When vibration monitoring vendors from the plant-floor world attempt to extend into utility distribution asset monitoring, they bring a toolset built for rotating machinery: ISO 10816 velocity bands, frequency-domain peaks indexed to shaft rotation rates, alarm thresholds calibrated on motors and pumps. The toolset works in the context it was built for. Applying it to distribution transformers and line reclosers produces results that oscillate between over-alarming on healthy assets and missing genuine incipient faults — an outcome that is arguably worse than no monitoring at all, because it trains field teams to dismiss anomaly flags as noise.
The root cause is not algorithmic sophistication — it is training domain mismatch. Generic vibration standards were calibrated on data from plant-floor machinery that has certain mechanical properties: rotating components, bearing frequencies, imbalance signatures. Distribution transformers have none of those. Their vibration physics are driven by electromagnetic forces (core magnetostriction, winding Lorentz forces), not rotation. The feature space is fundamentally different, and a threshold regime designed for one does not transfer to the other.
The Vibration Physics Are Different
A distribution transformer's vibration signature is governed by two primary electromagnetic phenomena. Core magnetostriction produces a strain cycle at twice the supply frequency — 120 Hz in a 60 Hz system — as the core laminations elongate and contract with the alternating magnetic flux. The winding structure adds load-dependent Lorentz force vibration, also at 120 Hz and its harmonics, with amplitude proportional to the square of the load current.
These forcing functions produce a signature that is fundamentally load-dependent and temperature-dependent. A transformer at 30% loading has lower winding force excitation than the same unit at 80% loading. A transformer with hot, expanded core laminations shows slightly different magnetostrictive response than a cold-start unit. This load-temperature coupling means that a single amplitude threshold — applied regardless of operating condition — will chronically alarm on healthy high-load operation while remaining silent during low-load operation with genuine mechanical looseness.
For line reclosers, the vibration signature is predominantly generated by the actuator mechanism during open/close operations, plus steady-state bus-bar and conductor vibration. The anomaly signatures of interest (contact wear, mechanism spring fatigue, hydraulic actuator degradation) manifest as changes in the time-domain shape of the operation transient, not as continuous broadband amplitude elevation. The ISO 10816 concept of a steady-state RMS velocity threshold has essentially no relevance to this detection problem.
What Per-Class Calibration Actually Involves
Fieldiq's anomaly models are calibrated at the level of asset subtype — not just "distribution transformer" as a monolithic class, but segmented by kVA rating tier and installation type. The key segmentation variables are:
- kVA rating: A 25 kVA single-phase unit and a 500 kVA three-phase unit have different core masses, different winding structures, and different absolute vibration levels at comparable loading percentages. They require separate baseline populations.
- Installation type (pole-mount vs. pad-mount): Pole-mounted transformers are subject to wind-induced mechanical excitation through the mounting structure. Pad-mount units are rigidly mounted and show different broadband noise profiles. Combining them in a single baseline population adds variance that obscures true anomalies.
- Core design (shell-form vs. core-form): Shell-form and core-form transformers have different magnetostrictive vibration patterns at the same loading level, particularly in the harmonic amplitude ratios above 120 Hz.
Calibration uses a population of confirmed-healthy historical telemetry from units of the same subtype — healthy meaning no DGA anomalies, no inspection findings, and no subsequent failure events within a two-year window following the calibration period. The baseline model learns the conditional distribution of vibration features (spectral amplitudes at key frequencies, harmonic ratios, broadband noise floor) as a function of the operating condition covariates (load current, ambient temperature). An anomaly flag is generated when the current observation falls outside the expected distribution by a statistically meaningful margin, conditioned on the operating state.
Quantifying the Precision Difference
The most direct way to evaluate this question is to run the same sensor data through a generic threshold model and a per-class calibrated model on the same held-out dataset and compare precision and recall.
On held-out data from our distribution transformer monitoring deployments, applying an ISO 10816 Zone B/C boundary (a commonly referenced generic threshold) as a single-number alarm criterion produces a false positive rate — healthy units flagged as anomalous — of 30–45% depending on the time window and loading conditions. Among high-load summer operating periods, healthy 500 kVA units exceed the generic velocity threshold regularly simply because their absolute vibration levels at high load exceed a threshold designed for smaller industrial equipment.
The per-class calibrated model on the same dataset produces false positive rates in the 4–8% range under comparable conditions, with recall on confirmed fault events above 85%. The improvement is not from a more complex algorithm — the model architecture is straightforward multivariate anomaly scoring on spectral features. The improvement is entirely from appropriate baseline conditioning.
We want to be precise about what these numbers represent: they reflect performance on Fieldiq's own validation datasets, which consist of distribution field assets in Texas Gulf Coast territory. Results in different climates, different fleet ages, and different loading profiles will differ. The claim is not that 94% precision is guaranteed across all deployments — it is that per-class calibration produces meaningfully better precision than generic thresholds on the same data, and the mechanism for that improvement is well understood.
The Nuisance Alarm Problem Is Not Academic
False positives are not just a nuisance — they actively degrade the value of a monitoring program. When field teams receive more anomaly flags that turn out to be healthy equipment than flags that represent real problems, the behavioral response is predictable: teams stop treating flags as actionable, investigation rates decline, and the monitoring program loses the organizational credibility needed to justify accelerated maintenance responses. This is the "cry wolf" failure mode, and it is more common in industrial monitoring deployments than the industry typically discusses publicly.
We are not suggesting that false positives are acceptable at any level — the engineering goal is to minimize them through better calibration. But the observation is that the cost of a false positive in a utility maintenance context (a truck roll to inspect a healthy unit, or worse, a reliability engineer who starts ignoring the alert queue) is higher than it might appear in the abstract precision-recall table. Calibration quality is not just an academic modeling exercise; it has direct operational consequences for whether the monitoring program functions as intended.
Per-Class Calibration Has Costs Too
Building per-class models requires enough labeled healthy data for each asset subtype to calibrate a reliable baseline. This is not free. A monitoring deployment that starts with 20 units of mixed subtypes may not have sufficient data for stable per-class calibration in the first month of operation. During that initial period, the anomaly model operates with wider confidence intervals, and the false positive rate is higher.
The calibration period also means that newly instrumented assets — which have no historical telemetry baseline — require time before the model produces reliable flags. Fieldiq's typical calibration window is 3–6 weeks of operating data across a range of loading conditions, including at least one peak-load period that exercises the unit above 70% of nameplate rating. Until that baseline is established, the model is in a warm-up state, and anomaly flag confidence is clearly indicated as reduced.
This is the honest tradeoff: per-class calibration produces better precision at the cost of a baseline acquisition period. Generic thresholds can flag immediately but flag badly. For a maintenance program managing a fleet of hundreds of units with defined criticality tiers, the calibration period cost is justified by the operational quality of the resulting anomaly signal.
