This paper presents an experimental study where a vision camera integrates coaxially into a laser beam welding tool to monitor beam deviations (beam offset) in laser stake welding of T-joints. The aim is to obtain an early detection of deviations from the joint centreline in this type of welding where the joint is not visible from the top side. A polynomial surface fitting method is applied to extract features that can describe the behaviour of the melt pool. A nonlinear autoregressive with exogenous inputs neural network model is trained to relate eight image features to the laser beam offset. The performance of the presented model is evaluated offline by different welding samples. The results show that the proposed method can be used to guide post weld inspection and has the potential for on-line adaptive control.