Benchmarking Traditional Machine Learning and Deep Learning Models for Fault Detection in Power Transformers
BHUVAN SARAVANAN and, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, PASANTH KUMAR M D, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, AARNESH VENGATESON, [email protected] National Institute of Technology, Tiruchirap · 2025
Read Benchmarking Traditional Machine Learning and Deep Learning Models for Fault Detection in Power Transformers by BHUVAN SARAVANAN and, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, PASANTH KUMAR M D, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, AARNESH VENGATESON, [email protected] National Institute of Technology, Tiruchirap (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-06295.txt.
Introduction
Power transformer faults can be classified by both classic and deep learning models, Bhuvan Saravanan and co-authors compare in a 2025 paper. The catalog creator field is garbled with emails and affiliations, while the head byline lists three authors. Accurate fault diagnosis is essential for the stability and safety of electrical power systems, since a missed fault can cascade.
A condition-monitored data set spanning ten months supplies gas concentration features, which are normalized for training. Five classic classifiers are trained, among them support vector machines, random forests, and gradient boosting. Four deep models are also evaluated, including recurrent and convolutional forms and TabNet. The study compares their fault classification performance on the same data.
Catalog record
| CREATOR | BHUVAN SARAVANAN and, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, PASANTH KUMAR M D, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, AARNESH VENGATESON, [email protected] National Institute of Technology, Tiruchirap |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 24 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2505-06295
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Benchmarking Traditional Machine Learning and Deep Learning Models for Fault Detection in Power Transformers BHUVAN SARAVANAN and, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, PASANTH KUMAR M D, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India, AARNESH VENGATESON, [email protected] National Institute of Technology, Tiruchirappalli, TamilNadu, India Abstract Accurate diagnosis of power transformer faults is essential for ensuring the stability and safety of electrical power systems. This research presents a comparative analysis of conventional machine learning (ML) algorithms and deep learning (DL) algorithms for fault classification of power transformer. Leveraging the condition monitored dataset which span for 10 months, various gas concentration features were normalized and used to train five ML classifiers—Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Random Forest (RF), XGBoost, and Artificial Neural Network (ANN) —as well as three DL models, including Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), One-Dimensional Convolutional Neural Network (1D-CNN), and TabNet. Experimental results demonstrate both conventional machine learning (ML) algorithms and deep learning (DL) algorithms performed at par: the highest accuracy among ML models was achieved by RF at 86.82%, whereas DL model 1D-CNN emerged very close at 86.30%. 0 0 footnotetext: —The source code used to implement the proposed model is publicly available at: https://github.com/BhuvanSaravanan/power_transformer_fault_detection I Introduction Power transformers are indispensable components in electrical power systems, serving as critical nodes for voltage regulation and energy distribution. Their operational integrity directly influences the stability and reliability of the entire power grid. However, transformers are susceptible to various faults due to factors such as insulation degradation, mechanical stresses, thermal overloads, and environmental conditions. These faults can lead to catastrophic failures, resulting in substantial economic losses and power outages [ 1 ] . Traditional diagnostic techniques, including Dissolved Gas Analysis (DGA), Partial Discharge (PD) monitoring, and Infrared Thermography (IRT), have been employed to detect and analyze transformer faults. While these methods have proven effective in identifying certain fault types, they often fall short in providing comprehensive, real-time diagnostics. For instance, DGA requires periodic oil sampling and laboratory analysis, which may delay fault detection [ 2 ] . PD monitoring is sensitive to external noise and may not accurately localize fault sources. IRT primarily detects surface temperatures and may miss internal anomalies. The limitations of conventional diagnostic methods have prompted the exploration of advanced techniques that leverage artificial intelligence (AI) and machine learning (ML) algorithms. These intelligent approaches can process vast amounts of data, identify complex patterns, and provide predictive insights into transformer health. For example, AI-based models have been developed to enhance the accuracy of fault classification by analyzing multiple parameters simultaneously, such as gas concentrations, temperature profiles, and vibration signals [ 3 ] . Moreover, the integration of multi-source data has emerged as a promising direction in transformer fault diagnosis. By combining electrical and non-electrical parameters, such as oil and gas characteristics, vibration signals, and thermal images, multi-source data fusion techniques can offer a holistic view of transformer conditions. This comprehensive analysis enables more accurate fault detection and prognosis, facilitating proactive maintenance strategies and reducing the risk of unexpected failures [ 4 ] . In summary, the evolution of transformer fault diagnosis is moving towards intelligent, data-driven methodologies that overcome the shortcomings of traditional techniques. By harnessing the capabilities of AI and multi-source data integration, these advanced diagnostic systems aim to enhance the reliability, efficiency, and safety of power transformers within the electrical grid. II Literature Survey Power transformers are critical components in electrical power systems, and their reliable operation is paramount. Traditional diagnostic techniques have been employed for decades to monitor and assess the health of these transformers. Among these, Dissolved Gas Analysis (DGA), Partial Discharge (PD) monitoring, and Infrared Thermography (IRT) are the most prevalent. Dissolved Gas Analysis (DGA): DGA is a widely used method for detecting incipient faults in oil-immersed transformers. It involves analyzing the types and concentrations of gases dissolved in transformer oil, which are byproducts of thermal and electrical faults. Various interpretation methods, such as the Key Gas Method, Duval Triangle, and Rogers Ratio, have been developed to diagnose specific fault types based on gas concentrations [ 5 ] . The Duval Triangle, introduced in the 1970s, utilizes the relative percentages of methane (CH 4 ), ethylene (C 2 H 4 ), and acetylene (C 2 H 2 ) to classify faults into categories like partial discharges, low-energy discharges, and high-energy discharges [ 6 ] . Despite its effectiveness, DGA has limitations, including the need for periodic oil sampling and potential delays in fault detection. Partial Discharge (PD) Monitoring: PD monitoring detects localized dielectric breakdowns within the transformer insulation system, which can be precursors to major failures. Techniques such as ultra-high frequency (UHF) sensors and acoustic emission detectors are employed to capture PD signals. However, PD detection is sensitive to external noise, and accurate localization of PD sources can be challenging. Moreover, the installation of PD monitoring equipment can be complex and costly, limiting its widespread adoption [ 6 ] . Infrared Thermography (IRT): IRT is a non-contact diagnostic technique that captures thermal images of transformer surfaces to identify hotspots indicative of abnormal heating. It is particularly useful for detecting issues like loose connections, overloading, and cooling system failures. Recent advancements have integrated IRT with machine learning algorithms to enhance fault classification accuracy . Nevertheless, IRT primarily detects surface temperatures and may not reveal internal faults. Additionally, environmental factors such as ambient temperature and emissivity can affect measurement accuracy [ 7 ] . While these traditional diagnostic methods have been instrumental in transformer maintenance, they each have inherent limitations. DGA requires manual sampling and may not provide real-time data. PD monitoring systems can be expensive and susceptible to interference. IRT offers surface-level insights but may miss internal anomalies. These constraints highlight the need for more advanced, integrated diagnostic approaches that provide comprehensive, real-time monitoring to ensure transformer reliability and longevity. III Methodology III-A Data Description Power transformers, integral to the stability and efficiency of modern power grids, require continuous monitoring to avoid catastrophic failures. Despite their generally robust construction, transformers are susceptible to both mechanical and dielectric failures , which may result from physical stress, insulation degradation, or electrical overloads. These failure mechanisms often manifest subtly, making early detection a critical element of predictive maintenance systems. In recent years, the integration of Internet of Things (IoT) devices has enabled the collection of fine-grained operational data for real-time transformer health assessment. This study uses a data set originally curated for fault detection research on power transformers [ 8 ] . The dataset spans from June 25, 2019, to April 14, 2020 , with time-stamped readings captured at 15-minute intervals . The data comprises a wide range of sensor inputs capturing vital transformer conditions such as voltage, current, oil temperature, winding temperature , and ambient environmental metrics. The dataset includes sensor readings from multiple critical points of the transformer infrastructure. These sensors monitor both electrical and thermal properties that are essential for identifying fault signatures. Table 1 summarizes the key sensors and their abbreviations, adapted from the original documentation in [ 8 ] Table I: Abbreviations and Full Sensor Name Abbreviations Sensor Name VL1 Phase Line 1 VL2 Phase Line 2 VL3 Phase Line 3 IL1 Current Line 1 IL2