Conventional denoising methods for terahertz time-domain spectral signals mainly use DCT (Discrete Cosine Transform) discrete cosine transform technology for denoising compression representation, which is vulnerable to the effect of complex redundant signal reconstruction, resulting in low time-domain signal-to-noise ratio. Therefore, a new denoising method for terahertz time-domain spectral signals needs to be designed based on filter machine learning. That is, based on filter machine learning, the time-domain spectral signal denoising parameters are extracted, the terahertz time-domain spectral signal denoising model is constructed, and the terahertz time-domain spectral signal denoising algorithm is designed, thus the terahertz time-domain spectral signal denoising is completed. The experimental results show that the designed THz time-domain spectral signal denoising method based on filter machine learning has high time-domain SNR(Signal-to-Noise Ratio) in different spectral ranges, which proves that the designed THz time-domain spectral signal denoising method has good processing effect, reliability and certain application value. In order to improve the quality of THz time-domain spectral image, It has made certain contributions to promoting the development of relevant fields.