DeepScan: deep learning driving sustainability in the photovoltaic industry

As Chile accelerates its commitment to shut down 50% of its coal-fired power plants by 2025, photovoltaic technology has moved to the center stage of the energy transition. To ensure the long-term sustainability of this booming industry, Fraunhofer Chile created DeepScan.

Funded by Corfo through the Crea y Valida program, this advanced monitoring system replaces slow, manual inspections with a high-tech combination of drones and deep learning. Its primary mission is to automate the early detection of failures, extending module lifespans and enabling smarter, timely maintenance decisions.

The inspection process begins in the air, where a drone equipped with a specialized camera scans the plant using electroluminescence (EL). Unlike standard visual checks or thermography, EL uses electricity to effectively "X-ray" the solar cells. Functioning cells emit light, while damaged areas appear dark, allowing the system to pinpoint microscopic irregularities and internal defects that remain invisible to the naked eye.

Once the flight is complete, the digital magic happens. Images are enhanced, cropped, and segmented for clarity before being processed by advanced software powered by Convolutional Neural Networks (CNNs). Using a pre-trained algorithm, the system automatically classifies panels as defective or functional with an accuracy exceeding 90%. By generating detailed reports automatically, DeepScan eliminates tedious manual data entry and drastically reduces human error, providing the scalable innovation required for the solar industry's continued expansion.

“ViTs have revolutionized computer vision”

DeepScan incorporates Visual Transformer (ViT) technology, which approaches image analysis holistically rather than through segmented processing. These models represent a significant advancement in computer vision, offering a superior capacity to process large volumes of visual information globally compared to traditional methods.

The application of ViTs is particularly advantageous for tasks such as solar module inspection, where the overall image context is as critical as individual details. However, the deployment of this technology presents specific challenges; ViT models are data-intensive and require substantially greater computational resources for training than conventional algorithms.

“The challenge was both the physical connection of the devices and their joint functionality"

Integrating the equipment for aerial electroluminescence capture marked a critical milestone in the project, presenting significant challenges in both physical connectivity and operational synchronization. The primary objective was to seamlessly link a specialized camera to the drone’s control system, enabling real-time remote adjustments of key parameters such as exposure time, light sensitivity, and focus. Furthermore, the camera optics required specific modification to perceive the near-infrared spectrum—light invisible to the human eye but essential for the accurate detection of internal solar panel defects.