ABSTRACT
Both military and civilian helicopters require routine maintenance to prevent problems while they are in flight One of the most dangerous problems is the failure of gearboxes, which can lead to a catastrophic crash and loss of life. Traditional methods of detecting future problems relied on vibrational analysis models of the gearbox, which were highly complex. Newer techniques focused on applying signal pre-processing techniques, such as the Short-Time Fourier Transform (STFT) or the Wigner-Ville Distribution (WVD) to the vibration time signal and using the results as inputs for neural network training. These methods have all reached a measured level of success. This paper will show that it is possible to have improved neural network learning speed as well as improved fault detection and classification utilizing a new signal pre-processing technique known as Enhanced Time-Frequency Distribution (ETF.) The helicopter gearbox vibration data was supplied by Accurate Automation Corporation and the network architecture used was a two-layer feedforward network using generalized backpropagation learning rule.
