Ex-situ porosity classification in metallic components by laser metal deposition: A machine learning-based approach

Abstract

Porosity in AM processes for metals is a recurrent problem which can lead to adverse effects such as crack initiation and ultimately to parts’ early-life failure. Daily tasks in the additive manufacturing industry are the metallographic analysis of the manufactured components. This guarantees to be able to define efficient process windows that guarantee the quality and repeatability of the components. Derived from this, it becomes necessary to have technological tools that facilitate the analysis of defects, particularly porosity. A new method for segmentation and classification of porosity based on image processing and machine learning techniques is proposed. The method is robust to any type of image, in RGB or gray-scale, acquired interchangeably with any microscope. The images are processed to eliminate noise and highlight defects. Then, an analysis based on the Hessian response is performed to segment the pores and calculate the numerical attributes that individually describe the defects. A new version of the random forest algorithm is proposed, which is used to build, train, and validate a porosity classification model. Our proposal is based on a feature reduction strategy, called guided regularization and it generates a subset of representative, non-redundant features. A new metric is also introduced to evaluate the importance of the extracted features. A comparative study is conducted to validate the accuracy of the proposed method. The results show that our proposal has better performance with a lower OOB error rate of 7.8% and outliers in a range of 11–13%. These results guarantee high accuracy in the porosity classification. Our proposal does not require large volumes of data or high computational capacity to train a classification model as other machine learning approaches require, which gives a competitive advantage. The proposed method may potentially be applied to other AM processes.