Prédiction de l’état de santé (soh) d’une batterie lithium- ion à l’aide de méthodes hybrides d’apprentissage profond

Abstract

Accurate estimation of the state-of-health (SOH) of lithium-ion batteries is critical for optimizing battery management and preventing premature failures. With the growing reliance on lithium-ion batteries in electric vehicles and energy storage systems, reliable SOH pre diction is more important than ever. However, SOH estimation remains challenging due to the evolving nature of battery characteristics over time and the limitations of existing approaches. Traditional methods, such as equivalent circuit models and electrochemical models, can be complex, time-consuming, and lack adaptability. Piloté par les données techniques offer promising alternatives but often require large labeled datasets and may struggle with generalization across varying conditions. To address these challenges, this thesis presents a hybrid deep learning model combining a masked autoencoder (MAE) and a Long Short-Term Memory (LSTM) network (without an attention mechanism) for SOH prediction. The model also incorporates a specific contrastive learning (CL) loss criterion to enhance representation learning. It is trained on sliding win dows of length 100 (batch size 16, 100 epochs) using normalized cycle_norm and capacity as input caractéristiques, derived from the charge-discharge cycle data of batteries B0005, B0006, and B0007. The model is evaluated on a separate test battery, B0018, where it achieves an RMSE of 1.51, an MAE of 0.81, and a coefficient of determination R² of 0.98. These results demonstrate that the combination of MAE for handling incomplete data, CL for improving generalization, and LSTM for temporal modeling forms an effective self-supervised architecture. Overall, the proposed approach confirms the potential of self-supervised learning for scalable and accurate battery health diagnostics.

Date	7 Jan 2026
Original language	French
Awarding Institution	École de technologie supérieure
Supervisor	Bassant Selim (Supervisor) & Waël Jaafar (Co-supervisor)