Due to the limitations of the acquisition environment and instrument performance, the collected seismic signals contain strong random noise, which presents great challenges for subsequent processing and interpretation. Multi-scale geometric analysis has attracted attention in recent years. This paper introduces non-local mean algorithm (NLM) into the Shearlet transform domain to denoise seismic signals. The algorithm firstly performs non-subsampled Shearlet transform on seismic signals, and approximates the generalized Gaussian distribution. The Shearlet coefficients are subjected to principal component analysis (PCA), and then the non-local mean processing Shearlet coefficients are used. Finally, the new Shearlet coefficients are inversely transformed by Shearlet to obtain the denoised seismic signals. The experimental results show that under low noise the proposed algorithm can achieve better denoising effect than the non-local mean algorithm. Therefore, the proposed algorithm is feasible for denoising seismic signals.