Multi-Fault Diagnosis in Solar Panels using Hybrid Feature Fusion Approach

Abstract

The use of photovoltaic (PV) systems is growing as a green energy source; but their output is highly dependent on environmental and operational faults including soiling and shading conditions. However, despite the successful application of previous deep learning approaches for single-fault detection, multi-faults (e.g., soiling and shading) are difficult to diagnose accurately in actual PV systems. In this research, we present a hybrid spatial feature fusion-based deep learning method for multiclass fault identification in PV panels. Our model is based on fusion of three state-of-the-art transfer learning models (EfficientNetB7, ResNet50, and MobileNetV2) to extract deep features that are then combined to provide a unified feature representation for classification. A custom dataset of 953 RGB images captured in Uttarakhand, India, representing three different states of the PV panel (clean, dusty, and multi-fault) is used to train and test the model. The proposed model outperforms each of the individual baseline models, with an average classification accuracy between 97.1% and 98.4%, precision of up to 96.7%, sensitivity between 92.3% and 100%, and specificity between 97.6% and 98.4%. It has also F1 value of 94.4% and a Matthews Correlation Coefficient (MCC) score of up to 0.965, demonstrating better predictive classification model. The improved accuracy is due to the amalgamation of various feature representations, providing increased capacity to discriminate among visually similar faults. Furthermore, the LIME-driven Explainable Artificial Intelligence (XAI) ensures interpretability by identifying key regions that contribute to the model's decision-making ability. The present study can be utilized in smart PV monitoring operations, inspection using drones and IoT-based smart solar farms, enabling early fault detection and improve energy production.