DEVELOPMENT OF A METHODOLOGY FOR SEARCHING FOR 3D PRINTING DEFECTS BASED ON THE RESNET MODEL

DOI 10.31673/2412-4338.2025.014545

Abstract

The article addresses the topical issue of automated defect detection in additive manufacturing (3D printing). It highlights the drawbacks of traditional quality control methods, such as visual inspection and specialized equipment usage, which are labor-intensive, costly, and subject to human errors. The necessity for developing automated, efficient, and universal approaches capable of operating in real-time and accurately identifying defects and their types is emphasized. This study proposes an approach based on deep convolutional neural networks, specifically utilizing the ResNet architecture, known for its high performance in computer vision tasks. The ResNet-50 model was chosen due to its optimal balance between network depth, fine detail recognition accuracy, and moderate computational resource requirements. A detailed description of developing a specialized dataset containing images of typical 3D printing defects such as under-extrusion, over-extrusion, layer shifts, model detachment from the print bed, and 'spaghetti' filament formation is provided. A clear methodology for organizing and annotating training data using specialized annotation tools (LabelImg, CVAT) was proposed, facilitating the creation of a high-quality dataset for model training. Modifications to the ResNet-50 architecture were also described, aimed at preserving as many initial image features as possible for effective detection of small defects. Key changes included reducing the convolution stride in the first layer (from stride=2 to stride=1) and excluding aggressive downsampling in subsequent layers, thus allowing the network to retain fine details and improve overall defect recognition accuracy. In conclusion, the proposed methodology has substantial potential for integration into automated 3D printing quality control systems, particularly suitable for resource-limited devices (such as Raspberry Pi equipped with NPU/GPU). Real-time quality monitoring, rapid defect detection, and reduction of wasteful reprints can significantly increase the economic efficiency of additive manufacturing processes.

Keywords: additive manufacturing, 3D printing, ResNet, machine learning, neural networks, automated quality control, printing defects..