A Hybrid Deep Learning Framework for Multi-modal PV Fault Detection

TY - GEN

T1 - A Hybrid Deep Learning Framework for Multi-modal PV Fault Detection

AU - Osei, Stepehen

AU - Bamisile, Olusola

AU - Christopher, Wisdom C.

N1 - ©2026 by the Institute of Electrical and Electronics Engineers, Inc. All rights reserved

PY - 2026/6/16

Y1 - 2026/6/16

N2 - The rapid global adoption of photovoltaic (PV) systems has intensified the need for reliable, efficient, and intelligent fault detection mechanisms to ensure sustainable energy production and grid stability. Traditional fault detection methods, such as manual inspections, I-V curve analysis, and threshold-based algorithms, often fail to generalize across diverse operational conditions, leading to delayed diagnosis and system inefficiencies. To address these limitations, this study proposes a novel deep learning framework for photovoltaic fault detection that integrates multi-modal data sources and hybrid model architectures. The proposed framework leverages electrical measurements (I-V curves, voltage, current, and power signals) and thermal imaging data for comprehensive fault analysis. A hybrid CNN-BiLSTM-Attention model is developed to capture both spatial and temporal fault patterns, while an attention-based feature fusion mechanism enables the model to focus on the most discriminative signal regions. The framework is trained and evaluated using both simulated and real-world PV datasets collected under diverse environmental and operational conditions. Extensive experiments demonstrate that the proposed model achieves superior accuracy, precision, recall, and F1-score compared to baseline architectures such as standalone CNN, LSTM, and traditional machine learning classifiers. Visualization through t-SNE and saliency maps confirms the interpretability of the fused latent features and the model's ability to localize fault regions effectively. Furthermore, robustness analysis under varying noise and irradiance conditions validates the generalization capability of the model. This research not only advances the state-of-the-art in PV fault detection but also contributes a scalable, interpretable, and real-time deployable solution for industrial applications.

AB - The rapid global adoption of photovoltaic (PV) systems has intensified the need for reliable, efficient, and intelligent fault detection mechanisms to ensure sustainable energy production and grid stability. Traditional fault detection methods, such as manual inspections, I-V curve analysis, and threshold-based algorithms, often fail to generalize across diverse operational conditions, leading to delayed diagnosis and system inefficiencies. To address these limitations, this study proposes a novel deep learning framework for photovoltaic fault detection that integrates multi-modal data sources and hybrid model architectures. The proposed framework leverages electrical measurements (I-V curves, voltage, current, and power signals) and thermal imaging data for comprehensive fault analysis. A hybrid CNN-BiLSTM-Attention model is developed to capture both spatial and temporal fault patterns, while an attention-based feature fusion mechanism enables the model to focus on the most discriminative signal regions. The framework is trained and evaluated using both simulated and real-world PV datasets collected under diverse environmental and operational conditions. Extensive experiments demonstrate that the proposed model achieves superior accuracy, precision, recall, and F1-score compared to baseline architectures such as standalone CNN, LSTM, and traditional machine learning classifiers. Visualization through t-SNE and saliency maps confirms the interpretability of the fused latent features and the model's ability to localize fault regions effectively. Furthermore, robustness analysis under varying noise and irradiance conditions validates the generalization capability of the model. This research not only advances the state-of-the-art in PV fault detection but also contributes a scalable, interpretable, and real-time deployable solution for industrial applications.

KW - CNN-BiLSTM-Attention

KW - Deep Learning

KW - Fault Detection

KW - Multi-Modal Data Fusion

KW - Renewable Energy

UR - https://www.scopus.com/pages/publications/105043648372

U2 - 10.1109/AEEES69423.2026.11556798

DO - 10.1109/AEEES69423.2026.11556798

M3 - Conference contribution

AN - SCOPUS:105043648372

SP - 983

EP - 989

BT - 8th Asia Energy and Electrical Engineering Symposium, AEEES 2026

PB - Institute of Electrical and Electronics Engineers Inc.

T2 - 8th Asia Energy and Electrical Engineering Symposium, AEEES 2026

Y2 - 27 March 2026 through 30 March 2026

ER -