Research On Fault Diagnosis Method Of Rolling Bearings Under Variable Load Based On Deep Learning

Rolling bearing is the key component of many rotating machinery,and its running state is influenced by many factors.The load of the rolling bearing is often changed in the actual work,and the change of the load will directly affect the change of the vibration features.Therefore,under variable load condition,if the running state of the rolling bearing and be accurately identified,it will be of great significance for ensuring the safe running of mechanical systems.Two fault diagnosis methods for rolling bearings under variable load are proposed in this research.(1)Based on ensemble empirical mode decomposition-Hilbert(EEMD-Hilbert)envelope spectrum and deep belief network(DBN),a fault diagnosis method for rolling bearings under variable load is proposed.At first,the vibration signals of each state of the rolling bearing are decomposed by EEMD.Then,the sensitive intrinsic mode function(IMF)components of each signal are selected and their envelope spectrum can be obtained using Hilbert transform.At last,the new high-dimensional data can be constructed by IMF envelope spectrum of each state vibration signal according to certain order,and then the new high-dimensional data are used as the input of DBN,whose each hidden layer node structure had been optimized using genetic algorithm,and multiple-state diagnosis of the rolling bearing under variable load can be achieved.The experimental results show that,if training data choose a certain load and testing data choose another load,the multiple-state features of the rolling bearing under different loads can be reflected better by using EEMD-Hilbert envelope spectrum than the time domain or frequency domain amplitude spectrum.And,comparing with the shallow learning support vector machine and BP neural network algorithm,DBN has a higher recognition rate and the recognition rate of each data set can reached more than 92.50%.However,the above methods still have complex signal processing process,and the ability of deep learning to extract deep features in high-dimensional data needs further improvement.Therefore,another fault diagnosis method is further proposed.(2)Based on convolution gauss deep belief network(CGDBN)and weighted mixed kernel joint distribution adaptation(WKJDA),a fault diagnosis method for rolling bearings under variable load is proposed.Firstly,the frequency domain amplitude spectrum signals of the rolling bearing are extracted features using CGDBN,and the deep generalized feature of the rolling bearing can be obtained.Then,the domain adaptation WKJDA algorithm is used to migrate the data features of the source domain and the target domain,and the difference between the two domains becomes smaller.Finally,the multiple-state recognition of the rolling bearing under variable load is realized by the K nearest neighbor(KNN)algorithm.The experimental results show that,using the same data set as the method(1),the deep generalized feature which extracted by the CGDBN has higher classification accuracy than using DBN,and the diagnosis method combining CGDBN and WKJDA can effectively overcome the disturbance of multiple-state recognition,the recognition rate of each data can reach more than 95.23%.