Onboard Engine FDI in Autonomous Aircrafts Using Compact Stochastic Nonlinear Modelling of Flight Signal Dependencies Dimitrios G. Dimogianopoulos, John D. Hios, and Spilios D. Fassois ∗ Abstract— An engine Fault Detection and Isolation (FDI) scheme for an autonomous aircraft is introduced. It relies on compact-in-size two-stage stochastic nonlinear modelling (valid for an entire flight regime) of the relationships among common flight variables. The key idea is that in the first stage, major nonlinearities in the dynamics are accounted for through the use of Constant Coefficient Pooled Nonlinear AutoRegressive with eXogenous (CCP-NARX) excitation representations. In the second stage any remaining information is modelled by low- order Constant Coefficient Pooled AutoRegressive Moving Aver- age (CCP-ARMA) representations. These relationships (and the corresponding identified two-stage model) are valid for aircraft engines in healthy state. Thus, purposely designed statistical hypothesis tests are used to detect changes in these relation- ships due to fault occurrence. The scheme’s performance and robustness are assessed with flights conducted under various external conditions and commanded attitude settings. Index Terms— Fault detection and isolation, stochastic non- linear modelling, aircraft systems, statistical decision making. I. INTRODUCTION In modern aircraft, and even more so in future pilot-less versions, high reliability and security should be obtained at the lowest possible cost. The replication of existing critical hardware (the “hardware redundancy” principle), coupled to a voting scheme to perform Fault Detection and Isolation (FDI), implies added weight and cost while its reliability for certain components (actuators) is often criticized [1]. Hence, creating additional redundancy for FDI purposes, requires the intelligent use of the existing hardware and the advanced processing of the collected information. A relevant FDI scheme applied to a gas turbine is reported in [2]. The turbine dynamics are modelled by means of a static linear representation (partly derived from physical principles), augmented by a nonlinear function of terms accounting for noisy input and output signals (which are a number of thermocouple measurements in the turbine exhauster), non-stationarities and so on. The identification of the model parameters is done with a two step procedure (first the nonlinear terms, then the linear ones). Statistical hypothesis testing is used to relate occurrence of the consid- ered faults to changes in the mean of a purposely designed Research supported by the European Commission [FP6 STREP project No. 503019 on “Innovative Future Air Transport System (IFATS)”]. The authors are with the Stochastic Mechanical Systems & Automation (SMSA) Group, Department of Mechanical & Aeronautical Engineering, University of Patras, GR 265 00 Patras, Greece, Tel/Fax: (++ 30) 2610 997 405, 2610 997 130, E-mail: {dimogian,hiosj,fassois}@mech.upatras.gr, Internet: http://www.mech.upatras.gr/∼sms *Corresponding author. Gaussian variable (statistical local approach principle). An engine FDI approach also based on this principle is proposed in [3]. However, the considered model features a purely nonlinear structure, the input and output measurements are typical turbine health parameters (pressures, shaft speeds and so on) and FDI is performed off-line (signals collected during the engine’s lifespan). A slightly different approach is presented in [4] and is inspired by the Interactive Multiple Model FDI scheme ap- plied to control surfaces sensors and actuators (see [5], [6]): Kalman filters provide linear models representing “healthy” and specific “faulty” engine dynamics with input and output signals being both engine control variables and other pa- rameters (pressures and various thermodynamic quantities). All model outputs are compared to that of the engine and the differences (referred to as error residuals) are evaluated by a supervising scheme. The model achieving the smallest (in some sense) residual “best” describes the current engine state, and FDI is concluded. Another scheme based on the same “multiple models” principle, but with Takagi-Sugeno fuzzy models instead of Kalman filters, is presented in [7]. Non-model based FDI schemes can be found in [8], [9]. The aim of this study is the design and the feasibility assessment of an onboard scheme for both conventional and autonomous aircraft. The scheme innovates in that engine FDI results are obtained without monitoring internal engine quantities, as in [2]-[7]. Instead, it relies on the modelling of relationships among common flight data (see [10], [11]) such as acceleration, surface moments, thrust and so on, for which a physics-based model may not be available (unlike [2]) or be very complex. Since no specific technical information (other than the available recorded data) is required, the proposed FDI scheme may be easily adapted to various aircraft types. The considered relationships are valid only for aircraft engines in healthy state, and are modelled by a two-stage stochastic nonlinear representation, conceptually similar to that in [2]. First, the nonlinear part of the dynamics is modelled using Constant Coefficient Pooled Nonlinear AutoRegressive with eXogenous (CCP-NARX) excitation representations. Then, any remaining information is modelled through low- order Constant Coefficient Pooled AutoRegressive Moving Average (CCP-ARMA) representations. Using such two- stage representations in the FDI scheme offers multiple ad- vantages: a) Due to the stochastic framework, they explicitly account for uncertainties during the modelling phase. b) Their pooled form (explained in section II) leads to the considered (nonlinear) aircraft dynamics being accurately Proceedings of the European Control Conference 2007 Kos, Greece, July 2-5, 2007 WeC08.6 ISBN: 978-960-89028-5-5 3422