Vision-Language Models for Transmission Line Fault Detection: A New Approach for Grid Reliability and Optimization.

Area of Science:

• Electrical Engineering
• Computer Vision
• Artificial Intelligence

Background:

• Effective transmission corridor fault detection is crucial for preventing widespread outages and operational costs.
• Existing methods often struggle with subtle, low-contrast defects and false positives outside the designated right-of-way.

Purpose of the Study:

• To improve the field reliability of an open vocabulary vision language backbone for power line fault detection.
• To enhance the detection of specific fault classes (e.g., insulator icing, conductor icing) without end-to-end retraining.

Main Methods:

• Integration of three domain-specific components with the Florence-2 vision language model: subclass-aware fusion, Power-Line Focus Then Crop normalization, and a corridor geo prior.
• Utilized training-free and parameter-efficient tuning modes for method evaluation.
• Developed a shared preprocessing and scoring pipeline for consistent assessment.

Main Results:

• Achieved higher accuracy in detecting skinny and low-contrast faults on unseen regions.
• Significantly reduced false alarms originating from areas outside the transmission corridor.
• Demonstrated improved score calibration within the critical triage confidence range.

Conclusions:

• The proposed domain-specific components effectively enhance an existing vision language model for robust power line fault detection.
• The methods maintain computational efficiency, making them suitable for deployment on unmanned aerial vehicles and edge devices.