Defect Detection in Fused Filament Fabrication With Artificial Intelligence

Fused Filament Fabrication (FFF) is a widely used additive manufacturing technology, but ensuring consistent print quality remains a challenge due to the frequent occurrence of defects. These defects arise from a combination of factors, including imperfections in printer control systems, inconsistencies in filament material properties, and variations in ambient environmental conditions like temperature and humidity. Existing defect detection methods often struggle with subtle geometric and dimensional deviations, which are particularly challenging to detect and classify.  Conventional, end-to-end AI approaches often require vast, manually labelled datasets of various defect types for effective training – a resource-intensive and often impractical undertaking.

This thesis addresses these limitations by developing a cost-effective and scalable automated defect detection and classification system in FFF. The research leverages the a prior knowledge of the intended design, enabling a direct comparison between the digital model and the physical printed part. This approach significantly simplifies defect identification. The research objectives were: 1) to develop an augmented reality (AR) based system for layer-by-layer detection of subtle geometric and dimensional deviations; 2) to create a machine learning framework for classifying the type of defect detected; and 3) to investigate the use of Large Language Models (LLMs) for defect classification without requiring extensive labelled data.

The developed system employed a novel AR-based approach for automated spatial alignment between the printed part (captured via images) and its digital design, eliminating the need for manual calibration. Image processing techniques were used to extract deviation data, from which engineered features are derived for machine learning-based defect classification (e.g., layer shift, stringing). The machine learning-based classifier exploits the non-linear mapping between observable deviations and underlying defect types. Finally, the research explores the use of LLMs, with a focus on prompt engineering, to further enhance classification accuracy, particularly in cases where the proposed classifier might struggle. This layered approach aims to leverage the strengths of each technique.

The AR-based deviation detection system achieved an accuracy of 86.66\% and a processing time of 1.35 seconds per layer. The machine learning-based defect classification framework achieved an accuracy of 93.41\%. The optimised LLM-based classification achieved an accuracy of 91.42\%. These results demonstrate that the integrated system can accurately detect and classify subtle defects in FFF-produced parts. The research contributes to the field by providing a novel, data-efficient, and automated FFF quality assurance solution. This can potentially reduce material waste, improve part quality, and increase production efficiency, ultimately facilitating the wider adoption of FFF technology.