Attention-Enhanced CNN-BiLSTM Model for Electric Vehicle Lithium-Ion Battery State of Health Estimation and Fault Classification

This paper analyzes the aging data of LIB (lithium-ion batteries) by combining the CNN (Convolutional Neural Network)- BiLSTM (Bidirectional Long Short-Term Memory Network) model with attention mechanism, realizes battery SOH (State of Health) estimation and fault classification, and improves the accuracy of electric vehicle battery state monitoring and the reliability of fault diagnosis. This paper introduces the CNN-BiLSTM-AM (Attention Mechanism) model, which uses multi-layer convolution kernels to perform fine-grained analysis on the signals of the charging and discharging process. It passes the features extracted by the CNN part into the BiLSTM to capture the dynamic change characteristics of the front-end dependency in the time series. The AM is used to perform weighted synthesis of the contribution of each CNN-BiLSTM module to achieve battery SOH estimation and fault classification. The experiment uses the LIB cycle life dataset from NASA (National Aeronautics and Space Administration). In SOH estimation, the MSE (Mean Squared Error) values of CNN-BiLSTM-AM on B0005, B0006, B0007 and B0018 are 0.015, 0.017, 0.017 and 0.015, respectively. The model in this paper still performs well in the battery fault classification task, with classification accuracy rates of 98.5% and 97.3% on B0005 and B0007, respectively. By comparing with CNN-BiLSTM, BiLSTM, XGBoost and Transformer models, the MSE data fluctuation of SOH prediction of this model under different battery temperatures is only 0.001, which is stable and better than other models. The CNN-BiLSTM-AM model introduced in this paper has significant SOH prediction performance and fault classification performance under different battery environments, providing technical support for the intelligentization of battery management systems.