Spacecraft Health Monitoring Using a Weighted Sparse Decomposition

Abstract

In space operations, spacecraft health monitoring and failure prevention are major issues. This important task can be handled by monitoring housekeeping telemetry time series using anomaly detection (AD) techniques. The success of machine learning methods makes them attractive for AD in telemetry via a semi-supervised learning. Semi-supervised learning consists of learning a reference model from past telemetry acquired without anomalies in the so-called learning step. In a second step referred to as test step, most recent telemetry time-series are compared to this reference model in order to detect potential anomalies. This paper presents an extension of an existing AD method based on a sparse decomposition of test signals on a dictionary of normal patterns. The proposed method has the advantage of accounting for possible relationships between different telemetry parameters and can integrate external information via appropriate weights that allow detection performance to be improved. After recalling the main steps of an existing AD method based on a sparse decomposition Pilastre et al (Sign Proc, 2019 [1]) for multivariate telemetry data, we investigate a weighted version of this method referred to as W-ADDICT that allows external information to be included in the detection step. Some representative results obtained using an anomaly dataset composed of actual anomalies that occurred on several satellites show the interest of the proposed weighting strategy using external information obtained from the correlation coefficient between the tested data and its decomposition on the dictionary.