KRISS team enhances pattern projection method with AI for 3D inspection of fuel cell surfaces during production

Researchers at the Korea Research Institute of Standards and Science (KRISS) have developed a technology for real-time detection of micro defects on the fuel cell surface during the production process. The technology is based on deep learning and enables real-time 3D measurement. It can detect defects on the surface shape in a single shot, allowing continuous monitoring of product quality without interrupting the manufacturing process.

Currently, a single-shot pattern projection method is employed for the real-time 3D measurement of the surface shape. A light with dense composite grid pattern is projected onto the surface of an object, and then the deformed pattern upon reflection is analyzed to obtain 3D information about defects or damages.

However, one limitation of this methodology is that it cannot measure surfaces with low reflectivity or various mixed patterns. For example, the metal separator, which is a core component of fuel cells, is challenging to be inspected using real-time 3D measurement because it is made of stainless steel (SUS) with an uneven surface.

To overcome this limitation, the Optical Imaging and Metrology Team of KRISS introduced an artificial intelligence (AI) algorithm to the pattern projection method. They used DYnet++, a new deep learning network developed by KRISS, which was trained with measurement data on thousands of surface shapes. This allows DYnet++ to perform real-time 3D morphology measurements of surfaces with low reflectivity or complex shapes.

DYnet++, the newly developed deep learning network by KRISS.

To apply this technology to fuel cell samples, the research team further trained the AI algorithm using data on metal separators with surface defects. The 3D morphology measurement results shows that dents and scratches on the sample surfaces, which are difficult to be determined through conventional 2D inspection, are successfully detected in a single shot because the AI algorithm acquired the application capabilities even with a small amount of data.

The 3D shapes predicted by the trained DYnet++ network

Thousands of image data with arbitrary complex patterns and their corresponding 3D shapes are acquired and used to train the designed DYnet++ network. Transfer learning is employed for the analysis of data obtained under the new environmental conditions, which requires only one-tenth of the amount of data to modify the trained model compared to what was used to build the initial algorithm.

Measurement results of a fuel cell sample with a rough surface and low reflectivity

Transfer learning is a machine learning technique that involves applying a pre-trained model to a new problem, particularly useful when there is limited data or time for training.

Based on the modified model, the 3D measurement results for various complex patterns can be predicted.

The technology developed through this research can be easily applied to the production line, regardless of the shape or size of the subject to be measured. It enables automatic inspection of defects even during a manufacturing process with external vibrations and significant temperature changes. This technology can contribute to the introduction of smart factories in various areas of the manufacturing industry, including fuel cells, by increasing productivity, improving product quality, and reducing costs.

KRISS will continue to conduct further studies to apply this technology to various industrial sites.

Resources

• M. T. Nguyen, Y. -S. Ghim and H. -G. Rhee, (2023) “DYnet++: A Deep Learning Based Single-Shot Phase-Measuring Deflectometry for the 3-D Measurement of Complex Free-Form Surfaces,” IEEE Transactions on Industrial Electronics, vol. 71, no. 2, pp. 2112-2121 doi: 10.1109/TIE.2023.3253940

steel