The tapping of metal components is a manufacturing task with great potential for automation, because the conditions affecting the industrial components are of limited variability. However, automation encounters two main problems: both the human- and the time-related costs associated with the manual classification of threads are excessive, and thread quality can vary greatly, due to tapping tool wear. In this study, the use of semi-supervised algorithms is proposed to improve the performance of machine learning–based models trained on real industrial datasets. The strategy was validated on a dataset of more than 7000 threads produced with 36 different tapping tools under the same working conditions involving nodular cast iron workpieces. Several algorithms were trained using datasets with different features and data processing. The best results were obtained with datasets using linear regression in which sinusoidal fluctuations in the raw signals were replaced by linear regressions and the slope of an 11-element rolling window was applied to extend the raw dataset. Algorithms were trained with different percentages of labeled datasets. The co-training-based algorithms almost systematically obtained the best results, yielding better results than the reference algorithms using a 100% labeled dataset. Besides, the proposed solution also achieved higher performance with 50% of labeled instances in the training dataset, drastically reducing the costs of manual labeling for that sort of industrial dataset.
