JEFFNet hybrid AI boosts solar panel fault detection accuracy by 94%

A team of researchers from multiple institutions, including Seyyedhamid Azimidokht and Mehdi Monemi, has developed JEFFNet (JEPA-EFFicientNet), a multibranch architecture for classifying faults in solar photovoltaic (PV) panels using thermal infrared (IR) images. The model fuses semantic representations from a JEPA-pretrained Vision Transformer with convolutional features from EfficientNetV2-S, enabling complementary learning that addresses challenges like class imbalance, limited texture, and subtle thermal differences. JEFFNet was evaluated on two public datasets: PVF-10 (10 and 2 classes) and InfraredSolarModules (ISM, 12 and 2 classes). On PVF-10 10-class, it achieves an F1-score of 93.21% and accuracy of 94.33%; on the binary task, 97.53% F1 and 96.41% accuracy. On ISM, the 12-class F1 is 72.60% and accuracy 83.88%, while binary performance reaches 94.69% F1 and 94.78% accuracy.

Notably, JEFFNet uses only 108.6 million parameters, a 47.2% reduction compared to the 205.91M parameters of GEPFNet, making it more parameter-efficient without sacrificing accuracy. The architecture's hybrid approach—self-supervised semantic features combined with supervised convolutional features—proves effective for thermal IR PV fault classification. The researchers have made the source code publicly available on GitHub. This work has significant practical implications: as solar PV systems scale rapidly, manual inspection becomes impractical. JEFFNet offers a scalable, non-contact solution that can accurately identify faults from thermal images, reducing downtime and maintenance costs for solar farms.