Context
Large photovoltaic plants require repeatable, standardised inspection approaches to maintain performance and availability at scale.
Challenge
Traditional drone‑imagery analysis is slow and difficult to standardise at scale. Reporting must also comply with IEC‑TS‑62446 requirements, increasing the need for faster, accurate, repeatable and cost‑efficient inspections while reducing manual analysis.
AI Solution
The solution combines drone-based thermal and RGB image capture, computer vision, cloud architecture, and deep learning to automatically detect, classify, and geolocate photovoltaic module defects. It uses image segmentation and object detection techniques, including three AI models: one to identify rows of PV panels, one to identify individual panels, and a fault classifier to detect and classify panel defects. Results are delivered through IEC-compliant reporting and an interactive dashboard with defect visualization, performance insights, and geopositioning.
Impact & evidence
Key metrics used to measure success include:
- Detection accuracy: precision and recall of defect identification and classification
- Operational efficiency: reduction in inspection time and manual analysis effort
- Asset and cost impact: improved plant availability, reduced downtime, and lower inspection cost per MW
Implementation / Adoption
Implementation requires high‑quality thermal and RGB drone imagery captured under suitable weather conditions, standardized DXF plant layouts for accurate geospatial mapping, and sufficient computing resources to support AI processing, while complying with drone regulations and data protection requirements and managing image‑quality variability.
Deployment followed a phased approach with system setup, drone data and DXF integration completed in 6–8 weeks, production rollout and user onboarding over 3–4 weeks, and realization over 2–3 months through operational use at scale, performance stabilization, and progressive efficiency and reliability improvements.
Data
Data used includes drone‑captured imagery, plant layout data (CAD floor plans (DXF)), operational metadata, historical and reference data (labelled defect datasets for AI model training and validation).
Data quality requirements include cleansing and augmentation, and new data collection is needed.
Future enhancements
Planned future enhancements include:
AI improvements: Increase robustness and accuracy of defect detection, refine the cell‑level detection approach, and extend analytics with historical trends and inspection comparisons.
Platform evolution: Enhance reporting, visualization, and human‑in‑the‑loop validation workflows to improve decision support.
