A novel fault diagnosis system for aircraft based on adaboost and five subsystems with different pattern recognition methods

The goal of this paper is to devise a fault diagnosis and decision system of aircraft and practice it in real flight condition. In order to face the real usage of system and keep its reliability and safety, the system contains two monitoring terminals according to different requirements and equipment conditions. One is on operation at the aircraft (on-line), which is based on AD-ABOOST with ten weak classifiers. It only needs to detect if there is any fault or not, and make two judgments of the health of aircraft - normal or mild dangerous (- fault tolerant system can resolve it) and dangerous (- back to airport) for pilot. Another one is for maintenance office (off-line), which is based on five subsystem with different recognition methods: Back-Propagation Neural Networks (BP), Probabilistic Neural Networks (PNN), Learning Vector Quantization Neural Networks (LVQ), Gaussian Mixture Models (GMM) and Decision Tree (DT). With the fusion of the diagnosis results of these five subsystems, the system can detail the diagnosis results and distinguish each fault.