Research On Incipient Fault Diagnosis Methodology Based On Weak Signal Feature Extracting And Its Application

The development of industrial automatization results in the increasing demand of early fault diagnosis. While the fault features may easily be contaminated for serious noise interference and complex working condition, how to detect weak fault features from low signal noise ratio (SNR) vbration signals becomes an urgent problem that need to be solved. Applying weak signal processing method and incipient fault diagnosis theory to conduct signal denoising and early fault diagnosis is a hot topic in mechanical fault diagnosis. This dissertation studied vibration signal processing, weak fault features detecting and incipient fault diagnosis method.(1) Vibration signal was decomposed into a set of low frequency parts and high frequency parts by adopoting the multi-resolution analysis method. Weak signal features at any point of the signal can be revealed with the changing of scale factor and translation factor. At each decomposition level, the wavelet coefficients were compared to the threshold that was preset according to threshold rules, and the denoising signal can be constructed by the new wavelet coefficients. Simulation results show that this method can effectively improve SNR of the orignal signal and reveal weak fault features.(2) Based on analysis local characteristics of the signal, vibration signal can be descomposed into a series of intrinsic mode functions (IMFs) ranging from high frequency to low frequency. Taking into account the defects of empirical mode decomposition (EMD) shifting processing and shortages of traditional extension methods, this dissertation proposed a novel methodology based on singular value decomposition (SVD) and support vector regression (SVR), periodic features can be detected effectively according to SVD, which provided an approach to determain signal extension length in SVR method. Meanwhile, a scanning method was employed to get rid of the redundant IMFs, simulation results show effectiveness of the proposed method.(3) Support vector machine (SVM) was employed to construct the optimal hyperplane, it can obtain the minimum structure risk and maximum generalization ability and outreach capacity. Meanwhile, introducing Gaussian kernel function, fault features can transform from low dimension space to high dimension space, which can significantly improve separability of training samples. This method was employed to conduct bearing fault diagnosis, realationship between fault categories, training samples and diagnostic accuracy was studied, results show that the proposed method can effectively identify early fault features even when the training samples are insufficient.(4) The dissertation developed a visual fault diagnosis system based on Matlab and Visual Basic, which provided an easy operation and high expansibility fault diagnosis tool.