State of health prediction of lithium-ion batteries in charging chambers based on multi-modal deep learning

In underground environments characterized by dust, humidity, and explosion risks, the degradation process of lithium-ion batteries (LIBs) often exhibits nonlinear and multistage characteristics, making it difficult for traditional single models to comprehensively capture their dynamic changes. To address this issue, a multi-modal deep learning-based method was proposed for predicting the state of health of LIBs in charging chambers. A multi-modal deep learning network model, TCN-BiLSTM-Transformer, was constructed, leveraging a multi-level feature extraction mechanism for efficient processing of temporal signals. The Temporal Convolutional Network (TCN), utilizing dilated convolutional kernels with an exponential expansion rate, captured multi-scale local features while preserving temporal integrity. The Bidirectional Long Short-Term Memory Network (BiLSTM) established bidirectional temporal dependencies through bidirectional Gated Recurrent Units (GRUs), effectively identifying both forward and reverse degradation features during battery deterioration. The Transformer layer dynamically allocated feature weights through a multi-head self-attention mechanism, focusing on key features of global degradation patterns. Multi-source sensory data, including voltage, current, and temperature, collected during the battery's operation, were employed as indicators of its health state. Pearson correlation analysis was conducted to evaluate the association between these health indicators and battery capacity, identifying five indirect health factors that served as inputs to the prediction model. Experimental results demonstrated that the proposed method achieved correlations exceeding 98%, with relatively low Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). In validation tests conducted under simulated operating conditions for explosion-proof LIBs in coal mines, the method attained a correlation of 99.47%, exhibiting smaller fluctuations and higher accuracy compared to predictions made using traditional methods.