Inspection Drone – Thermal Overheating Detection

Abstract

Timely inspection of telecom infrastructure is critical to ensuring service reliability and preventing equipment failure. Traditional manual inspection methods are costly, time-consuming, and often expose personnel to safety risks. This work presents an automated inspection framework based on thermal imagery captured by drones and computer vision techniques to detect overheating anomalies in telecom infrastructure. Using a YOLO-based object detection model, critical components are localized in thermal images, and temperature statistics are extracted at both global and object levels. The proposed system enables early detection of abnormal temperature patterns and provides actionable insights for preventive maintenance. Experimental results demonstrate the effectiveness of the approach in identifying high-risk components and supporting operational decision-making.

1. Introduction

Telecommunication networks rely on a wide range of physical infrastructure components, including antennas, transmission boxes, and power units, which are often deployed in hard-to-reach or hazardous environments. Overheating of these components can lead to performance degradation, service interruptions, or permanent equipment damage.

With the increasing availability of unmanned aerial vehicles (UAVs) equipped with thermal cameras, drone-based inspection has emerged as a promising alternative to traditional manual monitoring. Thermal imagery allows operators to observe temperature distributions that are invisible in standard RGB images, making it particularly suitable for detecting overheating faults.

However, interpreting thermal images manually remains labor-intensive and subject to human error. This motivates the development of automated systems capable of detecting critical components and analyzing their thermal behavior. In this work, we propose an AI-driven inspection pipeline that combines object detection and thermal anomaly analysis to support proactive maintenance of telecom infrastructure.

2. Related Work

2.1 Drone-Based Infrastructure Inspection

Previous research has explored the use of drones for inspecting infrastructure such as power lines, bridges, and telecom towers. UAVs offer flexibility, rapid deployment, and improved safety compared to ground-based inspections. Thermal cameras have been widely adopted to identify heat-related anomalies in industrial assets.

2.2 Thermal Image Analysis

Thermal image analysis has been applied to fault detection in electrical systems, mechanical equipment, and buildings. Common approaches rely on threshold-based methods or manual interpretation. While effective in controlled environments, these methods often struggle with complex backgrounds and varying environmental conditions.

2.3 Deep Learning for Anomaly Detection

Recent advances in deep learning, particularly convolutional neural networks (CNNs), have enabled robust object detection and anomaly detection in visual data. Models such as YOLO have demonstrated strong performance in real-time detection tasks. Integrating deep learning with thermal data has shown promising results for industrial inspection, but remains an active research area.

3. Methodology

The proposed inspection system consists of four main stages: thermal data acquisition, object detection, temperature extraction, and anomaly assessment.

3.1 Thermal Data Acquisition

Thermal images are captured using a drone-mounted infrared camera during routine inspection flights. The data covers various telecom infrastructure components under different environmental and operational conditions.

3.2 Object Detection

A YOLO-based object detection model is trained to localize critical telecom components within thermal images. Bounding boxes are generated for each detected object, enabling focused analysis on regions of interest rather than the entire image.

3.3 Temperature Feature Extraction

For each thermal image, two types of temperature measurements are computed:

• Global temperature statistics, representing overall thermal conditions
• Local temperature statistics, extracted from each detected bounding box

These measurements include maximum, minimum, and average temperatures, allowing precise assessment of component-level thermal behavior.

3.4 Anomaly Detection and Risk Assessment

Detected components are analyzed by comparing their local temperature statistics against expected operating ranges. Components exhibiting abnormal temperature levels are flagged as potential risks, enabling early intervention and preventive maintenance planning.

4. Experimental Setup

4.1 Dataset Description

The dataset consists of thermal images collected from telecom infrastructure inspections. Images include various component types and temperature distributions, reflecting real-world operational variability.

4.2 Model Training

The YOLO model is trained on annotated thermal images, where bounding boxes correspond to critical infrastructure components. Standard data augmentation techniques are applied to improve generalization.

4.3 Evaluation Protocol

The system is evaluated qualitatively and quantitatively by assessing detection accuracy and the reliability of temperature-based anomaly identification. The ability to distinguish normal operating conditions from overheating scenarios is a key evaluation criterion.

5. Results and Discussion

Experimental results demonstrate that the proposed system successfully detects critical telecom components in thermal images and accurately extracts localized temperature information. The comparison between global and bounding-box temperature statistics highlights the advantage of object-level analysis, which enables precise identification of overheating components that may not be evident from global temperature averages.

The automated pipeline reduces the need for manual inspection and provides consistent, repeatable assessments of infrastructure health. By identifying thermal anomalies early, the system supports proactive maintenance strategies and helps mitigate the risk of unexpected equipment failure.

Overall, the results confirm the feasibility and effectiveness of combining drone-based thermal imaging with deep learning for intelligent telecom infrastructure inspection.

Technology Stack