Font Size: a A A

Ultrashort Pulse Measurement Based On Convolutional Neural Network

Posted on:2023-04-07Degree:MasterType:Thesis
Country:ChinaCandidate:T WangFull Text:PDF
GTID:2530306617952059Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
Since the advent of ultra-short pulse lasers,because of the advantages of shorter pulse width and higher peak power,it has played an irreplaceable important application in many fields such as electrochemistry,laser processing,and material science.In the study of ultrafast events,it is very important to accurately measure the time-domain intensity envelope and phase of ultrashort pulses.However,the pulse width of ultrashort pulses is lower than the order of picoseconds,which exceeds the response limit of conventional electronic devices and cannot be directly measured by conventional electronic devices.The current ultrashort pulse measurement methods are divided into direct method and indirect method.The representative of the direct method is the photoelectric detection system composed of high-speed oscilloscope and fast photodiode.This method can perform effective photoelectric conversion for picosecond pulses,but cannot measure femtosecond pulses.Representative of indirect methods are intensity autocorrelation and frequency-resolved optical gating(FROG).The intensity autocorrelation can only measure the pulse width,and cannot give the time-domain intensity envelope and phase information of the pulse.FROG can measure the time-domain intensity envelope and phase of femtosecond pulses with high accuracy,but this method cannot adapt to low signal-to-noise ratio conditions,and the measurement process is very time-consuming.Neural networks combined with optical methods to achieve the measurement of ultrashort pulses have been shown to be feasible.However,the complex neural network model and data sets with a large number of redundant make the existing neural network-based pulse reconstruction algorithms fail to achieve good reconstruction effect and generalization ability(adaptability to fresh samples)in practical applications.In view of the above problems,this paper designs an ultrashort pulse measurement technology based on shallow convolutional neural network,which combines the self-designed second harmonic frequency-resolved optical gating(SHG-FROG)optical system and FROG neural network(FNN),intended to achieve fast and accurate measurement of ultrashort pulses.The specific research in this paper is as follows:1.The working principle of SHG-FROG,the generation of the second harmonic,the generation of phase matching and the calculation of the phase matching angle are studied in detail,the principle and implementation of the Principal Component Generalized Projection Algorithm(PCGPA)are deeply analyzed,a SHG-FROG optical system using barium metaborate(BBO)crystal as nonlinear medium is designed.2.An in-depth look at the architectures,algorithms and optimizations of fully connected neural networks and convolutional neural networks.The architecture and hyperparameters of the proposed FNN are designed and explored.The optimized FNN uses fewer convolutional layers,and uses a small simulated pulse training set that is closer to the real pulse to achieve model training and hyperparameter optimization,resulting in a lightweight model that is easier to train and has better generalization performance,and combined with the designed SHG-FROG system to form an ultra-short pulse measurement system.3.The ultrashort pulse measurement system was tested from three aspects:simulated pulse reconstruction,experimental pulse reconstruction and noise adaptation exploration.In the reconstruction of simulated pulses,the reconstruction performance of FNN for common pulses and irregular-shaped pulses is studied,the conjugate inversion and phase shift in the reconstruction process are analyzed,and the focus of PCGPA in reconstructing pulses from FROG-trace is explored.It is proved that FNN’s reconstruction performance and speed for simulated pulses are superior to PCGPA.In the reconstruction of experimental pulses,it is proved that the comprehensive reconstruction performance of FNN for experimental pulses is better than that of PCGPA,and a conjecture that FNN noise adaptability is stronger than PCGPA is proposed.In the noise adaptation exploration,we used the FROG-trace of real pulse with white Gaussian noise to explore the differences in the adaptability of FNN and PCGPA to noise,and found that the noise adaptability of FNN was significantly better than that of PCGPA,the above operations are repeated with simulated pulse to verify the accuracy of the obtained conclusion,and the trend of pulse width reconstruction of FNN and PCGPA with the increase of noise is analyzed.The difference of single pulse reconstruction efficiency between FNN and PCGPA is calculated,and it is concluded that FNN achieves fast and accurate reconstruction of ultrashort pulses.In conclusion,the proposed ultrashort ’pulse reconstruction system based on SHG-FROG combined with FNN has certain practical application value.
Keywords/Search Tags:ultrashort pulse laser, ultrashort pulse measurement and characterization, FROG, neural network, PCGPA
PDF Full Text Request
Related items