Abstract: Online monitoring of welding process and quality is of great significance for the quality assurance of aluminum alloy gas tungsten arc welding (GTAW). An online welding status monitoring method based on molten pool visual sensing and explainable machine learning was proposed. Firstly, a passive visual sensing system was used to acquire frontal molten pool images. Secondly, a image multi-region processing algorithm was proposed to extract geometric features of the molten pool's front-end area and straightness feature of the back-end area in real time, and the correlation between the extracted visual features and the welding status was analyzed. Finally, a CatBoost model was established to identify typical welding status, and SHAP (Shapley Additive exPlanations) method was introduced to explain the model. The experimental results show that compared with the geometric features of the molten pool's front-end area which are the focus of the existing methods, the straightness feature of the molten pool's back-end area is more effective in characterizing the molten pool state during the GTAW process. The proposed method can accurately identify the three typical welding status with an average accuracy of 96.15%, and can quantitatively analyze the influence of each feature variable on the recognition results.