Human-in-the-loop machine learning system for collaborative defect analysis in industrial 3D printing operations

The defect pattern complexity, limited availability of labelled training data, and real-time monitoring requirements are major challenges to quality assurance in industrial additive manufacturing. This paper provides a human-in-the-loop machine learning approach to collaborative defect detection in the 3D printing process. It integrates multi-modal sensing with uncertainty-directed active learning. The suggested HITL-MBCNN model uses a multi-branch convolutional neural network that integrates RGB imaging, thermal data, vibration data, and acoustic data as its inputs, where an attention mechanism is used to adjust the weight of sensor data. Monte Carlo Dropout measures uncertainty to select strategic samples, whereas explainable AI components assist in the decision-making process of operators. Evaluations were done on 56,600 samples in six different categories of defects, with an experimental assessment of 94.7% classification accuracy and 93.6% F1-Score, and active learning requiring 58.7% fewer samples to be annotated compared to random sampling. Edge deployment achieves 51.0 fps inference. Four-week industrial pilot proved that it was possible, showing a 72.9% decrease in defect rate, a 62.8% decrease in material waste and a 98.2% decrease in mean detection time.