Symbolic time series analysis based health condition forecasting in complex electronic systems

A major issue in system health forecasting emanates from the necessity of transforming instantaneous (fast time scale) health condition and/or performance related observations into slow time-scale estimates. Slow time-scale transformation refers to aggregation of information from observations within a time interval, and assigning a representative state or symbol to the whole interval. The sequence of such symbols can be used to track and forecast system performance/health condition in a reliable way. Symbolic time series analysis (STSA) that employs an entropy maximization approach towards observation partitioning and symbol assignment has been proven quite useful for this purpose. This paper presents an STSA based approach for forecasting performance/health conditions of complex electronic systems using outlier removal and information fusion based pre-processing, and non-linear dynamic Markov model-based post-processing schemes. The dynamic Markov model computes the probability of observing a word that is present in symbolic time series. The probability of transition from one state to another is estimated by traversing through the symbolic series transition probabilities. Thereby, a discrete state transition model is obtained that can serve as the estimator of a system's behavior (in terms of health or performance) over time. An advantage of Markov model is that it extends naturally to forecast the performance/health states and estimates the Remaining Useful Life (RUL). Under this work, a STSA-based forecasting scheme was developed and validated on a set of automotive GPS1,2.