A performance baseline for machinery condition classification by neural network

Abstract

This project develops a set of multi-layered perceptron neural networks to serve as performance baselines for the classification of the material condition of a representative helicopter intermediate gearbox by advanced neural network models currently under development.

The first half of a collection of machinery condition sensor data recording induced faults in a TH-1L helicopter intermediate gearbox is used to develop candidate network configurations and the second half of the data collection to test the candidate networks. The data is derived from three accelerometer sensor channels. The network with the lowest average machinery condition classification error is chosen as the baseline network for that sensor channel and described in "C" computer program code.

The error in gearbox machinery condition classification for these networks ranges from 2.2% for the sensor channel five network to 7.9% for the channel 5+6+7 combined network.