Gearbox vibration signal pre-processing and input values choice for neural network training

Vibration generated by a gear-set is a signal of its technical condition. Many papers have been published on relation between the vibration signal and the condition of a gear-set. It is convenient to investigate this relation by dividing the factors which have influence on vibration generated by a gear-set into four major groups: design, production technology, operational and change of condition. While considering all these factors for the neural network (NN) training we have to have many representations of vibration signal for different gear-set conditions. To overcome the problem with signal representations mathematical modelling and computer simulation (MM and CS) can be used. For vibration signal generation the model of a system consisting of a driving electric motor, a flexible coupling, two gear-sets (double stage gearbox) and a driven machine is used. The condition of a gear-set may be described by local as well as distributed faults. Fault models also have to be taken into consideration while using MM and CS when vibration signals are generated. The vibration signals have to be pre-processed and relation between vibration signals and gear-set condition is to be assessed. Vibration signal pre-processing transforms vibration signals into signal estimators as spectrum, cepstrum, envelope spectra, bispectrum and time-frequency spectrogram. There is a problem that how to select best suitable input values for the neural network training. The paper is going to consider the issues of gearbox vibration signal pre-processing and input values selection for the neural network training