心磁是一种人类心脏电生理活动激发的典型生物弱磁场，因其在诊断早期冠心病、心率不齐等心脏病上具有良好的应用价值而成为研究者的研究热门之一。心磁是一种幅值约100pT的极弱磁场信号，心磁探测的困难往往源于测量过程中各类强烈干扰。为了准确地测量心磁信号，高成本的测量装置和磁屏蔽设备成为标配。目前主流的心磁测量仪器为超导量子干涉仪，然而因为其体积巨大、成本高昂，这种仪器的应用受限。隧穿磁电阻（TMR）传感器有体积小、测量范围大、功耗低、灵敏度高等优势，具有无屏蔽环境下测量心磁，并发展为便携式或可穿戴设备的潜力。基于TMR测量心磁成为近几年心磁测量领域的重要研究方向。然而，TMR的本底噪声大且与心磁信号频域重合。目前的研究结果显示,由于信噪比较低，TMR只能测量到心磁信号中幅值最大的R波，而含有重要临床信息的T波则无法实现测量。本文通过集成新型TMR传感器与低噪声的信号调理电路，搭建了一套高精度、低噪声的TMR心磁测量装置，其噪声水平为18.7 。为了给TMR心磁仪所测得的心磁信号提供有效参考，本工作还使用了一款光泵浦原子磁力仪（OPM）实现了心磁测量，并提出两种降噪算法进一步提高OPM所测心磁信号的质量，合成心磁信号中加入实测噪声的实验表明在两种算法的辅助下，信噪比可超过9 dB。用TMR和OPM协同采集了四名成年男性的心磁，二者能同时观测到规律性的磁场波动，这证实了TMR心磁仪具有一般性的心磁测量能力。为了判断TMR所测心磁的准确性，使用TMR和OPM先后采集同一位置的心磁信号，并在多周期移动平均后进行比较。OPM和TMR所测心磁在波形特征和波形对应时间上一致性很高。特别是TMR首次清晰地测量到T波，T波在时间和幅值上的相对误差仅为3.4%和1.8%，意味着TMR在心磁测量领域具备了真正的实用意义。还搭建了一个合成TMR梯度计，在更大环境噪声下测量心磁，可有效降低工频干扰56.5%。论文搭建了一种递归神经网络（RNN），用深度学习的方式处理TMR测量的心磁中的噪声。针对仿真噪声和实际TMR电路噪声的处理中，经该神经网络处理后的心磁信号信噪比分别提高了13.06 dB和14.65 dB。此处理方式从噪声中还原了信号的主要特征，且保留了原始信号的长度，具有较强的实用价值。

Magnetocardiography (MCG) is a typical biological magnetic field generated by the electrophysiological activity of the human heart. MCG has been one of the hot spots of the researchers’ interest because of its applications in diagnosing early coronary artery disease, arrhythmia, and other heart diseases. As an extremely weak magnetic field signal with a magnitude of about 100pT, the difficulty of MCG detection often stems from various strong interferences during the measurement. High-cost measurement devices and magnetic shielding equipment have become routines for measuring the cardiac magnetic signal accurately. The mainstream instrument for MCG measurement is the superconducting quantum interference device (SQUID). However, the application of this instrument is limited because of its vast size and huge cost. With the advantages of small size, extensive measurement range, low power consumption, and high sensitivity, Tunneling Magnetoresistance (TMR) sensors have the potential to measure MCG in unshielded environments and be developed as portable or wearable devices. Measurement of MCG using TMR has become an important research direction in MCG detection in recent years.However, the noise of the TMR is significant and overlaps with the MCG signal in the frequency domain. The current study results show that due to the low signal-to-noise ratio (SNR), TMR can only measure the R-wave with the highest amplitude in the MCG signal. In contrast, T-wave, which contains important clinical information, is not available for measurement. In this paper, by integrating a new TMR sensor with a low-noise signal conditioning circuit, a high-precision MCG measurement device is built with a noise level of 18.7 pT/(Hz)1/2@9.8Hz. In order to provide a valid reference for the MCG measured by the TMR, an optically pumped magnetometer (OPM) was used to implement the MCG measurements. Two denoising algorithms are proposed to further improve the quality of the MCG measured by the OPM. The experiments of adding measured noise to the synthetic MCG show that the SNR can exceed 9 dB with the assistance of the two algorithms.The MCG of four adult males was collected in concert with TMR and OPM. Regular magnetic field fluctuations are observed simultaneously in the output of both measurement techniques, which confirms the general capability of the TMR magnetometer to measure MCG. To determine the accuracy of the MCG measured by the TMR, the TMR and OPM were used to collect the MCG signals at the same location. The MCG measured by OPM and TMR were in good agreement regarding waveform characteristics and correspondence times after the multi-period moving average. The relative errors of T-wave measured by TMR in time and amplitude are only 3.4% and 1.8%. The first clear and accurate measurement of T-waves by TMR means that TMR has practical significance in MCG measurement. A synthetic TMR gradiometer was also built to perform MCG measurements in environments with large residual magnetic fields. A gradiometer can effectively reduce powerline interference by 56.5%The paper proposes a recurrent neural network (RNN) to process the noise in the MCG measured by TMR using a deep learning technique. For the processing of simulated noise and actual TMR circuit noise, the SNR of the MCG signals improved by 13.06 dB and 14.65 dB after processing by the neural network, respectively. Deep learning restores the main features of the MCG signal from the noise without changing the length of the signal, which has a strong practical value.