IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 16, NO. 4, JULY 2008 789 A Statistical Method for the Detection of Sensor Abrupt Faults in Aircraft Control Systems Paraskevi A. Samara, George N. Fouskitakis, John S. Sakellariou, and Spilios D. Fassois Abstract—Aircraft sensors are important for proper operation and safety, and their condition is conventionally monitored based upon the hardware redundancy principle. In this work a statis- tical method capable of independently monitoring a single sensor, and thus enhancing reliability and overall system safety, is intro- duced. The method’s main advantages are simplicity, applicability to a wide variety of aircraft operating conditions, the handling of uncertainties, no need for additionally monitored signals, and no need for physics based aircraft dynamics models. The method is based on a statistical time series framework accounting for random effects and uncertainties, and exploits the fact that abrupt faults are characterized by time constants smaller than those of the air- craft. It employs monitored signal nonstationarity removal, signal whitening via novel pooled autoregressive modeling, statistical de- cision making, as well as electronic spike/glitch removal logic. The method effectiveness is demonstrated within the simulation envi- ronment of a small commercial aircraft via test cases and Monte Carlo experiments with abrupt faults occurring in an angle-of-at- tack sensor. Index Terms—Aircraft control systems, angle-of-attack sensor, fault diagnosis, sensor abrupt faults, statistical methods. I. INTRODUCTION S ENSORS are important aircraft instruments, as proper op- eration and overall system safety require their good and re- liable performance. In modern aircraft, reliability and safety are conventionally enhanced based upon the hardware redundancy principle. Yet, the stringent safety requirements imposed, and also the longer term objective of minimizing hardware replica- tion on board, make the additional use of software-based sensor monitoring and fault detection methods highly desirable. As a consequence, a number of aircraft sensor and actuator fault detection and isolation (FDI) methods have been devel- oped in recent years. Several of them are based upon linear or linearized models and Kalman filtering (KF)-type techniques. Hajiyev and Caliskan [1], [2] utilize analysis of KF-based inno- vations within a statistical framework. Multiple model adaptive Manuscript received November 16, 2005; revised September 12, 2006. Man- uscript received in final form December 17, 2006. Recommended by Asso- ciate Editor S. Kim. This work was supported by the European Commission via the Growth Project GRD1-2000-25261 (Affordable Digital Fly-by-Wire Flight Control Systems for Small Commercial Aircraft, Phase II—ADFCSII). P. A. Samara was with the Department of Mechanical and Aeronautical En- gineering, University of Patras, GR 265 00 Patras, Greece. She is now with the Department of Product Development, Frigoglass S.A.I.C., GR 252 00 Kato Achaia, Greece (e-mail: psamara@frigoglass.com). G. N. Fouskitakis was with the Department of Mechanical and Aeronautical Engineering, University of Patras, GR 265 00 Patras, Greece. He is now with the Department of Electronics, Technological Educational Institute of Crete, Crete GR 73133, Greece (e-mail: fouskit@chania.teicrete.gr). J. S. Sakellariou and S. D. Fassois are with the Department of Mechanical and Aeronautical Engineering, University of Patras, GR 265 00 Patras, Greece (e-mail: sakj@mech.upatras.gr; fassois@mech.upatras.gr). Digital Object Identifier 10.1109/TCST.2007.903109 estimation (MMAE) and interacting multiple model (IMM) esti- mation methods are postulated by Menke and Maybeck [3] and Zhang and Li [4], respectively. In these methods several par- allel extended KFs are designed, each one corresponding to a possible sensor or actuator fault. Zolghadri [5] utilizes a similar bank of extended KFs and decision making based upon the pa- rameter vectors or the innovations sequences for the detection of sensor faults. An alternative family of methods is based upon the concept of a bank of isolation estimators [6], [7], each one covering an area of the aircraft dynamics under specific failure scenarios. These methods are more involved and their design requires the analysis of the stability and learning properties of the fault iso- lation estimators approximating an unknown sensor fault, deter- mination of the adaptive decision thresholds for each estimator, and, determination of the fault isolability conditions. Within this framework, issues such as the use of nonlinear aircraft models and the effective treatment of uncertainties are under investiga- tion [6]. A third family of methods utilizes type observers for sensor fault detection and isolation [8]–[10]. The approach minimizes the influence of noise, disturbances, uncertainties, and commands on the residuals used for fault detection and iso- lation, and maximizes the effects of faults. Nevertheless, it relies on knowledge of a reliable model of the physical system and the computational burden may grow rapidly as the number of faulty sensors increases. Observer stability and robustness are theoret- ically guaranteed only in the neighborhood of the design points, although this can be mitigated using a multi-model approach. Neural network-based methods are also postulated in a number of studies [11]–[14]. They utilize identified neural network-type models for the estimation (“reconstruction”) of the signal of interest using other available measurements (the virtual sensor concept), while the detection of a failed physical sensor is based upon monitoring of the error between the signal obtained via the physical sensor and its estimated counterpart. The advantage of the neural network-based ap- proach basically lies with the model’s capability to capture the nonlinear dynamics of the aircraft (in contrast to KF-type methods that are based upon linearized models). Yet, a higher level of complexity is introduced, and training may be time consuming. A study comparing KF and neural network-based schemes for sensor fault detection in flight control systems is presented in Napolitano et al. [15]. A weak-model-based method, in the sense that only limited knowledge on the characteristics of the signals and the under- lying dynamics is required, for the detection of faults in aircraft sensors is presented in Golan et al. [16]. The faults are detected through nonlinear analysis using the wavelet transform and neural network-based classification. An alternative approach, which employs a virtual sensor in the form of a fuzzy model of 1063-6536/$25.00 © 2008 IEEE