Abstract: Aiming at the problems of feature redundancy and poor interpretability in the process of ultrasonic recognition of weld defects based on machine learning, an XGBoost model for ultrasonic recognition of weld defects was proposed based on SHAP interpretability. Four kinds of typical defects were machined on carbon steel weld samples. The ultrasonic reflection echo signal was collected by shear wave oblique incidence method, and 16 time-domain features, 16 frequency-domain features and 3 information entropy features were extracted respectively. SHAP values were calculated and 8 high-contribution features were selected to build a feature subset. Cross-validation and grid search were utilized to optimize the XGBoost model. The feature subset was used as input to identify the defect types. The results showed that average recognition accuracy of the four defects was 96.7%. Among them, the recognition effect of the transverse hole was the best, and the precision, recall and F_1-score all reached 100 %, followed by the triangular groove, and the square groove was slightly worse. The recognition result of the rectangular groove was relatively poor, and its precision, recall and F_1-score were all 93.3%. Finally, the correlation between high contribution features and defect categories was discussed, and the difference of feature contribution and its causes were analyzed.