Sugar content is one of the important indicators for watermelon grading, for the drawbacks of traditional watermelon detection methods, the feasibility of acoustic characteristics combined with machine learning for non-destructive detection and grading of watermelon was investigated. The acoustic detection system of watermelon was designed and the time domain signals of different batches of samples were collected. After the time domain signal was normalized, the frequency domain signal was obtained by fast Fourier transform and pre-processed by detrending. The principal components of the frequency domain signal were extracted by using principal component analysis, the cumulative contribution rate of the first three principal components was 95.32%, the samples with different levels were differentiable using the first and second principal components. Watermelon all-variable grading models were developed by using four different machine learning algorithms, and the prediction set classification accuracies all reached over 66%. Feature variables were extracted by using stability competitive adapative reweighted sampling algorithm, which reduced the number of variables by about 84%. The performance of the classification models developed using the extracted feature variables were all improved, with the support vector machine model achieved the highest prediction set accuracy (95.56%), F1 score (96%) and Kappa coefficient (93%). The results indicated that acoustic characterization combined with machine learning was feasible for non-destructive detection and grading of watermelons. The research result can provide a feasible technical solution for non-destructive detection and grading of watermelon, and provide a reference for non-destructive detection and grading of other similar fruits and vegetables.