| In recent years,neural networks have been developing rapidly with the improvement of computing processing capability,the breakthrough of algorithm and the development of big data.And they are widely used in various industries.Reservoir computing(RC)by a bio-inspired paradigm,is a unique and simple neural network.Compared to more complex conventional neural networks,RC simplifies the training processing greatly.And RC overcomes some obstacles in training period,such as slow convergence,local optimum and numerous training parameters.Besides,it is facilely implemented by available hardware and trained online.It is amusing to discover that RC can be realized by using existing optical devices and systems for high-speed real-time processing in the field of Optics.Therefore,RC has become one of the research hotspots in the field of Optics.And RC has been successfully applied in many cross fields.At present,most of the research on RC is focused on the RC structure with a single optoelectronic feedback loop based on a semiconductor laser.However,it is not sufficient to meet the actual needs to a certain extent faced with more complex time series prediction,speech recognition and nonlinear channel equalization problems.Therefore,it is necessary to further optimize the RC structure with a single optoelectronic feedback loop based on a semiconductor laser.In this thesis,we propose and characterize a novel reservoir computing system with double optoelectronic feedback loops based on a semiconductor laser,which is to be further improved in the performances of time series prediction,speech recognition and nonlinear channel equalization.Firstly,the theoretical models of the traditional RC structure,the RC with a single optoelectronic feedback loop and double optoelectronic feedback loops based on a semiconductor laser are simulated in this paper.Aiming at the task of 10-order nonlinear auto regressive moving average model(NARMA-10),the influence of system parameters of the new RC system with double optoelectronic feedback loops based on a semiconductor laser is analyzed in detail on its prediction performance.And the results obtained are as a valuable reference for optimizing the new RC system proposed.It is shown that the prediction performance of the system is preferable when the feedback intensity is in the range of [0.25,0.8].Especially the feedback intensity adjusted by about 0.4,the prediction performance of the system is best.And the acquired normalized root mean square error(NMSE)of the new RC system proposed is less than 10% when the time interval difference between short feedback loop and long feedback loop is 0.5ns,0.7ns and 1ns.Besides,the bias phases of two modulators roughly adjusted in the range of(?)the new RC system proposed can obtain preferable prediction performance.Three kinds of RC systems are applied to the tasks of NARMA-10,isolated spoken digits recognition and nonlinear channel equalization to compare their performances in this thesis.It is shown that the novel RC system proposed has superior advantages in performance of prediction,speech recognition and nonlinear channel equalization,compared to the other RC system.In the NARMA-10 task,the NMSE of the new RC system is less than 10% at least compared to the other RC structures which of NMSE is about 15%;in the task of isolated spoken digits recognition,the new RC system is realized at zero error recognition while other RC structures fail to achieve;in the task of nonlinear channel equalization,the symbol error rate of the new RC system is almost zero at the higher signalto-noise ratio,which is significantly better than others.In conclusion,the novel reservoir computing system based on a semiconductor laser with double optoelectronic feedback loops proposed in this paper is an effective scheme to improve the performances of RC in the aspects of prediction,recognition and nonlinear channel equalization.And it provides a certain reference significance for the related RC research. |
PDF Full Text Request