Prediction And Optimization Of Diesel Engine Performance Based On Machine Learning And Intelligent Algorithms

The development of marine diesel engine is in an era of opportunity and challenge,under the background of increasingly stringent emission regulations and the requirements of tough market,much more new technology were proposed and applied to improve the emission and performance of marine diesel engine,which has already shown its improvement.However,the introduction of new technology must come along with more controllable parameters and optimization targets,which would make the diesel engine more complicated.Moreover,this would also make the prediction and multi-objective optimization problem of diesel engine much more challenge.Thus,the traditional responses prediction and optimization method have been unable to satisfy the new diesel engine research.Meanwhile,the rise of machine learning and intelligent optimization method provides a new direction for prediction and optimization of diesel engine.This research introduced machine learning for the responses prediction of a marine high speed common rail diesel and made a comprehensive study for its predictive performance.Then,optimization study of two machine learning methods was conducted using intelligent optimization approach.In addition,in order to develop an approach for the multi-objective optimization,an accurate phenomenological simulation model of diesel engine was developed firstly,based on this simulation model,an efficient and accurate multi-objective optimization approach was developed using machine learning and intelligent optimization method,the optimal operating parameters of engine were determined using this optimization method.This optimization method provides a new way for engine parameters calibration and optimization.The specific contents and results of this research as follows:Firstly,the engine bench test was conducted,which covered 10%-100% load conditions of the propeller characteristics and contained 126 operating points.In addition,in order to conduct a multi-objective approach study for diesel engine,an accurate phenomenological simulation model of diesel engine was developed.This model made a comprehensive modeling of diesel engine fuel system,intake and exhaust system,turbocharger,governor,in-cylinder combustion process.After the calibration,60 cases were used to validation,the results confirmed the accuracy and robustness of the simulation model,and the reliability of using this model for the optimization approach study has also been confirmed.Then,based on the experimental data,two most widely applied machine learning methods in the engineering area,artificial neural network(ANN)and support vector machine(SVM),was investigated when applied to diesel engine responses prediction.The predictive performance affected by network’s structure,initial weights and threshold and the division of training samples was statistically analyzed,the results showed that the performance of ANN varies with the responses willing to predict.In general,the predictive performance had a certain regularity according to the network’s structure,but the effects of initial weights and threshold,the division of training samples were randomness.On the other hand,although there is some theoretical guidance about how the kernel function parameter and penalty factor affect the predictive performance of SVM,when comes to each response willing to predict,the effects should be studied specifically.In addition,based on the investigated results of machine learning methods,genetic algorithm and artificial fish swarm algorithm was employed to optimize these machine learning methods respectively.A novel optimization approach aiming at optimizing the division of training samples for ANN was developed,and it was compared to the network’s initial weights and threshold optimization approach,then a comprehensive optimization approach integrated these two approaches was conducted.The optimized results of the novel and comprehensive approach both showed an excellent performance.On the other hand,when artificial fish swarm algorithm was employed for SVM,its perfect efficiency and accuracy were confirmed.Finally,based on the study of machine learning and diesel phenomenological model,non-dominated sorting genetic algorithm(NSGA-II)was employed for the development of diesel engine multi-objective optimization approach,the optimization approach integrated the accuracy of phenomenological model and the efficiency of SVM.Moreover,enhancing training method was developed in order to ensure the accuracy of optimization results.Eventually,the optimization approach aiming at the Pareto optimal solutions of NOx and Soot emission was conducted and validated,and the optimization approach has shown an excellent performance.