Research On Mechanical Early Weak Signal Extraction And Fault Diagnosis

For incipient fault diagnosis of mechanical equipment, fault feature extraction is the most important and difficult task. Due to environmental impact, the measured signal is usually disturbed by strong noise which weakens or overwhelms the data features, and makes the information about equipment state hard to be obtained. It has been a hotspot that how to extract fault features from background noise. For sim, ple rotating machines, an algorithm based on EMD and Stochastic Resonance is presented by the studies of Morphological filtering and SVD denoising, and it proves to be useful through bearing fault diagnosis. For complex rotating machines, ICA is applied in this paper because of their multi signal sources, and this method proves to be useful through gearbox fault diagnosis. But for those non-rotating machines, such as cutting tool, the method based on detection of peculiar frequency is going to fail. So, in this paper, an intelligent diagnostic method combined with weak features extraction and improved b-spline fuzzy neural network is presented, and it is effectively used in tool wear monitoring. The main contributions are described as follows:The much de-noise method combined with empirical mode decomposition (EMD) were adopt for bearing fault diagnosis. Empirical mode decomposition is a self-adaptive time-frequency analysis method. The end effect of the EMD will be more serious and the quality of the decomposition will be worse when the fault features are extracted with EMD in heavy noise. So, this paper proposed a new wear fault feature extraction method based on morphological filter-singular value decomposition (SVD) and EMD. Firstly, the original signals are filtered by morphology filter and then de-noised by SVD. Finally, signals are decomposed by EMD into the intrinsic mode functions (IMFs) for fault feature extraction. The analysis results of the simulation and bearing fault data show that this method can improve the effectivity of EMD and extract weak fault features effectively.Stochastic resonance (SR) enhances the SNR of signals by noise, so it has an advantage to detect weak signals fault feature with a heavy noise. The bearing fault diagnosis experiment shows that the SNR is higher and the fault identification is more efficient by it when the failure frequency is close to the noise band. Traditional adiabatic elimination stochastic resonance in small parameters is not adapt to engineering weak signal detection in large parameters, so a new method based on frequency-shifted and re-scaling was proposed for weak signal detection in large parameters. At the same time, considering the deficiency of the single-parameter optimization in traditional stochastic resonance and in order to achieve the best stochastic resonance parameters, an adaptive stochastic resonance method combining with genetic algorithm is proposed to optimize the parameters of stochastic resonance system while the fitness function is the SNR, the output of the bistable system.Independent component analysis (ICA) is an effective method for weak signal detection for complex rotating machinery, which can extract isolated component from mixed signal. For single-channel ICA, a two-dimensional signal matrix is constructed by introducing virtual noise into the collected signal. Thus the underdetermined problem is solved and Fast ICA can be applied for signal denoising and feature enhancement. Furthermore, blind deconvolution based on signal in frequency domain is proposed for multi-channel blind source separation. By gearbox fault detection, the data shows that this method can be effectively uses for machine fault diagnosis.Tool wear monitoring is non-rotating, so it can't directly from the failure frequency to determine the fault. But it can be do by b-spline fuzzy neural network. In learning algorithm for fuzzy neural network with b-spline basis function, adaptive learning algorithm is often adopted with network parameters be calibrated according to the experience and it is easy to fall into local minimum during the learning process. So, genetic algorithms (GA) are adopted for global optimization. At the same time, it can improve the equipment fault diagnosis using the EMD, SR and Fast ICA to extract multi-parameter of tool fault feature. The tool wear diagnosis results show that the proposed method is effective.There are the summarization and expectation of the fault feature extraction technology development in the end of paper.