UAVs detect hazards with multi-directional Mamba on overhead transmission lines

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Overhead transmission line hazard detection is crucial for power systems and societal function.
• Unmanned Aerial Vehicles (UAVs) and deep learning offer advanced solutions but face challenges like complex backgrounds, small targets, and efficiency.
• Existing methods struggle with accurate and efficient detection in complex environments.

Purpose of the Study:

• To develop a novel deep learning model for enhanced overhead transmission line hazard detection using UAVs.
• To address the limitations of complex background interference, small-scale detection, and efficiency-performance balance.
• To introduce Mamba-based architectures for improved spatial modeling and detection accuracy.

Main Methods:

• Proposed the UAV-MDMamba model incorporating Mamba and State Space Models (SSMs) for linear complexity.
• Designed a Multi-Directional Mamba (MDMamba) block for superior image spatial modeling and background suppression.
• Implemented Patch-Level Inference Enhancement (PLIE) to boost the detection of small targets during inference.

Main Results:

• The UAV-MDMamba model demonstrated excellent performance on a newly collected and labeled dataset for complex hazard scenarios.
• The MDMamba block effectively captured hazardous areas, particularly in small-scale and complex background situations.
• PLIE significantly improved the detection accuracy for small targets.

Conclusions:

• The UAV-MDMamba model offers a significant advancement in overhead transmission line hazard detection.
• This research enhances both the efficiency and accuracy of detecting transmission line hazards, particularly in challenging environments.
• The proposed methods provide a robust solution for critical infrastructure monitoring using UAVs.