Fault detection and diagnosis of nonlinear system based on dynamic model analysis

Abstract

In order to locate the fault position and take available steps early for turbine-generator set operating under mal-condition, it is essential to build a reasonable system of condition monitoring and fault diagnosis. An effective method for vibration fault diagnosis based on integration of wavelet transform and neural network is presented. The advantage of the wavelet transform logarithmic time frequency bands is in achieving flexible frequency resolution, making it able to extract both high-frequency and low-frequency components from the original signal. The fault diagnosis model of turbo-generator set is established and the improved Levenberg-Marquardt optimization technique is used to fulfill network parameter identification. The wavelet neural network not only learns adequate decision functions and arbitrarily complex decision regions defined by the weight coefficients, but also looks for those parts of the parameter space that are suited for a reliable categorization of the input signals. By means of choosing enough samples to train the fault diagnosis network, the output result can determine fault mode in accordance with the input feature vector. The practical multi-concurrent fault diagnosis for stator temperature fluctuation and rotor vibration approves to be accurate and comprehensive.