State Identification And Prediction Of Microelectronic Device Dynamic System Based On Extreme Learning Machine

Microelectronic devices have been widely concerned since their appearance.It has been found in the researches that devices exhibit rich nonlinear dynamic phenomena such as single-period,multi-period or chaos depending on the selected parameters.In this paper,based on Extreme Learning Machine(ELM)algorithm,nonlinear dynamics of Duffing oscillator and Josephson junction are studied.The specific work is as follows:First,we studied the nonlinear dynamic state identification method of Duffing oscillator.As a common MEMS(micro-electro-mechanical System)component,Duffing oscillator is immune to Gaussian white noise but sensitive to weak periodic signals.It is widely used in weak signal identification and fault detection.How to identify the state of the system quickly and accurately is always a very key problem.Currently,the commonly used Duffing oscillator state recognition method is realized by calculating Lyapunov exponent.However,this method has some problems such as large amount of calculation and slow convergence speed,and the calculation result can only distinguish the periodic state and chaotic state,but can not realize the identification of various periodic states.Due to the above shortcomings,this paper proposes an accurate recognition method of Duffing oscillator state based on machine learning algorithm-Extreme Learning Machine.Through spectrum analysis of time-order signals of Duffing oscillator,we found that the number and size of peak values of spectrum can be used for feature extraction,and then combined with the bifurcation graph to generate data sets to train the Extreme Learning Machine model.Finally,we can achieve accurate identification of the periodic and chaotic states of Duffing oscillator.The numerical simulation results show that this method has a fast convergence speed and an average accuracy of more than 90%.This research has important significance for signal processing of Duffing oscillator.Secondly,We study the prediction method of bifurcation phenomenon of Josephson junction.Josephson junction is not only a common superconducting device,but also a typical nonlinear dynamic system.The prediction of the state change of Josephson junction has always been a very important research content because it should work in a stable state as far as possible during the use of the junction.In this paper,a prediction method of Josephson junction bifurcation interval based on Extreme Learning Machine algorithm and Noisy Precursors is proposed.In this method,the Josephson junction equation is firstly added with noise,and solved by the random Runge-Kutta method.Then,the noise precursor phenomenon can be observed by spectral analysis of the obtained time series.The phenomenon of the peak numbers and peaks position change characteristic feature extraction,generate training set training Extreme Learning Machine.After that,similar bifurcation points were found for the same feature extraction,and test sets were generated to test the Extreme Learning Machine.The simulation results show that this method can not only identify the state of the system before and after the bifurcation,but also recognize a small part of the bifurcation parameters before the bifurcation,which can be used as the indicators before the arrival of the bifurcation point,indicating that the system state is about to change.The above research combines machine learning algorithm with microelectronic devices to realize the identification and prediction of nonlinear dynamic states of microelectronic devices,which is of great significance to the study of device dynamics.