Reliability of AI in Predicting the State of Health of Li-Ion Batteries*

Lithium-ion (Li-Ion) batteries are widely adopted to power a large variety of mobile applications. Their capacitance degrades during lifetime, mainly with the increasing number of charge and discharge cycles. Such a capacitance degradation with respect to its nominal value is typically represented by the battery State of Health (SoH). Predicting the SoH of Li -Ion batteries is crucial to avoid the powered system’s service disruption. Artificial Intelligence (AI) can be adopted to predict the SoH of batteries. However, the hardware executing the AI may be affected by soft errors (SEs) during its operation in the field, possibly compromising the correctness of the AI prediction.In this paper, we analyze the impact of SEs on the reliability of AI in predicting the SoH of Li-Ion batteries. We will show that, due to SEs affecting the AI, unhealthy batteries may be erroneously classified as healthy, with possible dramatic consequences on the powered system’s operation. Moreover, healthy batteries may be wrongly classified as unhealthy, thus mandating unnecessary maintenance/replacement costs. A detailed analysis of the diverse effects of SEs, depending on the affected portions of the AI implementing hardware has been conducted, thus enabling to identify the parts that, if affected by SEs, are more likely to compromise the AI prediction ability.Based on the achieved results, the possible selective adoption of traditional fault tolerance techniques, such as ECCs, or duplication and comparison followed by recovery, has then been proposed, to enable to increase the AI reliability in predicting batteries’ SoH.