PNNL

Abstract

Developments in cable test instrumentation coupled with artificial intelligence and machine learning (ML) to aid in interpretation of cable test signals supports the feasibility for automated analysis of reflectometry tests for low voltage power cables. This work seeks to leverage prior ML work and success for low voltage cables to evaluate potential application to medium voltage (2kV to 10kV) installations. The Accelerated and Real-Time Environmental Nodal Assessment (ARENA) Cable Motor Test Bed at Pacific Northwest National Laboratory (PNNL) was used to test a medium voltage cable with several types of damage including thermal aging and low resistance conductor-to-shield faults. The cable was tested using an inductive clamshell coupler to protect the test instruments from the energized cable voltages that would damage the test instruments if coupled directly to the energized conductor. Both unsupervised and supervised ML methods were applied. Observations and conclusions include: • Extending models to assess reflectometry signals from cables of different lengths based on several normalization strategies was assessed and proved problematic. Visually, the alignment of cables using dynamic time warping (DTW) was successful. However, damage prediction accuracy fell below 50% when using the unsupervised method and from 22 to 61% using the supervised method. We suspect that the actual noise from damaged parts of the cables was aligned with the random noise that can be seen in all cables, making damaged cables indistinguishable from undamaged cables. This is particularly important for supervised learning approaches where good and damaged training data are likely to be from different cable lengths. This is less important for unsupervised approaches that will normally take baseline data when first connecting to the cable of interest and look for changes to that baseline. • The unsupervised model achieved a maximum weighted accuracy of 93.8% using the vector network analyser (VNA) Frequency Domain Reflectometry (FDR) Direct Connect data at 100 MHz with real preprocessing. The analysis indicated that Direct Connect and Isocoupled Unenergized methods produced significantly higher accuracy compared to other connection methods. Furthermore, higher frequencies correlated with better performance, especially when using the Direct Connect and Isocoupled Unenergized methods. Among preprocessing methods, real and imaginary preprocessing performed best, while the complex magnitude preprocessing method performed the worst. • Overall, this study demonstrates that the unsupervised ML methods previously developed for low voltage cables extend effectively to medium voltage cables, showcasing the robustness of the framework and approaches. These findings can serve as a guideline for commercial applications, aiding in the development of technologies that enable condition-based qualification of cables, ultimately enhancing the reliability and safety of nuclear power plant operations.