Performance improvement of CNN-based model for multiclass hotspot severity classification in photovoltaic thermal imagery

Intelligent diagnostic models that can identify and categorise thermal anomalies affecting efficiency and safety are essential for the reliable operation of photovoltaic (PV) systems. One of the most serious defects is thermal hotspots, which frequently result in energy loss and potential fire hazards. Convolutional Neural Networks (CNNs) have demonstrated promising performance when applied to aerial thermal images for PV fault detection. However, most existing models are limited to binary fault identification, thereby constraining their applicability for risk-based maintenance. Conventional CNN architectures also lack an optimised flatten-layer representation and fail to capture the full variability in hotspot severity. This study presents an enhanced CNN-based framework for multiclass classification of hotspot severity in PV modules. To improve feature abstraction and class separability across severity levels, the proposed model employs a staged architectural refinement strategy by progressively adding convolutional layers to a baseline CNN. Hotspot regions extracted from unmanned aerial vehicles (UAV)-acquired thermal imagery were categorised into low, medium, and high severity levels for model training and evaluation. The improved model achieved a 9.1% increase in accuracy and a 7.6% increase in F1-score, outperforming the baseline CNN, thus confirming its superior discriminative learning capability and diagnostic robustness. These findings demonstrate that architectural deepening and flattenlayer optimisation can advance automated thermographic inspection toward severity-aware predictive maintenance for PV systems.