Seismic signal denoising is the most representative link in seismic data processing, which is related to the accuracy of earthquake early warning system. Therefore, in order to solve the problem of low signal-to-noise ratio and waveform distortion of seismic signals after denoising by traditional denoising algorithms, this paper introduces a deep learning method and constructs a new time-frequency denoising model based on convolutional auto-encoder (CAE-ED) to train and test the synthetic dataset containing multiple noise types constructed based on the Stanford seismic dataset (STEAD), and compares the denoising results with traditional methods. Comparison. The experimental results show that compared with bandpass filtering and wavelet threshold denoising, the CAE-ED model has significant improvement in the average signal-to-noise ratio and average correlation coefficient, while the root mean square error value is greatly reduced, which indicates that the denoising model in this paper has a strong denoising ability, while the waveform distortion is smaller after denoising. The constructed seismic signal denoising model can provide strong technical support for the earthquake early warning system.