ARaNet: Attention and Residual Aware Network for Resilient Digital Twins in Rail Transit Equipment Manufacturing

INTRODUCTION: In rail transit equipment manufacturing, which encompasses locomotive body welding, metro vehicle assembly, and high-speed rail carriage production, high-fidelity 3D point cloud data acquired by industrial sensors serves as the foundation for digital twin modeling, automated quality inspection, and robotic guidance. However, harsh production environments characterized by metallic dust, welding spatter, mechanical vibration, and frequent occlusions by fixtures and tooling inevitably introduce severe data corruption, including missing regions and non-uniform point density. Such degraded sensor data undermines the reliability of downstream manufacturing processes that depend on accurate 3D representations. OBJECTIVES: This paper presents an effective 3D sensor data recovery method that reconstructs complete geometric representations from corrupted partial scans, specifically targeting the data integrity challenges encountered in rail transit equipment manufacturing. The proposed approach aims to simultaneously restore global structural completeness and local geometric precision, thereby enabling resilient digital twin systems that maintain operational continuity despite sensor-induced data loss. METHODS:We propose an Attention and Residual Aware Network (ARaNet) featuring a Multi-Scale Channel-Aware Convolution encoder and a Hierarchical Residual-Aware Decoder. The encoder performs Farthest Point Sampling at multiple resolutions (2048, 1024, 512 points) and applies a channel attention mechanism to dynamically weight feature channels, emphasizing geometrically salient structures such as sharp edges, curved surfaces, and mechanical joints that are critical in rail transit component geometries, including bogie frames, car body shells, and coupler assemblies. The decoder progressively generates point clouds from coarse resolution (64 points) to medium resolution (128 points) to high resolution (2048 points), incorporating a residual refinement module that predicts coordinate offsets to rectify local geometric errors. This capability is particularly valuable for precision metrology in component manufacturing. RESULTS: Experimental evaluation on the ShapeNet and ModelNet40 datasets demonstrates that ARaNet achieves substantial improvements over benchmark models PCN and PF-Net. Quantitative assessment shows average reductions of 12.9% and 23.8% in the Gt-to-Pre and Pre-to-Gt metrics, respectively. The generated point clouds exhibit more uniform distributions and superior detail restoration, particularly for complex mechanical geometries with curved surfaces and structural joints characteristic of rail transit equipment components. Conclusion:The effectiveness of integrating channel attention with residual refinement mechanisms for industrial 3D sensor data recovery is validated. ARaNet provides a robust data recovery foundation for resilient digital twin systems in rail transit equipment manufacturing, enabling sustained operational capability and rapid geometric reconstruction even when sensor data is compromised by production environment disruptions. However, it should be noted that ARaNet is currently validated exclusively on synthetic training data, and its performance under extremely high missing-region ratios or in real industrial deployment scenarios without domain adaptation warrants further investigation.