As one of the leading technologies of real-time dynamic deformation monitoring, global navigation satellite system (GNSS) has been widely used in deformation monitoring. Multipath error is the primary error source that limits the application of GNSS high-precision deformation monitoring, which cannot be eliminated by the double-differenced technique. It is a problem to separate and mitigate multipath in the GNSS deformation monitoring series in real-time. Therefore, we propose an approach to solve this problem, named time–frequency mask and convolutional neural network (TFM–CNN). The specific processes are as follows: (1) TFM–CNN network construction. We add a full-band deformation to the multipath and obtain the spectrogram of the signal by the short-time Fourier transform (STFT); meanwhile, the ideal ratio mask (IRM) is used to obtain the corresponding time–frequency mask matrix based on the spectrogram; furthermore, one-dimensional CNN mines the mapping relationship between the spectrogram and the time–frequency mask matrix. (2) Multipath separation. We obtain the spectrogram of the GNSS real-time deformation monitoring series by STFT. Then, we estimate its time–frequency mask matrix by the network. The matrix is multiplied by the spectrogram of the monitoring series to obtain the spectrogram of the multipath. Finally, we perform the inverse STFT to obtain the multipath in the GNSS monitoring series. The experimental results show that by training GNSS data in only a specific direction (such as the north direction), the mapping relationship between the spectrogram of multipath and the time–frequency mask matrix can be obtained, which can effectively separate multipath of the GNSS monitoring series in different observation environment in real-time. TFM-CNN significantly improves the monitoring accuracy and achieves millimeter level dynamic deformation monitoring.

作为实时动态变形监测的领先技术之一，全球导航卫星系统（GNSS）已广泛用于变形监测。多径误差是限制GNSS高精度变形监测应用的主要误差源，而双差技术无法消除这种误差。实时分离和缓解GNSS变形监视系列中的多路径是一个问题。因此，我们提出了一种解决该问题的方法，称为时频模板和卷积神经网络（TFM–CNN）。具体过程如下：（1）TFM-CNN网络建设。我们将全频带变形添加到多径中，并通过短时傅立叶变换（STFT）获得信号的频谱图。与此同时，理想比率掩模（IRM）用于根据频谱图获得相应的时频掩模矩阵；此外，一维CNN挖掘了频谱图和时频掩码矩阵之间的映射关系。（2）多径分离。我们通过STFT获得了GNSS实时变形监测序列的频谱图。然后，我们通过网络估计其时频掩码矩阵。将该矩阵乘以监视序列的频谱图即可得到多径的频谱图。最后，我们执行逆STFT以获得GNSS监视序列中的多径。实验结果表明，通过仅沿特定方向（例如北向）训练GNSS数据，可以获得多径频谱图与时频模板之间的映射关系，可以有效地实时分离不同观测环境下的GNSS监测序列的多径。TFM-CNN大大提高了监测精度，并实现了毫米级的动态变形监测。