Model of an automated product quality control system based on neural networks with unsupervised and semi-supervised learning

This paper addresses the problem of automated visual inspection of welded and hull structures in shipbuilding, where product reliability and operational safety critically depend on the timely and accurate detection of defects. Traditional optical inspection is limited by subjective human assessment and poor scalability, while conventional computer vision techniques suffer from data scarcity and variability of industrial environments. To overcome these limitations, an integrated mathematical framework is proposed that combines unsupervised and semi-supervised learning approaches. The architecture includes: (i) a convolutional autoencoder trained on defect-free reference samples for reconstruction-based anomaly detection; (ii) a two-stage sliding-window algorithm with dual thresholds for distinguishing weak defects from background noise under controlled false alarm rates; and (iii) a semi-supervised classification module that integrates contrastive learning with graph-based pseudo-labeling methods (k-NN and label propagation) to leverage large-scale unlabeled datasets. Joint optimization of reconstruction and discriminative representation aligns normality criteria with stable classification boundaries. Experimental validation confirms that the proposed method reliably detects both prominent and subtle defects, minimizes dependence on manual labeling, and can be seamlessly integrated into industrial quality assurance workflows. The main contribution lies in the development of a unified inspection model that fuses reconstruction-based, contrastive, and graph-driven approaches, demonstrating potential for improving reproducibility, reducing labor intensity, and enhancing the reliability of shipbuilding production.