Lightweight segmentation neural network for infrastructure maintenance in industrial digital twins

Digital twins are virtual representations of real-world objects, which offer a range of possibilities and potential benefits, especially in the case of industry, where such information can be used in a variety of ways. Simulation, optimization, or maintenance are among the most significant. In this study, we focus on an infrastructure maintenance model that may include a deep neural network model for detecting defects on metal surfaces, which are ubiquitous in the industry. Since digital twins must process data in real time and are often equipped with many different modules, it is important to prepare appropriately lightweight architectures. For this purpose, we present a lightweight neural network for fault detection, which is tested on the special case of welding faults. In our model, we used a bilateral architecture, equipping it with an attention module, thanks to which we managed to improve the performance of the bilateral network while maintaining a small number of parameters.