Multi-State Asynchronous Joint Estimation for Lithium-Ion Batteries - A Nature Inspired Multi-Timescale Approach

Multistate joint estimation enables more effective monitoring and assessment of battery operation conditions. This article proposes a multitimescale framework for asynchronous joint estimation of multiple coupled yet time-varying battery states, namely state of charge (SOC), state of energy (SOE), and state of health (SOH). First, an improved convolutional neural network (CNN) architecture is designed to enhance the model feature extraction capability. Specifically, inception modules are integrated into the CNN model for long timescale feature extraction, while the feature map stacking combined with attention mechanisms is employed for short timescale feature extraction, leveraging the inherent multiscale receptive fields of Inception to model global dependencies and the dynamic feature refinement capability of attention-enhanced stacking to emphasize local transient details. Second, a nature inspired r-GA optimization algorithm is proposed to enhance the training efficiency and performance of the CNN model by optimizing the network parameters. Besides, a Kalman filter is integrated to further improve the accuracy of state estimation by introducing physical information about the battery dynamics. Finally, the input dimensions and update strategies of the framework are tailored for different states to enable joint state estimation across multiple timescales. Experimental results demonstrate that the proposed framework produces much more precise estimation of SOC, SOE, and SOH with mixed-rate updates, while also exhibiting robustness when dealing with data corrupted with Gaussian white noise.