【摘要】 针对相干布局囚禁（Coherent Population Trapping,CPT）原子钟的比例积分微分（Proportional-IntegralDifferential,PID）参数优化问题，提出一种基于神经网络的全局寻优方法。首先通过数学建模模拟CPT原子钟的微波锁定过程，使用数字PID控制算法实现频率稳定。采用蒙特卡洛方法生成10 000组不同PID参数（K_p, K_i, K_d ）的系统响应数据，构建包含输入层、两层隐藏层及输出层的全连接神经网络模型，以均方误差（Mean Squared Error,MSE）为性能指标，建立参数与性能的非线性映射关系。试验结果表明，优化后的PID参数（K_p=0.496 0, K_i=0.005 4, K_d=0.001 6）使CPT原子钟的锁定时间从110 s缩短至30 s，提速73%；万秒Allan偏差（Allan Deviation,AD）提升1.5倍，验证了方法的有效性。与传统Ziegler-Nichols方法相比，神经网络通过全局映射避免了局部最优问题，且计算效率提升显著，为CPT原子钟的高性能控制提供了新途径。更多还原

【Abstract】 This paper addresses the Proportional-Integral-Differential parameter optimization problem of Coherent Population Trapping atomic clocks and proposes a global optimization method based on neural networks. First, the microwave locking process of the CPT atomic clock is simulated through mathematical modeling, and frequency stabilization is achieved by using the CPT demodulated signal as the error signal combined with a digital PID control algorithm. A total of10,000 sets of system response data with different PID parameters( K_p, K_i, K_d) are generated via the Monte Carlo method. A fully connected neural network model comprising an input layer, two hidden layers, and an output layer is constructed to establish a nonlinear mapping relationship between parameters and performance using the Mean Squared Error as the performance index. Experimental results show that the optimized PID parameters( K_p = 0.496 0, K_i = 0.005 4, K_d = 0.001 6)reduce the locking time of the CPT atomic clock from 110 seconds before optimization to 30 seconds, achieving a 73% speed increase; the ten thousand seconds Allan deviation is improved by 1.5 times, verifying the effectiveness of the method.Compared with the traditional Ziegler-Nichols method, neural networks avoid the problem of local optimality through global mapping and significantly improve computational efficiency, providing a new approach for high-performance control of CPT atomic clocks.更多还原