In the manufacturing of hot water tanks, welding quality evaluation and fault detection is a critical operation that is still frequently conducted visually or with the help of none destructive tests, which can generate a high consumption of time and resources. To overcome this problem, many methods have been proposed for defect detection based on the classification of the welding signals. However, most of the proposed methods do not consider the problems of defect localization and the generalization from small sample size. Moreover, studies on the interactions between welding signals seem to be absent in the literature despite its importance in the formulation of the defect detection problem. In this paper, we aim to address these gaps by presenting a study on the causalities between welding signals in the circular welding of hot water tanks. Based on the findings from the causality study, we propose a method that detects and localize welding defect with high accuracy and handles the problem of small sample size. Furthermore, we present a study on the defects root cause and show the possibility of early defect prediction.