The settlement deformation of subway tunnels during the construction and operation stages is a relatively complex nonlinear process. The original settlement deformation monitoring data will be affected by some noise during the collection process. In order to reduce the impact of noise and errors in the subsequent prediction process, a joint denoising model based on the improved empirical mode decomposition (ICEEMDAN) algorithm, wavelet threshold denoising and NLMS adaptive filtering was constructed. The model first uses ICEEMDAN to decompose the original data. After decomposing the intrinsic modal component IMF, it divides it into high-frequency and low-frequency components. Then the wavelet threshold is used to remove the components with high correlation coefficients in the high-frequency and low-frequency components. Noise and NLMS adaptive filtering are processed, and finally the additive reconstruction is performed to obtain the denoised data. Experimental analysis results show that compared with ICEEMDAN, wavelet threshold denoising and NLMS adaptive filtering algorithms, the joint denoising model has better denoising effect, higher correlation coefficient, signal-to-noise ratio and smaller mean square root error.