Research On Fusion Of Multiple Methods For Diesel Engine Fault Diagnosis

As the main power supply,diesel engine is widely used in our life.Due to the complicated structure and the poor working conditions,the possibility of failure is greatly enhanced.It is necessary to monitor and diagnose the condition of diesel engine to detect the faults as soon as possible and to guarantee the safety and reliability of whole power system.In this paper,4120 SG diesel engine is used as the object.The methods of features extraction,pattern recognition and fusion diagnosis are used to accurately diagnose diesel engine faults.The main contents include:(1)The vibration characteristics of cylinder head system are investigated.Time and frequency domain analysis,wavelet packet analysis,ensemble empirical mode decomposition(EEMD)and singular value decomposition(SVD)method are used to extract 3 kinds of features under normal condition,misfire,abnormal valve clearance,valve leakage and abnormal fuel supply advance angle.These kinds of features are time and frequency domain features,energy ratio of wavelet packet decomposition,singular value of EEMD decomposition and SVD decomposition.The features from fault states are compared with normal condition,and the variation regulation is analyzed.(2)The clustering parameters of fuzzy C mean(FCM)clustering are set,and the characteristic matrix which is composed of singular value feature vector is used as the input of FCM clustering algorithm after normalization.The iterative operation is performed and membership matrix and clustering center are updated in every iteration step.When meets the stop condition,iteration process is completed and standard model of fault diagnosis is obtained.20 groups of samples are used as testing samples and fault pattern recognition is achieved by calculating the Hamming nearness degree between the standard model and the testing samples.The results show that the FCM clustering algorithm can classify the test samples accurately.(3)Various training methods of BP neural network are compared,and the Levenberg-Marquardt method which has fastest convergence rate and least number of iteration is chosen to recognize fault pattern.3 neural networks are constructed to identify the samples which are composed of 3 kinds of features,identification results are obtained.The output of neural network is transformed into evidence body based on DS evidence theory and the basic probability assignment is obtained.The basic probability assignment and the confidence interval of 3 evidence bodies in joint action are calculated and diagnostic conclusions are drawn.In order to verify the feasibility of this method,the fusion diagnosis tests are carried out.The results show that,the value of basic probability assignment was increased and the uncertainty degree was reduced after DS synthesis,the results of 3 evidence bodies’ synthesis are better than the results of 2 evidence bodies’ synthesis.(4)For different kinds of operation state,100 working cycles are intercepted respectively.The cylinder head vibration signal in the circulation is decomposed by wavelet packet and energy feature vectors are obtained.The energy feature vectors are all divided into training set and testing set.The deep belief network(DBN)is constructed and the parameters are initialized.Training set is used to train DBN,and after the training process is completed,DBN test was performed using the testing set.With the change of DBN parameters,the variation of recognition error rate is observed and the rules for setting DBN parameters are obtained.