In short: A transformer health index combines condition evidence into an assessment that helps a utility prioritise further investigation or intervention. CIGRE TB 761 describes several ways to build Transformer Assessment Indices; it does not prescribe one universal 0–100 score, banding scheme or direction of improvement. Define the purpose, scoring rules and data confidence before comparing units, and assess failure consequences separately when ranking risk.
A transformer health index brings diagnostic information into an asset-management decision. CIGRE TB 761, Condition assessment of power transformers, was prepared by Working Group A2.49. It uses the broader term Transformer Assessment Index (TAI) because an index can support different purposes, including repair, refurbishment or replacement. The score is useful only when the user understands what it includes and how it relates to the decision.
The asset management problem
Power transformers represent substantial capital investment and support essential grid functions. Replacement cost, lead time and outage consequences vary with rating, design and location. A major failure can have effects beyond the unit itself, including:
- Unplanned outages causing grid instability
- Cascading failures across interconnected systems
- Fire hazards and safety incidents
- Environmental contamination from mineral oil release
- Extended downtime affecting power generation or distribution
Traditional diagnostic complexity
Transformer condition assessment can combine dissolved gas analysis, oil quality, electrical tests, thermal imaging, partial discharge measurements and operating history. The number of useful inputs depends on the equipment and the failure mechanisms being assessed. Expert interpretation remains necessary, but the evidence and rules can be recorded so that assessments are reproducible.
This complexity creates several critical problems:
Interpretation differences: Engineers can reach different conclusions when they use different assumptions, data-quality rules or operational context. Recording those choices makes the assessment reviewable; the use of a numerical score alone does not remove judgement.
Fleet workload: Reviewing every diagnostic result manually becomes demanding across a large fleet. Structured screening and traceable scoring can direct specialist attention to the units and uncertainties that matter most.
Validation: There is no universal accuracy percentage for traditional diagnostics or health indexing. A claim about performance needs a defined task, reference outcome, representative dataset and evaluation protocol. Fault-type classification accuracy is not the same measure as condition ranking or failure prediction.
A documented health index can make comparisons more consistent and reduce the time needed to screen a fleet. Comparability still depends on using an appropriate method and consistent evidence. Turning several measurements into one score does not by itself produce a calibrated measure of risk.
What is Health Index: technical definition
An index may use a weighted sum, a worst-case score, a hybrid representation or another documented aggregation method. TB 761 discusses these alternatives and their limitations. The starting point is to identify relevant failure mechanisms and the diagnostic evidence used to assess them, rather than selecting a formula and treating it as a universal measure of health.
Health Index scale and interpretation
A 0–100 scale can be a useful presentation choice, but its direction and condition bands are implementation choices. TB 761 does not standardise numerical bands or require a universal score-to-action timetable. A usable scoring specification should define:
- Which end of the scale represents better condition, and what each category means for the decision being supported.
- Which failure mechanisms and components contribute to the score, and which are assessed outside it.
- How incomplete, old or uncertain data affect the score and its confidence.
- How an urgent individual defect remains visible when other indicators are satisfactory.
- How engineers review the underlying evidence before assigning actions and timescales.
Action timescales should follow the assessed defect, its progression, available evidence and operating constraints. A score can help direct review, but a generic numerical band cannot safely prescribe immediate intervention or a twelve-month delay for every transformer.
What Health Index measures (and what it doesn’t)
The health index quantifies:
- Current condition state based on observable diagnostic evidence
- Relative condition comparison across appropriately assessed transformer populations; financial or system risk also requires the consequences of failure.
- Changes in the assessed condition over time, provided the scoring method and data coverage remain comparable.
The health index does NOT:
- Directly predict remaining useful life or annual failure probability without a separately validated model linking condition to those outcomes.
- Replace detailed root cause failure analysis when specific fault modes are identified
- Eliminate the need for engineering judgment
It gives decision makers a structured view of available evidence, with uncertainty made explicit, on which to apply operational context, risk tolerance and resource constraints.
Health Index methodology: from data to decisions
Developing a useful index requires a defined purpose, selection of failure mechanisms and evidence, an aggregation method, and checks that the resulting ranking supports the intended decisions. Validation should match the claim: a consistent condition score and a calibrated failure-probability model have different evidential requirements.
1. Feature selection
Input selection identifies diagnostic information relevant to the failure mechanisms and components being assessed. TB 761 encourages users to distinguish the mechanism from the measurement that indicates it. For example:
Dissolved Gas Analysis (DGA):
- Hydrogen and hydrocarbon gases can provide evidence of electrical or thermal activity, interpreted using the relevant fluid and equipment guidance.
- Carbon monoxide and carbon dioxide can contribute evidence about cellulose involvement, but also have other sources and require context.
- Gas ratios and graphical methods such as Duval can help identify possible fault types after analytical quality and application conditions have been checked.
Oil Quality Metrics:
- Moisture content affecting dielectric strength and cellulose aging rate
- Acidity reflecting oil degradation and oxidation byproduct accumulation
- Interfacial tension (IFT), interpreted with other oil tests as evidence of changes such as oxidation products or contamination.
Complementary Indicators:
- Furanic compounds, including 2-FAL, can support an assessment of cellulose ageing. An oil concentration is not a direct measurement of paper degree of polymerisation.
- Tan delta or power factor for dielectric property assessment
- Operational age and load history
Several measurements may reflect the same failure mechanism. Assessing them together can improve confidence, while blindly summing them can double-count evidence. Correlation alone is not a sufficient reason to discard a gas or test: its diagnostic role and the circumstances under which it provides independent information also matter.
2. Weighting methodology
The weighting methodology determines how individual parameters combine into the final index value.
Traditional approaches use linear weighted sums with weights derived from expert elicitation. This means asking experienced engineers to assign relative importance to different test results.
Data-driven methods are another option discussed in TB 761. They need sufficiently representative condition and outcome data, transparent evaluation and checks for bias. Weights learned from historical records are not automatically optimal, and reproducing labels generated by a scoring formula does not independently validate transformer condition.
Non-linear relationships can matter, but the chosen model must be justified by the intended assessment and available evidence. For example, moisture and thermal stress both influence cellulose ageing. Their interaction should be represented by a documented model or engineering rule, rather than assumed to validate a particular score.
3. Validation against ground truth
Validation should test the intended use of the index. Engineers can review whether known defects receive appropriate attention and whether rankings remain stable under plausible data uncertainty. If the score is also presented as a failure probability, that additional claim requires outcome data, a defined time horizon and calibration; a score of 45 does not intrinsically establish a particular probability.
This requires:
- Large datasets linking historical diagnostic measurements to observed outcomes
- Statistical analysis of performance and uncertainty, with explicit attention to selection bias, missing data and the definition of each outcome.
- Evaluation on assets and periods held apart from model development, and on other populations where a claim of transferability is made.
Health Index calculation process
The transformation from raw diagnostic data to health index involves several processing stages:
Data acquisition: Capture laboratory results, inspections, electrical measurements and online observations at intervals appropriate to the asset and the issue being investigated. There is no universal annual or semi-annual schedule for every index.
Data normalization: Accounts for measurement unit differences and scales parameters to comparable ranges.
Missing data handling: Addresses the practical reality that complete diagnostic datasets rarely exist, particularly for older transformers with incomplete testing history.
Derived indicators: Ratios, graphical fault identification and gas trends can add useful information when the data meet the relevant method’s application conditions. Interpret acetylene with its trend, analytical uncertainty, equipment design and possible sources. Age alone does not turn a given acetylene concentration into a benign result.
Aggregation: Combines normalized, weighted features into the final index score using linear or non-linear mathematical models.
Output interpretation: Use the score and its confidence to select cases for review. Review the contributing failure mechanisms before choosing maintenance, additional testing or replacement.
Condition-based maintenance with Health Index
A health index can support condition-based maintenance alongside statutory duties, manufacturer recommendations, operating experience and scheduled inspections. Fixed intervals remain appropriate for some activities; the purpose of condition information is to improve the timing and targeting of decisions where it can do so reliably.
CBM paradigm shift
Condition-based maintenance uses measured or assessed condition to help determine when an intervention is justified. A health index can support that workflow when its triggers are tied to documented failure mechanisms and reviewed by the responsible engineers.
For example, an organisation could establish a policy with the following steps, using its own validated categories:
- Review units with significant defects promptly to decide whether continued operation, further testing, repair or replacement is appropriate.
- Increase diagnostic attention where credible trends or unresolved uncertainty warrant it.
- Continue the applicable monitoring programme for satisfactory assessments, while keeping independent alarms and protection requirements active.
Fleet-level capital planning
At fleet level, consistent assessments can reveal concentrations of defects or missing evidence. A cluster associated with a design or production period is a reason to investigate common causes. It is not proof of a manufacturing defect until differences in age, duty, maintenance and sampling have been considered.
Comparing index distributions can help target review across voltage classes, designs and operating environments, provided the method and data coverage are comparable. Apparent differences can also reflect different measurement coverage or scoring assumptions.
Risk-based investment prioritization
A condition index can contribute to a risk assessment, but condition and consequence are separate inputs. TB 761 discusses probability-of-failure estimation and the issues to consider before acting on an index. Any conversion from score to probability must be established for the relevant population and time horizon.
Where a defensible failure probability and consequence estimate are available, their combination can support expected-loss calculations. This article does not assign annual failure probabilities to particular health scores. A replacement decision also needs operational alternatives, outage consequences, uncertainty and the possibility of repair.
Technical limitations and practical considerations
Despite its utility, the health index faces inherent technical limitations that practitioners must understand.
Information loss from dimensionality reduction
The index provides a single-number summary of complex, multidimensional condition data. Information loss is inevitable. Two transformers with identical index values may have markedly different diagnostic profiles requiring different maintenance approaches.
One unit might show elevated thermal gases with normal moisture, while another has elevated moisture with little gas generation. An equal aggregate score does not make their failure mechanisms, consequences or best interventions equivalent.
Temporal resolution constraints
An index based on periodic tests is a snapshot of the evidence available at those dates. New or rapidly developing defects can emerge between samples, and some failure modes are poorly captured by routine oil tests. The update frequency and stale-data handling should therefore be visible to the user.
This limitation motivates integration of online monitoring systems that continuously track key parameters, though incorporating high-frequency online data with low-frequency offline diagnostic results introduces methodological complexity.
Data quality dependencies
The index assumes diagnostic data accuracy and representativeness. Sources of error include:
- Dissolved gas samples contaminated during collection
- Post-event measurements interpreted without recording the event, elapsed time, sampling location and relevant operating conditions.
- Measurements from non-representative tap changer compartments
Data quality assurance represents an essential but often overlooked aspect of reliable health indexing. This includes validating measurement procedures, detecting outliers, and confirming sample chain-of-custody.
Root cause analysis limitations
An aggregate index summarises the assessed evidence, while the component and failure-mechanism results explain the causes of an unfavourable score. The same overall category can result from very different combinations of ageing, moisture, electrical activity or mechanical concerns.
Detailed diagnostic interpretation remains necessary for developing effective remediation strategies.
CIGRE TB 761: guidance for assessment indices
CIGRE Technical Brochure 761, prepared by Working Group A2.49 and published in 2019, is guidance on Condition assessment of power transformers. It develops the TAI concept around the user’s purpose, available information, failure mechanisms and uncertainty. Among its subjects are:
Defining the purpose of an assessment index and a scoring matrix suited to the decision being supported.
The strengths and limitations of alternative scoring methods, including the need to validate data-driven approaches.
Using assessment results alongside risk, practical intervention options and other asset-management considerations.
Handling missing or uncertain information and interpreting the diagnostic evidence for transformer components.
Users can compare their methodology with TB 761’s recommendations, but should describe the actual method and its limitations. The brochure does not certify software, prescribe one score or guarantee comparability between utilities using different methods.
Implementation: traceable assessment and decision support
Software can automate data handling, calculate documented indices and make the evidence behind a fleet ranking easier to review. RONIN AI is Seetalabs’s transformer assessment project. A useful implementation connects each reported result to its inputs, supported method and equipment scope so that the responsible engineer can inspect the assessment.
An implementation should show which inputs it uses, which failure mechanisms it covers, how it treats missing data and how an engineer can inspect the result. Processing speed or a compact input set is not evidence of CIGRE certification, diagnostic accuracy or safe continued operation.
Evaluate performance against a defined decision: fault identification, condition ranking and prediction of future failure are different tasks. Use distinct assets, reliable reference outcomes, separation between development and test data, and explicit treatment of inconclusive cases. This gives engineers an interpretable basis for comparing methods and choosing where automated screening adds value.
Conclusion: the future of transformer asset management
A transformer health index can help organise condition evidence and prioritise engineering attention. TB 761 places that work within the broader design of Transformer Assessment Indices, with a defined purpose and explicit treatment of uncertainty.
A carefully specified and reviewed implementation can support:
- Consistent assessment records that different engineering teams can review and reproduce.
- Fleet-wide condition screening, followed by a separate assessment of consequences and operational priorities.
- Condition-based maintenance proposals with a clear link between diagnostic evidence and the intervention being considered.
- Risk discussions that distinguish observed condition, estimated failure probability, consequence and uncertainty.
The index complements engineering judgement and detailed diagnostic analysis. Its value comes from making evidence, assumptions and priorities visible, while allowing individual urgent defects to override an otherwise reassuring aggregate.
For an ageing and diverse fleet, the practical aim is a consistent, reviewable assessment process. Use an index to identify where attention is needed, then return to the underlying evidence to decide what action is justified.
Technical references: CIGRE TB 761 (2019), WG A2.49, executive summary; Chapters 1–4 and 6–7; Table 2-2; and Section 8.1. IEEE C57.104-2019 and IEC 60599:2022 support the distinction between DGA screening, fault identification and condition assessment.
