Aerospace and Electronic Systems Magazine April 2017 - 26
Feature Article: DOI. No. 10.1109/MAES.2017.150049 Supervised Learning Algorithms for Spacecraft Attitude Determination and Control System Health Monitoring Bassem Nassar, Wessam Hussein, Egyptian Armed Forces, Cairo, Egypt Mohamed Medhat, NARSS, Cairo, Egypt INTRODUCTION The main concern of any space mission operation is to ensure the health and safety of spacecrafts. The worst case in this circumstance is probably the loss of a mission but the more common interruption of spacecraft functionality can result in compromised mission objectives. Spacecraft telemetry holds information related to state-of-health (SOH) of its subsystems. Each parameter has information that represents a time-variant property (i.e., a status or a measurement) to be checked. Moreover, the tremendous increase in telemetry data volume and its complexity directs the need for more efficient and scalable data processing systems. As a consequence, there is a continuous improvement of telemetry monitoring applications in order to reduce the time required to respond to changes in a spacecraft's SOH. So, a fast conception of the current state of the spacecraft is thus very important in order to respond to failures. Furthermore, the progressive growth in spacecrafts leads to increase of the level of standardization in spacecraft operations. Anomaly detection algorithms, also known as outlier detection algorithms seek to find portions of a data set that are somehow different from the rest of the data set. Some good surveys of anomaly detection algorithms can be found in [1], [2]. A supervised anomaly detection algorithm requires training data consisting of a set of examples of anomalies, and a set of examples of nominal data. From the data, the algorithm learns a model that distinguishes between the nominal and the anomalous data points. Supervised anomaly detection algorithms typically require tens or hundreds of labeled examples of anomalies, plus a similar number of labeled examples of nominal data points, in order to obtain adequate performance. Unsupervised anomaly detection algorithms are trained using a Authors' current address: MTC, Mechatronics, Kobry Elkobbah, Cairo, Egypt 11646. E-mail: (bassemamo@yahoo.com). The primary version of this article was partly presented at the 2015 IEEE Aerospace Conference, March 7-14, 2015, Yellowstone, Big Sky, MT. Manuscript received April 9, 2015, revised August 25, 2015, February 17, 2016, May 27, 2016, and ready for publication June 1, 2016. Review handled by M. Jah. 0885/8985/17/$26.00 © 2017 IEEE 26 single set of unlabeled data, most or all of which is assumed to be nominal. It learns a model of the nominal data, which can be used to signal an anomaly when new data fails to match the model. The data-driven approach seeks to build a model for detecting anomalies directly from the data, rather than building it based on human expertise. Data-driven process monitoring or statistical process monitoring (SPM) applies multivariate statistical methods and machine learning techniques to fault detection and diagnosis for space operations, which has become one of the richest areas for research and practice over the last two decades. Based on methods from multivariate statistical analysis, SPM has found wide application in various fields including space operations. So, due to the data-based nature of the SPM methods, it is relatively easy to apply to real processes of rather large scale compared with other methods based on systems theory or rigorous process models. In this article, we propose three supervised learning algorithms to detect anomalies in space mission operations. The first algorithm is our supervised statistical algorithm based on the projection to latent structure discriminant analysis (PLS-DA) technique. The second is multivariate statistical software called soft independent modeling for class analogy (SIMCA-P) developed by Umetrics. The third is applying the nonlinear support vector machines (SVMs) algorithm to the spacecraft telemetry for the first time. We describe a novel supervised PLS-DA statistical dimensionality reduction algorithm coded in Matlab. The algorithm presents a practical approach for fault management that includes: (i) fault detection; (ii) fault identification; (iii) diagnosis and quality monitoring. The task of the algorithm is to model, analyze, and classify the telemetry as well as to identify key contributors to anomalous events automatically which lead to further details about the attitude determination and control subsystem (ADCS) behavior. However, this approach doesn't build a causality direction between variables, rather it builds a correlation trend inside a bounded region. The algorithm represents an efficient tool for ADCS anomaly detection, which gives useful information for the flight controllers and analysts to support decision making. Also, it is the starting point to enhance our algorithm to be coupled with other unsupervised and supervised learning techniques to solve the relevant problems, which can't be done in commercial software like SIMCA-P. The contributions of this design work can be enumerated as follows. IEEE A&E SYSTEMS MAGAZINE APRIL 2017