Measurement Aided Training of Machine Learning Techniques for Fault Detection Using PLC Signals

The re-use of channel estimation performed by power line communication (PLC) modems for monitoring of cable health conditions has recently been investigated in several works. In particular, cable diagnostics solutions based on machine learning techniques have been shown to process the PLC channel-estimation samples intelligently to differentiate fault conditions from the benevolent load changes. Previous studies have been based on synthetically generated training and test signals to optimize and validate the machine learning models. To deal with the mismatches between the purely synthetically generated signal samples and those encountered in a real implementation, in this paper, we propose S-parameter measurement aided generation of channel estimation samples. Specifically, we describe the behaviour of our device under test (DUT) through its S-parameter measurement and synthetically generate varying terminal load conditions. Then we train and use machine learning models to determine the health of the DUT. We describe the proposed approach and apply it to data obtained from laboratory measurements.

I. Introduction

Cable diagnostics constitutes an indispensable part for the asset monitoring to ensure the safe operation of smart power grids [1] . Various cable diagnostics solutions have been developed in the past to help preventing cable in-service failures, which could lead to potentially dangerous situations and severe economic losses [ 2 , Ch. 6]. One approach is to determine the cable health conditions by analyzing the high-frequency signals propagated along the cable. This is based on the principle that the inception of a cable anomaly alters the electric signal propagation [ 2 , Ch. 6].