Research On Intelligent Fault Diagnosis And Prediction Method For Large Engine

Engine is a kind of power machinery widely used in industrial and agricultural production,social life and military equipment.Its running state directly affects the working performance of the whole mechanical equipment.Therefore,it is of great theoretical and applied significance to study the prediction and diagnosis of engine fault.12150 ZL engine is a kind of large one which is mainly used for heavy special machinery and third generation armored combat readiness.It has the characteristics of large tonnage,complex structure,large dynamic load and high lifting power,but the frequency that “three mechanisms and four systems” are out of order is very high.Thus,it is tricky to ensure the high stability and reliability of the engine in the given working conditions.Affected by nonlinear factors such as dynamic load,contact force,clearance and stiffness,the measured vibration signals of large-scale engines present nonlinear and non-stationary mixed signals under different operating conditions.Taking 12150 ZL engine as the research object,the thesis,by studying feature extraction,diagnosis,recognition and prediction of fault signals,solved the problems of low accuracy of fault signal extraction and fault diagnosis and inaccuracy of fault recognition.The thesis presents a fault signal feature extraction method based on deep migration compressed sensing.Firstly,the deep migration learning and sparse self coding network model are studied to reduce the difficulty of training when the fault samples are multi-source and heterogeneous,and solves the problems of the lack of early fault training samples,and solve the problems of long time and poor stability of feature learning in each layer of the network.Secondly,the SL0 compressed sensing method is studied to measure the fault signal sparsity by solving a nonlinear optimization problem,the signal is reconstructed and the accurate classification result is output.Finally,the DTLNet-SL0 feature extraction method is proposed to improve the efficiency and accuracy when the fault samples are multi-source and heterogeneous.This method has higher efficiency and accuracy of feature extraction compared with other traditional methods such as Entropy-SVM,Wavelet-SVM and BPNet.This thesis also proposes a fault diagnosis method based on time series fusion network.This thesis studies the fault diagnosis method of convolution neural network model and bidirectional long-term and short-term memory network fusion,the thesis puts forward to CNN-Bi LSTM fusion method,establishes the feature correlation in a specific time window,and realizes the goal of long-term and short-term memory of time-series features.It solves the common problems of “three mechanisms and four systems” of large-scale engines,such as many kinds of faults,large fault signal data and the poor effect of fault diagnosis when unstable fault characteristics of different time sequences occur.The experimental results show that the accuracy of CNNBi LSTM method is over 96%,which is higher than that before fusion,and achieves better fault diagnosis effect.What’s more,this paper presents a fault recognition method based on Alex Net and kernel optimized SVM.When the fault sample data is large,and the traditional fault recognition methods have the problems of high computational cost and low recognition accuracy,the methods are proposed,which are the alexnet neural network training based on Re LU Alex Net and improved SVM of HPSO-PDT core.A multi-classifier,suitable for large-scale engine fault recognition,is constructed,which sends the feature vectors trained by Alex Net to HPSO-PDT-SVM for classification.Through the experimental comparison,the fault recognition method based on Alex Net and kernel optimization has higher accuracy than TW-SVM,PSO-SVM and HPSO-PDT-SVM,and the fault recognition accuracy is over 95%,which verifies that the fault recognition method based on Alex Net-SVM has better effect and improves the fault signal recognition accuracy.This thesis constructs a fault prediction model based on SAPSO big data feature attribute selection.Due to the time lag and low accuracy of traditional engine fault prediction methods,it is difficult to predict engine fault.Firstly,the feature attributes of 12150 ZL engine condition monitoring and health assessment big data are analyzed in the paper,the fuzzy constraint model of fault signal feature attributes and the method of attribute selection based on particle swarm optimization are proposed,and the fault prediction model of big data attribute selection based on adaptive particle swarm optimization is constructed.Through the experimental comparison,the accuracy of this method is improved by about 16.5% on average compared with the traditional feature attribute selection method,which verifies the fault prediction model based on SAPSO big data attribute selection,and realizes the condition based maintenance and autonomous targeted maintenance of large engine.The methods in this thesis proposed in this thesis can improve the fault diagnosis and prediction effect of large-scale engine with variable working conditions and heterogeneous signals,expand the engine intelligent fault diagnosis method,and improve the engine fault diagnosis technology and level,which has good theoretical research significance and high engineering application value.