An increasing attention is paid to fault detection in manufacturing. Researches are carried out in order to improve the quality and productivity in such systems. Machine Learning (ML) techniques are often used for fault detection tasks. Besides, ontology and semantic web technologies have a great potential to represent, organize and reuse the expert knowledge. In this paper, a Knowledge-based fault detection method for manufacturing processes is proposed, relying on Ontology and Machine Learning techniques. The approach is iterative in the sense that new faults can be detected by ML and added as new knowledge into the ontology periodically. It eases fault detection in industrial contexts, where faults are generally rare. Experiments conducted with seven fault-detection oriented UCI datasets have shown the effectiveness of our proposal. It is composed of a real-time classifier with a reject option, to enable the detection of new defects based on the existing knowledge described in the ontology. When new defect signature is discovered, it is added into the ontology as new knowledge through a semantic mapping. As a result, we have shown the ability of the proposed architecture to detect new faults and to increase the overall accuracy as new faults are included in the ontology. It also conducts to an evolving ontology that will be used in further research to support a generalization process to enable the detection of known defects in new contexts.