A Hybrid LSSVR-HMM Based Prognostics Approach

Prognostics is an important aspect in health management system and condition-based maintenance (CBM). It is a task to predict the future health of systems, which has few researches so far. In this paper, a hybrid approach for failure prognostics is proposed. The approach combines least squares support vector regression (LSSVR) with hidden Markov model (HMM). Features extracted from sensor signals are used to train HMMs, which represent different health levels. LSSVR algorithm is used to predict feature trend. According to the probabilities of each HMM, it can determine the future health state and estimate the remaining useful life (RUL). To evaluate the proposed approach, test was carried out using bearing vibration signals. Simulation results show that the LSSVR-HMMs method is efficient in prognostics. It can forecast long before failures occur and predict the RUL.