A new study titled ‘A novel approach for fault location and defect quantification in large-scale photovoltaic plants,’ has revealed a new inspection method combining drone thermography, georeferenced mapping and transformer based deep learning for operations and maintenance on large photovoltaic plants. Tested on operating sites in Spain and the United Kingdom, the approach delivers module level fault location, objective severity scoring and a direct inspection cost of around €202 per campaign.
The workflow is built in three phases. First, it uses flight telemetry such as GPS position, altitude and gimbal angle to model the camera field of view and keep image acquisition consistent. Second, a transformer based detection model processes thermal frames to locate defects and extract pixel level temperature data. Third, it maps the plant with georeferenced visual imagery using OpenDroneMap, the Segment Anything Model version 2 and a Unidirectional Histogram technique to define panel boundaries and tie each thermal anomaly to a physical module.
In validation across two plants with 12 and 20 strings and a total of 2 460 modules, the system achieved module detection accuracy between 95% and 98.55%. Fault detection using a RoboFlow Detection Transformer model reached an overall accuracy of 96.6%, with the best case module detection at 98.55%. Across 85 identified faults, 34.1% were classified as severe, 27.1% as moderate and 38.8% as low severity, using indicators that consider intensity, spatial distribution and affected area.
Measurement quality was checked against an external infrared sensor, showing an overall accuracy of ±2 °C or ±2% of the measured value. The authors describe the result as a scalable, layout independent methodology that can be deployed in real industrial settings to improve confidence in maintenance decisions and reduce downtime.
For asset owners and O&M contractors in Africa, the method points to a practical path for routine aerial inspections that combine high diagnostic performance with predictable costs. By producing precise, georeferenced defect maps and an objective severity ranking, the approach can help prioritise module replacements, string level troubleshooting and performance recovery across large solar farms.
Link to the full paper HERE
Author: Bryan Groenendaal
June 29, 2026
July 22, 2026
July 13, 2026
July 15, 2026
June 27, 2026
July 8, 2026
Disclaimer | Privacy Policy | Terms & Conditions | Returns Policy | Intellectual Property | Cookie Policy
© 2019 – 2026 GBA Digital Media Group. All Rights Reserved | Site Credit
Copyright Green Building Africa 2026.
Subscribe to our weekly Top 5 Stories
"*" indicates required fields
You must be logged in to post a comment.