Defect detection in laser-based powder bed fusion process using machine learning classification methods

The aim of this study is to deploy machine learning (ML) classification methods to detect defective regions in additive manufacturing, colloquially known as 3D printing, particularly for the laser-based powder bed fusion process. A custom-designed test specimen composed of 316L was manufactured using EOS M 290 machine. Multinomial logistic regression (MLR), artificial neural network (ANN), and convolutional neural network (CNN) classification techniques were applied to train the ML models using optical tomography infrared images of each additively manufactured layer of test specimen. Based on the trained MLR, ANN, and CNN classifiers, the ML models predict whether the manufactured layer is standard or defective, yielding five classes. Defective layers were classified into two classes for lack of fusion and two classes for keyhole porosity. The supervised approach yielded impeccable accuracy (>99%) for all three classification methods, however CNN inherited the highest degree of performance with 100% accuracy for independent test dataset unfamiliar to the model for unbiased evaluation. The high performance and low cost of computing observed in this work can have the potential to detect and eliminate defective regions by tuning the processing parameters in real time resulting in significantly decreased costs, lead-time, and waste. The proposed quality control can enable mass adoption of additive manufacturing technologies in a vast number of industries for critical components that are design- and shape- agnostic.