Anomaly detection of spacecraft inertial measurement units based on multi-scale dependency neural network

Abstract

Inertial Measurement Units (IMUs) are critical for spacecraft attitude control, navigation, and positioning. Anomaly detection is essential to ensure the safe and stable operation of these systems. However, existing anomaly detection methods face challenges in terms of precision, generalization, and the interpretability of captured features. To address these issues, this paper proposes an anomaly detection method based on Multi-Scale Dependency (MSD) neural network, leveraging an unsupervised deep learning approach focused on diverse anomaly characteristics. The proposed method employs Temporal Convolutional Network (TCN) to extract local temporal features, which enables the identification of point anomalies, local collective and contextual anomalies. The method also employs Graph Convolutional Networks (GCN) with Maximal Information Coefficient (MIC) to model multivariate dependencies, effectively capturing cross-variable relationship. Additionally, multi-head self-attention mechanism is utilized for global temporal modeling, enhancing the detection of collective and contextual anomalies. Finally, anomaly states are detected by combining reconstruction and prediction. Experimental results demonstrate that the proposed method achieves F1 scores exceeding 95 % across IMU datasets and three public datasets, highlighting its precision and generalization capabilities. Furthermore, the interpretability of the model is enhanced through Shapley value analysis and Monte Carlo sampling-based quantification, promoting the practical application of the method.