Evolving Fuzzy System Applied to Battery Charge Capacity Prediction for Fault Prognostics

TL;DR

This paper introduces an evolving fuzzy system approach for predicting battery charge capacity degradation, enabling accurate fault prognostics and uncertainty quantification using data-driven models adaptable to individual units.

Contribution

The paper presents a novel methodology using Evolving Fuzzy Systems for battery RUL prediction, incorporating recursive error tracking for uncertainty estimation in fault prognostics.

Findings

EFS models effectively predict battery RUL using historical and streaming data.

The approach quantifies uncertainty in long-term predictions.

Experimental results on NASA battery data validate the method's accuracy.

Abstract

This paper addresses the use of data-driven evolving techniques applied to fault prognostics. In such problems, accurate predictions of multiple steps ahead are essential for the Remaining Useful Life (RUL) estimation of a given asset. The fault prognostics' solutions must be able to model the typical nonlinear behavior of the degradation processes of these assets, and be adaptable to each unit's particularities. In this context, the Evolving Fuzzy Systems (EFSs) are models capable of representing such behaviors, in addition of being able to deal with non-stationary behavior, also present in these problems. Moreover, a methodology to recursively track the model's estimation error is presented as a way to quantify uncertainties that are propagated in the long-term predictions. The well-established NASA's Li-ion batteries data set is used to evaluate the models. The experiments indicate that generic EFSs can take advantage of both historical and stream data to estimate the RUL and its uncertainty.