IoT Edge level solution for analyzing fast occurrences, utilizing oscilloscope smart sensor: Partial discharge detection and analysis utilizing oscilloscope smart sensor

Abstract

This master’s thesis designs and evaluates an Internet of Things (IoT) edge-level prototype solution for detecting and analysing fast occurrences, focusing on online Partial Discharge Detection (PDD) in medium-voltage electrical grids. The proposed solution combines an oscilloscope-based smart sensor with an edge computing unit to enable automated event capture, local data filtering and feature extraction without relying on continuous cloud streaming. The prototype uses a PicoScope 2406B as a programmable oscilloscope sensor and a Raspberry Pi Compute Module 4 as an edge device. In laboratory testing, the oscilloscope was configured to acquire four-channel waveform blocks at 62.5 MHz over 40 ms windows, producing 2.5 million samples per channel per capture. With this configuration, the solution reliably detects PD events with rise times up to 32 ns and frequency content up to 31.25 MHz. Events with faster transients or higher-frequency components are also detected to around 40–60 MHz with varying amounts of aliasing. Depending on the discharge type, this detection threshold can be greater, depending on how the discharge permeates in the powerline or insulation. Due to limited access to the originally planned real PD datasets from the TUNI INGA research project, the evaluation was performed using simulated PD waveforms with partial-discharge characteristics and realistic grid noise, replayed via an arbitrary waveform generator to a smart sensor oscilloscope. The edge-side analysis pipeline applies PRPD-based feature extraction and statistical metrics (e.g., skewness, kurtosis, and peak characteristics) to classify discharge type as PD/no-PD. Classification accuracy was high for several fault types with clear signal and type features (e.g., internal discharge 98.9%, particle discharge 92.0%, tracking 88.7%, corona 87.0%) but lower for more irregular, random or overlapping classes (e.g., electrical treeing 47.5%, water treeing 41.5%, tree contact/arcing 23.6%). Classification for data with no PD was close to perfect with 98.7%. The macro-average F1 score across all classes was 68.1%. Performance measurements show feasible acquisition and buffer transfer, while persistent storage and classification dominate processing time (file write 7803 ms; statistical classification 5970 ms on average). The proposed solution is suitable for detecting fast occurrences, in this case PDs, and is capable of analysing and classifying them.