Towards intelligent online diagnosis and degradation prognostics of lithium-ion batteries: A mechanism–data fusion approach

Lithium-ion batteries experience complex degradation governed by multiple interacting mechanisms, posing challenges for real-time aging-mode identification. To overcome this issue, we propose a mechanism–data fusion framework that couples an extended single-particle model (SPM) with a multi-task learning (MTL) architecture. The electrochemical model explicitly incorporates solid–electrolyte interphase (SEI) growth and lithium plating side reactions, and employs a multi-swarm cooperative adaptive particle swarm optimization (MSCPSO) algorithm to achieve accurate parameter identification across different temperatures and C-rates. A three-branch MTL framework is then constructed to jointly predict key degradation indicators—including the loss of lithium inventory (LLI), loss of active material (LAM), SEI and plating layer thicknesses, and plating-induced capacity loss—while also classifying the occurrence of lithium plating. Experimental validation demonstrates strong physical consistency and robustness of the proposed framework under various operating conditions. Among the tested architectures, the MT-LSTM model exhibits the best overall performance, achieving a lithium-plating detection accuracy of 99.63 % and an R² exceeding 0.97 for multi-target regression tasks. This unified and scalable framework enables quantitative identification of multiple degradation mechanisms directly from charge–discharge data, offering a practical, real-time, and physically interpretable tool for next-generation battery health management systems.