Real-Time Partial Discharge Detection on Edge Devices

Background and Motivation

Partial discharges (PD) are early indicators of insulation degradation in medium-voltage cables and cable joints. Detecting PD events and tracking their severity and frequency in real time allows grid operators to intervene before catastrophic failures occur. Currently, PD detection relies on offline analysis of collected data. Since PD signals are typically high-frequency and short-lived, continuous transmission of raw waveforms is impractical and expensive. Moving detection capability to edge devices would enable immediate alerting while significantly reducing the volume of data that must be transmitted and stored.

Alliander is currently developing Vulcan, a multi-functional edge sensor for the mediumvoltage grid. Vulcan is designed to measure and analyze grid signals on-device. A key ambition is to perform PD detection directly on the sensor. Running PD detection algorithms on edge hardware introduces significant constraints. Edge devices have limited computational resources, memory, and power budgets. The algorithms must therefore be lightweight enough to execute in real time while maintaining sufficient accuracy to reliably distinguish PD events from noise and other transients.

Promising solutions come from TinyML, the field concerned with deploying machine learning models on low-power microcontrollers. Techniques such as model quantization, pruning, and knowledge distillation can reduce the computational footprint of neural networks by orders of magnitude. Classical signal processing methods, such as matched filtering, wavelet transforms, and spectral analysis, also remain relevant due to their computational efficiency and interpretability. This project will evaluate candidate algorithms for edge-based PD detection, implement the most promising approaches, and validate their performance on representative data.

Objectives

1. Establish a benchmark dataset and performance metrics: Curate a labeled dataset from real cable measurements containing confirmed PD events and normal operating conditions. Define performance metrics (e.g., detection rate, false positive rate, latency, computational cost) that reflect both detection quality and edge deployment feasibility.
2. Conduct a literature review and algorithm survey: Identify candidate algorithms suitable for real-time PD detection on edge devices. Approaches to investigate include matched filtering, quantized neural networks and other TinyML methods, reduced-order basis methods, wavelet transform-based detection, and spectral analysis techniques. Evaluate candidates based on accuracy, computational efficiency, memory footprint, and compatibility with target hardware.
3. Implement and optimize selected algorithms: Develop implementations of the most promising algorithms. Iterate on the design to improve detection performance, reduce resource requirements, and ensure robustness. Validate performance against the benchmark dataset and defined metrics.
4. Prepare field deployment: Ensure that mature algorithm implementations are ready for deployment on the Vulcan sensor. Document the code and provide guidance for integration and field testing.

What we would like to see

Do not take these as a strict requirement, a sufficiently motivated student that does not match the points below should most definitely still apply:

• Background in physics, electrical engineering, computer science, or math
• A strong interest in machine learning
• Experience with Python, should preferably be able to write code without assistance

Contact: Alex Kolmus (Alliander) & Sander Rieken (Alliander)