Trustworthy battery state of charge estimation enabled by multi-task deep learning

The accurate estimation of state of charge (SOC) is critical for the reliable operation and safety of lithium-ion batteries. However, unexpected operating conditions, such as dynamic temperature variations and sensor faults, pose significant challenges to the reliability of data-driven estimation methods. This work addresses these challenges by proposing a novel multi-task deep neural network (DNN) framework that integrates a plug-and-play anomaly detection module to enable trustworthy SOC estimation. The proposed model employs an encoder to capture temporal dependencies from input voltage and current signals, followed by two parallel modules for SOC estimation and anomaly detection. The anomaly detection module reconstructs the input voltage using the learnt features and current sequence, identifying unreliable SOC estimations based on large reconstruction errors. Experimental validation using 188,157 discharging data points demonstrates the model's robustness under diverse unexpected conditions. The SOC estimation root mean squared error (RMSE) is reduced from 4.38 % to 0.86 % under temperature variations, and significant improvements are also observed in the presence of sensor faults. Additionally, its generalisation capability is demonstrated through validation on a public dataset featuring a different battery type. These results highlight the proposed method's effectiveness in enhancing both the accuracy and reliability of SOC estimation. The predictive performance is further rationalised by visualising the extracted latent features, providing deeper insights into the model's capabilities. The proposed multi-task learning paradigm has the potential to be adapted for trustworthy estimation of other battery states. • Trustworthy SOC estimation by integrating a plug-and-play anomaly detection module into deep neural networks (DNNs). • A multi-task DNN for efficiently integrating the SOC estimation and anomaly detection. • The method is validated under various unexpected operating conditions using collected discharging data. • The developed model can detect abnormal input under unexpected temperature and sensor fault conditions. • The method is highly flexible in both structure and application.