Most maintenance departments running their own high-voltage distribution already do acoustic partial-discharge detection. A technician walks the switch room once or twice a year with an ultrasonic probe and a TEV instrument, listens at each panel, and writes down what they hear. The method works: it is non-contact, it needs no outage, and the physics is sound. Its weakness is not sensitivity but sampling. Insulation degrades continuously and the survey happens twice a year. This article is about closing that gap, and about what actually stands in the way, which is rarely the detector.
Why the survey is the right starting point
There is an argument in industrial AI that says a new method has to displace an old one to be worth buying. In switchgear the opposite is true, and it is the reason we lead with this application.
The plant already believes that partial discharge radiates acoustically, because its own technicians hear it through a probe. The plant already has a decision process for what to do when discharge is found. And the plant already accepts that the measurement is indirect. None of that has to be sold. The only question on the table is whether the same measurement can be made continuously, without a technician, at a false-alarm rate that a maintenance planner will tolerate.
That is an engineering question with a measurable answer, which is a much better place to start than a discussion about whether AI works.
What continuous changes, technically
The environment stops being controlled
A technician surveys when the room is quiet and stands where the reading is best. A permanent sensor sits in one position through every shift, every season, next to whatever else is running. Compressed-air leaks, contactor operations, cleaning, drilling in the next bay and the switchgear fans all radiate ultrasound. Detection under those conditions is the whole problem, and it is why a threshold on ultrasonic amplitude fails within a week of installation.
The representation carries more weight than the classifier
Work our founder led and published at INTER-NOISE 2021 (DOI 10.3397/in-2021-2373) compared several time-frequency representations of airborne acoustic emission before training deep networks to detect and classify discharge events. The finding that transfers most directly into a deployment is that the choice of representation did as much work as the choice of network. Discharge pulses are short, broadband and repetitive with a relationship to the power cycle; representations that preserve that structure make the learning problem easy, and representations that smear it make it hard regardless of model capacity.
The evaluation has to be written around maintenance, not accuracy
A detector that is right 99 percent of the time and wrong twice a week will be ignored by the third week. The quantity that matters is alarms per panel per month at the sensitivity the plant wants, and the severity trend that lets a planner schedule rather than react. We design the evaluation around that before training anything.
What a deployment looks like
- Ultrasonic-capable microphones at panels and bays. Air-insulated components only: airborne acoustics reaches switchgear, bushings, insulators and cable terminations, and does not reach inside an oil-filled transformer tank. Transformers are monitored with tank-mounted acoustic-emission sensors instead, which is a different sensor and a different installation.
- An edge device per switch room. Detection and classification run locally and continuously. Only events, their acoustic fingerprint and a severity trend are forwarded. Raw audio never leaves the room, which keeps the IT security review short and the bandwidth requirement near zero.
- Integration with the system the planner already uses. An event that does not appear in the maintenance system does not exist. This is usually a small amount of integration work and the single biggest determinant of whether the installation is still in use a year later.
- A baselining period. Every switch room has its own background. A few weeks of normal operation before the detector goes live is what separates a system that alarms on the cleaning trolley from one that does not.
For groups with several sites, federated learning lets detectors improve across substations without centralising recordings, which matters when the recordings would otherwise document the operating pattern of the plant.
Who this is actually for, in India and in Europe
Utilities are the obvious buyer and the hardest one to reach: transmission and distribution utilities procure through state tenders and registered-vendor lists that take years to enter. The reachable first buyers are industrial: steel, cement, chemical and large automotive sites running their own high-voltage distribution, captive power plants, and independent producers. They own the asset, they carry the outage cost directly, and they can sign a project without a tender.
The argument in those plants is simple and does not need an AI framing at all. An unplanned transformer or switchgear failure stops the site. The US Department of Energy puts predictive maintenance generally at 35 to 45 percent less downtime and 25 to 30 percent lower maintenance cost (FEMP O&M Best Practices Guide); electrical assets are where that arithmetic is least contested, because the failures are expensive and the warning period is long.
The honest limits
Airborne acoustic detection is a screening method, not a diagnosis. It tells a planner that discharge activity is present at a location and how it is trending. Locating the defect inside a panel, deciding whether it is surface tracking or a void, and judging remaining life still need a specialist with the right instruments and, usually, an outage.
It is also limited to what radiates into air. Inside sealed, oil-filled or gas-insulated equipment, the acoustic path is different and so is the sensor. Any vendor who offers one microphone for all of it is selling past the physics.
The value of continuous monitoring is not that it sees more than the technician. It is that it never goes home.
Key takeaways
- Handheld ultrasonic and TEV surveys already work; the deficiency is sampling frequency, not sensitivity.
- Permanent sensing has to survive an uncontrolled acoustic environment, which is what defeats fixed thresholds.
- The time-frequency representation matters as much as the network, per the INTER-NOISE 2021 work.
- Evaluate on alarms per panel per month and severity trend, not on classification accuracy.
- First buyers are industrial sites with their own HV distribution, not utilities; airborne acoustics covers air-insulated assets only.
If any of this matches a problem on your line, the fastest way to find out what is possible is a free discovery call followed, where it makes sense, by a feasibility study of two to ten days.