Machine Learning Approach for Classification of Faults in Cable via Compressed Sensing Time-Domain Reflectometry

In this paper, a new method for cable diagnostics is proposed based on compressed acquisition of Orthogonal Multi-tone Time Domain Reflectometry (OMTDR). In this study, the Random Demodulator (RD) is used to sample the reflected signals, and we propose a Convolutional Neural Network (CNN) and a K-Nearest Neighbors (KNN) for classifying hard and soft faults in cables from compressed reflected signals without reconstruction. In our study, the training and test sets are obtained using a wireline diagnostic simulator. The classification results of the machine learning models demonstrate the high efficiency and accuracy of performing fault detection directly from the compressed data, with all fault types classified with an accuracy of over 99%.