Industry 4.0-Compliant Artificial Intelligence-based power transformer fault classification method during data missing conditions

Abstract

Power transformers are fundamental components in electrical systems, responsible for the efficient transfer of energy between different voltage levels. Despite their robustness, these devices are subject to failures over time, such as internal electrical faults that can compromise not only the transformer itself but also the stability of the entire interconnected system. In this context, the use of intelligent solutions that enable continuous monitoring and accurate fault classification becomes increasingly relevant, contributing to faster diagnostics and, thus, reducing equipment downtime. This work proposes an innovative approach for classifying internal faults in power transformers, with a particular focus on those occurring in the bushings. The proposed method is based on the development of a classification system that combines advanced mathematical techniques with supervised machine learning models. A key feature of the approach is its ability to operate effectively even with incomplete data, a common condition in real industrial environments where communication failures, sensor defects, or signal noise can compromise measurement integrity and, consequently, the diagnostic process. The methodology integrates the Real-Time Bounded Stationary Wavelet Transform (RT-BSWT) with machine learning models. After applying this transform to the current signals, energy values are extracted from the coefficients and used as the input features for the classification stage. Three supervised algorithms were implemented: Decision Tree, Random Forest, and Logistic Regression. These models were selected for their strong performance in classification tasks and for their interpretability. The system was evaluated using a simulated dataset containing different types of internal faults in transformer bushings, including single-phase, two-phase, three-phase, and phase-to-ground faults, applied to both the primary and secondary sides. The tests also considered variations in fault resistance, fault angle, load conditions, noise, and, most importantly, the absence of one phase current, aiming to simulate realistic operational scenarios. The results demonstrated that the proposed approach is highly effective, achieving accuracy rates above 95% under ideal conditions. Among the tested models, the Random Forest algorithm performed the best. The effectiveness of the method was also validated through a direct comparison with a traditional wavelet-only approach, without machine learning. While the conventional method showed reasonable performance in ideal conditions, with up to 85% accuracy, its effectiveness was severely affected by data loss and noise, dropping to 36.9% accuracy in adverse scenarios. It also struggled to identify specific fault types, such as secondary-side two-phase-to-ground faults or three-phase faults under missing data. In contrast, the machine learning-based method showed higher accuracy and adaptability, correctly classifying all 20 fault types evaluated, even under signal loss and different noise levels. Beyond its technical contribution, this work aligns with the principles of Industry 4.0 by integrating intelligent data analysis, real-time responsiveness, and robustness against data loss. These characteristics significantly expand the potential for applying the proposed methodology in modern industrial environments. The system can be integrated into existing protection and monitoring architectures, providing direct support for operational decision-making. In summary, this work presents a significant contribution to the field of diagnosis and classification of internal faults in power transformers. Its ability to perform under adverse conditions and maintain high accuracy even with incomplete data makes it a promising solution for both academic research and practical applications in the electric power sector. This approach may serve as a foundation for the development of smarter protection systems, contributing to increased safety and operational efficiency.