Automated Keypoint Estimation for Self-Piercing Rivet Joints Using micro-CT Imaging and Transfer Learning

2025-02-23

AI Summary

Non-destructive, precise evaluation of structural integrity in self-piercing riveted (SPR) joints remains challenging in the automotive industry. Method: This paper proposes a keypoint localization method integrating micro-CT imaging and transfer learning, adopting a two-stage paradigm: synthetic-data pretraining followed by fine-tuning on limited real-world samples. We adapt a U-Net–based semantic segmentation architecture with domain-adaptation strategies to mitigate inter-scan variability in micro-CT data. Contribution/Results: To our knowledge, this is the first work to introduce keypoint localization for non-destructive SPR assessment, enabling fully automated quantification of critical parameters—including rivet head height, interlock length, and bottom sheet thickness. On real test samples, the mean localization error across three keypoints is <0.015 mm, and parameter measurement accuracy reaches 98.2%, substantially outperforming conventional destructive testing. The framework demonstrates strong potential for inline, non-destructive quality inspection in industrial production environments.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Intelligent Robots: Localization, Mapping, and Navigation
• Computer Vision: Segmentation

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Security and Privacy: Large-scale security measurements
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

The structural integrity of self-piercing rivet (SPR) joints is critical in automotive industries, yet its evaluation poses challenges due to the limitations of traditional destructive methods. This research introduces an innovative approach for non-destructive evaluation using micro-CT imaging, Micro-Computed Tomography, combined with machine vision and deep learning techniques, specifically focusing on automated keypoint estimation to assess joint quality. Recognizing the scarcity of real micro-CT data, this study utilizes synthetic data for initial model training, followed by transfer learning to adapt the model for real-world conditions. A UNet-based architecture is employed to localize three keypoints with precision, enabling the measurement of critical parameters such as head height, interlock, and bottom layer thickness. Extensive validation demonstrates that pre-training on synthetic data, complemented by fine-tuning with limited real data, bridges domain gaps and enhances predictive accuracy. The proposed framework not only offers a scalable and cost-efficient solution for evaluating SPR joints but also establishes a foundation for broader applications of machine vision and non-destructive testing in manufacturing processes. By addressing data scarcity and leveraging advanced machine learning techniques, this work represents a significant step toward automated quality control in engineering contexts.

Problem

• Automated keypoint estimation for SPR joints
• Non-destructive evaluation using micro-CT imaging
• Transfer learning for real-world data adaptation

Innovation

• Micro-CT imaging for non-destructive evaluation
• Transfer learning with synthetic data
• UNet architecture for keypoint localization