An LSSVR-based algorithm for online system condition prognostics

Online machine condition prognostics are useful for condition based maintenance decision making in order to prevent unexpected machine breakdown, human injuries, and costs due to loss of productivity. Noise present in measured condition indicators which can represent machine conditions may, however, adversely affect prognostic results. In literature, machine condition prognostics considering noisy observations of condition indicators are rarely reported. In this work, we propose an algorithm to predict true values of condition indicators based on noisy observations. The proposed algorithm jointly uses least square support vector regression (LSSVR), genetic-algorithm-based optimization, and cumulative sum (CUSUM) technique. LSSVR is used to predict true values of condition indicators. To handle noise effects in observations, parameters of LSSVR are selected using an optimization process of which model is specially developed. Genetic algorithms (GA) are used to solve the optimization problem. To accommodate changes of indicator values and noise effects, CUSUM is employed to trigger re-determination of LSSVR parameters when current predictions are not acceptable. The proposed algorithm is compared with five reported methods using two simulation datasets and two experimental datasets.