In tunable diode laser absorption spectroscopy (TDLAS) gas detection, the accuracy of the detection system is reduced due to the impact of nonlinear noise on the second harmonic. In this paper, a noise reduction method CEEMDAN-EP-WTD was proposed. Combining wavelet threshold method and complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) via evaluation parameter (EP) and sample entropy (SE). The CEEMDAN-EP-WTD can evaluate the noise level of the (intrinsic mode function) IMF component. It calculates the unknown noise standard deviation of the IMF component in the actual signal, which is related to the wavelet threshold. Based on simulation and experimental analysis, the signal-to-noise ratio of the harmonic signal is improved from 4.41 to 28.8596. The minimum detection limit of the experimental system is 4.72 ppm and the detection accuracy is 4.4 %. The results confirm that the algorithm successfully suppresses system noise and enhances the performance of the detection system.