Research On Composite Fault Diagnosis Method Of Rotating Machinery Based On Deep Convolution Neural Network

Bearing-rotor system failures of rotating machinery occur frequently,and the presence of a single fault is likely to induce other faults.The failures of bearing and rotor in rotating machinery account for a large proportion,while bearing-rotor composite failures also occur.If failure is allowed to develop arbitrarily,it will cause irreparable loss of life and property.Therefore,the research of fault diagnosis methods for rotating machinery is of great significance.The condition monitoring of the equipment can make a reasonable plan for the maintenance of the equipment,which could not only avoid the situation of "insufficient maintenance" and "excessive maintenance",but also provides a reference for the maintenance decision.The fault diagnosis method of rotating machinery based on signal processing usually uses vibration acceleration as an important research carrier,and the collected signal components are very complicated,and there are often many interference components.It is difficult to obtain useful information from the original signal without pre-processing.In response to these problems,this paper proposes a feature extraction method combining instantaneous frequency mean curve and VMD-EES.Firstly,simulating bearing fault signal based on bearing vibration model.A set of IMFs was obtained by VMD decomposition,it was found that there were still many noise components in IMFs' envelope spectrum,so the fault characteristics were still not obvious.The autocorrelation function and extending the Shannon entropy operation in the envelope spectrum can not only effectively reduce the noise that cannot be eliminated,but also enhance the fault characteristics.In order to verify the effectiveness of the proposed method,the fault characteristics were extracted successfully from the simulating signal and the SKF6203 rolling element weak fault test signal.And compared with the method of the EMD-EES,it is found that the IMFs obtained by the VMD decomposition has a reasonable center frequency and bandwidth,and the fault information is relatively concentrated.Due to the coupling mechanism of bearing-rotor composite faults is not sufficient,so the existing time domain and frequency domain statistical indicators cannot be used to effectively reflect the fault state.In particular,the manual extraction of frequency domain features lacks theoretical guidance,and in-depth learning does not require prior knowledge.Therefore,this paper proposed a method combining Hilbert envelope spectrum and DCNN to realize multi-level,multi-fusion,multi-scale and differentiated fault feature extraction and identification process for fault feature construction.The average identification accuracy of a single fault of the rolling bearing in CRWU electrical laboratory reached 100%.At the same time,this method has been used to identify the N205 rolling bearing-rotor composite fault(7 kinds of equipment status including bearing-rotor composite failure and normal conditions),and the average recognition accuracy is 90.11%.Through detailed research and visual representation of feature distribution,it was found that the characteristic of complex faults overlapped to some extent,which resulted in the unbalanced distribution of misdiagnosis in Hilbert envelope spectrum and DCNN method and seriously affected the generalization ability of the model.Aiming at the misdiagnosis distribution imbalance problem of the above model and the inherent shortcomings of DCNN network,this paper combines the feature extraction ability of deep learning with the classification ability of SVM in small sample space,and proposes a multi-domain and multi-input semi-DCNN rotating mechanical compound fault.This method combines the clear theoretical basis time-domain index with the frequency-domain index extracted by DCNN to identify the fault types,it forms a diagnosis mechanism that combines human successful experience with machine in-depth learning.In N205 rolling bearing-rotor experiment,16 frequency domain features extracted from envelope spectrum by DCNN are dimensionally reduced by PCA,and then combined with peak value and kurtosis index to form feature set,which is fed into PSO-SVM.The experimental result shows that multi-domain?multi-input semi-DCNN effectively solves the problem of DCNN misdiagnosis distribution unbalanced.The recognition accuracy of the rolling element and rotor composite faults was increased by 32.43% and 33.80%,respectively,and the overall accuracy of the overall recognition was increased to 99.16%.