LiDAR-Simulated Multimodal and Self-Supervised Contrastive Digital Twin Approach for Probabilistic Point Cloud Generation of Rail Fasteners

Abstract

This study presents a novel deep-learning framework designed to efficiently generate high-fidelity three-dimensional (3D) point clouds of rail fasteners. The proposed method overcomes limitations associated with traditional light detection and ranging (LiDAR) technology, including high cost and incomplete data acquisition. The framework utilizes transformers and Bayesian neural networks for self-supervised contrastive learning, enabling real-time 3D point cloud generation from depth and grayscale images. This approach fosters automation and efficiency in maintenance tasks by producing accurate geometric reconstructions of rail fasteners. Beyond 3D generation, the framework learns latent representations of two-dimensional (2D) data, potentially impacting downstream tasks. This research contributes significantly to the development of digital twins in infrastructure monitoring by providing a cost-effective and efficient pathway for real-time 3D data generation. The proposed methodology holds promise for enhancing safety, reliability, and cost-effectiveness across a broad spectrum of infrastructure applications.