Research On Output Collinearity Of Echo State Network

The traditional artificial neural network generally adopts gradient and evolutionary learning algorithms,which makes it converge slowly,easy to fall into local minima and high computational complexity.Echo State Network(ESN)has fast learning speed and good generalization ability,which has attracted extensive attention in academic circles and has been successfully applied in various fields.However,the core of the ESN has a sparse connection and a huge reservior,and too large a reservior will lead to the output collinearity of the ESN,which will lead to the inaccuracy of the output weights,and lead to the instability of the network and poor prediction performance,which limits its application and popularization.Therefore,the research on the output collinearity of ESN can not only promote the development of ESN theory,but also have high practical application value.In this paper,the output collinearity of ESN is comprehensively analyzed and the effectiveness of regularization method to solve the output collinearity of ESN is studied.An ESN with log penalty and an ESN based on smooth log penalty are proposed respectively.Finally,aiming at the problem of NOx emission in the process of cement clinker calcination,a NOx concentration prediction model is established to realize the online prediction of NOx emission concentration.The main research work of this thesis is divided into the following six chapters:In Chapter 1,the research background,significance and research status of echo state network are introduced.In Chapter 2,firstly,the basic model of ESN is introduced and the echo state property are analyzed.Then,the reason and influence of output collinearity of ESN are analyzed theoretically.Finally,the influence of key parameters of ESN on performance is analyzed experimentally.In Chapter 3,firstly,it analyzes how regularization can reduce generalization error and improve the performance of ESN.Then the application of different regularization in ESN is introduced.Finally,the regularization can effectively solve the output collinearity problem of ESN from the experimental aspect.In Chapter 4,an ESN with log penalty is proposed and it is proved that the log penalty is unbiased.Then the simulation results show that the regularization with log penalty can improve generalization performance and prediction accuracy and solve the problem of ESN.In Chapter 5,an ESN based on smooth log penalty is proposed and the convergence of smooth log penalty is analyzed.Then through experimental comparison,it is concluded that the smooth log penalty can solve the problem of output collinearity of ESN more effectively.Finally,the ESN based on smooth log penalty is applied to the on-line prediction of NOx concentration release in cement clinker sintering system.In Chapter 6,summarize the important achievements of the whole paper and look forward to the future research work.