SWAN: Sliding Window based Attention Network for Early Battery Anomaly Detection

Electric bicycles have become essential tools for addressing the last-mile delivery challenge, particularly in developing countries. The advent of battery swapping services has partially resolved common issues faced by electric bicycles, such as limited battery life and inefficient charging. However, battery swapping service providers continue to face significant challenges, including battery loss, theft, and overdue returns—all of which negatively impact service quality and result in economic losses. In this study, these issues are collectively referred to as battery abnormalities. Early detection and analysis of battery abnormalities are critical for enabling proactive interventions. In this paper, we introduce the sliding window-based attention network (SWAN) for early detection of battery anomalies. By designing a joint loss function, our model effectively balances the timeliness and accuracy of anomaly detection. Furthermore, we present innovative feature expansion and extraction algorithms specifically designed to address the unique characteristics of battery time series data. Experimental results show that SWAN substantially outperforms existing models in the early detection of battery anomalies.