Solar photovoltaic module multi-class anomaly detection using MobileNet architecture with skip connections

Reliable detection of photovoltaic (PV) module anomalies is essential for improving the operational reliability and maintenance of solar energy systems. Infrared thermographic imaging provides a non-invasive means of identifying abnormal thermal patterns associated with module faults. However, accurate multiclass classification remains challenging because of similarities between fault signatures, variations in thermal characteristics, and imbalanced distributions of defect categories. This study investigates a lightweight deep-learning framework based on MobileNet with conditional skip connections for automated PV module anomaly classification using infrared thermographic images. The architecture employs depthwise separable convolutions to support computational efficiency while learning discriminative features from thermal images. Conditional skip connections are incorporated to facilitate feature propagation where the input and output dimensions are compatible. The framework is evaluated across binary and multiclass classification settings, with its performance examined using classification accuracy, class-wise evaluation metrics, confusion matrices, and training and validation behaviour. The results indicate the potential of the proposed architecture to distinguish normal operating conditions from anomalous modules and classify multiple fault categories using a compact network. Nevertheless, similarities between certain defect classes and variations in class representation remain important challenges for reliable fault identification. This study contributes to the development of lightweight image-based diagnostic methods for PV inspection. Further validation using independent datasets, class-balanced evaluation metrics, controlled architectural ablation, and standardized computational benchmarks is necessary to establish generalizability and practical deployment suitability.