Logically optimized and probabilistic integrated photovoltaic fault finding package based on machine learning

Introduction Artificial intelligence (AI) has been widely used to detect faults and failures in photovoltaic (PV) systems, particularly those that conventional protection devices fail to identify. However, previous AI-based approaches still face major limitations, including neglecting critical detection conditions, relying on large and complex datasets, and lacking simultaneous and accurate multi-fault detection and classification. Methods To address these challenges, a novel PV fault detection framework is proposed by combining a fuzzy logic (FL) system with a particle swarm optimization (PSO) algorithm. An initial dataset is generated from the current–voltage (I–V) curve of a PV array. Manhattan distance (MD) and Chebyshev distance (CD) features are extracted from the I–V characteristics. A wide set of machine-learning classifiers is evaluated, and the FL system nominates the most reliable models based on mean accuracy, F1-score, and standard deviation. PSO is then used to determine the optimal subset of classifiers and to assign optimized weights for ensemble prediction. Several output-combining techniques are also examined to obtain the most accurate final classification. Results Model verification is performed using a dataset that includes normal operation as well as line-to-line (LL), open-circuit (OC), and degradation (DEG) faults under various environmental (irradiance, temperature) and electrical (mismatch, impedance) conditions. The proposed FL+PSO-based model achieves outstanding accuracy in detecting and classifying multiple PV faults and outperforms recent state-of-the-art approaches. Discussion The integration of distance-based feature extraction, fuzzy-driven classifier selection, and PSO-optimized weighting significantly enhances robustness and reduces sensitivity to environmental variations. These improvements enable reliable multi-fault detection even when fault signatures closely resemble normal conditions. Conclusion The proposed FL and PSO-based ensemble provides a highly accurate and reliable solution for multi-fault detection in PV arrays. Its performance surpasses existing approaches, making it a strong candidate for practical implementation in real PV monitoring systems.