Research On Fault Diagnosis Of Marine Diesel Engine Based On Improved Multi-scale CNN

With the rapid development of computer technology,the fault diagnosis technology of mechanical equipment is becoming more and more automatic and intelligent.The concept of "intelligent engine room" proposed by China Classification Society in the code for intelligent ships will promote the rapid development of intelligent diagnosis technology of ship equipment.As the main power source of ships,the failure of diesel engine and its system will endanger personnel safety and cause economic losses.Moreover,the diesel engine has a complex body structure,and faults occur frequently in harsh working environments such as high temperature,humidity and severe vibration.It is of great significance to make full use of artificial intelligence technology to diagnose diesel engine faults quickly and accurately.Traditional intelligent diagnosis methods rely on artificial selection of features and expert knowledge base.It is difficult to extract the features of vibration signals,which is difficult to meet the development needs of intelligent fault diagnosis.With excellent learning ability,the fault diagnosis method based on deep learning can adaptively extract features from vibration signals containing rich information,so as to complete fault diagnosis efficiently and accurately.Taking the vibration signal of diesel engine as the experimental data,this paper studies the fault diagnosis of diesel engine under strong noise and variable working conditions by using the method of deep learning.The main research work of this paper is summarized as follows:(1)This paper expounds the composition,structure and working principle of diesel engine,and probes into the main fault modes and characteristics of diesel engine.The fault tree analysis method is used to analyze the causes of fuel system faults with high incidence of faults.By analyzing the excitation source of vibration signal and its transmission path in the body,the position of measuring point is determined.(2)Aiming at the random error caused by equipment defects and environmental interference,the Least Square Method and Five-spot Triple Smoothing Method are used to preprocess the vibration signal.Explore the principle of Convolution Neural Network model,improve the gradient dispersion and over fitting problems in CNN model by using Batch Normalization and Dropout algorithm,and build a Onedimensional Convolution Neural Network diagnosis model to study the fault diagnosis of one-dimensional vibration signal.(3)Aiming at the problems of slow training speed of the model and the inability of a single size convolution kernel to completely extract the characteristic information in the vibration signal,the optimal optimization algorithm is selected to improve the model,and a fault diagnosis model based on Multi-scale One-dimensional Convolution Neural Network is proposed.Through comparative experiments,it is verified that the Adabound optimizer can combine the advantages of Adam optimizer and SGD optimizer,improve the training speed while maintaining high accuracy,and the multi-scale feature fusion framework can extract the complementary features of different scales in the vibration signal.Set up comparative experiments to verify the recognition accuracy and stability of the proposed Multi-scale One-dimensional Convolutional Neural Network(MS-1DCNN)model compared with other commonly used diagnostic models.(4)Aiming at the problems of strong noise interference and variable working conditions in the actual working process of diesel engine,a fault diagnosis model based on Multi-scale Improved Residual Network is proposed.Firstly,Dilated Convolution and Residual Block with jump connection line are used to improve the classical Residual Block,and then two attention mechanism modules,Squeeze and Excitation Block,Convolution Block Attention Module are introduced.They are embedded into the improved Multi-scale One-dimensional Residual Network model respectively,and two Multi-scale Attention One-dimensional Dilated Residual Network(MSA-1DDRN)diagnosis models are constructed.The optimal void ratio and connection coefficient of the model are obtained through experiments,and the evaluation index is introduced to evaluate the diagnostic performance of the model.Finally,the anti noise and variable condition experiments are set to verify that the proposed model still has good stability and diagnostic performance under the above conditions.