Layer inspection via digital imaging and machine learning for in-process monitoring of fused filament fabrication

We present a solution for layer inspection based on digital imaging and machine learning (ML) suitable for application to in-process monitoring of fused filament fabrication. Top-down images of the layer are captured in-process via a digital camera, decomposed into patches representing specific types of topographic patterns, and processed through a binary classifier, trained to recognize acceptable and out-of-control states in relation to the presence/absence of topographic defects. Classifiers implementing different types of ML technologies (support vector machines on dense image features, convolutional neural networks of different depths, and convolutional autoencoder) are investigated and compared in terms of performance at detecting layer defects. The generalizability of the approach to different part geometries is also discussed. A prototype implementation is illustrated through application to selected test parts. Research achievements as well as open challenges are highlighted.