This report documents work done for the Turbine Engine Fault Detection and Isolation Program. A gas path performance algorithm has been developed which can be used to trend engine module health. The Maintenance Information Management System was developed for the integration of data into the maintenance framework of the services. These tools have been applied to test data from the F100/EDS, TF34/TEMS and TF41/IECMS data acquisition systems. (Author).
Maintenance decision analysis models for evaluation of the TF34 maintenance process, both with and without the Turbine Engine Monitoring System (TEMS), are formulated. These models form the foundation for the U.S. Air Force to establish techniques for determining optimal policy for troubleshooting and maintenance on its aircraft engines using decision analysis methods. Technical background is provided and models presented. Model structure and parameters, as well as input and output, are treated. A preliminary plan for model evaluation is given, including methods for data collection, model evaluation criteria, as well as solution techniques and algorithms for the actual model evaluation. Conclusions are drawn and directions for future activity are suggested. (Author).
Widely used for power generation, gas turbine engines are susceptible to faults due to the harsh working environment. Most engine problems are preceded by a sharp change in measurement deviations compared to a baseline engine, but the trend data of these deviations over time are contaminated with noise and non-Gaussian outliers. Gas Turbine Diagnostics: Signal Processing and Fault Isolation presents signal processing algorithms to improve fault diagnosis in gas turbine engines, particularly jet engines. The algorithms focus on removing noise and outliers while keeping the key signal features that may indicate a fault. The book brings together recent methods in data filtering, trend shift detection, and fault isolation, including several novel approaches proposed by the author. Each method is demonstrated through numerical simulations that can be easily performed by the reader. Coverage includes: Filters for gas turbines with slow data availability Hybrid filters for engines equipped with faster data monitoring systems Nonlinear myriad filters for cases where monitoring of transient data can lead to better fault detection Innovative nonlinear filters for data cleaning developed using optimization methods An edge detector based on gradient and Laplacian calculations A process of automating fault isolation using a bank of Kalman filters, fuzzy logic systems, neural networks, and genetic fuzzy systems when an engine model is available An example of vibration-based diagnostics for turbine blades to complement the performance-based methods Using simple examples, the book describes new research tools to more effectively isolate faults in gas turbine engines. These algorithms may also be useful for condition and health monitoring in other systems where sharp changes in measurement data indicate the onset of a fault.
Automated engine monitoring has emerged as an important element in the Air Force's strategy to reduce propulsion system support costs and to improve aircraft operational availability. There has been a long history of development activity directed towards engine monitoring. These systems have demonstrated that sensor and automated data acquisition can be implemented effectively in both prototype and operational applications. Historically, however, no Air Force system has resulted in validated improvement in the engine maintenance and logistics process nor in a substantial cost savings. This situation is due in part to the fact that the performance data were not reduced to a concise, usable format relevant to the decision process of the maintenance personnel. Moreover, there was no procedure developed for integrating the performance data into the maintenance framework. This report presents the results of an intensive study of the Air Force maintenance/logistics process based on a selected sample of tactical bases, depots, and major commands. The objective is to define the requirements that the Air Force engine management structure imposes on automated data integration, in general, and engine performance monitoring, in particular. Such an automated integration of turbine engine monitoring system data with current data systems requries coordination between a variety of sources, both manual and automated. The results of this study are the requirements for such integration based on typical Air Force maintenance needs. (Author).
Aircraft engine diagnostic methods are reviewed. The role of computer-aided diagnostic procedures for current and future engines is discussed from the aspects of performance monitoring, trending, and fault detection/isolation. Development of advanced maximum likelihood or regression algorithms for each of these is presented. A methodology is developed for applying these algorithms to models derived from engine test stand or flight data. Specific computational results are given for a high performance turbofan engine. (Author).
An effective fault detection and isolation (FDI) technology can play a crucial role in improving the system availability, safety and reliability as well as reducing the risks of catastrophic failures. In this thesis, the robust sensor FDI problem of gas turbine engines is investigated and different novel techniques are developed to address the effects of parameter uncertainties, disturbances as well as process and measurement noise on the performance of FDI strategies. The efficiencies of proposed techniques are investigated through extensive simulation studies for the single spool gas turbine engine that is previously developed and validated using the GSP software. The gas turbine engine health degradation is considered in various forms in this thesis. First, it is considered as a part of the engine dynamics that is estimated off-line and updated periodically for the on-board engine model. Second, it is modeled as the time-varying norm-bounded parameter uncertainty that affects all the system state-space matrices and third as an unknown nonlinear dynamic that is approximated by the use of a dynamic recurrent neural network.In the first part of the thesis, we propose a hybrid Kalman filter (HKF) scheme that consists of a single nonlinear on-board engine model (OBEM) augmented with piecewise linear (PWL) models constituting as the multiple model (MM) based estimators to cover the entire engine operating regime. We have integrated the generalized likelihood ratio (GLR)-based method with our MM-based scheme to estimate the sensor fault severity under various single and concurrent fault scenarios. In order to ensure the reliability of our proposed HKF-based FDI scheme during the engine life cycle, it is assumed that the reference baselines are periodically updated for the OBEM health parameters. In the second part of the thesis, a novel robust sensor FDI strategy using the MM-based approach is proposed that remains robust with respect to both time-varying parameter uncertainties and process and measurement noise. The scheme is composed of robust Kalman filters (RKF) that are constructed for multiple PWL models. The parameter uncertainty is modeled by using a time-varying norm bounded admissible structure that affects all the PWL state space matrices. The robust Kalman filter gain matrices are designed by solving two algebraic Riccati equations (ARE) that are expressed as two linear matrix inequality (LMI) feasibility conditions. The main objective is to propose a robust filter that satisfies the overall performance requirements and is not affected by system perturbations. The requirements include a quadratically stable filter that ensures bounded estimation error variances having predefined values. In the third part of the thesis, a novel hybrid approach is proposed to improve the robustness of FDI scheme with respect to different sources of uncertainties. For this purpose, a dynamic recurrent neural network (DRNN) is designed to approximate the gas turbine engine uncertainty due to the health degradations. The proposed DRNN is trained offline by using the extended Kalman filter (EKF) algorithm for an engine with different levels of uncertainty, but with healthy sensors. The convergence of EKF-based DRNN training algorithm is also investigated. Then, the trained DRNN with the fixed parameters and topology is integrated with our online model-based FDI algorithm to approximate the uncertainty terms of the real engine. In this part, the previously proposed HKF and RKF are integrated with the trained DRNN to construct the hybrid FDI structure.
