A Multiscale Fusion DNN With Adaptive Gradient Optimization for Lithium-Ion Battery Pack Fault Diagnosis and SoC Prediction

Fault diagnosis and state-of-charge (SoC) estimation for battery energy storage systems (BESS) are traditionally performed independently, requiring separate deep neural networks (DNNs) and leading to significant computational and storage overhead. To address this, we propose M3Net, a multi-scale multimodal fusion deep neural network for multi-task learning, incorporating an adaptive gradient algorithm. M3Net enhances feature extraction by integrating convolutional kernels of varying sizes, capturing diverse temporal and spatial patterns. To mitigate task conflicts, we introduce a model-independent adaptive gradient updating mechanism, optimizing gradient flow between tasks. Experimental results demonstrate that M3Net significantly improves both efficiency and accuracy in lithium-ion battery pack fault diagnosis and SoC estimation.