Research On Model Selection For Support Vector Regression And Its Application In Prediction Of Combined Dynamics Environment

Effective prediction of structural response always plays a important role in the fields ofaeronautics, astronautics, car-manufacturing and so on. One of the key questions is how toanalyze and simulate the true mechanical environment of structure. Because structures inworking condition are regularly excited by different kinds of loads, for example, vibration,shock, noise and so on, traditional numerical methods such as finite element method,statistical energy analysis, etc could not predict structural response effectively becausecombined dynamical environment is hard to reproduce. Moreover, direct experimentalmeasurement is hard to repeat many times because of the complexity of environment.There exists a mapping relationship between responses of one linear system under differentboundary conditions. Therefore, the structural vibration response under one boundarycondition can be predicted by the response under another condition. The key part ofenvironmental prediction based on mapping relationship is the predictive power for new data,i.e. generalization ability. Besides, considering the small sample size, this thesis utilizessupport vector regression(SVR) to establish mapping relationship. Aiming at improving thepredictive precision of mapping relationship model, this thesis studies the model selectionalgorithms of SVR. It is of great importance to improve generalization ability as well aspredictive performance of mapping relationship model. The main works and contributionsare listed as follows.1. To withdraw the model information from response data distribution, a weighted SVRsolution path algorithm is proposed. According to the sample’s importance, this algorithmsets different weight of error penalty parameter on each sample and optimizes the model bymeans of modifying the value of weights. Furthermore, a heuristic weight-settingoptimization algorithm is proposed to compute the optimal weights by using particle swarmoptimization(PSO). The idea of this algorithm is to transfer the primal problem to a globaloptimization problem with two variables. After solve this optimization, the optimal solutionpath model is determined. Experimental results show that the proposed algorithm canimprove the generalization ability of mapping relationship model effectively.2. Aiming at choosing the optimal SVR parameters, a new decimal-coding small-world optimization algorithm is proposed. This algorithm employs tabu search to construct localsearch operator and has good global convergence. Furthermore, based on the analysis ofeffectiveness of leave-one-out bound of LS-SVM on regression problems, a new modelselection algorithm based on small-world strategy is proposed for LS-SVM regression.Experimental results show that the proposed algorithm can avoiding premature of parametersand has better generalization ability than traditional methods.3. To improve the predictive performance from noisy experimental data, a new regressiontransfer learning algorithm based on principal curve is proposed from aspect of transferlearning. This algorithm utilizes non-parametric approach to seek the common-across-tasksrepresentation among multiple related regression tasks by computing the principal curve, andthen weights the samples by means of this curve. Numerical results demonstrate that thisalgorithm can withdraw the model information of mapping relationship from low noisy data,and improve the model performance from high noisy response data.4. Accoring to the transfer characteristic of multiple information sources, a newmultiple-input multiple-output(MIMO) modeling method is proposed based on SVR featureselection. This method treats the modeling procedure of mapping relationship as MIMOproblem, and reconstructs input feature by adopt whole-brand frequency or time responsedata. For eliminating the negative influence of redundant response data, the feature selectionalgorithm based on SVR generalization error bound is integrated into MIMO modelingprocedure. Experimental results show this method can improve the generalization abilityeffectively.5. To formulate experimental condition by means of response data, a new dynamical loadidentification approach is proposed. This approach utilizes SVR solution path algorithm asmodeling algorithm, and can be applied to identify inversely a wide variety of mechanicalloads which have different forms. Numerical results demonstrate that this approach canavoid error disturbances caused by ill-posed matrix in traditional frequency domain method,and outperforms greatly the common methods in terms of identification accuracy andnumerical stability.