Beyond the Surface: Inferring Internal Battery Pressure and Temperature Using Optical Fibre Sensing and Machine Learning

Abstract Virtual sensing of internal battery parameters is essential for improving safety and control without invasive instrumentation. We combine embedded optical fibre sensors with deep temporal models to reconstruct internal temperature and pressure evolution of lithium-ion (LIB) and sodium-ion (SIB) batteries. The physical sensors data provide ground truth targets, while surface and ambient temperature, voltage, current, and charge are used as features. Bayesian optimisation identifies chemistry-dependent temporal convolutional network configurations, with LIB favouring shorter input windows and larger convolutional kernels, while SIB benefits from longer windows and smaller kernels. The resulting virtual sensors achieve high-fidelity reconstruction across both chemistries. Internal temperature and pressure reached RMSE of 0.075 °C and 0.288 °C, and 0.048 bar and 0.013 bar, for the LIB and SIB, respectively. Residual, cycle-level, and cycle-resolved analyses demonstrate stable generalisation, and permutation-importance profiles confirm reliance on physically meaningful electro-thermal features, enabling non-invasive, chemistry-adapted monitoring for battery management systems.