A Two-Step Sub-Sampling Approach for a Computationally Efficient Particle Filter-Based Prognosis

Abstract

Since their introduction, prognostic algorithms based on Particle Filtering (PF) have secured a leading position wherever reliable, grounded, and explainable predictions are required. If this success led to a plethora of research aimed at further enhancing accuracy and explainability, comparatively less attention has been paid to addressing the high computational costs associated with particle propagation. Against this backdrop, this paper approaches the idea of sub-sampling the prediction steps, allowing the propagation of the system state directly between non-consecutive time instants, thus obtaining an optimized trade-off between accuracy and computational effort. In particular, driven by the idea that predictions should be more accurate towards the end of life, this research proposes a novel strategy based on sub-sampling with two different sampling frequencies. The two-phase sub-sampling scheme divides the prognostic horizon into two phases, identifying the ideal set of the three driving parameters: the frequency for the first phase, the switching time, and the higher frequency for the last phase. The computationally intensive tasks are delegated to an offline calibration phase, during which an XGBoost-based recommendation model is trained. In contrast, the online phase involves a single, rapid inference to recommend the optimal configuration, enabling the particle filter (PF) to operate with the selected two-rate propagation strategy while adhering to a user-defined computational time budget. On an experimental Li-ion battery-discharge case study, the framework reduces online computational time by up to 95 percent while keeping the relative error below 3 percent and achieving MAPE lower than 0.52 percent with respect to the standard PF. Comparative analysis demonstrates the framework’s capacity to identify an optimal time-error trade-off, thus effectively merging those qualities that have made PF-based approaches the go-to solutions for explainable prognostic with the computational efficiency required for online implementation.