Session

Technical Poster Session 1: Ground Systems & Operations

Location

Utah State University, Logan, UT

Abstract

In this paper, we propose a ground-based automated novelty detection system for a small satellite attitude dynamics control system using a one-sided learning algorithm: One-Class Support Vector Machine (OC-SVM) method. This fault-detection system was designed to only learn from nominal behavior of the satellite during the commissioning phase and to identify and detect anomalies when there was a subtle behavioral failure in the attitude control system. The detection system was trained by only observing the nominal attitude dynamics behavior of a small satellite for a period of time. Training data was obtained from reaction wheel outputs in a healthy attitude control system, and reaction wheel currents and angular velocities were selected as training features. A one-class classifier was built from a hyperplane decision function during training. An adaptive Sequential Minimal Optimization (SMO) method was utilized to solve the quadratic problem in the application of OC-SVM algorithm to provide an optimal solution for the hyperplane decision function. Two tests were performed on the system to validate its feasibility and detection accuracy. Untrained reaction wheel bearing failures were added into the attitude control system validation tests to examine whether the fault-detection system was capable of detecting and diagnosing the reaction wheel failures. Training and testing performance for the fault-detection system are presented with discussion.

Available for download on Saturday, August 07, 2021

Share

COinS
 
Aug 7th, 12:00 AM

Automated Fault-Detection for Small Satellite Pointing Control Systems Using One-Sided Learning

Utah State University, Logan, UT

In this paper, we propose a ground-based automated novelty detection system for a small satellite attitude dynamics control system using a one-sided learning algorithm: One-Class Support Vector Machine (OC-SVM) method. This fault-detection system was designed to only learn from nominal behavior of the satellite during the commissioning phase and to identify and detect anomalies when there was a subtle behavioral failure in the attitude control system. The detection system was trained by only observing the nominal attitude dynamics behavior of a small satellite for a period of time. Training data was obtained from reaction wheel outputs in a healthy attitude control system, and reaction wheel currents and angular velocities were selected as training features. A one-class classifier was built from a hyperplane decision function during training. An adaptive Sequential Minimal Optimization (SMO) method was utilized to solve the quadratic problem in the application of OC-SVM algorithm to provide an optimal solution for the hyperplane decision function. Two tests were performed on the system to validate its feasibility and detection accuracy. Untrained reaction wheel bearing failures were added into the attitude control system validation tests to examine whether the fault-detection system was capable of detecting and diagnosing the reaction wheel failures. Training and testing performance for the fault-detection system are presented with discussion.