Artificial intelligence and machine learning are appearing throughout the power industry, and partial discharge (PD) analysis is no exception. Software now claims to classify PD patterns, reject noise and predict failures. For engineers responsible for real decisions, the useful question is not whether AI is fashionable but what it genuinely changes in the workflow – and where engineering judgement still rules.
Where AI genuinely helps
1. Pattern classification
Phase-resolved PD patterns (PRPD) are rich in information: internal voids, surface discharge, corona and interface activity produce characteristic shapes. Machine learning trained on labelled datasets can classify these patterns faster and more consistently than a junior engineer, flagging the pattern family for a senior review.
2. Noise rejection
Substation environments are noisy. Algorithms that learn the difference between PD and interference – periodic noise, switching transients, communication signals – improve the signal quality of automated measurements, a real gain for online monitoring systems that run unattended.
3. Multi-sensor correlation
Where several sensors watch one asset, AI can combine channels – UHF, acoustic, HFCT – to raise confidence in a genuine PD source and discard instrument artifacts.
4. Trend and anomaly detection
For fleets of monitored assets, algorithms detect deviations from each asset’s baseline automatically, prioritising which units deserve attention rather than flooding engineers with alarms.
What AI does not change
Measurement quality still comes first
An AI classifier is only as good as the data it receives. Sensor placement, grounding, background correction and measurement discipline remain the foundation – no algorithm rescues a badly measured signal.
Maintenance decisions remain human
AI can flag a pattern or a trend; deciding what it means for a specific transformer – whether to schedule an outage, sample oil, or add monitoring – stays engineering judgement informed by context.
Training data quality is the constraint
Classification accuracy depends on labelled training data. Suppliers must be able to explain what data their models were trained on and how they validate performance.
How AI enters the PD toolkit
- Instrument-side intelligence – handheld detectors and monitoring units with built-in classification and noise rejection.
- Software-side analysis – analysis platforms that process survey data, maintain trends and flag anomalies.
- Fleet-level platforms – where monitored fleets are aggregated for central asset management.
What to ask a supplier claiming AI
- What exactly does the AI do – classification, noise rejection, prediction – and how is it validated?
- What data trained the model, and does it cover our asset types and environments?
- Can we see the raw signals behind any AI conclusion, so an engineer can verify?
- Does it degrade gracefully – when uncertain, does it ask for review rather than assert?
About HUWOR
Zhuhai Huawang Technology Co., Ltd. (HUWOR) manufactures PD detection instruments and online monitoring systems, and follows the industry trend toward smarter analysis software while keeping measurement quality and engineering transparency at the centre. Instruments and systems are supplied worldwide with OEM/ODM options and application guidance. Contact the HUWOR sales team for the current analysis software capabilities.