The tea production is developing towards large-scale currently, and the quality evaluation of tea after processing still relies on highly subjective artificial sensory evaluation, which is not suitable for the large-scale development of tea. Near-infrared spectroscopy (NIRS) has rich structural and compositional information, which is suitable for the detection of physicochemical parameters of hydrogen-containing organic substances. So it is widely used in the detection of biochemistry components of tea and the classification, such as authenticity discrimination and origin traceability. The work aims to take 'Yinghong 9' black tea as the research object, and propose a rapid grading method for tea quality based on NIRS. Firstly, a total of 42 samples of black tea processed from the fresh tea leaves of various grades were collected, a sub-sample was taken from each sample and ground into powder. A NIR spectrometer was used to scan tea powder to collect the spectrum of each sample. Secondly, quantitative models for the biochemistry components were constructed based on NIRS to gain the biochemistry component information of black tea. Thirdly, 5 professional tea tasters were invited to conduct sensory evaluation on all samples. Based on the opinions of the tea tasters, the quality grade of tea samples were determined. Finally, the relationship between sensory evaluation results and biochemistry components were established to achieve the quality grading of 'Yinghong 9' black tea. In particular, when establishing the grading model, only the black tea processed from the second grade fresh leaves was selected and divided into three grades according to the sensory evaluation results. The quantitative models of four biochemistry components including tea polyphenol, soluble sugar, free amino acid and caffeine in black tea were established while these four quantitative models were preprocessed by combination data correction and normalization to reduce noise, drift as well as other interference and improve the difference between samples. These quantitative models were uniformly built using Partial Least Squares algorithm after using Genetic Algorithm, Successive Projections Algorithm, Variable Combination Population Analysis combined with Genetic Algorithm and other algorithms respectively to extract features. In order to ensure the reliability and stability of the model, Kennard-Stone algorithm was used to divide the samples into calibration set and test set before modeling, and K-fold verification was used in the modeling process. The principal components of the four quantitative models were all less than 10. The coefficients of determination on calibration set were tea polyphenol 0.974 5, soluble sugar 0.887 6, free amino acid 0.963 6 and caffeine 0.860 6 and the Root Mean Squared Error were 0.630 0, 0.298 3, 0.045 6, 0.162 6, respectively. The grading model based on sensory evaluation and biochemistry components had an accuracy of over 85%, which was built using Random Forest algorithm with 35 trees. The research results provide a feasible scheme for rapid grading of processed black tea based on specially graded fresh tea leaves, and effectively improve the interpretability of black tea grading.