Fault Diagnosis in Hybrid-Fiber Coaxial Networks Using Sliding Window-Based Features

Proactive Network Maintenance (PNM) is a cornerstone for cable network reliability. Accurate fault detection and diagnosis of faults in hybrid-fiber coaxial (HFC) networks are significant for providing customers high service quality, optimizing network operations and minimizing related costs. Fault detection and diagnosis in this industry has been widely explored via labeling and data-driven techniques. Yet, academic contributions lack of studies focusing on cable network fault diagnosis via Full-Band Capture (FBC) downstream data analysis. Another criticality in the field is the absence of ground truth and expertise uncertainty around the correct fault labeling of raw impaired data. With basic expertise knowledge about the fault types and driven by the assumption of single cable modem (CM) signal representative of a single fault state, this paper offers a fault diagnosis scheme using FBC downstream data. At first, a data matrix is constructed by concatenating 78 distinct features computed for a series of empirical sliding windows and steps. Next, we apply augmentation techniques to balance the classes considering different augmentation ratios. Following, we employ Pearson Correlation for the reduction of highly correlated features and Genetic Algorithm (GA) for the final feature selection. Random Forest is used as surrogate model in GA. Through different experimental runs, our approach shows high classification accuracy across nine classes of network fault states, establishing a foundation for state-of-the-art diagnosis results in the field.