Exploration of Multi-output Gaussian Process Regression for Residual Storage Life Prediction in Lithium Ion Battery

The shift from conventional war to troubled peace has seen the reduction of full-scale deployment of military equipment for operations. Conversely, the increased occurrences of natural disasters and epidemics necessitate the provision of contingency equipment, to readily surge for responsive support. Long term storage presents a more cost-effective approach to own, maintain and operate such military or civil equipment. Recent developments in artificial intelligence and machine learning have fueled interest in equipment prognosis and proactive predictive maintenance to enable prognostics and health management (PHM). However, this is not prevalent in the domain for equipment under long term storage. This could be attributed to slower, and possibly different degradation modes, coupled with lack of large data sets for prognosis in the storage scenario. In this paper, we explore the use of multi-output Gaussian Process (MOGP) regression, on lithium ion battery units in storage, under various state-of-charge and temperature conditions, and discuss empirical effects of training data selection and test data parameters on the accuracy of the residual storage life (RSL) prediction. This will pave way for the development of RSL prediction methods for equipment fleet under storage, considering the scenario of limited or no knowhow of the degradation model.