Research On Multi-level Support Vector Regression Machine And Its Application In Converter Noise Prediction

Rail transportation is a kind of travel mode with high capacity,high speed and convenience,low energy consumption and less pollution.As the "heart" of the rail vehicle,the converter is also the main source of noise in the vehicle.The commonly used methods for noise analysis of converters are noise testing and numerical simulation,but they have the limitations of high professionalism,time and effort,and cannot get the noise response quickly.In recent years,data-driven proxy modeling techniques can greatly save the cost of noise testing and numerical simulation.With the advantages of simple construction and ease of use,proxy modeling techniques are widely used in modeling and optimization problems in industry.Support Vector Regression(SVR)is a popular proxy modeling technique with the characteristics of small sample requirement and suitable for non-linear problems.However,how to efficiently and accurately perform the agent model selection is another challenge before scholars.Therefore,based on the support vector machine regression study,this paper proposes a model selection mechanism and applies it to the variator noise prediction of rail vehicles.The results show that this method can predict the results more accurately than other commonly used SVR methods.The specific research of this paper is as follows.(1)Three commonly used support vector machine regression models are studied,and a multi-level SVR automatic model selection mechanism is proposed accordingly.Firstly,the mapping relationship between data features and SVR,kernel functions is constructed by the dataset of benchmark functions.Secondly,the data features of the training data of the solution problem are extracted and the combinations of SVR and kernel functions are screened.Finally,the relevant parameters are optimized by the particle swarm algorithm.The algorithm can "analyze" the problem and select the optimal data model adaptively to achieve higher accuracy modeling results.(2)An intelligent design software based on support vector machine regression models has been developed for rapid modeling and prediction of data table information.The software integrates the basic steps of the agent modeling technique: data analysis,model selection,model validation,and model approximation,into a software format.The "model recommendation" module,which is based on an automatic model selection algorithm,provides a more intelligent service to users.(3)A converter noise test was conducted,and the noise data obtained were used to assist in the converter noise reduction design.Firstly,the noise test was conducted,followed by analysis of the noise characteristics of the noise data,while simplifying the design variables to facilitate the construction of the noise model.Finally,the noise model is constructed based on the noise data,and then the test data can be predicted.In this paper,support vector machine regression theory is studied and the proposed model selection algorithm can be better applied to the prediction of converter noise,which is of great significance for extending to other engineering fields.