Optimization On Rotary Machinery Multi-fault Classification Based On HVD And RVM

With the increasing degree of automation of modern rotating machines,one device failure can cause great economic losses because of the large scale suspension.It's critical to recognize the fault type and puts forward solutions accordingly as soon as possible when a device is out of order.Turbine rotor is one of the most important power plant production equipment,when it comes to failure,it's urgent to identify the fault type based on vibration signals.This article studies in rotor common fault diagnosis and proposes a rotor vibration fault feature extraction method based on HVD method approximate entropy.Besides,it also optimizes the multi-class relevance vector machine system based on binary trees.First of all,an improved method is proposed for solving the false components problem of Hilbert vibration method.This method combines KL divergence,mutual information and correlation methods and cluster these indexes for false components automatic identification,which improves the usability of HVD.Secondly,fault signals feature extraction methods based on approximate entropy and HVD method has been studied.Vibration simulation data come from the Bentley rotor test bench.The discriminatory abilities of different fault feature extraction methods are examined through the difference between the different samples heart distance and the sum of each radius.Experimental results show that compared with approximate entropy and fuzzy entropy,approximate entropy has obvious advantages.Finally,the research results show that the normal binary tree structure of the multi-class system displays higher efficiency and lower time cost.Because positive and negative sample type selection in the classifier can severely affect the performance,a high dimensional distance metric index is proposed based on fractional norm.The index can effectively measure different samples separability.After optimizing the type selection of positive and negative samples based on this index,experiment results show that optimized system has remarkable improvement in classification accuracy.