Towards a transparent weakly supervised framework for thermographic PV defect classification and anomaly localisation

Severe class imbalance and lack of expert pixel-level annotations hinder automated PV module inspection via thermography. This study presents a transparent weakly supervised deep-learning framework for module-level defect classification and coarse anomaly localisation. It employs a SpectraNet-based encoder with a Focus Gate Module (CBAM-style attention), Residual Refinement Units, Spatial Compression Modules, and Channel-Selective Filter Blocks. The framework addresses the rarity of defective modules and missing expert masks in public datasets. Classification uses the original ~1:60 imbalance with focal loss and inverse-frequency weighting (no synthetic resampling). It achieves Defect Detection Rate 0.96, ROC AUC 0.998, and PR AUC 0.997. Localisation relies on artificial pseudo-masks from thermal intensity thresholding and yields F1 0.65, IoU 0.48, and Dice 0.65 (agreement with pseudo-labels, not expert boundaries). Results indicate utility for module-level triage and coarse visual evidence of thermally salient anomalies under limited annotations, but reflect internal dataset performance rather than cross-environment robustness. The work supplies a practical baseline while acknowledging pseudo-label noise, a small defective class, normalised 8-bit inputs, and absence of expert validation. Future efforts will add calibrated radiometric data, expert masks, leakage-aware validation, and cross-site testing.