PaperlayerView exact revision on arXiv

Source-linked AI summary

Power Transformer Health Index and Life Span Assessment: A Comprehensive Review of Conventional and Machine Learning based Approaches

Syeda Tahreem Zahra, Syed Kashif Imdad, Sohail Khan, Sohail Khalid, Nauman Anwar Baig

arXiv:2504.15310v1cs.LGcs.AI

TL;DR

Power-transformer health assessment and remaining-life prediction matter for maintenance planning, but aging factors are interconnected and conventional diagnostic approaches have limitations. This review examines conventional techniques and supervised ML approaches for transformer health assessment and prognosis. It discusses their strengths, limitations, and future research directions.

  • Problem

    Transformer aging and interconnected factors complicate remaining-life prediction, while conventional assessments can be reactive and periodic, leaving degradation undetected between inspections.

    Source: §3 · p. 9, §5 · p. 14

  • Method

    The review consolidates literature on transformer health assessment and prognosis using supervised ML, considering operational data, DGA, temperature, and other parameters.

    Source: §1 · p. 2

  • Results

    The review discusses the strengths and limitations of different supervised ML approaches and identifies potential avenues for further research and improvement.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The review provides researchers and practitioners with a synthesis of supervised ML approaches for transformer health assessment and life-span prediction.

    Source: §1 · p. 2

    Training data often contain few fault examples because transformers predominantly operate under normal conditions, limiting data diversity for ML.

    Source: §5 · p. 15

Abstract

from arXiv · show

Power transformers play a critical role within the electrical power system, making their health assessment and the prediction of their remaining lifespan paramount for the purpose of ensuring efficient operation and facilitating effective maintenance planning. This paper undertakes a comprehensive examination of existent literature, with a primary focus on both conventional and cutting-edge techniques employed within this domain. The merits and demerits of recent methodologies and techniques are subjected to meticulous scrutiny and explication. Furthermore, this paper expounds upon intelligent fault diagnosis methodologies and delves into the most widely utilized intelligent algorithms for the assessment of transformer conditions. Diverse Artificial Intelligence (AI) approaches, including Artificial Neural Networks (ANN) and Convolutional Neural Network (CNN), Support Vector Machine (SVM), Random Forest (RF), Genetic Algorithm (GA), and Particle Swarm Optimization (PSO), are elucidated offering pragmatic solutions for enhancing the performance of transformer fault diagnosis. The amalgamation of multiple AI methodologies and the exploration of timeseries analysis further contribute to the augmentation of diagnostic precision and the early detection of faults in transformers. By furnishing a comprehensive panorama of AI applications in the field of transformer fault diagnosis, this study lays the groundwork for future research endeavors and the progression of this critical area of study.

1. Introduction

Transformer aging and failures make health assessment and remaining-life prediction important for maintenance planning and asset management. The review examines supervised ML approaches alongside transformer-health literature and discusses their strengths, limitations, and future directions.

1. Introduction

  • Power-transformer aging and unexpected failures can cause utility and consumer losses, making health assessment and remaining-life prediction important for maintenance and replacement planning.The paper notes that transformers can have potential lifespans of up to 60 years.Source: §1 · p. 1
  • Condition monitoring supports asset management by detecting potential faults early and tracking equipment performance during operation and maintenance.The paper links proactive maintenance with reduced failure and outage risks and optimized resource allocation and costs.Source: §1 · p. 1
  • The review consolidates supervised ML research on transformer health assessment and prognosis using operational data, DGA, temperature measurements, and other parameters.It considers accuracy, precision, recall, and F1-score when comparing predictive models.Source: §1 · p. 2
  • The review’s findings and discussion aim to help researchers and practitioners understand supervised ML approaches’ strengths and limitations and identify research directions.The paper also describes demand for non-invasive diagnostic and monitoring techniques and intelligent analytical tools.Source: §1 · p. 2

2. Power transformer health assessment

Transformer health assessment centers on oil-paper insulation and combines health-index scoring with tests of oil quality, dissolved gases, and furanic compounds. For DGA, the review describes conventional sampling and interpretation alongside ML diagnosis and emerging online monitoring approaches.

2. Power transformer health assessment

  • The health index combines operational observations, on-site inspections, and laboratory tests, with oil-paper insulation state particularly important when service-history or design data are limited.The scoring-weighting approach assigns factor scores using expert judgment; cited HI inputs include aging, loading, pollution, and on-site test data.Source: §2 · pp. 2–3
  • HI computation uses oil quality assessment, DGA, and furanic-compound tests; DGA detects internal faults, while furan content indicates paper-insulation aging.Oil-quality tests include acidity, breakdown voltage, interfacial tension, water content, color, and dielectric dissipation factor.Source: §2 · p. 3

2.1. Dissolved gas analysis

  • DGA proceeds from quantifying extracted dissolved gases to identifying gas components for fault diagnosis, with severity assessed from gas types and concentrations.IEC 60567 standardizes gas sampling, while IEC 60599 guides dissolved-gas interpretation using basic gas ratios.Source: §2.1 · p. 3
  • Conventional DGA interpretation methods rely on expert decisions; the review notes poor key-gas recognition efficiency and accuracy, inadequate coding in three-ratio methods, and possible fault misjudgment.These shortcomings can hinder detection of hidden and unusual faults.Source: §2.1 · p. 4
  • ML-based DGA methods include EMD- and ITD-derived features and XGBoost classification, while a Sequential Kalman filter with sensor fusion was reported to improve estimation accuracy to over 90%.The review reports that ITD achieved better F1-scores than the previously discussed EMD technique; the fusion result was validated with Majority Voting, Dempster Shafer Evidence Theory, and IEC-TC-10 data.Source: §2.1 · p. 4
  • Online DGA options include laser spectroscopy, fiber-optic sensors, portable chromatography, and continuous gas monitors, but these devices and methods can require substantial computational resources and ML expertise.The review describes these approaches as enabling real-time or on-site monitoring, including at remote or hard-to-access transformers.Source: §2.1 · p. 4

2.2. Partial discharge

Partial discharge (PD) signals insulation deterioration, but their volume and complexity challenge human interpretation. The section describes how machine learning supports online fault identification and offline assessment of insulation degradation.

2.2. Partial discharge

  • PD indicates insulation deterioration, and its pulse, timing, discharge interval, and rate can characterize the activity and its source.PD also produces electrical, acoustic, light-emission, and electromagnetic effects, which can be detected using different sensor types.Source: §2.2 · p. 4
  • PD information can be too vast and complex for humans to interpret fully, motivating AI-based diagnostic frameworks.ML approaches extract hierarchical characteristics from input data, which the passage associates with more accurate and reliable outputs.Source: §2.2 · pp. 4–5
  • ML can diagnose online and offline faults, reduce reliance on expert interpretation, and rapidly track insulation deterioration after training on diverse fault data [Karthikeyan et al., 2008; Majidi et al., 2015].The passages describe potential reductions in maintenance costs and improved transformer dependability [Karthikeyan et al., 2008; Majidi et al., 2015].Source: §2.2 · p. 5
  • ML monitoring uses an online phase to identify problems and an offline phase to evaluate insulation-degradation rate [Karthikeyan et al., 2008; Majidi et al., 2015].The offline phase repeatedly trains and tests the algorithm, which the passage says contributes to lower detection time and greater diagnostic accuracy.Source: §2.2 · p. 5

2.3. Breakdown voltage test

The breakdown voltage (BDV) test measures transformer-oil insulation strength through the voltage at which disruptive discharge occurs. BDV declines with oil aging, but low readings may also reflect severe impurity, and acceptable thresholds vary by transformer rating.

2.3. Breakdown voltage test

  • BDV measures the voltage at which transformer-oil insulation becomes conductive, indicated by a spark, and is reported in kV/mm.The test indicates the oil insulation’s ability to withstand electrical stress.Source: §2.3 · p. 5
  • BDV is negatively correlated with transformer-oil aging, but a low reading does not primarily imply aging and can indicate severe oil impurity.The test uses an oil sample in a cell with hemispherical electrodes; the voltage producing disruptive discharge is the breakdown voltage.Source: §2.3 · p. 5
  • The test applies an electric field until gas production exceeds bubble formation and disruptive discharge occurs.The passage names the BA75 analyzer as a widely used instrument for characterizing oil samples.Source: §2.3 · p. 5
  • 30 kV/mm is the minimum acceptable BDV for transformers operating at 230 kV; the threshold is 28 kV/mm for 69–230 kV ratings and 23 kV/mm at or below 69 kV.The passage identifies these as rating-dependent minimum acceptable values.Source: §2.3 · p. 5

2.4. Photoluminescence (PL), Fourier-transform Infrared spectroscopy (FTIR) and Ultraviolet-Visible Spectroscopy (UV–VIS)

The section covers oil-characterization techniques that probe chemical or optical properties. FTIR tracks aging-related changes in functional-group absorbance, while photoluminescence (PL) shows better correlation with dissolved-gas-analysis-derived degree of polymerization than UV–Vis spectroscopy.

2.4. Photoluminescence (PL), Fourier-transform Infrared spectroscopy (FTIR) and Ultraviolet-Visible Spectroscopy (UV–VIS)

  • FTIR characterizes transformer-oil composition through infrared spectra, providing information on chemical composition, aging, and degradation.The section describes FTIR as a valuable tool for offline oil characterization and links its information to maintenance and replacement decisions.Source: §2.4 · pp. 5–6
  • Methane and carbonyl peak absorbance tends to increase with transformer-oil aging, while hydroxyl peak absorbance decreases [Alshehawy et al., 2017].FTIR identifies functional groups from spectral peaks at frequencies corresponding to atomic vibrations; the section also describes oxidation and thermal decomposition as sources of oil chemical change.Source: §2.4 · p. 6
  • PL spectroscopy has better correlation with dissolved-gas-analysis-derived degree of polymerization than UV–Vis spectroscopy [Alshehawy et al., 2021a].UV–Vis measures optical absorption by wavelength and has faced criticism for cost and limited sensitivity to fluorescence; PL measures emitted-light intensity and can use online sensors [Alshehawy et al., 2021a; Alshehawy et al., 2021b].Source: §2.4 · p. 6

2.5. Total acid number

Total acid number (TAN), also called neutralization number, quantifies the base needed to neutralize acid in transformer oil. The section describes standardized titration procedures and notes that oil below the marginal Oil Quality Index Number reference is generally considered unsafe for continued use.

2.5. Total acid number

  • TAN is the milligrams of potassium hydroxide (KOH) required to neutralize acid in one gram of transformer oil, expressed as mgKOH/g.It is also called the Neutralization Number (NN), and the cited BS 62021-1 procedure defines the titration endpoint as pH 11.5.Source: §2.5 · p. 6
  • Oil with a TAN below the marginal Oil Quality Index Number reference is generally considered unsafe for continual use and should be reclaimed.BS 62021-1 specifies automatic potentiometric titration, while BS 62021-2 addresses colorimetric titration.Source: §2.5 · p. 6
  • In automatic potentiometric titration, the oil sample is dissolved in solvent and titrated with indicating and reference electrodes; base volume at pH 11.5 is reported as NN.The method determines the endpoint by measuring potential rather than primarily requiring an indicator.Source: §2.5 · p. 6

2.6. Frequency domain spectroscopy and frequency response analysis

FRA and FDS reveal winding and insulation conditions through frequency-dependent measurements, while furan tests help assess insulation aging. The review describes ML-based methods for interpreting these diagnostic signals and estimating furan content.

2.6. Frequency domain spectroscopy and frequency response analysis

  • FRA traces are compared with healthy-winding fingerprints or equivalent-circuit responses, while FDS records insulation dielectric response across frequencies from mHz to MHz.FRA detects winding changes through impedance, transfer-function, or admittance patterns; FDS supports assessment of moisture, aging, and faults.Source: §2.6 · p. 7

2.6. Frequency domain spectroscopy and frequency response analysis; 2.7. Furan tests

  • ML-based FRA interpretation claims over 98% accuracy, while in-service furan-content prediction achieved 90% accuracy.The cited studies report these results for different diagnostic tasks: FRA fault interpretation and transformer-oil furan prediction.Source: §2.6 · p. 7, §2.7 · p. 7

2.6. Frequency domain spectroscopy and frequency response analysis

  • FRA/FDS interpretation requires expertise and lacks a widely accepted automatic algorithm, but ML has been used to identify fault presence, type, and location (Samimi et al., 2016; Picher et al., 2020).The described ML-based FRA system is trained on data from failure scenarios derived from experiments or models.Source: §2.6 · p. 7

2.7. Furan tests

  • Furanic compounds indicate cellulose-insulation aging, but their measurement can be influenced by oil parameters and their correlation with DP can be weak.Furan estimation can avoid sampling insulation paper to measure DP, which may require draining oil and can be impractical near winding tops.Source: §2.7 · p. 7

2.8. Fiber optic sensors

Optical fiber sensors provide transformer temperature and strain measurements, including winding-temperature monitoring, while Fiber Bragg Grating sensors offer multiplexing and broad sensing range. The review notes that sensor data typically need analytical methods such as ML for fault interpretation.

2.8. Fiber optic sensors

  • A temperature-rise test on a single-phase prototype transformer with ODAN cooling used 20 optical fiber sensors along the windings to collect temperature data.The approach combined experimental and numerical methodologies.Source: §2.8 · p. 7
  • Fiber Bragg Grating sensors use periodic refractive-index variations in single-mode fiber and offer multiplexing and broad sensing range for high-voltage monitoring.Their integration in transformers is still advancing, with potential improvements in measurement accuracy and equipment monitoring; Fig. 9 presents their operating principle.Source: §2.8 · p. 8, Figure 9 · p. 8
  • Optical fiber sensors provide precise temperature and strain measurements but typically need additional analysis for fault diagnosis, where ML can classify complex patterns.Optical fiber sensing systems include interferometric, distributed, spectroscopy, and FBG sensors; Fig. 8 depicts a typical system.Source: §2.8 · p. 8, Figure 8 · p. 8

3. Power transformer remaining life assessment

Transformer life assessment is complicated by interacting aging factors, particularly the degradation of solid insulation. The review describes models that use insulation, oil, monitoring, and statistical data to estimate remaining life or failure likelihood.

3. Power transformer remaining life assessment

  • Temperature, moisture, oxygen, mechanical stress, and corrosive elements all contribute to aging, and their interconnected effects complicate remaining-life estimation.Temperature degrades cellulose, moisture disrupts its chain, oxygen promotes oil oxidation, and mechanical stress weakens oil-paper and pressboard components.Source: §3 · p. 8
  • Thermal, hydrolytic, and pyrolytic stress degrade cellulose insulation, whose degree of polymerization falls from 1100–1600 in new transformers to about 200 at the end of useful life.Aging also produces water, furan, carbon dioxide, carbon monoxide, and acids that further accelerate insulation aging.Source: §3 · p. 8
  • A regression model estimates furan content from moisture, temperature, capacity, oxygen, breakdown voltage, and age, then derives DP to calculate remaining lifespan.The model uses multiple linear and nonlinear regression techniques.Source: §3 · pp. 8–9
  • Full simulation accounts for long-term variations in load and ambient temperature but is computationally intensive, whereas renewal modeling is simpler and must include long-term shifts for accurate reliability assessment.Both approaches use estimated load factors and ambient temperatures with IEEE life-consumption models to assess insulation life and failure time.Source: §3 · p. 9
  • A condition-ranking approach combines winding degradation, a health index using monitoring data and expert judgment, and decommissioned-transformer statistics to estimate individualized, condition-dependent failure probability.The approach supports transformer comparison and maintenance or replacement recommendations.Source: §3 · p. 9
  • Monte Carlo-generated data supplement limited prototype-transformer experiments for DTCM, while oil-test ML models estimate remaining life and can use alternative inputs when data are unavailable.The DTCM work validates generated data with statistical significance tests and uses empirical regression to estimate elapsed and remaining life; oil-test models are reported to support timely asset decisions.Source: §3 · p. 9

4. Machine learning models

The review describes supervised learning models for transformer ageing, health-index assessment, and fault diagnosis, then surveys their reported strengths and limitations. It covers neural, kernel, tree-based, instance-based, and optimization approaches, alongside other diagnostic techniques.

4. Machine learning models; 4.2. Machine learning techniques; 4.10. Additional techniques

  • Regression predicts sensor-linked ageing properties, classification assigns OQIN severity, and the review surveys supervised, unsupervised, reinforcement, and additional diagnostic algorithms.The additional approaches listed include GMM, tensor-based classifiers, BSS, MLR, SVDD, SRC, Bayesian networks, and rough set theory.Source: §4 · pp. 9–10, §4.2 · pp. 11–12, §4.10 · p. 14

4.1. Deep learning models

  • After autoencoder compression, logistic regression paired with MLP AE achieved the best accuracy among tested models and outperformed existing models.The study used 31 input parameters from 608 transformers. Separately, a Bayesian wide-and-deep model categorized health index using expert-ordered categories and quantified parameter and prediction uncertainty; KPCA augmentation with a deep residual network was reported to expand samples and improve diagnosis.Source: §4.1 · p. 11

4.3. Artificial neural network

  • Abu-Elanien et al. (2011) report 97% test-sample identification across three classes; Jin et al. (2023) report 92% diagnostic accuracy for residual BPNN.The first result used actual data from 59 transformers; the second method deepened BPNN with residual network modules for gas-feature analysis.Source: §4.3 · p. 12

4.4. Support vector machine

  • SVM-GA outperformed the IEC three-ratio method, standard SVM, and ANN for diagnosis; a cascade used GA to tune the SVM kernel.SVM is described as useful with limited data, but performance depends on selecting suitable parameter values.Source: §4.4 · pp. 12–13

4.5. Random forest

  • RF performed better than ANN across data sizes in a 91-sample oil-transformer comparison.RF combines decision trees and predicts by majority vote; reported real-transformer tests also found higher diagnostic accuracy than SVM.Source: §4.5 · p. 13

4.6. K-nearest neighbour

  • KNN classifies using neighboring labels and regresses by averaging neighbors’ outputs; it requires no training and is presented as suitable for real-time data streams.The review also describes KNN as non-parametric and applicable to small and medium-sized datasets.Source: §4.6 · p. 13

4.7. Genetic algorithm; 4.9. Particle swarm optimization

  • GA optimizes weighted health-index specifications and, with SVM, improves diagnosis accuracy; PSO-based classifiers report accuracy and efficiency gains and redundant-input removal.The cited GA work concerns health-index calculation and fault diagnosis; PSO is also applied to classifier parameter optimization.Source: §4.7 · p. 13, §4.9 · p. 14

4.8. Convolutional neural network

  • After oversampling balanced the training data, CNN outperformed optimized ML classifiers and remained reliable with noise up to ±25% across full and reduced feature sets.The review identifies Table 6 as a comparison of CNN and other methods’ performance matrices.Source: §4.8 · p. 13, Table 6 · p. 14

5. Author’s opinion and future work

The review contrasts conventional transformer assessments with ML approaches and identifies unresolved challenges for future work.

5. Author’s opinion and future work

  • Conventional DGA and insulation-resistance assessments can be reactive and periodic, leaving degradation between inspections undetected and potentially contributing to unexpected failures.The review presents these limitations as motivation to explore more advanced assessment techniques.Source: §5 · p. 14
  • ML can analyze high-dimensional data and identify patterns not apparent with conventional methods; the review describes neural networks, SVMs, decision trees, and random forests as serving different assessment needs.The review associates neural networks with complex patterns, SVMs with clear separation margins, and tree methods with feature-importance insights.Source: §5 · p. 14
  • Future research challenges include scarce failure data, transformer variation, imbalanced datasets, real-time detection, on-field assessment complexity, and fault localization.The review notes that normal-operation data are more abundant than failure data and that entirely error-free predictions are impractical given transformer variation and uncertainty.Source: §5 · p. 15

6. Conclusion

The conclusion reviews conventional transformer diagnostics and emphasizes AI and ML approaches for condition assessment and life prediction. It frames data-based prediction as useful for processes and data patterns that conventional analytical approaches struggle to handle.

6. Conclusion

  • The review suggests OQIN and TAN characterize transformer-oil aging more accurately than other methods that may not directly correlate with aging.It also reviews conventional diagnostic methods and their challenges, including BDV, spectroscopy, DGA, and acidity measurement.Source: §6 · p. 15
  • The review highlights AI, particularly neural networks, for condition assessment and life prediction, and describes ML as useful for large datasets with diverse spatial and temporal characteristics.It presents data-based predictions as offering insight into processes that were previously challenging to analyze.Source: §6 · p. 15

CRediT authorship contribution statement

The statement assigns conceptualization, methodology, drafting, supervision, validation, review, and funding acquisition across five authors.

CRediT authorship contribution statement

  • Zahra is credited with conceptualization, methodology, and the original draft; the other four authors’ listed roles include supervision, validation, review, editing, or funding acquisition.Baig is credited with funding acquisition.Source: CRediT authorship contribution statement · p. 16
Loading 2504.15310v1…