Weak Signal Detection And Research On Mechanical Fault Diagnosis System

Signal processing algorithms are very important to mechanical fault diagnosis. These are important research fields to study how to correct signal from noise-polluted signals, and how to detect weak characteristic signals of mechanical fault from background noise. Fault diagnosis system could be used to acquire signals and analyze signals, and the system could guide maintenance personnel to make decision of fault diagnosis. Thus the development of fault diagnosis system directly affects fault diagnosis field. Aiming at the electric-mechanical equipments, this dissertation focuses on the research on filtering, weak feature extraction and fault diagnosis system.If useful signal is contaminated by colored noise which is formed from white noise, and the white noise could be measured, using Adaptive Neuro-Fuzzy Inference System (ANFIS) can approach that colored noise through nonlinear training. Empirical Mode Decomposition (EMD) could decompose mixed signal into a series of Intrinsic Mode Functions (IMFs), then the non-stationary mixed signal is decomposed into approximate stationary signals. This dissertation introduces a new signal processing method which proved effective in improving filter precision. Firstly, the method using EMD to decompose mixed noise-polluted into IMFs; Secondly, the method using ANFIS to approach colored noise which be contained in each IMF. The filter's entire adaptability of the new method will not be damaged, because EMD also is an adaptive method which is independent from signal itself.Non-uniform sampling could break the limit of sampling theorem. But if amplitude of one signal is more than 10% smaller than the other's when two signals are sampled by non-uniform sampling, the weak signal couldn't be detected from the mixed signal's spectrum because of sample time's pseudo-randomness. Using a character that Independent Component Analysis (ICA) is immune to signal's amplitude but sensitive to orthogonality, the weak signal's frequency is detected by FastICA through construct virtual signal.This dissertation also gives another method to solve the non-uniform sampling's problem. The method was also used to detected weak sinusoidal signals embedded in a uniformly distributed white noise. This dissertation introduces two portable fault diagnosis systems, one is based on System on Chip (SoC) and Real Time Operation System (RTOS), and it has higher performance price ratio; the other is based on DSP, FPGA and LabVIEW, it has several interfaces, the ability of wireless communication, and it has on-line reconfigurable software functions.This dissertation finally realizes EMD and step-changed Stochastic Resonance on SoC and FPGA, respectively. The realization of those two algorithms lays a good foundation for engineering application. During the realization of EMD, new drop order decomposition is introduced to meet the limit of storage space in embedded system.