Research On Rolling Bearing Fault Diagnosis Of Petrochemical Units Based On BP Neural Networks And D-S Evidence Theory

In industrial petrochemical factories,the structure and composition of rotating systems have been becoming more and more complex.At present,petrochemical units increasingly tend to be large-scale,high-speed,and automated.In the field of fault diagnosis of rolling bearings,advanced signal processing,pattern recognition,artificial intelligence techniques have greatly improved the stability and safety of industrial systems;however,due to the high complex in industrial environment,there still exists misdiagnosis and low accuracy factors.Especially,in rotating machinery,rolling bearings are the most widely applied components as well as the easily faulty components.The performance deterioration of rolling bearings will not only lower down the stability of the whole system,but also introduce the unexpected downtime and personnel injury,which results in negative influence on the whole manufacturing process.Hence,the study of rolling bearing fault diagnosis in petrochemical systems is able to detect,pinpoint,and eliminate potential failures,which have great practical and theoretical significance.This paper takes the rolling bearing of petrochemical unit as the research object,1)introducing the vibration characteristics of the rolling bearing,the failure mechanism,and the single-point damage fault diagnosis theoretical model;2)analyzing the application of wavelet transform,BP neural network,DS evidence theory and other methods;3)studying the four commonly seen bearing failures in petrochemical rotating machinery and meanwhile providing a theoretical model for rolling bearing fault diagnosis.Considering the inaccuracy and misdiagnosis in fault diagnosis of roiling bearings,this paper proposed a bearing diagnosis approach based on BP neural networks and D-S evidence theory.For fault diagnosis of rotating machinery bearings,there are problems such as indiscernibility and low accuracy.In the proposed method,D-S evidence theory is applied to fuse the results obtained respectively from time-domain and frequency-domain analysis using wavelet analysis as signal processing method and BP neural networks as classifiers.Specifically,the obtained bearing signals are first decomposed and reconstructed by applying wavelet packet decomposition,and the first four wavelet coefficients are selected to be reconstructed.Then,seven time-domain features are extracted from each reconstructed wavelet coefficient,which are fed into BP neural network for classification.Afterwards,seven frequency-domain features are extracted from coefficient and then are also fed into BP neural network for fault recognition.Based on the obtained probability outputs from BP neural networks,bass functions are obtained by normalizing these outputs,which are used for information fusion using D-S evidence theory.In order to prove the effectiveness of this proposed method,different conditions of bearing data are analyzed contributed by the Engineering Laboratory of Case Western Reserve University and the Key Laboratory of Fault Diagnosis of Petrochemical Equipment in Guangdong Province respectively.The experimental results have demonstrated that the proposed method is efficient in detecting and classifying different states of rolling bearings with a higher classification accuracy rate than the use of time-domain and frequency-domain analysis individually.