Fault Diagnosis of Helicopter Gearboxes Using a Neural Network Based Novelty Detector

The detection of novel or anomalous patterns has a wide range of applications including condition-based maintenance, system monitoring and fault detection, medical diagnosis and the detection of suspicious behaviour such as fraud or computer network intrusion. The objective of the proposed Phase I effort is to demonstrate the feasibility of developing a neural-network based novelty detector and apply the developed system for helicopter gearbox fault detection from vibration data. The neural-network based novelty detector will be designed and analyzed based on statistical pattern recognition ideas. The Phase I effort includes the analysis of current neural-network-based and immune-system-based novelty detectors, the development and implementation of a neural-network based novelty detector with desirable statistical properties, and the implementation and testing of the developed architecture for the detection of faults in helicopter gearboxes using vibration data. The anticipated innovations in the proposed approach are: 1) the development of a neural network architecture that has desirable properties for novelty detection, such as low false alarm and high hit rate, 2) the development of an approach for predicting the network expected performance based on statistical pattern recognition principles, and 3) the implementation and testing of the developed system for helicopter gearbox fault detection. The SSC project team will be supported by Bell Helicopter for providing data and knowledge about Helicopter gearbox failures and Prof. Michael Jordan, who has done pioneering work in the area of statistical interpretation of neural networks.