Defect detection by in-process monitoring plays a key role in the traceability and optimization of production. Many fault detection algorithms are trained on known faults. However, industrial data is generally unlabeled and certain faults are unknown or missing in the training dataset. This paper presents an original approach for the incremental discovery of new manufacturing defects, by Bayes rule and distance rejection. Rejects are analyzed periodically to determine the possible appearance of new defect cluster among them. Visualization then supports the cluster interpretation by a manufacturing expert. The approach was successfully applied to a screwing database from automotive industry.