We evaluate the use of legacy power line modems (PLMs) for fault diagnostics, and in particular, focus on short-circuit faults in underground power cables. Prior works have shown that broadband power line communication channel estimates that are computed within the PLMs can be used to gain insight into the health of underground cables. However, several legacy PLM chip-set implementations do not provide access to the estimated channel frequency response in its entirety. Therefore, to facilitate and accelerate a practical roll-out of a PLM-based diagnostics solution, we investigate if readily extractable parameters, such as the estimated signal-to-noise ratio values and/or the computed precoding matrices in case of multiple-input multiple-output (MIMO) transmission, provide sufficient indication into the cable health status. By extracting suitable features from this raw data, we show through simulations that our machine learning based automated cable diagnostics solution achieves satisfactory results in predicting faults, and near-perfect performance in fault identification.
