Detecting anomalies in space-object light curves using temporal-attention LSTM
Weixiao Li, Yu Zhang, Jianfeng Wang, Chen Guo, Hui Zhi, Xianqun 显群 Zeng 曾, Peng Qiu, Lei Wang, Jihao Yin · Space habitation. · 2025
The increasing density and activity of space objects have raised significant concerns regarding space situational awareness. Light curves offer a non-intrusive, cost-effective means to monitor these objects and detect anomalies such as rapid attitude maneuvers. In this paper, we present TA-LSTM, a novel anomaly detection framework that employs a temporal-attention LSTM (Long Short-Term Memory) network to identify anomalous events in space-object light curves. Our approach is structured into three primary modules: (1) an input module that segments light curves using a sliding window to generate fixed-length sequences, (2) a prediction module that utilizes an LSTM network enhanced with a temporal attention mechanism to forecast subsequent magnitude values, and (3) an anomaly detection module that computes prediction errors and flags anomalies using statistical thresholds derived from normal light curve data. To mitigate the challenge of limited labeled data, we developed a large-scale simulated light curve dataset covering a diverse range of anomaly scenarios. Experimental results show that TA-LSTM consistently outperforms mainstream methods in anomaly detection and exhibits strong potential for generalization to real-world applications, offering an effective solution for advancing space object monitoring and situational awareness.