Research On Fault Diagnosis Of Rotating Machinery Method Based On Machine Learning

Rotating mechanical equipment has been widely used in metallurgy,aviation,transportation,chemical,energy and other industries,with the development of maximization,heavy type,complication,precision and high speed.It is important to monitor the status and diagnose the fault of the equipment so as to ensure its safe operation.In recent years,signal processing and machine learning have been improved and applied in many fields,which provide solid foundations for fault-feature extraction and intelligent fault diagnosis based on mechanical vibration signals.This thesis makes an in-depth study on the early fault recognition and diagnosis of the bearing,a key and vulnerable component of the mechanical equipment.This work is a part of the national natural science foundation project "research on the early fault sparse feature recognition of low-speed heavy-duty machinery".The main contents are described as follows.The thesis proposed an adaptive Complementary Ensemble Empirical Mode Decomposition(CEEMD)and complete CEEMD method to detect the impact signal caused by the early bearing fault and to extract the feature frequency of the fault,aiming at solving the problem of early fault feature extraction of the bearing,Three inherent defects of empirical mode decomposition(EMD)are analyzed,including the end effect,overfitting/underfitting,and the mode mixing.Firstly,a mesh search algorithm based on least squares mutual information is embedded to the CEEMD to adaptively set the amplitude of the added white noise in the EMD decomposition,so as to adaptively suppress the modal mixing problem of the EMD.Then,on the basis of adaptive CEEMD,the homotopy least-squares support vector double regression and shape preserving piecewise cubic spline interpolation algorithm are proposed to suppress the end effect and the overfitting/underfitting problem of EMD,respectively.Finally,a complete CEEMD method is constructed by suppressing the inherent defects of EMD.By analyzing and processing the early fault signals of the bearing,this method isproved to be superior to other methods in detecting micro-impulse signals and extracting characteristic frequency of early faults.An early fault feature extraction method based on Teager energy operator and optimal variational mode decomposition(VMD)is proposed to solve the problem of early fault feature extraction for large-scale,low-speed and heavy-duty mechanical equipment.In the early failure of such an equipment,the impact component is extremely sparse and weak.So it is difficult to directly use the signal processing method for feature extraction.The impact components of the original vibration signals were enhanced by using the Teager energy operator.Then,the optimal VMD was constructed by using the variable-dimension chaotic pigeon group optimization algorithm,and the enhanced vibration signal was reconstructed and its envelope spectrum was extracted by using the optimal VMD method.Finally,the envelope spectrum that represents the fault characteristic is effectively extracted from the simulated signals and the measured signals.In addition,a diagnosis strategy that is based on the optimal VMD and relevance vector machine(RVM)is proposed to diagnose the fault types and damage degrees of the bearing while working under the variable load conditions.Firstly,the quantum chaotic mapping is introduced to extend and improve the fruit fly optimization algorithm(FOA),which effectively improves the global search ability and convergence speed of the standard FOA.Then,the improved FOA is used to search the optimal combination value of the key parameters of the VMD.Thus,an optimal VMD is presented.It is used to extract the two-dimensional marginal entropy of bearings that are working under variable load conditions.The two-dimensional marginal entropy is used as the training sample of RVM.Then,two different multi-classification extension strategies are adopted to improve the standard RVM: a multi-classification strategy named nested one-to-one for improving the multi-classification performance of RVM,and the other strategy named variational RVM for improving the multi-classification performance of the standard RVM.Finally,the fault signals of the bearings with different degrees of the faults and damage under variable load condition are collected.Their two-dimensional marginal spectral entropy is extracted by using the optimal VMD.One part is used as the learning sample set of the improved RVM,the other part is used to be tested.The experimental results show that different types of faults and their degrees of damage are effectively identified by the improved VMD.