Research On Deep Learning Network Model For Fault Feature Extraction And Diagnosis Of Flexible Thin Wall Bearing

Harmonic drive is one of the common mechanisms in robot joints,and flexible thinwall bearing is a key component of harmonic drive.It is different from general bearings in that it has largescale radial deformation during motion,when the shape of flexible bearings is elliptical.This characteristic leads to the increase of failure probability of flexible bearings,and it will cause great harmfor robots with high transmission accuracy.Early detection of faults and diagnosis of fault types are very important for the reliability and safety of robotic systems.Aiming to solve this problem,fault diagnosis methods of flexible bearings is studied.Internal load distribution of flexible bearings is quite different from that of general bearings.In order to determine the load distribution law of flexible bearings after installation and adding load.Firstly,deflection mechanism of bearing under load is analyzed theoretically and relevant calculation formulas are established.Then,results of formulasare tested by finite element analysis to determine load distribution law of flexible bearings in a specific posture,and internal load of roller support point is verified.Then,in order to study the effect of fault feature based methods in fault diagnosis of flexible bearings,this paper compares the accuracy of time and frequency domain statistics,fault feature frequency,wavelet energy moment and the proposed singular value energy moment in fault diagnosis of flexible bearings.Singular value energy moment combines variational mode decomposition and singular value decomposition.Experimental results show that singular value energy moment has a good effect in bearing vibration signal with multiple rotational speeds.Fault diagnosis methods in the previous chapters all need to construct fault features,so fault feature based method will be affected by the relevant knowledge and experience of researchers,so it is difficult to guarantee the accuracy.In order to solve this problem,this paper studies how to achievefullpotential of deep learning,using unprocessed time-domain vibration signal as the input of network,compares between convolutional neural network,deep belief network,stacked sparse self-coding network and long short-term memory network are made,these four kinds of models are applied to fault diagnosis of flexible bearings.Next,in order to make further effortsin improving the accuracy of diagnosis,this paper synthesizes the advantages and experimental results of the four network models,and finally proposes the LCD network model for flexible thin-walled bearings.This model uses the structure of longshort-term memory network,convolutional neural network and deep belief network,and can be divided into two parts: LC feature extractor and DBN classifier.In order to train this new model,the corresponding training methodis designed.In the end,this paper analyses the learning effect of the new model,divides it into forward and backward researches,then a series of operation methods in backward research are proposed,andstudies of what the model learns at each layer and what evidence fault diagnosis bases on is given.