Machine Learning Case Studies for Quality Evaluations
Marc Harris, Timken Steel
The rapid boom in big data generation, data science, and machine learning has led to a massive opportunity for continuous improvement, optimization, and automation across virtually all industries. The focus here is on applications of machine learning tools within the steel industry, demonstrating the capabilities, speed, and accuracy using MathWorks® Deep Learning Toolbox. Two steel specific case studies are presented: automated non-metallic inclusion classification and coarse dimensional measurement of in-process steel production. The technique of transfer learning was employed to reduce computational overhead, and it was found that a proper ground truth training dataset with intelligent image pre-processing yielded results with better-than-human accuracies at vastly superior speeds. Implementation of finished models in a standardized dynamic-link library format provided seamless integration with other common programming languages, which led to a straightforward, easily scaling, production roll out. Advantages were immediately apparent with regard to task specific man-hours and evaluation consistency. Basic architectures, pre-processing steps, training parameters, and model performance are described for each case study.
Published: 16 Nov 2020