RAMAN: Robust Approaches for Multimodal Anomaly Detection in Mars Rover Power Systems

This paper proposes RAMAN, a framework of approaches for multimodal anomaly detection that is robust to multiple anomaly types, input data, and domain constraints for the Mars Science Laboratory (MSL) power subsystem. Existing anomaly detection systems focus on a subset of anomalies and require a human expert for verification, making the process prone to bias and scaling issues. RAMAN combines the strengths of unsupervised feature extraction methods with the domain knowledge of Martian power systems. It uses thresholding methods based on architectures like autoencoders and spline approximations to find anomalies in different data modalities. The system is evaluated on the power systems data of the Curiosity rover and preliminary results show that it outperforms existing baselines while also scoring high on qualitative evaluation from field experts. The main strengths of RAMAN are its unit composition that helps detect anomalies in subsystems, easy extension to other Martian and non-Martian systems, and easy usability through interactive interfaces.