Abstract: Due to the defect of judging fermentation quality by human observation and difficulty in detecting biochemical components quickly Taking WIP(work in progress) of the Congou black tea in the fermentation process as the research object, in this study, we established quantitative analysis models of sensory quality scores and physical and chemical quality indicators (theaflavin, thearubigin, theaflavin, catechin and phenol to ammonia ratio) based on near-infrared spectroscopy (NIRS). In the process of model establishment, the influence of the feature variable selection method on the prediction model was discussed. Firstly, the acquired data of near-infrared spectral was preprocessed by SNV (standard normal Z transformation), and then combined Si-PLS (synergy interval partial least squares), SFLA (shuffled frog leaping algorithm), CARS (competitive adaptive reweighted sampling) and SPA (successive projections algorithm), the optimal characteristic wavelength variable of each quality indicators was then selected. Finally, the PLS (partial least squares) linear prediction model and the SVR (support vector machine regression) nonlinear prediction model of each fermentation quality indicators were established based on the preferred wavelengths. The comparison of model results showed that the variable screening methods such as Si, CARS, SFLA and SPA can effectively compress the variables compared with the full-band variable model and the number of wavelength variables was greatly reduced under the premise of stabilizing the performance of the model. Among them, nine characteristic wavelength variables closely related to the sensory score were optimized by the SPA method, from which the variable compression rate was as high as 98.88%. The best selection methods for the correlation characteristic wavelength variation and the numbers of sensitive characteristic wavelengths screened of TFs (theaflavins), TRs (thearubigins), TBs (teabrownine), catechin and TP/FAA (Ratio of tea polyphenol and three amino acids) were SPA-PLS and 16, SPA-PLS and 10, CARS-PLS and 33, SPA-PLS and 9, SFLA-PLS and 12, respectively and sensitive features wavelengths were screened. The best screening method for the correlation characteristic wavelength variation of total amount of catechin was SPA-PLS, and a total of nine extremely sensitive characteristic wavelengths were screened. The best screening method for the correlation characteristic wavelength variable of TP/FAA (Ratio of tea polyphenol and three amino acids) was SFLA-PLS, and a total of 12 extremely sensitive characteristic wavelengths were screened. The RPD values of SVR prediction models that were established by the optimal variable of sensory quality, TBs and catechin were all greater than 3, which were 3.923, 3.234, and 5.462, respectively. The RPD values indicated that the results of NIR quantitative analysis were accurate and reliable, and the model prediction performance was extremely high, which can be used for quality control. The RPD value of TP/FAA model was 2.815 (>2) indicating that the model has a good predictive performance and can be used for quantitative analysis. The RPD value of TRs was 2.223 (>2) indicating that the model had good predictive performance. The RPD value of TFs was 1.770 (between 1.4 and 1.8), indicating that the model has poor predictive performance and can be used to make rough prediction and correlation assessment for examples. The predictive performance of SVR model of TFs was poor, because the content of TFs was lower (<1%) and the components was extremely complex, which affected the difficulty and precision of the model construction. The study established a quantitative rapid detection method for key chemical constituents in black tea fermentation, and completed dynamic monitoring of fermentation quality status. The research results set a theoretical foundation for the practical application of the near-infrared spectroscopy rapid detection of the fermentation quality of black tea.