Research And Application Of Surrogate Model Construction Method Based On Extreme Learning Machine

The optimization of complex products design based on surrogate model can greatly reduce both development cost and design cycle,which is a research hotspot in the field of modern engineering design.Due to the mutual coupling of many factors,there is a strong nonlinear relationship between the design variables and target variables of complex products.Extreme Learning Machine（ELM）is a kind of stochastic feedforward neural network that has been proven to approximate any training data with zero error,showing good approximation ability and generalization performance for strong nonlinear models.As ELM served as the research carrier,the surrogate model construction of complex products and their subsequent improved products was studied in this thesis.The main research work was as follows.（1）For the problem of complex product surrogate model construction with accumulated rich simulated data,an improved ELM modeling method based on genetic programming（GP）was proposed.The limited number of neurons,random input weights and fixed activation function limit the nonlinear representation performance of ELM.In order to improve the approximation ability of ELM,the symbolic regression（SR）method realized by GP was introduced to optimize the activation function of ELM.As a result,a GP-ELM modeling method for single-layer ELM and a DGP-ELM modeling method for deep-layer ELM were proposed.The experimental results shown that the proposed methods can effectively improve the approximation and generalization performance of ELM surrogate model.（2）For the problem of planning a small amount of simulated data to construct an ELM surrogate model for the first-generation product,sequential sampling algorithms based on determinantal point processes（DPPs）were proposed.To achieve spatial diversity sampling,a k-DPPs（SDS）sequential sampling algorithm using the mutual exclusion characteristics of similar samples in DPPs sampling was proposed.Further,on the basis of studying the characteristics of ELM online learning algorithm,two uncertainty evaluation indexes were derived,including model uncertainty based on E-optimal design and sample uncertainty based on residual prediction.Thus,a k-DPPs（US）sequential sampling algorithm was proposed,which combined spatial diversity and uncertainty.The experimental results shown that the proposed sequential sampling algorithms improved the generalization performance of the ELM surrogate model.（3）For the problem of constructing an ELM surrogate model of improved product in the iterative design of complex product,which have the same series of old as reference,the transfer learning modeling methods by reusing the simulated data of old product were proposed.Multi-fidelity approximation modeling is a modeling method that fuses high and low precision samples.The idea was extended to transfer learning,and an Additive Multifidelity Extreme Learning Machine（AMF-ELM）modeling method was proposed.Then,the shortcomings of AMF-ELM were analyzed,and an Iterative Additive Multi-fidelity Extreme Learning Machine（IAMF-ELM）modeling method was proposed.Furthermore,according to the online learning algorithm characteristics,the proposed k-DPPs（SDS）and k-DPPs（US）sequential sampling algorithms were introduced into the AMF-ELM and IAMF-ELM methods.The experimental results shown that the proposed transfer learning modeling methods can construct a high precision ELM surrogate model with very few samples.If the simulated data were planned by the proposed algorithm,the generalization performance of the model can be further improved.（4）The proposed GP-ELM and DGP-ELM modeling methods were used to construct the global deflection surrogate model of the I core type metal sandwich panel,and the surrogate models of C core type and V core type metal sandwich panels were constructed by AMF-ELM and IAMF-ELM modeling methods.The proposed k-DPPs（SDS）and k-DPPs（US）sampling methods were used to construct the surrogate model of the maximum cavity pressure of luffing cylinder of the telescopic handler with the rated load of 6klb,and the surrogate model of the telescopic handler with the rated load of 10 klb was constructed by AMF-ELM and IAMFELM modeling methods.