摘要
由于超导量子干涉仪(SQUID)在实际测量过程中易受到多种干扰源的干扰,加上微弱磁场自身的敏感性,导致许多信号异常数据接近于正常数据,对信号异常数据的检测带来了挑战。为了解决这一问题,设计了一种基于深度学习的框架来提升SQUID传感器在磁场信号异常检测中的精准度。模型主体框架由三维卷积神经网络(3DCNN)模块搭建,整体采用U型架构来优化信息流动。在注意力方面,以压缩和激励(SE)通道自注意力模块代替目前最优(SOTA)的注意力编码器Transformer来增强特征提取能力,在有效提取特征的同时大幅降低计算复杂度。使用三折交叉验证的方法进行训练与验证,与目前SOTA的Transformer编码器对比,实验结果显示,改进后的模型在不需要多层编码器堆叠情况下,有效降低了计算量,提高了异常信号检测的平均精准率。
Due to the fact that superconducting quantum interference device(SQUID)is susceptible to interference from various interference sources in the actual measurement process and the sensitivity of weak magnetic field itself,many signal anomaly data are close to normal data,which brings challenges to the detection of signal anomaly data.In order to solve this problem,a deep learning based framework was designed to improve the accuracy of SQUID sensor in the detection of magnetic field signal anomalies.The main frame of the model was built by the 3D convolutional neural network(3DCNN)module,and the overall structure was Ushaped to optimize information flow.In terms of attention,squeeze and excitation(SE)channel self-attention module is used to replace the current state-of-the-art(SOTA)attention encoder Transformer to enhance feature extraction capability,effectively extract features and reduce computational complexity.The training and verification were carried out by the three-fold crossvalidation method.Compared with the current SOTA Transformer encoder,the experimental results show that the improved model effectively reduces the amount of calculation and improves the average accuracy of abnormal signal detection without the need for multi-layer encoder stacking.
作者
陈佳钰
赵凯明
张斌珍
Chen Jiayu;Zhao Kaiming;Zhang Binzhen(Key Laboratory of Instrumentation Science&Dynamic Measurement of Ministry of Education,School of Instrument and Electronics,North University of China,Taiyuan 030051,China)
出处
《微纳电子技术》
2025年第2期142-149,共8页
Micronanoelectronic Technology
基金
山西省基础研究项目(202203021212120)
山西省回国留学人员科研资助项目(2022-143)。
关键词
磁场信号
异常检测
信号分类
深度学习
超导量子干涉仪(SQUID)传感器
magnetic field signal
abnormal detection
signal classification
deep learning
superconducting quantum interference device(SQUID)sensor
