As the accuracy and stability of a blood glucose level model is affected by the noise in near infrared non-invasive blood glucose detection process, an improved complete ensemble empirical mode decomposition method with adaptive noise was proposed for denoising of near infrared spectroscopy signals. Meanwhile, a mode selection method based on Frechet distance combining with the feature of curve curvature was proposed for the selection of Intrinsic Mode Functions(IMFs). Firstly. the complete ensemble empirical mode decomposition method with adaptive noise was introduced in the denoising processing of near infrared spectroscopy, and the basic principles and concrete realization processes of empirical mode decomposition, ensemble empirical mode decomposition, complementary ensemble empirical mode decomposition and the complete ensemble empirical mode decomposition based on adaptive noise were described. Then, an improved complete ensemble empirical mode decomposition method with adaptive noise based on curvature and discrete Frechet distance was applied in denoising for simulation signals and spectral signals, and their standard deviation and the Signal to Noise Ratio(SNR) were taken as the evaluation indexes. The simulation and experimental results show that the standard deviation of the improved method based on curvature and discrete Frechet distance in the near infrared spectral signal is 0.179 4, and the SNR is 19.117 5 dB, which extracts useful information, realizes the separation of signal and noise, and improves the quality of reconstructed signals. The proposed method has a good adaptability to effectively identify and separate the signal and noise components.