An non-destructive diagnostic method for complex wired networks based on distributed reflectometry and digital-twin-based transfer learning

Reflectometry is a non-destructive diagnostic technique for wired networks, where injected signals with specific time-frequency characteristics are analyzed through their reflections to detect faults. A common method for analysing these reflected signals is data-driven approaches, which involve training a fault diagnosis model using faulty network data. However, the scarcity of fault samples often makes it challenging to train effective fault diagnosis models. This paper proposes a fault diagnosis method for wired networks that integrates distributed reflectometry and digital-twin-based transfer learning, enabling improved fault diagnosis performance even when fault samples are limited. First, distributed reflectometry is enhanced with a generalised reflectometer arrangement to address issues of localisation ambiguity and signal attenuation. Next, an automated fault injection system is designed to collect actual fault data from RG58 coaxial, two-core RVV and multi-core RVV cables, while simultaneously generating two simulation datasets using a digital twin model based on LTSpice. Then, a deep transfer learning model, WiredNetworkNet, is introduced, demonstrating strong performance on real fault data by training on large-scale simulated data combined with limited real-world data. Finally, the model is tested on fault datasets from the three real cable types, showing significant improvement over current transfer learning methods.