The first skeptical question we hear from grid engineers when we explain electromagnetic signature monitoring is a good one: substations are full of radiated interference. Switching events, VFDs on adjacent equipment, corona discharge from high-voltage bus, inverter harmonics from renewable generation interconnects. How do you get a clean signal from the transformer when the measurement environment looks like that?
It is a real problem, and the honest answer is that it took us longer to solve than we expected. This piece explains what the interference sources actually look like, how we approach the signal separation problem, and where the limits of the method are.
What We Are Measuring and Why the Noise Problem Is Hard
The transformer's electromagnetic output we are trying to characterize is the near-field magnetic flux measured at the tank wall, typically using a toroidal coil or hall-effect sensor installed at the terminal bushings or low-voltage secondary connections. The signal of interest is the harmonic content of the 60 Hz fundamental: specifically the amplitudes and phase relationships of the 2nd, 3rd, 5th, and higher harmonics, along with any interharmonic components that appear under specific fault conditions.
In a laboratory environment with a single transformer and a controlled supply, these harmonic levels are small but measurable. The 3rd harmonic in a healthy distribution transformer typically runs at 1 to 3 percent of the fundamental. A developing winding fault may shift this by 0.3 to 0.8 percentage points over the first several weeks. That is a signal-to-noise ratio problem even before external interference enters the picture.
In a working substation, the sources of external interference in the relevant frequency range are substantial. Variable frequency drives generate significant harmonic currents at frequencies that appear in the measurement band. Capacitor bank switching creates transients that ring at frequencies overlapping with the harmonic analysis window. Arc furnaces and other nonlinear industrial loads on the same distribution bus inject harmonic currents that arrive at the transformer through the supply network. In substations with co-located renewable generation, inverter-based sources add their own harmonic and interharmonic profiles to the ambient spectrum.
The core difficulty is that most of these interference sources are also variable in time: load cycles, switching events, and operational changes at other facilities all modulate the interference level. A naive approach that just measures harmonic amplitudes at fixed intervals and tracks trends will confound transformer fault signatures with supply-side harmonic changes.
The Separation Problem: Our Approach
The key insight that guides our signal processing architecture is that transformer-originating harmonic changes have a different relationship to load than supply-side or external harmonic injection does. When a supply-side harmonic source changes, the harmonic content at the transformer measurement point changes in ways that are correlated across multiple measurements in the substation. When a transformer-internal fault develops, the change appears specifically in the flux at that transformer's measurement location and it has a characteristic load-normalized signature.
Our processing architecture has three main layers:
Load normalization. We continuously estimate the expected harmonic profile at each transformer based on the present fundamental current magnitude and the load history. The transformer's harmonic generation is load-dependent even in healthy conditions: core saturation level, magnetizing current harmonic content, and load-dependent thermal effects all modulate the healthy-state harmonic profile. A raw 5 percent increase in 3rd harmonic amplitude at high load is not the same anomaly as a 5 percent increase at the same load point seen three months earlier. We normalize measurements to a consistent operating point before computing deviations from baseline.
Supply-side harmonic reference. At each installation we place a reference measurement on the supply bus, upstream of the transformer. This gives us a real-time characterization of the ambient harmonic environment that is not specific to the transformer under study. When supply-side harmonics change, they appear in both the supply reference and the transformer measurement. Fault-originated changes appear in the transformer measurement but not in the supply reference. The differential between the two, after load normalization, is a much cleaner signal than either measurement alone.
Statistical baseline with seasonal and operational context. We maintain a rolling statistical model of the transformer's harmonic signature as a function of load and time of day, updated continuously. Deviations are flagged when they are statistically inconsistent with the historical population of measurements at similar operating points. This avoids the false alarm problem of absolute thresholds: a transformer in a highly distorted supply environment may have elevated baseline harmonics that would trigger a fixed threshold, while the relevant information is whether its own specific harmonic profile is changing, not whether it exceeds some absolute level.
Sources of False Positives and How We Handle Them
Honest reporting on this requires acknowledging that we have had false positives in field conditions, and understanding their causes was essential for improving the approach.
The most common source in early deployments was capacitor bank switching. A large capacitor bank switching on the distribution feeder creates a transient that includes significant harmonic content in the relevant frequency range, and if it happened to occur near a measurement sample interval, it would produce an apparent spike in harmonic amplitude. We addressed this by implementing transient exclusion logic: measurement samples captured within a defined window of a detected voltage transient above a threshold magnitude are excluded from the baseline analysis. The window size is tunable; we set it based on the time constant of the local supply impedance at each site.
The second major source was VFD harmonic injection. Variable frequency drives generate characteristic harmonic families, typically centered on the 5th, 7th, 11th, and 13th harmonics for a standard 6-pulse front end. When a large VFD on a distribution feeder is started or its load changes, the harmonic injection into the supply changes with it. If the supply reference measurement is placed at a point that does not see the same VFD injection as the transformer measurement, the differential tracking will misattribute a VFD change as a transformer-internal change. The fix was requiring co-location of the supply reference at the same bus as the transformer low-voltage measurement, not upstream at a common bus with other feeders branching off.
Temperature-driven measurement drift in the sensor electronics is a third source. Substation environments can swing 40 to 60 degrees Celsius between summer peak and winter low. We addressed this through temperature-compensated sensor calibration and periodic automated self-calibration against the fundamental signal magnitude.
What the Signal Looks Like in Practice
In a field case from a municipal substation pilot: a 2.4 MVA distribution transformer showed a gradual increase in normalized 3rd harmonic deviation starting in week 4 of monitoring. The deviation was initially within the normal statistical range but at the upper boundary. By week 7, the 5th harmonic showed a correlated increase, which is a pattern that does not appear in the supply reference variations we had seen for that site over 6 weeks of baseline. The pattern was consistent with a partial winding involvement: the ratio of 5th to 3rd harmonic deviation matched the signature profile in our fault pattern library for early turn-to-turn fault onset, not for a supply harmonic injection event.
A DGA sample drawn at week 8 showed hydrogen at 72 ppm, which is within the normal range per IEEE C57.104 but represented a 28 ppm increase from the 6-month prior sample. An accelerated sampling 3 weeks later showed hydrogen at 89 ppm, and the CO trend had also begun. That trajectory triggered a maintenance inspection, which found degraded winding paper on the secondary side consistent with the EM signature pattern. The transformer was scheduled for refurbishment.
Where the Method Has Limits
We have found consistently that EM signature monitoring is most reliable for winding insulation degradation and developing turn-to-turn faults. These produce the characteristic harmonic pattern changes described above, and the signal is persistent and load-correlated in a way that distinguishes it from interference.
For bushing failures not involving significant winding current redistribution, the EM signature change is smaller and less discriminating. A bushing degrading through moisture ingress without yet producing significant dielectric current leakage may not produce a detectable EM signature change until the fault is further advanced than the PD detection would catch it. We are direct about this: Magnefy is not a replacement for bushing-specific diagnostics such as power factor and capacitance measurement. It is a layer that catches winding fault onset earlier than DGA, while the bushing-specific methods remain appropriate for that failure family.
Highly distorted supply environments, where total harmonic distortion on the supply bus is above 8 to 10 percent continuously, present a harder working environment for the differential tracking approach. The supply reference approach still works, but the statistical baseline needs more data to achieve equivalent sensitivity because the variance of the ambient harmonic floor is higher. In our experience with industrial substation sites near large arc furnace or nonlinear manufacturing loads, we achieve equivalent detection sensitivity but with a longer baseline accumulation period of 8 to 12 weeks versus the 3 to 4 weeks needed at clean utility substations.
The separation problem is real and the solution is not simple, but it is tractable. The key is that it cannot be solved at the sensor level alone. It requires processing that is aware of the operating context, the supply environment, and the transformer's own load-dependent behavior. Building that context is the work; the sensor is just the starting point.