Remaining useful life prediction in embedded systems using an online auto-updated machine learning based modeling

Abstract

Systems on Chips are increasingly involved in critical equipment in the fields of aeronautics, transportations, and energy. Therefore, monitoring their life cycle is a crucial issue for safety and hazard-prevention. This paper deals with a data-driven method for online prediction of the Remaining Useful Life (RUL) of the safety-critical System-on-Chips (SoC). This method is based on the detection and prediction of drifts in their operating temperatures. The work starts with a description of the formal relationships between temperature drifts and the degradation process of SoCs to justify the choice of the temperature as an indicator of the level of the degradation in the system. Then, temperature-based physical health indicators are constructed using data-driven analytical redundancy. Since temperature varies not just according to the degradation state of the system, but also according to its various normal operating points, data-driven analytical redundancy makes it possible to obtain a health indicator that has a well-defined physical meaning, and which is only sensitive to the SoC degradation process. To predict the remaining useful life of the chip, the trend of the drift is modeled using an auto-regressive neural (NAR) network. The latter is updated online according to the evolution of the temperature drift and the state of the system. Finally, forecasts of the remaining useful life of the SoC are obtained using a combination of temporal projection and threshold data. Simulations and experimental results highlight the effectiveness and accuracy of the proposed approach.