A 2026 paper in Scientific Reports reports a multi-task learning framework, PD-IntelliFusionNet, that fuses acoustic-emission and UHF features to diagnose discharge type and assess severity — 93.66 % accuracy at 10 dB signal-to-noise ratio, holding 91.82 % on 20 % of the training samples. It is a reminder of what one sensor cannot do. PRPD pattern analysis stays the reference view because its horizontal axis carries what no amplitude plot holds: the point in the power cycle at which each pulse occurred.
What the horizontal axis of a PRPD plot really is
A phase-resolved partial discharge plot puts pulse magnitude on the vertical axis and something on the horizontal axis that is not time: position within the mains cycle, folded across many cycles into a 0–360 ° window. The folding is the point. A discharge that always fires near the voltage peak is a different defect from one that fires at the zero crossing, even when both produce identical amplitudes.
That creates a dependency that is easy to forget: phase resolution requires a reference, and without a voltage or current reference from the test circuit or the unit, the same data can be plotted at any phase and the pattern means nothing. Reconstructing that reference is what non-conventional methods under IEC TS 62478 have to solve, and why a calibrated coupler under IEC 60270 remains the reference measurement.
Three-phase comparison is the cheapest discriminator
A three-phase transformer gives the analyst two controls one channel cannot. External interference — from the substation, a neighbouring unit or a welding set — reaches all three phases through a shared path. A defect inside the tank belongs to one winding, or has a geometry that makes it phase-specific.
So the first question is never how large the discharge is, but whether it appears on all three phases at the same phase position and comparable amplitude. A simultaneous, near-identical signature across three channels points to something arriving from outside, and in utility substations that test removes most of what fills an alarm log.
Four patterns three phases reveal
| Pattern across the three channels | What it points to | What to do next |
|---|---|---|
| Same shape, amplitude, phase position | External interference or common-mode source | Correlate with an acoustic sensor; find the source |
| Same shape, differing amplitude, consistent phase | Internal defect, path attenuates per phase | Trend the strongest phase; locate with UHF |
| Marked phase offset between channels | Defect near a lead-end or tap | Inspect the lead-end and tap changer |
| Present on one phase only | Defect belongs to that winding or its accessory | Baseline that phase; shorten the interval |
None is a diagnosis on its own; each is a direction. The value of the three-channel view is that it removes the largest single cause of alarm fatigue before any classifier is asked a question.
What does channel synchronisation have to do with PRPD pattern analysis?
Everything, once more than one channel is involved. If two channels are not referenced to the same time base and the same phase reference, a genuine phase offset between phases is indistinguishable from a sampling delay, and the table above becomes unreadable.
In practice this is hardware, not software. The TPD-400 acquires through a 125 MS/s, 14-bit IED with a multi-channel synchronisation skew under 100 ns, which holds the relative phase stable; it is designed to IEC TS 62478 and GB/T 42287, with reporting over IEC 61850 (DL/T 860). Figures are the manufacturer’s datasheet values. A system that timestamps each channel independently can still draw a clean-looking PRPD plot whose phase relationships are an artefact.
Does PRPD pattern analysis work on DC equipment?
Not in the same way. A phase-resolved plot assumes a periodic reference, and DC offers none to fold over — the axis is defined by the 50 Hz or 60 Hz cycle. On HVDC equipment the productive views are pulse-sequence analysis and pulse repetition rate against amplitude, which is why converter-station practice leans on those displays instead. Ask what an analysis actually plots before quoting a phase-resolved figure for a converter station.
What the 2026 methods change, and what they do not
Multi-sensor fusion does not replace the phase-resolved view; it consumes it. The Scientific Reports framework combines acoustic and UHF features with wavelet and physics-informed terms, and its headline claim is about how little training data a classifier needs — not about which display an engineer should trust. CIGRE’s 2026 session work on digital-twin diagnostics makes a comparable point about placement: three sensors positioned from an attenuation model matched a conventional four-sensor layout. Both argue for better input, and the best input remains three synchronised channels read against the unit’s baseline.
PD monitoring and DGA answer different questions about the same transformer, and a fleet running both can cross-check a PRPD trend against gas generation; the DGA-900 covers the gas side. For the pattern library, see PRPD and PRPS pattern analysis.
Send us the three-phase records and we will read them against the unit’s baseline.
Sources
- Zhang, J. “A multi-task learning framework for diagnosing partial discharge types and assessing severity.” Scientific Reports 16 (2026).
- CIGRE Session 2026, paper A2_11489_2026 — digital-twin partial discharge diagnostics and sensor placement.
- IEC 60270:2025, Charge-based measurement of partial discharges, Ed. 4.0.
- IEC TS 62478; GB/T 42287; IEC 61850 (DL/T 860).
- TPD-400 figures: manufacturer datasheet (vendor data).