Evaluating Machine Learning Algorithms for Fault Detection in Solar Cells: Performance and Limitations

Abstract

The extensive distribution, the penetration and popularity of PV have turned fault diagnostics from a general inspection into an operation necessity. With a larger PV system that becomes ever more complex, undetected fault can cause cumulative power failure with lower system performance and hence cost, especially in situations where real‑time monitoring is a prerequisite. Whilst infrared thermography, electroluminescence (EL) imaging, and current–voltage (I–V) curve analysis are common diagnostic methods, their automated implementation in industrial‑scale PV plants remains limited by practical and computational bottlenecks. We propose a multimodal fault diagnosis framework that incorporates EL image analysis with infrared thermal measurements and electrical I–V data collected by operating PV systems. Instead of depending purely on a single diagnostic method, this framework studies the performance metrics across the learning paradigms for identical experimental conditions. Based on the same dataset, a convolutional neural network (CNN) is applied and compared with established machine learning models, i.e. support vector machines (SVM) and extreme gradient boosting (XGBoost). In experimental testing, we use 2,624 real EL images. The results are as follows: CNN models have the best level of classification at 95.2%, whereas inference speed should be compatible with edge‑based processing needs. Additionally, CNN architectures are more conducive to pattern recognition (spatial fault prediction) for image data, whereas XGBoost is competitive for the structured numerical features used in predictive maintenance. The presented framework is a practical option with a high sensitivity to photovoltaic system accuracy with a low computational cost. Its results are in favor of real‑time operation and large‑scale industrial deployment, which needs both detection performance and resource constraints, respectively.