1.Key Laboratory for Quantum Optics and Center of Cold Atom Physics, Shanghai Institute of Optics and Fine Mechanics, Chinese Academy of Science, Shanghai 201800, China 2.Center of Materials Science and Optoelectronics Engineering, University of Chinese Academy of Sciences, Beijing 100049, China
Fund Project:Project supported by the National Natural Science Foundation of China (Grant No. 11704391).
Received Date:22 February 2019
Accepted Date:30 April 2019
Available Online:01 July 2019
Published Online:05 July 2019
Abstract:Magnetic shielding plays an important role in magnetically susceptible devices such as cold atom clocks, atomic interferometers and other precision equipment. The residual magnetic field in a magnetic shield under a varying external magnetic field can be calculated by the Jiles-Atherton (J-A) hysteresis model and magnetic shielding coefficient. According to the calculation results, the variation of internal magnetic field can be compensated for the active compensation coils. However, it is difficult to practically obtain the exact values of the five magnetic-shielding-related parameters in the J-A hysteresis model and the other two magnetic-field-attenuation-related parameters. It usually takes a lot of time to match the parameters manually according to the measured hysteresis loop and it is difficult to ensure that the final parameters are the global optimal values. The machine learning method based on artificial neural network has been used as an efficient method to optimize the parameters of complex systems. Owing to the powerful computing capability of modern computers, using the artificial neural network to optimize parameters is usually much faster than manual optimization method, and has a greater probability of finding the global optimal parameters. In this paper, the five J-A parameters and the other two parameters relating to magnetic field attenuation are optimized by the method of online learning based on artificial neural network, and the residual magnetic field in the magnetic shield is predicted under the simulated satellite magnetic field environment. By comparing the measured residual magnetic field with the predicted value, it is found that the machine learning method can optimize the magnetic shielding characteristic parameters more quickly and accurately than the manual optimization method. This result can not only help us to compensate for the magnetic field better and optimize the parameters of our cold atom system, but also validate the application of neural network in a multi-parameter physical system. This proves that the in-depth learning neural network can be conveniently applied to other physical experiments with multi-parameter interaction, and can quickly determine the optimal parameters needed in the experiment. This application is especially effective for remote experiments with slow response to parameter adjustment, such as scientific experiments carried out on satellites or deep space. Keywords:artificial neural networks/ magnetic shielding/ hysteresis/ cold atom clock
5.参数优化结果与验证使用神经网络调参的结果如图6所示, 首先用一个正弦变化的环境磁场进行调参. 图6(a)横坐标为总循环次数, 纵坐标为我们设定的标准差, 蓝色点为初始随机地用于训练神经元的12组参数, 可见在第20轮左右, 即初始训练后预测的第8组参数就已经开始明显收敛于最优值, 最终得到的J-A参数为Ms = 540074, c = 0.58, a = 19817, k = 9.50, SE = 36.37, C = -2.37. 标准差值为4.90, 在该组参数下计算得到的磁滞回线和实测值如图6(b)所示, 红线为预测值黑线为实测值. 图 6 (a)一个典型的求参过程图, 预测参数值的标准差随实验轮次逐渐降低; (b)相应参数计算得到的磁滞回线与实测回线的比较 Figure6. (a) A typical continuation process graph, the standard deviation of the predicted parameter values is gradually reduced with the experimental round; (b) comparison of hysteresis loop and measured loop calculated by corresponding parameters.
表1手动调参和自动调参得到的参数值 Table1.Parameter values obtained by manual tuning and automatic tuning.
另外, 初始随机参数的不同可能导致不同的预测速度, 但通常都能够收敛到理想的参数点. 图7展示了另两组初始随机参数进行预测的标准差和循环轮次图, 同样蓝色点为初始随机的用于训练神经元的12组参数, 可以看出由于初始随机值不同, 在本次训练中更快达到了标准差小于5的期望值. 图 7 另外两组不同随机初始训练参数下, 标准差随实验轮次的收敛图 Figure7. Convergence graph of standard deviation with experimental rounds under two different random initial training parameters.
接着将该组参数应用于接近实际卫星在轨经历的环境磁场, 得到的结果如图8所示. 图8(a)和图8(b)分别是在环境磁场随时间变化情况下, 利用使用神经网络调参后得到的参数来计算屏蔽内磁场和实测屏蔽内磁场的对比图, 横坐标分别为时间和环境磁场; 图8(c)表示计算值和实测值的差值随时间变化曲线; 图8(d)表示计算值和实测值的差值随环境磁场变化曲线. 图中标准差越小说明计算越准确. 其中黑线是仅考虑衰减项$ {B_{SE}}$不考虑磁化项$ {B_{\rm M}}$的计算结果差值, 另外两组则分别用手动得到的参数和神经网络得到的参数来计算, 其中手动找到的参数标准差为9.33, 神经网络调参标准差为7.01. 可见使用J-A模型计算屏蔽内参数是非常必要的, 仅考虑衰减项而忽略磁滞项将会导致很大误差, 另外利用神经网络获取J-A模型的参数比手动参数更加快速准确, 为我们对空间钟磁屏蔽补偿参数的选择提供了有效的帮助. 图 8 在近地轨道磁场环境下实测磁场与计算磁场的对比图 (a)横坐标为时间; (b)横坐标为环境磁场; (c) 使用手动调节参数, 自动求得的参数, 以及仅考虑衰减项不考虑磁化项, 计算磁场与实测磁场的差值, 横坐标为时间; (d)横坐标为环境磁场 Figure8. A comparison of the measured magnetic field with the calculated magnetic field in a near-Earth orbit magnetic field environment; (a) The x axis is time; (b) the x-axis is the ambient magnetic field; (c) use manual adjustment parameters, automatically obtained parameters, and simple $ {B_{SE}}$ to calculate the difference between the magnetic field and the measured magnetic field, the abscissa is time; (d) the x-axis is the ambient magnetic field.