Research Of Engine State Detection Based On Vibration Data Driven Method

The engine plays an irreplaceable role in industrial and agricultural productions as a power source.The engine is developing to integration under the increasingly severe energy and environmental crisis,but the following complexity brings a high frequency of failures.Among fault detection methods,condition-based maintenance driven by vibration data is potential in current research because of its simple structure and low cost.However,the engine vibration signal has strong nonlinearity and randomness,and the background noise easily covers the weak features caused by early faults.Therefore,appropriate algorithms should be researched to recognize faults accurately and efficiently.This paper takes a diesel engine as the research object to select typical abnormal fuel injection quantity,abnormal fuel injection advance angle,abnormal rail pressure,and valve clearance fault for the bench experiment under various working conditions.The signal decomposition,data dimensionality reduction,pattern recognition,and time series prediction algorithms are optimized deeply based on collected vibration signals.The main research contents are shown as follows:(1)The framework of the variational mode decomposition(VMD)is optimized to propose a recursive VMD(RVMD)algorithm.The decomposition level of the VMD should be selected manually,which is hard to decompose complex vibration signals collected in various working conditions under typical faults.Therefore,the VMD is implemented in the recursive frame to decompose signals level-by-level and controlled by stop condition based on the energy difference tracking method.Then,a dynamic quadratic penalty is proposed by analyzing gradients of the residual signal energy to match the recursive frame.The comparison shows that the proposed RVMD can extract feature components with little noise adaptively and efficiently.(2)The multi-verse optimizer(MVO),the semi-supervised local Fisher discriminant analysis(SELF),and the echo state network(ESN)are researched to propose a stepby-step typical engine faults detection method under various working conditions.An improved MVO(IMVO)is developed by optimizing the traveling distance rate(TDR)and accelerated search mechanism.Combined with the IVMO,the SELF based on the nearest neighbor(NNSELF)is proposed to select labeled samples appropriately.A multilevel structure is designed to optimize its performance in significant dimensionality reduction of the big-scale data.The ESN with multiple output matrices(MOM-ESN)is built for recognizing complex data.Comprehensively,the vibration signal is decomposed by the RVMD,and 24-dimensional feature set is extracted.The MOM-ESN is employed to analyze the feature set compressed by the multilevel NNSELF for fault detection,and a recognition rate of 91.3% is obtained.Results show that the proposed method has higher accuracy and efficiency than traditional algorithms.(3)The fixed convolution kernel ESN(FCK-ESN)is researched to propose an endto-end typical engine faults detection method under various working conditions.Based on the comparison among several time-frequency representation algorithms,the bispectrum is selected as the pre-treatment to show more information about the fault feature in the vibration signal.In the FCK-ESN,the fixed convolution kernel is designed by edge detection operators and the Gaussian low pass filter,the autoencoder(AE)is employed to improve generalization and reduce the data dimensionality further,and the analyzed data with three time-steps is finally recognized by the ESN.Compared with traditional algorithms,the proposed FCK-ESN obtains the best result of 92.3%and shows ideal stability.(4)An improved ESN(IESN)is researched for engine vibration trend long-term prediction.A condition of the echo state property(ESP)related to the Lipschitz constant of reservoir activation function and the maximum structured singular value of reservoir is firstly researched.A smooth composite reservoir activation function is then designed,and its Lipschitz constant is solved.Finally,a reservoir structure with eigenvalues distributing uniformly is built by the decoupling method.The result shows that the IESN can predict the vibration signal intensity in the next five time-steps and has better accuracy and long-term prediction performance than traditional methods.This paper comprehensively researches the VMD,SELF,and ESN to propose a stepby-step method and an end-to-end method for typical faults detection in various working conditions and a prediction algorithm for the vibration trend.A methodology system for engine state detection is built.The analyzing results based on experimental data show that the proposed methods have advantages both in accuracy and efficiency.The study has certain instructional significance for engine health evaluation theoretical research and is expected to apply in engineering practice.