Ultrasonic feature-driven interpretable state-of-health estimation for lithium-ion batteries using a hybrid attention-enhanced learning framework

Abstract

Accurate estimation of the state of health (SOH) of lithium-ion batteries is essential for ensuring the safe and reliable operation of electric vehicles and energy storage systems. Most existing SOH estimation methods rely primarily on electrical features, which have limited capability to characterize internal degradation information and often lack sufficient interpretability. To address these limitations, this study proposes an ultrasonic feature-driven SOH estimation method based on an Improved Sparrow Search Algorithm (ISSA)-optimized Convolutional Neural Network-Squeeze-and-Excitation-Bidirectional Long Short-Term Memory (CNN-SE-BiLSTM) model. An ultrasonic testing platform was established to acquire response signals during battery aging, and candidate ultrasonic features, including Initial amplitude, Maximum amplitude, Minimum amplitude, and Initial time of flight (TOF), were extracted. Pearson correlation analysis was used to select initial amplitude, maximum amplitude, and initial TOF as model inputs. The CNN-SE-BiLSTM model was then developed to capture degradation-related information and temporal variations in ultrasonic features, while ISSA was employed to optimize key hyperparameters. Experimental results show that the proposed method achieves an RMSE of 0.0180%, an MAE of 0.0148%, and an R2of 0.9987 under the final feature set. Compared with CNN-SE-BiLSTM without ISSA, the RMSE is reduced by 78.18%. With only 30% training data, the RMSE remains 0.0461%, indicating good limited-sample performance. SHapley Additive exPlanations (SHAP) analysis further confirms the dominant roles of Initial amplitude and Initial TOF. The results demonstrate that ultrasonic features provide effective non-electrical degradation information for accurate, stable, and interpretable battery SOH estimation.

Keywords: State-of-health estimation, Ultrasonic feature, Improved sparrow search algorithm, SHapley Additive exPlanations