A novel asymmetric loss function for deep clustering-based health monitoring and anomaly detection for spacecraft telemetry

Aerospace systems essentially require health monitoring and anomaly detection to enhance system safety and reliability and to avoid system failure in spacecraft and satellites operating in remote and harsh environments. Telemetry data is the cornerstone of analyzing, monitoring, and estimating the health, safety, and performance of spacecraft and satellites. Moreover, telemetry data plays a crucial role in enabling better-informed decision-making and faster response in the event of malfunctions or emergencies in spacecraft and satellites, ultimately contributing to the safety, reliability, and success of space missions. Currently, various deep learning models have become extremely popular due to the frequent utilization of conventional symmetric loss functions. However, conventional symmetric loss functions may not precisely capture or measure the complex nonlinearity patterns and relationships present in spacecraft and satellite telemetry data. Besides, symmetric loss functions may not effectively capture the temporal dependencies and dynamics in spacecraft telemetry data. To address these issues, a novel asymmetric loss function named Linear-Log-cosh (Lin-Log-cosh) is proposed. The effect of the proposed Lin-Log-cosh asymmetric loss function is investigated using the Deep Clustering-based Local Outlier Probabilities approach (DCLOP), which is a deep clustering-based health monitoring and anomaly detection approach for spacecraft and satellite telemetry. This paper uses actual CubeSat telemetry data to evaluate whether and how the performance of the DCLOP approach is affected by incorporating the proposed Lin-Log-cosh asymmetric loss function. The experimental findings prove that the new presented asymmetric DCLOP for deep clustering can evidently enhance accuracy and F 1 -Score by more than 1.0% compared to those achieved by the DCLOP with a symmetric Dynamically Weighted Loss (DWL) function. For anomaly detection, the asymmetric DCLOP can evidently enhance accuracy and F 1 -Score by an average of 0.25% and 2.0%, respectively, compared to those obtained by the DCLOP with a symmetric DWL function. This competitive performance obviously and positively affects the monitoring of the health status of spacecraft and satellites, guaranteeing earlier warnings of on-orbit failures.