APPLICATION OF ARTIFICIAL INTELLIGENCE AND EXPERT SYSTEMS IN POWER TRANSFORMER DIAGNOSTICS

INTERNATIONAL SCIENTIFIC JOURNAL SCIENCE AND INNOVATION SPECIAL ISSUE “MODERN PROBLEMS AND PROSPECTS FOR THE DEVELOPMENT OF DIGITAL TRANSFORMATION IN ENERGY” SEPTEMBER 24, 2025 34 APPLICATION OF ARTIFICIAL INTELLIGENCE AND EXPERT SYSTEMS IN POWER TRANSFORMER DIAGNOSTICS Аbdullabekova D.R., Qutbidinov O.M., Shamsiyev J.X. Tashkent University of Information Tech-nologies named after Muhammad al-Khwarizmi, Uzbekistan,Tashkent https://doi.org/10.5281/zenodo.17557981 Abstract. This paper presents a study on the application of artificial intelligence methods (machine learning, deep learning) and expert systems for power transformer diagnostics. A multilevel architecture is proposed that integrates classical diagnostic channels (DGA — dissolved gas analysis, thermography, partial discharges, vibration diagnostics) with machine learning models and a fuzzy aggregator to calculate an integrated Health Index. The study includes statistical data on transformer failures, synthetic and interpretable model outputs, as well as example visualizations: distribution of failure causes, membership functions for DGA, and thermogram analysis. Keywords: artificial intelligence, machine learning, expert system, transformer diagnostics, DGA, thermography, Health Index, predictive maintenance Introduction. Relevance and problem: Power transformers are key elements of power systems, and their reliability determines the stability of power supply. Traditional diagnostic methods (DGA, thermography, partial discharges, vibration analysis) provide local information that requires expert interpretation. With the development of AI and data availability, it becomes possible to create automated diagnostic systems that enhance accuracy and predictability of failures. Objective of the study: To develop the structure of an integrated diagnostic system combining expert rules and machine learning models to improve the accuracy of condition classification and early fault detection. Methods. System аrchitecture 1. Data acquisition: DGA (concentrations of H2, CH4, C2H2, C2H4, CO, CO2), thermography (IR), partial discharge (PD) signals, vibration signals, electrical parameters (currents/voltages, resistance). 2. Preprocessing: normalization, noise filtering of thermograms (median filtering), data augmentation for thermography. 3. Feature extraction: gas ratios (Rogers' ratios, Duval triangle features), thermogram statistics (maximum temperature, hotspot contour), spectral features of PD and vibrations. 4. Modular classifiers: Random Forest (RF) for DGA, Convolutional Neural Network (CNN) for thermograms, SVM for PD features. 5. Expert fuzzy aggregator: converts probabilistic outputs of modules into linguistic categories (Normal, Warning, Critical) and calculates the Health Index (weighted average). Dataset and Statistics (open and synthetic data used) INTERNATIONAL SCIENTIFIC JOURNAL SCIENCE AND INNOVATION SPECIAL ISSUE “MODERN PROBLEMS AND PROSPECTS FOR THE DEVELOPMENT OF DIGITAL TRANSFORMATION IN ENERGY” SEPTEMBER 24, 2025 35 For demonstration purposes, aggregated statistical reports (international surveys), public DGA datasets, and synthetically generated thermograms were used. Approximate figures and diagrams are provided in the paper (Figs.1–3). Results. Failure Statistics (aggregated/synthetic data) Average annual transformer failure rate: 0.1–0.3% per unit per year (depending on voltage class and region). Most serious failures are associated with: windings and internal connections (~35–45%), bushings (~15–25%), tap changers (~10–20%), operational errors and other factors (remaining share). This fig.1 shows the statistical distribution of the main causes of power transformer failures. The chart demonstrates that winding and internal connection failures account for the largest share (~40%), followed by bushing failures (~20%) and on-load tap changer (OLTC) issues (~15%). Other categories include cooling system problems and operational errors. The figure highlights the critical areas that should be prioritized in preventive diagnostics. Fig.1 distribution of causes of transformer failures Effectiveness of modular classifiers (sample performance). DGA (Random Forest): accuracy 86–92%, F1-score 0.84–0.90. Thermography (CNN): accuracy 82–90% in hotspot detection and classification. PD (SVM): accuracy 75–85% depending on signal quality. Aggregated Health Index. Example of HI calculation for a transformer: 𝐷𝐺𝐴 𝑠𝑐𝑜𝑟𝑒 = 0.9, 𝑇ℎ𝑒𝑟𝑚𝑜𝑔𝑟𝑎𝑝ℎ𝑦 𝑠𝑐𝑜𝑟𝑒 = 0.6, 𝑃𝐷 𝑠𝑐𝑜𝑟𝑒 = 0.4, 𝑤𝑒𝑖𝑔ℎ𝑡𝑠 = [0.4, 0.35, 0.25] ⇒ 𝐻𝐼 = 0.90.4 + 0.60.35 + 0.4 ∗ 0.25 = 0.36 + 0.21 + 0.10 = 0.67 (𝐶𝑜𝑛𝑑𝑖𝑡𝑖𝑜𝑛 − "𝑊𝑎𝑟𝑛𝑖𝑛𝑔"). INTERNATIONAL SCIENTIFIC JOURNAL SCIENCE AND INNOVATION SPECIAL ISSUE “MODERN PROBLEMS AND PROSPECTS FOR THE DEVELOPMENT OF DIGITAL TRANSFORMATION IN ENERGY” SEPTEMBER 24, 2025 36 This fig.2 illustrates fuzzy membership functions for dissolved gas analysis (DGA) parameters, including hydrogen (H₂), methane (CH₄), and acetylene (C₂H₂). Each curve represents the degree of membership of the gas concentration to linguistic categories such as “Normal,” “Warning,” and “Critical.” The figure demonstrates how fuzzy logic enables smooth transitions between states, ensuring interpretability and robustness of diagnostic decisions. Fig.2 — membership functions for DGA Discussion. Advantages: integration of multi-channel data improves diagnostic accuracy and reliability; fuzzy logic ensures interpretability and compatibility with expert rules. Limitations: model performance depends on the size and labeling of datasets; it is necessary to consider transformer age, operating conditions, and design features. Future prospects: implementation of online monitoring (IoT), continuous model updating (on-the-fly learning), and explainable AI (XAI) for increasing engineers’ trust. Conclusions. The proposed architecture demonstrates the practical applicability of combining expert systems and AI for power transformer diagnostics. The suggested aggregation scheme and illustrative results show potential for reducing unexpected failures and optimizing preventive maintenance. REFERENCES 1. Mirowski P., LeCun Y., "Statistical Machine Learning and Dissolved Gas Analysis", IEEE Trans. Power Delivery, 2012. 2. CIGRE Reports — Transformer reliability surveys, 2013–2020. 3. Sutikno H. et al., "Machine learning based multi-method interpretation to enhance DGA...", 2024. 4. Elmuradovich E. D. et al. ENERGIYA TIZIMLARIDA INTELLEKTUAL O ‘LCHASH VOSITALARINI QO ‘LLASH MASALALARI //Science and innovation. – 2023. – Т. 2. – №. Special Issue 8. – С. 430-434. 5. Abdelwahab S.A.M., "Transformer fault diagnosis intelligent system", Sci Rep, 2025. 6. Recent reviews on AI in transformer diagnostics (2023–2025). 7. Эшмурадов Д. Э., Айтбаев Т. А. Разработка технологии изготовления полупроводниковых датчиков температуры //Texnika yulduzlari ilmiy jurnali. – №. 1-2. – С. 86-90.