Intelligent health state diagnosis of lithium-ion batteries for electric vehicles using wavelet-enhanced hybrid deep learning integrated with an attention mechanism

Abstract The increasing integration of lithium-ion batteries in electric vehicles has spurred extensive research aimed at enhancing the safety and efficiency of battery management systems. A fundamental component of battery management system functionality is the accurate estimation of the state of health (SOH), which is critical for ensuring the dependable and safe operation of electric vehicles. To address this challenge, we introduce FCM-CNN-WNN-WBILSTM-AM, a time-series prediction framework specifically designed to improve SOH estimation accuracy for Li-ion batteries. The proposed framework starts with a preprocessing phase using fuzzy c-means (FCM) clustering to group batteries with similar characteristics, enabling more precise and customized predictions. It then employs a convolutional neural network (CNN) for initial feature extraction, followed by a wavelet neural network (WNN) layer to handle the non-stationary nature of battery degradation. A wavelet bidirectional long short-term memory (WBILSTM) layer further enhances time-series analysis by capturing both past and future dependencies. To refine feature selection and improve predictive accuracy, an attention mechanism (AM) is integrated, ensuring the model focuses on the most relevant information. To improve computational efficiency and ensure global optimization, the FCM-CNN-WNN-WBILSTM-AM framework employs the RMSprop optimizer, replacing the commonly used Adagrad optimizer. The experimental validation, utilizing multi-battery datasets from the National Aeronautics and Space Administration and the Center for Advanced Life Cycle Engineering repositories, demonstrates the model’s effectiveness in accurately capturing SOH trends and degradation patterns. The results indicate a significant enhancement in SOH estimation accuracy, achieving a root-mean-squared error of 0.0013 and a mean absolute percentage error of 0.0015. These outcomes highlight the model’s superior predictive performance and reduced error rates compared to existing methodologies. Ultimately, the proposed advancements contribute to improving the reliability and longevity of electric vehicle batteries, promoting the widespread adoption and sustainability of electric mobility solutions.