Applied use of AI for stator winding insulation diagnosis using online PD monitoring

Traditionally, maintenance personnel have performed the PD measurement during offline tests conducted during scheduled shutdowns. Despite advancements in technology and adoption of online techniques, offline testing remains prevalent due to the complexity of diagnosing PD under normal operating conditions, where electrical noise, interference from other equipment and multiple simultaneously present PD sources complicate analysis.

Online PD monitoring avoids the need to measure during machine stoppage, contributing to better maintenance planning and downtime reduction. However, it introduces the challenge of distinguishing machine-originated discharges from external noise coming from connected power electronics, and discharges from surrounding assets, such as cables or terminations.

Accurate diagnosis requires isolating each defect’s contribution and interpreting both phaseresolved patterns and activity trends.

While current methods allow for such analysis, they typically demand experienced qualified technicians capable of performing data clustering in a multidimensional space extracted from measurements and isolating PD signatures related to defects, for the purpose of interpreting complex PD patterns.

The most advanced systems offer permanent continuous online PD monitoring, enabling early detection of defects and tracking of defect evolution under varying operational conditions.

However, these systems often require regular and constant expert oversight, which is rarely feasible in practice, either because of resources unavailability and/or because of unaffordable oversight costs.

This paper presents a novel approach to online PD monitoring that leverages a combination of automated clustering algorithms with neural network-based models for PD defect identification and classification: while automated clustering algorithms separate different PD pulses sources, the artificial intelligence (AI) model perform the PD pattern classification and defect identification. By using this approach, the system significantly reduces the need for expert intervention, enabling rapid and user-friendly deployment of online monitoring solutions.

Diagnostic alerts are generated only when critical defect patterns are detected, making both periodic and continuous monitoring viable.

The development process involved designing, implementing, and validating artificial intelligence (AI) tools specifically tailored for rotating machine insulation. A critical part of the implementation was the creation of a training dataset compiled from real-world PD measurements provided by manufacturers and maintenance teams: as raw data of PD measurements is not normally retained once the maintenance report is done, the training data was generated mainly from PRPD diagrams contained in the available maintenance reports. On the other hand, PRPD diagrams do not contain other information regarding physical characteristics of the PD pulses, which prevented proper PD pulses clustering. To fill this lack of input information, clustering technics used in other high-voltage assets were applied and tailored to the rotating machine solution.

Following successful laboratory validation, the method was applied to machines during field measurements, demonstrating promising results and confirming the effectiveness of AI-driven online PD diagnostics in predictive maintenance.