With the deepening of the concept of “green tires”, the effective composition of tire rubberis directly related to the qualification of rubber. But tire rubber is a “black” analysis system for the inspection department, and it’sexceedingly crucialto accurately detect rubber components by the existing methods. Terahertz time-domain spectroscopy (THz-TDS) technology has been successfully applied to material detection and analysis, but the terahertz spectral data observed from a complex sample of rubber represents the comprehensive results of several interrelated components or interaction of characteristic components in many cases, where as the actual information contained in the raw data may overlap, which will conversely affect the analysis of the components in the rubber mixture. In order to solve the problem of terahertz spectral overlap, the characteristics of continuous smoothing of terahertz spectral matrix and sparse concentration matrix are combined this paper, then the 2 norms with smoothing characteristics and the 1/2 norm with sparsity characteristics into the non-negative matrix factorization method is introduced, which are applied to the separation of terahertz aliased spectra, so as eparation method of terahertz aliasing spectral based on spectral feature Constrained Non-negative Matrix Factorization (CNMF) is proposed. Firstly, nitrile-butadiene rubber combined with vulcanization accelerator 2-Mercaptobenzothizzole(MBT) to form a binary mixture in diverse proportions, and it combined with vulcanization accelerators MBT and tetramethy1 thiuram monosulfide (TMTM) to form a ternary mixture in different proportions. Then the terahertz time domain spectrum of all samples ismeasured by terahertz spectroscopy system, which the measured data is subjected too btain a corresponding absorbance spectrum. Further, principal component analysis is performed on the obtained spectral matrix to initially determine the number of components of the mixture. Finally, the Non-negative Matrix Factorization (NMF), Non-negative Matrix Factorization based on pure variables initialization(PNMF) and CNMF methods are used to the decomposition of the mixture data matrix and spectral analysis of the aliased spectrum. The results show that the separation effect of the CNMF algorithm is better than that of NMF and PNMF method, and the corresponding results of the characteristic absorption peak are accurate. In addition, the correlation coefficients of separation results for different component mixtures are higher than 89%, and the spectral angles are less than 0.5 with a higher reduction degree of purity spectrum. Therefore, the constrained non-negative matrix factorization algorithm is introduced into the separation of terahertz aliasing spectra, which is preferable to extract the characteristic information of single components in complex mixtures and provides a better foundation for the qualitative analysis and quantitative calculation of subsequent terahertz multi-component mixtures as well as the considerable research prospects in the field of quality testing of green tires and rubber.