Research On Intelligent Diagnosis Methods For Typical Diesel Engine Faults Based On Analytic Single Channel Vibration Signals

As a kind of general power machinery,diesel engines are widely used in industry,agriculture,national defense,transportation,etc.However,diesel engine malfunctions now and again,due to its complex structure and harsh working environment.To ensure the safety and reliability of diesel engine and mechanical system using diesel engine as power source,and to reduce the loss caused by sudden failure,diesel engine health condition assessment and non-disassembly fault diagnosis have become one of the leading research topics at the present stage.Although the vibration signal of diesel engine contains abundant state information,it cannot be directly used to assess health condition and diagnose fault because of the changeable working conditions and the relatively weak power of the characteristic signals in early stage of engine fault.Therefore,further analyzing of engine vibration signal and extracting significant parameters which can characterize the operating state become the key issues in the research of fault diagnosis.Base on the practical problems of diesel engine faults detection,this thesis took valve clearance and fuel injection faults with high probability as objects and presented a new fault diagnosis method on the single channel vibration response signal.The main achievements are as follows:(1)This thesis proposed a collaborative signal processing method using optimized Variational Mode Decomposition(VMD)and Independent Component Analysis(ICA)to quickly and accurately detect multiple independent source signals through a single channel vibration signal.Firstly,this thesis took advantage of VMD to decompose single channel signal and obtained several Intrinsic Mode Functions(IMFs).This algorithm has obvious advantages in accuracy comparing to recursive algorithm such as Empirical Mode Decomposition(EMD).Secondly,this thesis utilized ICA to process the results of VMD in order to solve the problem that VMD cannot decompose the signals with the same frequency but different source.In the process of VMD analysis,this paper optimized the decomposition level K and the quadratic penalty term ? to get the best combination of parameters.The analysis result showed that ?=8400,K=6 is the best combination in this test.By comparing and analyzing a variety of ICA algorithms,including Fast Independent Component Analysis(Fast ICA),Robust Independent Component Analysis(Robust ICA)and Kernel Independent Component Analysis(KICA)in accuracy and computational efficiency of the secondary processing of VDM results,this thesis found that the Robust ICA had relatively good comprehensive performance.After above optimization and comparison,the collaboration-based method obtained expected results.(2)This thesis proposed an optimized Bispectrum diagonal projection analysis method for fault feature extraction.This method overcomes the disadvantage that Bispectrum diagonal slice method would lose a large amount of significant information of Bispectrum 3D-results so that the signal characteristics cannot be fully reflected.After the decomposition of the signal,this thesis proposed two feature extraction methods based on Bispectrum,diagonal projection and diagonal accumulation.By comparison,the diagonal projection method can obtain more comprehensive characteristic information.Based on the collaborative VMD-ICA method and diagonal projection method,comparing and analyzing 10 characteristic parameters,including mean,variance,skewness,kurtosis,peak-to-peak value,square root amplitude,root mean square,Shannon entropy,maximum singular value,fourth-order cumulant,this thesis proposed a visualized method to get the priority of the characteristic parameters,which shows the within-class scatter and within-out scatter simultaneously.The 10 characteristic parameters sorted by this method provides a selection basis of the feature parameters for fault mode intelligent recognition.(3)For distinguishing multiple faults,this thesis designed a multiple-sample classifier based on Deep Belief Network(DBN).The four highest priority parameters are used as the input of the classifier,and then the working state of the diesel engine was identified.Firstly,instead of original signals,this thesis input the characteristic parameters to compress the structure of DBN and improve the diagnostic efficiency.Next,this thesis built the DBN to recognize six working states(five fault states and one normal state)of valve clearance faults and fuel injection faults.Lastly,comparing with the genetic algorithm-BP neural network(GA-BP)results show that the accuracy of single fault diagnosis and the accuracy of multiple fault diagnosis are significantly improved,and the accuracy of multi-fault identification is improved by 87% rose to 95.3%,which proved the superiority of the classifier designed in the paper.In conclusion,based on VMD,ICA,Bispectrum,and DBN algorithms,through a lot of optimization and comparative analysis,this thesis established an accurate and efficient method for engine fault diagnosis with single channel vibration signal.The results showed that this method works well in the diagnosis of early engine faults and can be applied to different working conditions.Although this method was established for typical faults of diesel engine,it is appropriate for most mechanical structures because of its certain universality,and has guiding significance for technology development and engineering application of mechanical mechanism health evaluation and early fault diagnosis.