Diagnosis Of Rolling Bearings Based On Feature Parameters Of Alpha Stable Distribution

Rolling bearings, as one of the most commonly used parts of machinery, whose running status is a key factor to decide whether a machine is well-performing, so the importance of performance monitoring and fault diagnosis of rolling bearings is self–evident. When there is pitting in rolling bearings, or other failures, it's very clear that there would be fault pulse in sampling signal. However, in the early stage, the amplitude of fault pulse is so small that it is always covered by sampling signal noise; in the late stage, the amplitude of fault pulse becomes stronger and stronger, and the sampling signal itself will clearly show the pulse characteristics.This article starts from the angle of statistical signal processing; by analyzing the non-Gaussian characteristics and the pulse characteristics in simulation fault signal of rolling bearings as well as the actual fault signal, this article introduces the concept of Alpha stable distribution: a statistical model that can be used to describe the signals with obvious pulse feature. As a generalized Gaussian distribution model, Alpha stable distribution model is more accurate and reasonable in fitting probability density distribution of the fault signal of rolling bearings.By the mathematical expression of simulation bearing fault signal, this article analyze some key factors that can effect fault signals of rolling bearings, such as bearing fault level, and how can they effect the feature indexαof Alpha stable distribution and kurtosis. Then the author concluded that theαindex and kurtosis just reflex the pulse feature of signals, they have no direct relation with the actual fault pulse strength, this article also compared advantages and disadvantages of theαindex and kurtosis in the performance evaluation of rolling bearings, then proposed a new evaluation parameter based on Alpha stable distribution, and through simulation and experimental analysis, this article also compared it with theαindex and kurtosis, then find out that it has a higher level of sensitivity to the fault pulse and the sensitivity can be adjusted according to different conditions.Based on the feature parameters of Alpha stable distribution, this article also proposed an early-staged fault signal detection method combined with self-adapted wavelet and a fault classification method combined with neural network. Analysis of the experiment data indicates a satisfied result of using combination of modern signal processing methods and Alpha stable distribution.