Aug 2026· Electrica· Vol 26· 0 citations· 29 references
TL;DR
This article explores machine learning techniques (MLTs) as a modern alternative to enhance the interpretation of DGA data for early-stage fault detection in service transformers, and demonstrates that random forest and gradient boosting outperform others, achieving up to 98% accuracy.
Abstract
The service transformer stands as a pervasive and essential component within the energy infrastructure. Service transformers are critical elements in power distribution, and their longevity hinges on effective monitoring. Traditional fault detection methods, especially dissolved gas analysis (DGA), though widely used, often suffer from delays due to manual sampling and lab-based analysis. This article explores machine learning techniques (MLTs) as a modern alternative to enhance the interpretation of DGA data for early-stage fault detection in service transformers. The method relies on the use of multiple DGA datasheets to investigate the fault typing capabilities and suitability of different MLTs. It evaluates multiple algorithms—logistic regression, support vector machines (SVM), random forest, and gradient boosting—using actual DGA datasets. To validate the best class algorithms, this article also looks at performance accuracy and then evaluates the top-performing algorithm. The results demonstrate that random forest and gradient boosting outperform others, achieving up to 98% accuracy, and are especially useful for condition monitoring professionals dealing with insulating oil analysis, as compared to the accuracy achieved of 81.4% with SVM and 76% with artificial neural network (ANN) in the case of previously published work.
This study presents a novel diagnostic methodology for oil-filled transformers, leveraging dissolved gas analysis (DGA) and a no-code machine learning framework. Conventional diagnostic methods such as Duval Triangle and gas ratio techniques are constrained by limited adaptability to modern transformer designs, susceptibility to misdiagnosis, and reliance on expert interpretation. Meanwhile, artificial intelligence (AI) approaches often demand significant programming expertise, large datasets, and computational resources. To address these challenges, the research introduced a unified machine learning model developed using Microsoft Azure Machine Learning, enabling field engineers to build and deploy diagnostic models without coding. The model utilized concentrations of 5 key gases: methane, ethylene, acetylene, ethane, and hydrogen, as input features to classify transformer conditions into 6 fault classes, including a no-fault state. 3 algorithms, namely Gradient Boosting, Random Forest, and Logistic Regression were evaluated, with Gradient Boosting consistently outperforming other models and Duval Triangle, achieving the highest accuracy of 81.3% with 2 false negatives. When the output labels were consolidated to 4 fault types, diagnostic accuracy improved to 95.6% without any false negative. Model performance was optimized through experiments involving train-test ratio variation and feature scaling. The model was further validated through field trials on actual transformers, demonstrating high reliability and alignment with forensic findings. This no-code approach empowers engineers to integrate domain expertise directly into model development, enhancing diagnostic precision and operational efficiency. The findings support scalable, interpretable, and proactive maintenance strategies, contributing to resilient infrastructure and sustainable energy systems.
Chee-Ying Chan, Salmey Abdul Halim, N. M. Mohd Nor et al.· Jurnal Kejuruteraan· 0 citations
Transformers play an indispensable role in any power system. The health condition of these devices should be the top priority. Early fault detection of these devices is essential to have sustainable power flow. There are many routinary transformer tests like winding test, furan analysis, insulation resistance tests, but dissolved gas analysis stands to be one of the most critical tests among others. This is to the fact that the DGA test can evaluate the major condition of the transformer. Dissolved Gas Analysis (DGA) methods, while widely used, often struggle with accuracy and scalability under complex fault scenarios. This paper proposed a novel ML-based DGA framework that integrates the IEEE standard with Principal Component Analysis (PCA) and Gradient Boosting Machine (GBM) to enhance transformer fault diagnosis. PCA captures 95% of the variance with five principal components. The framework showed a test accuracy of 87.5% and a cross-validation accuracy of 86.05%, outperforming traditional methods such as the Duval Triangle (83.08%) and IEC Ratio Method (82.05%), as well as other machine learning models, including Random Forest (77%) and Support Vector Machines (37%). These findings demonstrate the effectiveness of the framework as a soft sensor in Transformer diagnostics.
Apolinario Awit, Joseph Jay Brañanola, Maria Vina Presbitero et al.· international journal of eng...· 0 citations
The diagnosis of faults in power transformers (PT) plays a crucial role in ensuring the reliability and stability of power systems. However, traditional fault detection techniques are prone to low detection accuracy and are not stable enough under complex operating conditions. To address these challenges, this paper proposes a novel hybrid Variational Mode Decomposition (VMD), Graph Attention Network (GAN) with Adaptive Extreme Learning Machine (AELM) method for a transformer fault diagnosis framework. VMD can efficiently extract discriminative frequency bands of the transformer signals, GAN can dynamically learn the importance and relationship of the extracted feature, and AELM can classify rapidly and accurately with low complexity. The dataset was obtained from simulations of various fault conditions in the Matlab/Simulink. The experimental results demonstrate that the proposed method has an accuracy of 99.5%, a precision of 99.66%, a recall of 99.33%, and an F1-score of 99.5% compared to existing methods. The proposed hybrid framework enables efficient classification capability, feature learning with attention, and adaptive feature extraction, which contributes to improved performance.
Problems with power transformers reduce grid reliability and can lead to large financial losses. Traditional Dissolved Gas Analysis (DGA) methods, such as the Key Gas Method, Ratios, and Duval Triangle, often give unclear results when faults are complex or occur together. This study introduces a Statistically Guided Decision Tree (SGDT) framework, a combined approach that uses statistical DGA analysis and decision tree learning to identify transformer faults. This approach creates rule-based fault categories using advanced statistical analysis of real DGA data and tests how well these categories work with a Decision Tree model. Advanced statistical techniques such as dispersion and association metrics, confidence intervals, and distribution characteristics were used on a wide range of DGA records collected from a 275 kV transformer to define threshold values. Thereafter, fault classification rules were developed, and finally, a decision tree algorithm was developed to evaluate whether the gas concentration-based rules for fault labelling aligned with real data behaviour. The classification accuracy of 0.993 was achieved, indicating a high rate of correctly identified fault types. The F1-score, representing the harmonic mean of precision and recall, was 0.980, confirming both high precision and recall. Specifically, the recall was 0.980, meaning that 98% of real fault cases were correctly found, while the precision was 0.981, showing that 98.1% of predicted fault cases were correct. The Area Under the Curve (AUC) was 0.987, showing the model could clearly tell the difference between fault and non-fault cases. This work demonstrates the effectiveness of the current proposed SGDT framework, and this will help utilities that want to digitise their transformer maintenance and diagnostics for better decision-making.
Bongumsa Mendu, O. Oni, O. M. Longe· Energies· 0 citations
An enhanced diagnostic approach integrating the Sparrow Search Algorithm with PNN is developed, which confirms that SSA can substantially strengthen the anti-disturbance capability of PNN, thereby providing reliable technical support for power transformer fault diagnosis and offering certain reference value for enhancing the operational reliability of power grid equipment.
Bo Liu, Jianghong Dong, Shuyu Ren et al.· Journal of Physics, Conferen...· 0 citations
Abstract. The explosive growth of grid-connected renewable energy systems (RES) has increased the complexity of the operation of the modern power infrastructure, making the detection of the faults reliably an inevitable condition of the stable functioning and safety. Traditional single-algorithm and rule-based monitoring systems are not sufficiently flexible or discriminatory to distinguish between the many varieties of faults that occur in photovoltaic (PV) arrays, wind turbines, battery management systems, and grid-tie inverters. The paper suggests a new stacking-ensemble hybrid machine learning model that will combine a one-dimensional convolutional neural network with a bidirectional long short-term memory (CNN-BiLSTM) module, a Random Forest classifier, and an XGBoost gradient booster as base learners under the guidance of a logistic regression meta-learner. Publicly available SCADA and lab bench data were used to create a curated multi-source dataset of 8,400 labelled samples to represent five operational states. The proposed framework achieved an accuracy of 97.8, a macro-averaged F1-score of 97.1, and a Matthews Correlation Coefficient (MCC) of 0.972, outperforming all the compared baseline methods at least by 3.3 percentage points. The findings verify the effectiveness of the hybrid stacking paradigm in identifying faults in real-time and multiple classes in heterogeneous renewable energy settings.
A. Gopalakrushna· Materials Research Proceedin...· 0 citations