A Unified Nonlinear Adaptive Approach for Detection and Isolation of Engine Faults

A challenging problem in aircraft engine health management (EHM) system development is to detect and isolate faults in system components (i.e., compressor, turbine), actuators, and sensors. Existing nonlinear EHM methods often deal with component faults, actuator faults, and sensor faults separately, which may potentially lead to incorrect diagnostic decisions and unnecessary maintenance. Therefore, it would be ideal to address sensor faults, actuator faults, and component faults under one unified framework. This paper presents a systematic and unified nonlinear adaptive framework for detecting and isolating sensor faults, actuator faults, and component faults for aircraft engines. The fault detection and isolation (FDI) architecture consists of a parallel bank of nonlinear adaptive estimators. Adaptive thresholds are appropriately designed such that, in the presence of a particular fault, all components of the residual generated by the adaptive estimator corresponding to the actual fault type remain below their thresholds. If the faults are sufficiently different, then at least one component of the residual generated by each remaining adaptive estimator should exceed its threshold. Therefore, based on the specific response of the residuals, sensor faults, actuator faults, and component faults can be isolated. The effectiveness of the approach was evaluated using the NASA C-MAPSS turbofan engine model, and simulation results are presented. Introduction National Transportation Safety Board accident data covering 7,571 U.S.-registered aircraft from 1980 to 2001, categorized by accident cause, reveal that 52 percent of the hardware-induced accidents were related to aircraft system malfunctions and 36 percent of these were caused by propulsion system component malfunctions (Ref. 1). Therefore, a real-time fault diagnosis scheme for aircraft engines might significantly improve flight safety by enabling accurate and early detection and isolation of incipient fault conditions. An important area of engine health management (EHM) is sensor validation. A sensor fault may lead to poor regulation or tracking performance, or even affect the stability of the control system. Moreover, a faulty sensor output may cause inaccurate diagnostics/prognostics, resulting in unnecessary replacement of system components or mission abortion. Therefore, it is important to correctly assess the health of onboard sensors. In addition to sensors, certain propulsion system components and actuators may fail as a result of aging or damage due to harsh operating conditions or combat. Existing nonlinear EHM methods often deal with component faults, actuator faults, and sensor faults separately. Specifically, when dealing with sensor validation, people usually assume there are no component or actuator faults; when dealing with component and actuator faults, it is often assumed that there are no sensor faults. In the former case, a NASA/TM—2010-216360 1 component or actuator fault may be misinterpreted as a sensor fault; in the latter case, a sensor fault may be misinterpreted as a component or actuator fault. Both cases may potentially lead to incorrect fault diagnostic decisions and unnecessary maintenance. Therefore, it would be ideal to address sensor faults, actuator faults, and component faults under one unified framework. Several researchers have investigated the development of such a unified fault diagnostic framework (Refs. 2 to 7). However, most of those results are based on linear engine models. In addition, the desire for future propulsion systems to perform over an extended range of operating conditions, characterized by dramatic variations in dynamic pressure and nonlinear thermal dynamics, requires enhanced performance and new functionalities. Consequently, the dynamics of aircraft engines are usually highly nonlinear and rapidly changing. Many existing fault detection and isolation (FDI) methods are based on the assumption that the system exhibits linear behavior in the neighborhood of a steady-state operating point, and therefore linearization-based methods are used (e.g., the Kalman filter-based methods developed by Merrill et al. (Ref. 8) and Kobayashi and Simon (Refs. 9 and 10). Approaches that involve linearizing the engine dynamics about several steady-state operating conditions and blending parameters and controllers for these operating points tend to be rather complicated. Moreover, when the effect of various faults is taken into account, the size and complexity of the scheduling and calibration tables are significantly increased, which makes design and real-time implementation very difficult. Therefore, future EHM designs will benefit significantly from new methods that are directly based on intrinsic nonlinearities of the engine dynamics. To address this issue, a unified nonlinear adaptive framework for detecting, isolating, and estimating sensor faults, actuator faults, and component faults for aircraft engines is developed in this paper. The presented approach is based on a bank of nonlinear fault diagnostic estimators that generate estimated measurements using advanced nonlinear adaptive estimation/learning techniques. The approach employs a nonlinear adaptive estimation architecture that is capable of directly dealing with nonlinear dynamic system models and nonlinear faults (Refs. 11). By extending the FDI logic to handle transient conditions as well as steady-state conditions, false alarms can be reduced, and early detection and isolation of sensor faults, actuator faults, and component faults can be achieved. With the presented adaptive nonlinear FDI techniques, unstructured modeling uncertainty is taken into account to improve performance over a wide range of operating regimes. Unstructured modeling uncertainty refers to the case where the modeling uncertainty function appears possibly in all state equations without being pre-multiplied by a known distribution matrix that satisfies certain conditions. This allows us to formally introduce adaptive thresholds for both the fault detection and isolation tasks. In general, adaptive thresholds have advantages over fixed thresholds because they enhance fault sensitivity and robustness with respect to modeling uncertainty (Ref. 15). The purpose of this paper is to highlight these features using a realistic gas turbine engine simulation and to demonstrate how sensor faults, actuator faults and component faults are dealt with under a unified framework. Preliminary simulation results are presented to show the effectiveness of the approach using a transient operating scenario. The remainder of the paper is organized as follows. First, the FDI approach is presented. Second, a neural networkbased adaptive engine model is introduced to enhance fault sensitivity and diagnostic robustness with respect to modeling uncertainty resulting from normal engine degradation. Then, the application of the techniques to a high-fidelity engine simulation is presented. Simulation results for an example transient test case are provided. Finally, conclusions and possible future work are presented.