The operational problem that condition monitoring creates — somewhat counterintuitively — is a prioritization problem. Once you have continuous telemetry from a large transformer fleet, you will have more anomaly flags than you have available maintenance crews on any given week. The question is not whether a transformer is showing an anomaly; it is which anomaly-flagged transformer warrants immediate dispatch, which can wait two weeks, and which can be deferred until the next scheduled inspection window.
Getting this triage wrong in either direction is costly. Over-prioritizing minor anomalies on non-critical assets dispatches crews unnecessarily, consuming maintenance budget that should go toward confirmed high-risk situations. Under-prioritizing genuine incipient faults on critical feeders allows those faults to progress to unplanned failures — exactly the outcome the monitoring program was designed to prevent.
The answer is a composite risk score that combines the anomaly signal strength with the consequence of failure for each specific asset. This post explains the scoring architecture Fieldiq uses and the calibration decisions that matter most for each utility's specific context.
The Four Scoring Dimensions
Fieldiq's composite Asset Health Index (AHI) combines four distinct dimensions into a single prioritization score. Each dimension addresses a different aspect of failure risk.
Dimension 1: Anomaly Severity
The anomaly model's output is not a binary flag — it is a continuous score representing how far the current asset signature deviates from its calibrated baseline, normalized to a 0–1 range. A score of 0.3 represents a modest, statistically elevated deviation that warrants monitoring attention but not immediate dispatch. A score of 0.85 represents a substantial, persistent deviation across multiple telemetry streams (vibration, temperature, and load-response all simultaneously anomalous) that has been sustained for more than 72 hours — a situation warranting escalated response.
The anomaly score is the foundation of the composite index, but it cannot stand alone as the prioritization criterion. A 0.85 anomaly score on a transformer serving 8 customers on a low-consequence rural lateral is less operationally urgent than a 0.55 anomaly score on a transformer serving 400 customers on a critical commercial feeder with no normally-open switching path to an alternate source.
Dimension 2: Consequence of Failure
Consequence quantifies what happens if this specific transformer fails unplanned. The primary inputs are customer count on the secondary circuit, customer criticality tier (residential, commercial, industrial, critical facility), and the availability of alternate supply. These inputs produce a consequence multiplier that scales the anomaly score into a risk-prioritized ranking.
Utilities often already have a transformer criticality tier in their asset management system — a GIS-based analysis that classifies each transformer by customer count, voltage sensitivity, and switching flexibility. If that analysis exists, it can be imported directly as the consequence input. If it does not exist, a simplified version can be approximated from outage management system history: transformers that appear frequently in major event restoration sequences, or that have historically required extended restoration times, tend to be high-consequence assets.
Dimension 3: Asset Thermal Age
Two transformers can show the same anomaly severity score but have very different failure timelines depending on how much thermal aging they have accumulated. An anomaly signal on a 30-year-old transformer that has experienced multiple heat-event loading exceedances represents a different risk level than the same signal on a 12-year-old unit with similar load history but no major heat events.
The thermal age dimension incorporates the cumulative loss-of-life calculation from the IEEE C57.91 model, which accounts for both chronological age and thermal history. For transformers where continuous temperature and load telemetry is available, the thermal age is calculated from the actual operating history. For transformers that are newly instrumented, a thermal age estimate is derived from installation year and the feeder's historical loading profile.
Dimension 4: Geographic Outage Impact
The fourth dimension addresses the SAIDI contribution structure of the specific circuit. A transformer failure that would cause a sustained outage within a high-SAIDI-weighted geographic zone (dense population, regulatory sensitivity, or utility-defined critical corridor) warrants higher prioritization than a comparable failure in a zone with lower weighted impact.
This dimension maps onto utilities' existing reliability performance management structures. If a reliability engineering team already uses a feeder or zone-level SAIDI weighting for planning purposes, that weighting directly informs this scoring dimension. The key is that the geographic impact dimension amplifies the risk score for assets in areas where failure consequences extend beyond the immediate secondary circuit to system-level reliability performance.
Combining the Dimensions
The composite AHI is a weighted product of the four dimensions:
AHI = W1 × Anomaly + W2 × Consequence + W3 × ThermalAge + W4 × GeoImpact
The default weights in Fieldiq's model are calibrated for a general Texas IOU distribution context, but the weights are configurable. A utility with a large rural cooperative territory where restoration time is the dominant reliability cost should weight the Consequence and GeoImpact dimensions more heavily relative to Anomaly severity. A utility in a dense urban territory where customer count per transformer is high and crew response time is short may weight Anomaly severity more heavily, because the consequence dimension is relatively uniform across the fleet.
Calibration of the weights is one of the structured discussions that happens during a pilot program deployment. The starting point is the default weights applied to the pilot fleet, followed by a retrospective review of the dispatch decisions that were generated: were the dispatched assets the ones the reliability engineering team's experience would have predicted? If the model is consistently surfacing assets that field teams consider lower-priority than other flagged assets they know about, the weights need adjustment.
Handling Ties and Edge Cases
Any fleet risk ranking will produce situations where multiple assets have similar composite scores. The tiebreaking heuristic that works best in practice is recency of anomaly onset: among assets with similar composite AHI scores, dispatch first the one where the anomaly score has increased most rapidly in the last 7 days. A deteriorating trend is more urgent than a stable anomaly at the same current level, because the rate of change carries forward-looking information about failure timeline that the current score alone does not capture.
Edge cases worth handling explicitly: transformers in planned outage windows (scheduled for replacement in the next 30 days) should be explicitly flagged as "deferred — planned replacement" rather than appearing in the active dispatch queue. An anomaly flag on a transformer that is already scheduled for replacement is not a maintenance trigger — it is confirmation that the planned replacement was correctly prioritized. Allowing it to consume dispatch capacity from the active anomaly queue would be a workflow error.
What the Scoring Model Does Not Replace
Fleet risk scoring is a decision-support tool, not an automated dispatch system. We are not suggesting that the composite AHI should trigger automatic crew dispatch without human review. The role of the scoring model is to surface the right asset at the right time in the reliability engineer's attention queue — not to replace the engineer's judgment about the appropriate response action.
There are factors that the scoring model cannot capture: recent local intelligence from field crews who have observed something unusual, planned switching work that makes a specific circuit inaccessible, spare unit inventory constraints that affect what maintenance actions are executable this week. These factors belong in the engineer's decision, not in an automated scoring model. The model's job is to make sure the engineer is looking at the right 10 assets, not at 60 equally undifferentiated anomaly flags.
