Machine Learning Applied to Thermally Aged Cable Reflectometry

Recent developments in instrumentation have demonstrated that it is feasible to monitor the condition of energized cables online using frequency domain reflectometry (FDR) and spread spectrum time domain reflectometry (SSTDR). However, the response spectra from these measurements are complex to interpret and do not lend themselves to simple threshold alarms. To address this challenge, machine learning (ML) techniques have been used, demonstrating high accuracy in predicting normal or anomalous cable behavior when using binary normal and anomalous training and test data. A more plausible scenario for cable insulation damage, however, involves a slowly developing material change due to long-term exposure to thermal or radiation stresses, subtly altering material properties and resulting in an altered reflectometry response. The practical challenge lies in recognizing when such changes are significant enough to raise concern. The Pacific Northwest National Laboratory (PNNL) Accelerated and Real-Time Experimental Nodal Analysis (ARENA) cable motor test bed was used to measure the FDR and SSTDR responses of an energized cable as a section of it was thermally aged over 70 days. Online reflectometry spectra were collected and post-processed to simulate real-time analysis, aiming to distinguish normal from anomalous behavior. These spectra were also contrasted with off-line and direct conductor-coupled reflectometry tests. For use with ML algorithms, it was necessary to label the reflectometry results as either normal or anomalous. This was accomplished with witness samples, which were aged alongside the main cable and periodically tested for elongation at break and tensile strength-two offline destructive tests indicative of cable damage. These destructive tests revealed a natural breakpoint for differentiating the two conditions at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\sim 35$</tex> days. Both supervised and unsupervised ML methods were applied. The results indicated that the ML methods could effectively classify cable data as normal or anomalous and that there was a strong correlation between the reflectometry results and the elongation at break and Fourier transform infrared spectroscopy destructive off-line tests of the witness samples. In addition, the online testing was nearly as clear and effective as off-line tests. These findings suggest that the integration of ML techniques with reflectometry can provide a reliable method for monitoring and early detection of cable insulation damage.