Abstract: The vibration energy of the roadheader during it excavates coal and rock is treated as the source in seismic while heading in coal mines,which avoids the hazards and hindering normal heading operation caused by the explosive source.Therefore,it has the advantage of carrying out underground seismic exploration while coal mine is in normal heading operation.However,due to the low energy of coal-rock excavated by the roadheader,the SNR (signal-to-noise ratio) of the pulse-like source obtained by cross-correlation is low compared with employing explosive source.In addition,because various equipment are installed in the narrow heading space of coal mine,the frequency components of environmental noise are complex and the energy is usually large,which further reduces the SNR of the signals along the excavation and affects the quality of the data processing such as the migration and imaging.The noise attenuation in seismic while heading in coal mines is seen as solving an inverse problem.Based on that the coefficient distribution of the signal in the time-frequency domain is sparse relative to the ambient noise,it was proposed to embed the constraint of L1-norm in the inverse problem.Subsequently,the inverse problem was solved through a fast iterative algorithm to achieve the purpose of noise attenuation.Compared with the traditional wavelet threshold denoising method,the proposed method reduced the influence of diffusion of time-frequency coefficient and improved the resolution of time-frequency coefficient.Therefore when the frequency component of signal was complex,the fidelity of effective signal after noise attenuation was higher.The method proposed was verified by analog and actual data respectively.The frequency distribution form of the analog data was comprised of two situations:the independent distribution of dominate frequencies (40 Hz and 60 Hz) and the overlapping distribution of dominate frequencies (20 Hz and 60 Hz),then three types of noisy signals with different SNRs where synthesized (the SNR were 10.8,6.8 and 1.2 respectively).The denoising results of the traditional method and proposed method confirmed that when the SNR was high and the dominate frequencies were independently distributed,the denoising results of the two methods where comparable,and for signals with low SNR and overlapping dominate frequencies,the SNR of the denoising signals of the proposed method could reach about 4 times than that of the traditional method.However,since the frequency components of the actual seismic exploration data were more complicated,the comparison of the two denoising results of the actual signals showed that the proposed method had more advantages in maintaining trends and details of the signal.