基于深度学习技术，提出一种适用于重力数据去噪的U-Net深度神经网络方法，并使用大范围样本数据对网络结构进行训练，构建了去噪处理的网络模型。实验结果表明，该方法能够学习重力数据的内部特征，进而实现重力数据的信噪分离，去除噪声更加彻底，处理含有3 mGal噪声幅度的重力数据，去噪结果的均方误差相对巴特沃斯滤波方法和小波分解滤波方法分别减小了73.8%和51.6%。该方法具有较强的鲁棒性，处理含有更高幅度噪声的数据时仍然表现出良好的去噪性能，所得结果的信噪比可达到13.216 db，相对另外2种对比方法分别提高了73.8%和51.6%。

In order to improve the accuracy of processed gravity data affected by noise error, we proposed a U-Net deep neural network method based on deep learning technology, and constructed a network model suitable for gravity data denoising by training the network structure with a large range of sample data. Experimental results show that the proposed method can learn the internal characteristics of gravity data, thus achieving signal-to-noise separation of gravity data, and removing noise more thoroughly. In the case of processing gravity data containing 3 mGal noise amplitude, the mean square error of denoising results is decreased by 73.8% and 51.6% respectively compared with Butterworth filtering and wavelet decomposition filtering. Moreover, the proposed method has a strong robustness, as it still shows good denoising performance when processes the data with higher noise amplitude. The signal-to-noise ratio of obtained results can reach 13.216 dB, which is 73.8% and 51.6% higher than the other two methods, respectively.