Physics-Informed Photovoltaic Fault Diagnosis Using Hard-Clamped Feature Engineering and Progressive Fine-Tuning

Limited labeled fault data and discrepancies between simulated and measured signals complicate photovoltaic (PV) fault diagnosis. This study develops a transfer learning framework that combines hard-clamped electrical features with progressive fine-tuning. Four measured channels are expanded into a ten-dimensional (10D) representation containing power, an equivalent-resistance indicator, and consecutive-sample differences. A multi-scale convolutional neural network (MSCNN), bidirectional long short-term memory (BiLSTM), and self-attention form the sequence classifier. The model is pretrained on 12 simulated fault classes and adapted to five measured classes in the Costa PV dataset through partial backbone freezing and subsequent layer-wise fine-tuning. Target signal blocks are partitioned before preprocessing and window generation. With 250 training windows per class, the proposed schedule achieves 86.16% ± 0.16% accuracy and 86.18% ± 0.16% macro F1-score over three random seeds, compared with 83.49% ± 0.28% and 83.54% ± 0.28% for standard fine-tuning. Feature ablation increases accuracy from 78.45% ± 0.57% for the four measured channels to 86.16% ± 0.16% for the complete pipeline. Across 50–1000 training windows per class, progressive fine-tuning improves accuracy over standard fine-tuning by 2.45–4.80 percentage points. These results support the use of electrical feature construction and staged parameter adaptation for PV diagnosis with limited labeled field data.