Inspection of the Assembly Integrity of Components in Complex Mechanical Structures From a 3-D Model to the Real World

Abstract

The assembly integrity of components in complex mechanical structures is critical for ensuring the performance and operational safety of mechanical equipment in fields such as aerospace and transportation. Although deep learning technology has shown significant potential for use in automatic visual inspection tasks, the complexity of the mechanical structures leads to data collection challenges, and the massive number of required data annotations is both labor-intensive and time-consuming. In this article, an assembly integrity inspection method is proposed in this article through a synthetic image style-transfer network (SISTN), which establishes a three-in-one technical framework of “model-driven, domain adaptation, and object detection.” In the model-domain stage, a model-driven synthetic image generation method (MDSM) is designed. By leveraging dynamic equations to adjust parameters such as 3-D model parameters, environmental lighting conditions, and background interference and integrating the constraint relationships between components, an annotated synthetic image dataset with physical consistency is automatically generated. To address the cross-domain challenges between the model domain and the physical domain, a style-transfer-based synthetic image-domain adaptation method is proposed. By incorporating a self-attention mechanism and a perceptual loss function into the CycleGAN, generalization is enabled from the model domain to the physical domain through a small number of images derived from real-world scenes, thereby enhancing the realism of synthetic images. Finally, an object detection network is employed to achieve component recognition and assembly integrity inspection. The experimental results validate the effectiveness of the SISTN, which contributes to improving the assembly quality and efficiency of complex mechanical structures.