Integrating Aging Data and Machine Learning for RealTime Battery State of Safety Monitoring

Abstract

Lithium-ion batteries play a crucial role in modern energy systems. However, safety assessments still rely mainly on destructive abuse testing, in which batteries are intentionally driven into potentially hazardous failure modes. Such testing is costly, requires extensive safety precautions, and is impractical for covering the wide range of operating conditions ultimately managed by the BMS. The presented study investigates whether battery safety can be predicted non-destructively using only aging data. A Long Short-Term Memory (LSTM) model is trained on discharge cycle data, including voltage, current, capacity, and empirically calculated actual SoS values, to predict future SoS for battery cells. Three different (LiFePO4, LFP)/graphite pouch-type battery cells with LiFePO4 (LFP) as cathode active material and graphite as anode) are cycled under selected charge–discharge protocols (1C–2C and 0.5C–1C) at 100% depth of discharge. They were analyzed to develop the LSTM model across their full first-life and second-life aging ranges, down to a 60% state of health (SoH). Predicted capacities were mapped to a continuous State-of-Safety (SoS) metric using an analytical degradation–safety relationship and subsequently translated into failure probability and European Council for Automotive R&D (EUCAR) severity levels to quantify hazard evolution. The results show that LSTM reliably captures both linear and nonlinear aging trends, with close agreement between predicted and measured capacities. Derived SoS and hazard metrics indicate that all three cells maintain high intrinsic safety throughout most of their lifetime, with only mild increases in lowseverity event probabilities near end-of-life and extremely low probabilities of severe outcomes. Our findings demonstrate a feasible, non-destructive framework for integrating safety assessment into battery aging studies and real-time monitoring.