Using Bayesian Networks in Gear Fault Diagnosis

Abstract

In rotary machinery, the symptoms of vibration signals in the frequency domain have been used as inputs for neural networks and diagnosis results can be obtained by network computation. However, in gear or rolling bearing systems, it is difficult to extract symptoms from vibration signals in the frequency domain where shock vibration signals are present, and neural networks do not provide satisfactory diagnosis results without adequate training samples. Bayesian networks provide an effective approach for fault diagnosis in cases given uncertain knowledge and incomplete information. To classify the shock of vibration signals in the gear system, this study uses statistical factors of vibration signals. Based on these factors, the fault diagnosis is implemented by using Bayesian networks and the results of the two methods, namely, back-propagation neural networks and probabilistic neural network in gear train systems, are compared.