Parameter-Efficient Fine-Tuning for Photovoltaic Cell Defect Classification: A Systematic Comparison of LoRA, QLoRA,...

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Automated visual inspection is crucial for photovoltaic (PV) module quality assurance.
• Adapting deep learning models for PV defect classification faces challenges like high memory requirements, imbalanced datasets, and computational constraints for edge deployment.
• Parameter-efficient fine-tuning (PEFT) methods are underexplored in PV defect classification despite their success in natural language processing.

Purpose of the Study:

• To benchmark Parameter-efficient fine-tuning (PEFT) methods, specifically Low-Rank Adaptation (LoRA) and Quantized LoRA (QLoRA), against full fine-tuning for photovoltaic (PV) cell defect classification.
• To evaluate the performance and resource efficiency of different PEFT strategies on a real-world PV defect dataset.

Main Methods:

• A ConvNeXt-Tiny backbone was used for defect classification on the POLY dataset (17,377 images) of polycrystalline PV cells.
• Four adaptation strategies were compared: full fine-tuning, LoRA (rank 8 and 16), and 4-bit QLoRA (rank 16).
• Experiments preserved natural class imbalance and used a group-aware StratifiedGroupKFold protocol to ensure robust evaluation.

Main Results:

• All PEFT variants achieved slightly higher macro-F1 scores than full fine-tuning while training significantly fewer parameters (26-52x less).
• QLoRA with rank 16 (QLoRA_r16) yielded the highest macro-F1 score (79.92 ± 0.75%), outperforming full fine-tuning (78.26 ± 0.94%).
• QLoRA_r16 demonstrated reduced peak GPU memory usage (~30% lower than full fine-tuning) and improved F1 scores for intact and surface-diffuse defect classes.

Conclusions:

• Parameter-efficient fine-tuning (PEFT) presents a promising, resource-efficient approach for photovoltaic (PV) cell defect classification.
• QLoRA, in particular, shows potential for improving classification accuracy and reducing computational demands.
• Further research is needed to isolate the effects of quantization and adapter rank, and to validate performance on embedded devices.