Title: Comparative analysis of machine learning algorithms for steel plate defect classification

Authors: Ioannis D. Kordatos; Panorios Benardos

Addresses: School of Mechanical Engineering, Section of Manufacturing Technology, National Technical University of Athens, Heroon Polytechniou 9, Zografou, Athens, GR15780, Greece ' School of Mechanical Engineering, Section of Manufacturing Technology, National Technical University of Athens, Heroon Polytechniou 9, Zografou, Athens, GR15780, Greece

Abstract: In manufacturing, defect detection is typically performed manually to ensure the required quality of the produced parts; however, this is a labour-intensive and time-consuming process, and therefore a lot of research focuses on its automation. This work presents an investigation of four machine learning algorithms (logistic regression, support vector machines (SVM), k-nearest neighbour (k-NN), and Random Forest (RF)), as they have been applied to defect classification in steel plate manufacturing using the Faulty Steel Plates public dataset. These algorithms have been used in combination with different techniques to determine the most appropriate model inputs and parameters. Out of the investigated combinations, the highest accuracy, equal to 82.5%, was achieved by a RF classifier with polynomial feature engineering, hyperparameter tuning and a grid search method. Finally, a graph-based, serial, three-stage custom classifier has also been developed, which has achieved an accuracy of at least 92% in every stage.

Keywords: automated part inspection; industry 4.0; multiclass classification; imbalanced learning; synthetic data generation.

DOI: 10.1504/IJMMS.2022.127211

International Journal of Mechatronics and Manufacturing Systems, 2022 Vol.15 No.4, pp.246 - 263

Received: 05 May 2022
Accepted: 20 Aug 2022
Published online: 28 Nov 2022 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article