Research On Extreme Learning Machine And Its Applications To Continuous Galvanizing Process

The importance of intelligent modeling methods and algorithms is addressed along with the development of the information technology.These advanced modeling methods present promising performance when applying in the measuring and control processes of modern industries.The data-driven modeling methods are receiving increased attention,especially in the complicated processes where the setup for measurement units is not an easy task;thus,High-quality real-time measurements are not available in this case.In this context,the academia and industry are devoting more efforts to researching and developing advanced modeling methods.The theories of Extreme Learning Machine(ELM)are discussed in this dissertation.Furthermore,the ELM based intelligent modeling method is developed and utilized under the background of steel strips industries in depth,and the main results include:(1)Hybrid Chaotic Optimization Search Algorithm based ELM with Affective Interaction(HCOAAI-ELM)algorithm is proposed.The Hybrid Chaotic Optimization Search Algorithm with Affective Interaction(HCOA-AI)is employed,which works to combine the global searching ability and local searching ability of Particle Swarm Optimization with Constriction Factor based on Levy Flight(LFPSO-CF)and Chaotic Optimization Algorithms(COA),respectively.Moreover,the searching modes of the method can be switched according to the method of Affective Interaction(AI).The proposed algorithm decreases the number of the hidden nodes;consequently,the model complexity of ELM is reduced.The HCOAAI-ELM improves the learning efficiency of Incremental Extreme Learning Machine(I-ELM).(2)Multi-learning Clonal Selection Algorithm(MLCSA)based ELM algorithm is proposed.We adopt the Baldwinian Learning and Lamarckian Learning to form the multi-learning clonal selection strategy,which enhances the reproduction ability of antibody population with high affinity.The training speed of ELM is improved because the optimal parameters of hidden-layer neurons is reproduced in clone and updated,and an over-complex model tends to be avoided.The proposed algorithm can improve the real-time performance of the model in industry process.(3)Orthogonal Convex Incremental Extreme Learning Machine(OCI-ELM)algorithm.Considering the solution of the least squares problem,the Gram-Schmidt orthogonalization method is employed to find the optimal solution of the descriptor linear system.Further,the OCI-ELM model is developed by proofing the reliability and convergence.The simulation results show that the proposed method improves the prediction accuracy and generalization ability compared to the baseline cases.The proposed algorithm deals with the modeling issues in complicated working conditions.(4)The prediction model for continuous galvanizing coating weight is investigated utilizing the deep ELM method.Because the zinc gauge is not capable to provide the timely information of coating weight,the measurement delays impacts the control performance of the continuous galvanizing process.We analyze the factors related to the coating weight,as well as the process parameters.The deep structure with the multi-layer nonlinear mapping capability is implemented to extract the principle variables influencing coating weights.The method improves the accuracy in the process of the coating weight prediction.Further,the prediction results are sent to the air knife control system by means of hierarchical method.The coating weight can be controlled accurately and promptly.(5)The elongation prediction model of steel-strips in annealing furnace is developed with Deep Network based on Stacked OCI-ELM-AEs(DOC-IELM-AEs).The delay of welding inspections causes the excessive use of galvanized sheet in the process,and the air knife may even bedamaged in abnormal situation.We implemented the proposed OCI-ELM combining with the mechanism model of elongation in annealing furnace,and a hybrid prediction model is constructed.The model takes advantage of the deep structure created by auto-encoder;thus the prediction accuracy for the position of welding can be significantly improved.The proposed hybrid model is implemented into the welding tracking system,and the ideal tracking effect is obtained.The improved method ensures the security and efficiency of the main components in the process.