Welding plays an important role in the metallurgic process, being a critical part of continuous processes. The early detection of welding defects is a key aspect to guarantee productivity. There are factories in which the welding testing is performed visually by an operator. In this scenario, the physiological and psychological aspects of the operator can determine the productivity due to unnecessary repetitions of welds. This paper proposes an on-line intelligent system for operator support. The goal is to reduce the unnecessary repetitions of welds. The proposed method uses data mining and machine learning techniques fed by the information extracted from the process data and from the data obtained by an infrared camera, creating an objective model that estimates the weld reliability. Flexibility and adaptability are two key concepts in the proposed design.