EXPLORING A PIX2PIX MODEL FOR INFRARED TO ELECTROLUMINESCENCE IMAGE TRANSLATION IN PHOTOVOLTAIC MODULES: A FEASIBILITY STUDY

Abstract

This study explores the feasibility of using the image-to-image translation model, Pix2Pix, in translating images of solar cells from the infrared domain to the electroluminescence domain. Even though electroluminescence imaging is a powerful fault detection tool, it can be expensive and time-consuming; in contrast, infrared imaging is more accessible and cheaper but lacks direct fault visualization. Hence, the Pix2Pix model is proposed to bridge this gap. This study focuses on fine-tuning hyperparameters of the baseline model, such as lambda weight, discriminator update frequency, and learning rate, to effectively succeed in the image translation task. Seven scenarios were tested and evaluated qualitatively, through visual inspection, and quantitatively using adversarial loss, discriminator loss, L1 loss, and total generator loss. The results conclude that the best performance is achieved at a lambda value of 20 with full discriminator update frequency and a learning rate of 0.0002, generating visually realistic electroluminescence images that preserve structural details and revealed potential fault regions while achieving the lowest generator adversarial loss of 0.9174, L1 loss of 0.1183 and total generator loss of 3.2829. Comparisons with the original Pix2Pix model trained on the Facades dataset confirmed that, despite the increased complexity and subtle feature mapping in photovoltaic imagery, the fine-tuned model generated electroluminescence images with superior structural fidelity. Although loss values did not always align with ideal reference points, this was attributed to the subtle correlation between thermal and electroluminescence imaging. Overall, the findings confirm the feasibility of electroluminescence image generation from infrared input using Pix2Pix, opening new possibilities for non-invasive fault detection in solar cells.

Date of Award	2025
Original language	American English
Awarding Institution	HBKU College of Science and Engineering