Before I co-founded Magnefy, I spent 15 years designing protection relay systems and substation automation for investor-owned utilities. In that work, I was on the receiving end of several vendor pitches for predictive analytics systems. The pitch was usually some variation of: feed us your historical SCADA data, our models will identify patterns associated with past failures, and we will flag transformers that look like they are trending toward a future failure.
Most of those systems never made it past a trial deployment in our operations. Some were purchased, used for a year or two, and quietly de-provisioned. The engineers who had to work with them were not unintelligent or resistant to change. They were skeptical for good reasons, and understanding those reasons is worth taking seriously if you are building or evaluating fault prediction tools for grid operations.
The Well-Founded Skepticism About Statistical Prediction
The core objection from grid engineers to SCADA-data-derived predictive models is not obscurantism. It is based on a structural feature of the problem: the training data for any model built on historical failure events is dominated by normal operation. Transformers fail infrequently. A fleet of 200 transformers with an annual unplanned failure rate of 2 to 4 percent produces 4 to 8 failure events per year, spread across different failure modes. For a model that needs to learn discriminating features, this is a thin dataset. The ratio of normal-to-failure examples means that a model that simply predicts "no failure" for everything achieves 96 to 98 percent accuracy while being useless for the actual purpose.
Beyond dataset size, there is a deeper problem: the features that correlate with past failures in one fleet's SCADA history may not generalize to a different fleet, or even to a different time period in the same fleet as the equipment mix ages. Load growth patterns change. Climate changes. The vintage of equipment changes as replacements occur. A model that fit well to 2010-2018 failure patterns in one utility's distribution network may underperform on a different utility's network or on that same network in 2024.
Maintenance engineers who have watched a high-confidence model flag an asset, scheduled a crew to inspect it and found nothing, then watched a different asset fail without any prior flag, learn a specific lesson: the confidence score on these predictions is not operationally calibrated. A system that says "87% probability of failure in the next 90 days" means nothing if you have no empirical basis for interpreting what 87 percent means in practice for your fleet.
What Physical Signals Offer That Statistical Models Do Not
When a protection relay operates and isolates a fault, the reason the engineer trusts it is not that it matched a historical pattern of fault-like SCADA data. It is that the relay measured a specific physical quantity, the current through a protected zone, compared it against a threshold derived from the physics of the protected equipment, and operated when that measurement exceeded the threshold. The trust comes from the physical causality chain being intact.
DGA is trusted by maintenance engineers for the same reason. Dissolved acetylene in transformer oil is a specific chemical compound that is produced by specific physical processes, namely high-energy electrical arc discharge. When DGA shows elevated acetylene, the interpretation is not "this matches the pattern of past failures with elevated acetylene." The interpretation is "high-energy arc discharge has occurred inside this transformer, and arc discharge is a known precursor to catastrophic failure." The physical mechanism is understood, and the measurement is of a direct byproduct of that mechanism.
The reluctance to trust statistical predictions is not reluctance to trust data. It is a reasonable preference for measurements that track physical quantities whose relationship to failure is understood, over correlations derived from historical records whose generalizability is uncertain.
Why EM Signature Fits the Physical Signal Category
When we designed Magnefy's approach, this distinction was central to the architecture decisions. We are not building a model that learns what failing transformers look like from historical failure records and then pattern-matches against your fleet. We are measuring a physical quantity, the harmonic content of the electromagnetic field at the transformer terminals, that changes in predictable ways when the winding current distribution changes due to a developing fault.
The physical mechanism is direct: a turn-to-turn fault creates a low-impedance parallel current path within the winding. That changed current path produces a changed harmonic composition in the external EM field. The change is deterministic: a given fault geometry produces a specific harmonic signature change. The detection approach is to track that change relative to the transformer's own baseline, load-normalized so that the natural variation from load cycling does not mask the fault-induced change.
A grid engineer who asks "why did it flag this transformer?" gets an answer grounded in physics: the normalized 3rd harmonic amplitude has increased by X percent relative to baseline at comparable load points, and that specific harmonic shift has a known physical cause in partial winding involvement. The answer is not "the model's confidence score exceeded 85%." It is a specific measurement of a specific physical quantity, interpreted against a known physical mechanism.
This matters operationally because it changes the maintenance decision calculus. When a system flags a transformer with a physical measurement, the maintenance team has something to investigate: they can draw a DGA sample and look for the expected gas evolution, they can request an FRA comparison to check for winding geometry changes consistent with the EM anomaly, and they can schedule a field inspection with a specific question in mind. A statistical flag from a black-box model gives the team a confidence number but no direction on what to look for.
Where Physical Signal Measurement Also Has Limits
I want to be direct about where this argument does not hold. Saying that EM signature measurement is a physical signal is accurate. Saying that the EM measurement is infallible or that false positives do not occur is not accurate.
The measurement is real, but the interpretation layer involves modeling choices. The baseline model for each transformer is fit from observed data, and if the baseline fitting period included a developing fault that went undetected, the baseline itself is contaminated. The load normalization model is fit empirically and may not perfectly capture the transformer's behavior under unusual loading conditions. The threshold for flagging a deviation as anomalous rather than normal variation involves a statistical decision with a false positive rate that depends on the threshold chosen.
What the physical signal approach buys is that the measurement itself is not opaque. When a trained engineer wants to interrogate a Magnefy alert, we can show the raw harmonic amplitude history, the load-normalized deviation curve, the comparison with the baseline distribution, and the specific features that triggered the alert. The engineer can look at that data and make their own judgment about whether the pattern is convincing. That is a different situation from being handed a confidence score from a model that ingested 50 SCADA channels and produced an output through layers of nonlinear transformations.
Building Tools That Engineers Can Trust
The trust problem in predictive maintenance is not fundamentally a technology problem. It is a credibility problem. An engineer who has been burned by overpromised and underdelivered predictive tools will not give a new system the benefit of the doubt. Credibility is earned incrementally by being right when the system says something is wrong, and by having an accountable explanation for why it flagged what it flagged.
We have been deliberate about not overstating the confidence levels in our alerts and about being specific about what fault families the EM anomaly detection is and is not designed to catch. An engineer who trusts our system for the things it is genuinely good at, namely early winding fault onset detection, is more valuable as an operator of the system than one who thinks it will catch every possible transformer failure and then loses trust when it does not.
The engineers I worked with in protection relay design trusted the tools they trusted because those tools had been right repeatedly, and when they were wrong there was an understandable reason. That is the only path to operational adoption in critical infrastructure. Transparency about what is measured, what the measurement means physically, and where the method has limits is not a weakness. It is the foundation of durable credibility.