The weld pool contains abundant information about the welding process and can thus be utilized to accurately monitor the weld penetration. This paper addresses the dynamic estimation of the weld penetration in GTAW process. A machine vision based weld pool sensing system is utilized and the 3D weld pool surface is reconstructed in real-time. Various dynamic experiments under different welding conditions are conducted to acquire data pairs for establishing the correlation between the front-side weld pool characteristic parameters and the weld penetration specified by its back-side bead width. Due to the substantial inertia of the welding process, the weld penetration can be more accurately estimated if the adjacent weld pools are used. Hence, a nonlinear dynamic Adaptive Neuro-Fuzzy Inference System (ANFIS) model is developed to estimate the weld penetration in real-time. It is found that the weld penetration can be estimated with satisfactory accuracy by the proposed online monitoring system.