Few-Shot Class-Incremental Continual Learning-Based Fault Diagnosis of Photovoltaic Arrays Using Prompt-Guided Knowledge Distillation and I-V Curves

Zheng, H, Wu, L ORCID: 0000-0003-0468-3294, Chen, Z ORCID: 0000-0002-3471-6395, Cheng, S, Lin, P, Long, C ORCID: 0000-0002-5348-8404 and Jiang, L ORCID: 0000-0001-6531-2791 (2026) Few-Shot Class-Incremental Continual Learning-Based Fault Diagnosis of Photovoltaic Arrays Using Prompt-Guided Knowledge Distillation and I-V Curves IEEE Internet of Things Journal, 13 (13). pp. 29395-29407. ISSN 2327-4662, 2327-4662

Text bare_jrnl_new_sample4(1)(2).pdf - Author Accepted Manuscript Available under License Creative Commons Attribution. Download (4MB) | Preview

Abstract

Photovoltaic (PV) arrays usually operate in harsh outdoor environments for decades, exposing them to complex and evolving faults and thus necessitating continuous fault diagnosis over their full lifecycle. However, new fault classes incrementally emerge with extremely limited labeled samples, posing the few-shot class-incremental learning (FSCIL) challenges for continuous fault diagnosis. To tackle the issues, this article proposes a novel prompt prototype network (PPN), which integrates prompt learning with knowledge distillation to address both the stability-plasticity dilemma and the sample scarcity in FSCIL. The PPN framework employs global task-invariant and variational task-specific prompts (T-Prompts) to modulate feature extraction through parallel distillation and prototype branches, simultaneously directing the distillation branch to preserve prior fault knowledge and guiding the prototype branch to rapidly adapt to new fault patterns. Experiments are performed to emulate real-world PV lifecycle fault increment scenarios and fault evolution processes across normal operating conditions and 17 fault types that progress from basic single faults to complex compound failures. The results show that the proposed approach achieves 97.15% accuracy (ACC) with only 0.48% catastrophic forgetting under six-way five-shot settings, surpassing TagFex by 3.75% ACC while reducing forgetting by 87.27%. Moreover, our method maintains robust performance across varying incremental granularities and data scarcity levels, while achieving accelerated real-time inference and lightweight model updates on edge hardware. Therefore, this article establishes a practical FSCIL solution for PV fault monitoring, enabling continuous fault diagnosis and rapid adaptation to new emerging fault patterns in the PV system.

Item Type:	Article
Uncontrolled Keywords:	46 Information and Computing Sciences, 40 Engineering, 4602 Artificial Intelligence, 6 Clean Water and Sanitation
Divisions:	Faculty of Science & Engineering Faculty of Science & Engineering > School of Engineering Faculty of Science & Engineering > School of Engineering > Electrical Engineering and Electronics
Depositing User:	Symplectic Admin
Date Deposited:	29 Apr 2026 09:55
Last Modified:	14 Jul 2026 16:11
DOI:	10.1109/JIOT.2026.3686213
Related Websites:	Publisher
URI:	https://livrepository.liverpool.ac.uk/id/eprint/3198212
Disclaimer:	The University of Liverpool is not responsible for content contained on other websites from links within repository metadata. Please contact us if you notice anything that appears incorrect or inappropriate.