Postharvest quality defects are critical concerns to the food quality, lower product value and growers’ revenue. Conventional quality management relies heavily on empirical approaches and is facing challenges. This research investigates the feasibility of using fluorescence fingerprinting technology as a rapid, nondestructive method for detecting quality defects in Concord and Niagara grape juice caused by downy mildew and grape berry moth damage. Grape samples were collected from Cornell Lake Erie Research and Extension Laboratory to prepare juice samples with 0% to 5% quality defects to mimic the industrial practice. A workflow including grape homogenization, filtration and dilution was developed, and the optimal dilution factors were determined for both cultivars. Fluorescence absorbance transmittance excitation and emission matrix (A-TEEM) of the samples was measured at 200-600 nm (2 nm interval) excitation wavelength, 245-827 nm (2.33 interval) emission wavelength, and 0.2 s integration time. The A-TEEM data was subjected to Raman normalization, interpolation, inner filter effect correction, and Rayleigh masking as pre-processing. The pre-processed data was analyzed using parallel factor analysis (PARAFAC) to identify the key wavebands contributing to the fluorescence fingerprints of each sample, and a convolutional neural network (CNN) model was developed for binary classification of quality defects in both cultivars following an industry guideline (≤ 3% as clean, ≥ 4% as defective) with 10-fold validation. Untargeted metabolomic analysis was performed using an UHPLC-MS system was employed to compare the chemical composition difference among the clean and defective grapes and identify the key compounds that contributed to the fluorescence fingerprints. The optimal dilution factors for Concord and Niagara cultivars were determined to be 500-fold and 2000-fold, respectively. Grapes with different cultivars and defect levels showed different fluorescence fingerprints. Two major fluorescence regions Ex. 200–220 nm /Em. 310–330 nm and Ex. 270–280 nm /Em. 320–350 nm. Fluorescence intensities of the key wavebands identified from PARAFAC showed an increasing trend with the increase in the quality defect levels but showed nonlinear and un-monotonical trends. The CNN models with ResNet-18 backbone showed an average accuracy about 80% for Concord and 68% for Niagara, suggesting good performance. Chemical analysis results showed opposite trends of changes of the key chemical compounds in the Concord and Niagara grapes. Tyrosine, tryptophan, catechin and epicatechin’s autofluorescence were identified as major contributors to the fluorescence difference. The findings from this study suggest that the fluorescence A-TEEM assisted by advanced chemometric and machine learning techniques can be used as a novel, accurate method to detect quality defects, and potentially help the juicing grape processors to improve the efficiency and consistency of quality management at postharvest.