Unsupervised Anomaly Detection for Time Series Data of Spacecraft Using Multi-Task Learning

Although in-orbit anomaly detection is extremely important to ensure spacecraft safety, the complex spatial-temporal correlation and sparsity of anomalies in data pose significant challenges. This study proposes new multi-task learning-based time series (MTAD) method, which captures learn generalized normal patterns hence facilitates detection. First, four proxy tasks are implemented for feature extraction through joint learning: (1) Long short-term memory-based prediction; (2) autoencoder-based latent representation learning reconstruction; (3) variational (4) representation-based prediction. Proxy Tasks 1 4 capture temporal by fusing space, whereas 2 3 fully spatial data. The isolation forest algorithm then detects from extracted features. Application a real dataset reveals superiority our method over existing techniques, further ablation testing each task proves effectiveness multiple tasks. proposed MTAD demonstrates promising potential effective spacecraft.