基于卷积神经网络智能识别吸收峰的激光稳频方法

Laser Frequency Stabilization Method Based on Intelligent Identifying Absorption Peaks with Convolutional Neural Network

针对饱和吸收光谱激光稳频技术中多吸收峰难以自动识别、误差信号灵敏度低的问题,提出了一种基于卷积神经网络(CNN)智能识别铷原子吸收峰的激光频率锁定方法。首先设计了大、小卷积核相结合的5个卷积-Re LU-最大池化模块+两层全连接层的一维卷积神经网络;然后对激光器进行线性扫频,获得了包含24个饱和吸收峰的铷原子的饱和吸收光谱信号S_(sa)(t),提取每个吸收峰的位置序号,将其与饱和吸收光谱信号作为卷积神经网络训练的标签和数据,利用训练后的卷积神经网络实现吸收峰的智能识别;最后采用正交解调技术精确提取饱和吸收光谱信号的相位,使本振信号相位与其自动匹配,提高误差信号的灵敏度。设计了基于上位机卷积神经网络智能寻峰+现场可编程门阵列(FPGA)实时信号处理的激光频率锁定系统,锁定后的激光频率与光频梳拍频结果表明:在7500 s内,激光频率波动标准差为7.94 k Hz;平均时间为64 s时,相对Allan方差为3.50×10^(-12)。所提出的方法可被广泛应用于饱和吸收光谱激光稳频领域。

Objective To address the challenges of automatically identifying absorption peaks and the low sensitivity of error signals in the saturated absorption spectrum laser frequency stabilization technique,a method using convolutional neural network(CNN)is proposed for recognizing rubidium atomic absorption peaks.This approach is highly applicable in the realm of saturated absorption spectroscopy laser frequency stabilization.Traditional techniques are limited to identifying and locking onto specific absorption peaks within a narrow laser tuning range,necessitating manual pre-adjustment of the laser frequency close to the absorption peak.However,in practice,the initial laser operating point is often unknown,requiring broad frequency scans to locate the target absorption peak signal.This can result in detecting multiple groups of absorption peaks.Moreover,the process of deriving error signals is complicated with respect to the phase delay between the saturated absorption signal and local oscillator signal,impacting error signal sensitivity.Typically,phase adjustment of the local oscillator signal is manually performed and monitored with an oscilloscope to capture the most sensitive error signal.This method is inefficient and inaccurate,and thereby,fails to satisfy the demands of high-precision automatic laser frequency stabilization.Consequently,a CNN-based laser frequency stabilization method,which intelligently recognizes rubidium atomic absorption peaks and automatically adjusts for phase delay,is introduced to realize long-term precision stabilization of laser frequency.Methods Initially,a one-dimensional convolutional neural network(CNN)was designed,incorporating a combination of five large and small convolution kernels.This design included“convolution-ReLU-maxpooling”modules followed by two fully connected layers.A linear sweep of the laser frequency was then performed to acquire a spectrum signal from rubidium atoms,containing 24 saturated absorption peaks.The sequence number of each absorption peak was extracted,and these numbers,along with the rubidium atomic spectral signals,were used as labels and data for CNN training,respectively.The trained CNN was then employed for the intelligent identification of absorption peaks.The quadrature demodulation technique was adopted to accurately extract the phase of the saturated absorption spectrum signal and match it with the phase of the local demodulation signal,thereby improving the sensitivity of the error signal.A laser frequency stabilization system,based on CNN intelligent peak search and integrating computer and real-time signal processing with a field-programmable gate array(FPGA),was developed.Locking tests and frequency stability experiments were conducted on this system.It was demonstrated through experimental results that the method of laser frequency stabilization,based on CNN intelligent recognition of rubidium atomic absorption peaks and automatic phase delay matching,can achieve long-term precision stabilization of laser frequency.Results and Discussions In response to the challenges of automatically identifying multiple absorption peaks and the decreased sensitivity of error signals due to phase delay,a laser frequency stabilization method that utilizes convolutional neural network(CNN)to identify rubidium atomic absorption peaks is proposed.The designed one-dimensional CNN model(Fig.4)converges(Fig.5),enabling intelligent recognition of multiple absorption peaks within saturated absorption spectral signals(Fig.6 and Fig.7).Through automatic phase delay matching,the phase delay significantly reduces from 100.93°to 0.02°,leading to increases in both the zero-crossing slope and amplitude of the error signal.This in turn substantially enhances its sensitivity.A laser frequency stabilization system,incorporating CNN-based intelligent peak search with computer and real-time signal processing via FPGA,is developed(Fig.1).This system locks onto the cross peak 85Rb F=3→F′=CO3-4 for a locking test(Fig.11).The locked laser undergoes beat frequency experiments with an optical frequency comb to assess frequency stability.Experimental outcomes reveal that the minimum relative Allan variance over a span of 7500 s is 3.50×10^(-12)@τ=64 s(Fig.14).This illustrates that the proposed CNN-based approach for intelligent identification of rubidium atomic absorption peaks,coupled with automatic phase delay matching for laser frequency locking,facilitates long-term precise stabilization of laser frequency.Conclusions In this study,a laser frequency stabilization technique utilizing CNN is introduced for the intelligent recognition of rubidium atomic absorption peaks.This approach not only facilitates the intelligent identification of multiple absorption peaks across a broad tuning range of lasers but also supports long-term precise laser frequency stabilization.Experimental evidence shows that a specially designed one-dimensional CNN model is capable of accurately identifying 24 absorption peaks within the rubidium atomic spectrum signal.Automatic phase delay adjustment from 100.93°to 0.02°significantly enhances the error signal’s sensitivity.Following the application of laser frequency stabilization,the minimum relative Allan variance decreases to 3.50×10^(-12),when average time is 64 s.Consequently,this method holds potential for broad application in areas such as saturated absorption spectroscopy and laser frequency stabilization.