Adaptive Mamba-LSTM Hybrid Network with Joint State Estimation and Fault Diagnosis for Lithium Batteries

Accurate State of Charge (SOC) estimation and fault diagnosis are the bread and butter of safe lithium-ion battery management systems. Unfortunately, conventional approaches hit a wall when trying to tackle both SOC and fault estimation simultaneously under faulty conditions, not to mention the headache of parameter optimization. To crack this nut, this paper rolls out a Mamba-LSTM multi-task learning model powered by Adaptive Architecture Search. The ball gets rolling with the initialization of network structural parameters’ search space, followed by leveraging differentiable gradient propagation to automatically fine-tune parameters, ultimately cooking up a hybrid architecture that marries Mamba’s selective feature extraction prowess with LSTM’s temporal modeling chops. To be precise, the Mamba module locks onto global dependencies and keeps noise at bay via its state space model, while the LSTM network hones in on extracting local temporal features. What’s more, the model features a dual-branch output setup that spits out both SOC estimates and fault classification probabilities in one fell swoop. Finally, the feasibility of the proposed method is validated using a lithium-ion battery dataset containing various fault states.