DF-3DNet: A Lightweight Approach Based on Deep Learning for 3D Telecommunication Tower Asset Classification

Abstract

The transition from 4G to 5G communication systems and the phase-out of 3G equipment have increased the demand for efficient telecommunication tower inspection and maintenance. Traditional manual methods are time-consuming and risky, prompting the adoption of unmanned aerial vehicles (UAVs) equipped with LiDAR sensors. This research introduces a framework for telecommunication tower asset inspection, utilising a lightweight, deep learning-based 3D classifier called DF-3DNet. The process involves raw 3D point cloud data collection using DJI's Zenmuse L1 LiDAR, optimal flight planning, data pre-processing, augmentation, and classification. The study focuses on two key asset classes—radio frequency (RF) panels and microwave (MW) dishes—which are prevalent in telecommunication towers. DF-3DNet, an enhanced version of PointNet, incorporates advanced data augmentation methods and class balance compensation to optimise performance, particularly when working with limited datasets. The model achieved classification accuracies of 0.6613 on ScanObjectNN, 0.8171 on ModelNet40, and 0.869 on the telecommunication tower dataset, demonstrating its effectiveness in handling noisy, small-scale data. By streamlining inspection workflows and leveraging AI-driven classification, this framework significantly reduces costs, time, and risks associated with traditional methods, paving the way for scalable, real-time telecommunication tower asset management.