Abstract: [Purposes] This study proposes a denoising method based on the deep residual shrinkage network (DRSN)to solve the challenges of suppressing non-stationary noise in GNSS coordinate time series. [Methods] Firstly,one-dimensional convolutional neural network architecture with channel attention mechanism is const,ructed,and residual connections are utilized to avoid the gradient vanishing problem.Subsequently,the feature channel thresholds are dynamically calculated and an adaptive threshold processing mechanism is designed to achieve high-fidelity noise suppression while preserving signal details by combining global average pooling and fully connected layers. [Findings] Simulation experiments show that compared with the traditional variational modal decomposition,complementary integrated empirical modal decomposition,and deep neural network methods,the root-mean-square error of this paper's method is reduced by 41.6%,35.8%,and 15%,respectively,the signal-to-noise ratio (SNR)increases to 29.82 dB,and the correlation coefficient reaches 0.998 5.In the real-world data experiments,the method reduces the RMSE by an average of 42.3%at the three monitoring stations compared to the optimal comparative method,and the average SNR increases to 98.18 dB. [Conclusions] This proposed method significantly outperforms the other traditional methods.This study provides a new technical pathway for high-precision GNSS positioning and surface monitoring.

Key words: deep residual shrinkage network, GNSS coordinate time series, adaptive denoising, channel attention, residual learning