Vibratory Data Fusion for Gearbox Fault Detection Using Autoassociative Neural Networks

This paper deals with gearbox fault detection by vibration analysis. A new processing procedure is proposed which uses the information from several acquisition channels. The approach is based on the assumption that there is a non-linear relationship among the instantaneous vibration magnitude registered for each measurement location. This relationship is captured in the connection weight matrix of an Autoassociative Artificial Neural Network (AANN), which is trained to provide an output vector equal to the input one. In this work, the time synchronous average signal (TSA) for each channel corresponding to the no fault condition is used to train an AANN. Once the AANN is trained, it is used with new data registers as a linear prediction error filter. If the new register contains the same data structure as the training set the prediction error will be low and the machine is working properly. Otherwise, when the new register differs from the training set, as a consequence of a fault, prediction error will be increased in each channel. In this way the information from not only one channel but more than one is used for fault detection and diagnosis as the error signal depends on the TSA signal from all channels. The proposed approach provides a new tool for gear fault detection that is compared on the basis of experimental registers with the most traditional gear processing tools based on TSA such as residual and regular signals. The possibility of generalizing the net prediction capabilities using a training data set that contains several load cases is also explored.