

In order to learn the performance degradation mode of machines for diagnostics and prognostics in Cyber-Physical System (CPS), it is necessary to analyze observed sensor data to find the internal run-to-failure states of a system. In this paper, the research goal focuses on learning the internal state and transition of the degraded states from original data, which has some advantages to reveal the working dynamics of the system. Due to the existence of unlabeled data, the paper proposes an unsupervised framework based on clustering and Hidden Markov Model, named Cluster-based Hidden Markov Model (cHMM). The cHMM aims at converting raw sensory stream into a sequence of symbols as the initial observation and hidden state sequences, and then an extended Viterbi algorithm based on Hidden Markov Model (HMM) is used to discover the final stable hidden states and transitional rules in a dynamic programming way. Based on the learned model and expert's knowledge, performance degradation failure and the root-cause could be predicted and reasoned. Finally, experiments and proof-of-concept demonstration are given to validate the feasibility and effectiveness of the framework based on C-MAPSS turbofan engine dataset.