Researchers from the South Dakota School of Mines and Technology have developed a smart detection system that automatically identifies surface defects on aluminum gas meter lids. By combining three different deep learning models, the ensemble approach achieved over 97% accuracy in detecting cracks and cold flows, offering a faster, more reliable alternative to manual inspection that could improve safety and reduce waste in manufacturing.
In gas meter lid manufacturing, even a small defect can lead to a dangerous gas leak. Currently, in most small-and-medium aluminum die cast manufacturers, inspectors still examine these critical components manually. The inspection process is slow, expensive, facing serious labor shortage, and prone to human error.
Now, researchers from the South Dakota School of Mines and Technology have developed a smart detection system that automates this process. By combining three state-of-the-art, off-the-shelf deep learning algorithms, the system spots surface defects, such as cracks and cold flows, with over 97% accuracy and 0.95 F1-Score, and with the pre-trained model, the detection can take place within one second, significantly outperforming traditional manual checks.
“Our work shows that by integrating existing AI architectures into an ensemble, we can build a robust and practical tool for in-line quality control in traditional manufacturing," highlighted by the authors. "This approach not only improves the safety and quality of these critical components but also helps small-and-medium manufacturers reduce costs and address labor shortages in an industry that relies heavily on skilled inspectors."
The research, published in the journal Artificial Intelligence and Autonomous Systems , tackled the problem of detecting flaws in aluminum die-cast lids for residential natural gas meters. The team used a dataset of over 3,300 images, captured with inexpensive webcams and labeled by expert inspectors.
They compared three state-of-the-art AI models: a standard convolutional neural network, a residual convolution network, and a vision Transformer. While all three performed well, the team found that combining them into an ensemble model produced the best results, achieving a peak accuracy of 97.1%. This approach significantly reduces the risk of missing a defect, which is vital for products like gas meter lids where failure is not an option.
The next step is to expand the system to detect a wider variety of defect types and to integrate this technology directly into the manufacturing line, enabling real-time quality control.
The research was funded by the Competitive Research Grant from the South Dakota Board of Regents.
This paper “An Ensemble Deep Learning Approach for Surface Defect Detection in Aluminum Die-Cast Gas Meter Lids” was published in Artificial Intelligence and Autonomous Systems .
Qian W, Ayorinde O, Chen S, Guo L, Jensen D. An ensemble deep learning approach for surface defect detection in aluminum die-cast gas meter lids. Artif. Intell. Auton. Syst. 2026(1):0006, https://doi.org/10.55092/aias20260006.
Artificial Intelligence and Autonomous Systems
Experimental study
Not applicable
An ensemble deep learning approach for surface defect detection in aluminum die-cast gas meter lids
26-Jun-2026