Unsupervised Fault Detection in Transmission Lines: Development and Evaluation on a CPS Testbed

Accurate and timely fault detection in transmission lines is critical for maintaining the reliability and stability of power systems. Traditional methods predominantly rely on supervised machine learning techniques that require extensive labeled datasets, which are challenging and costly to obtain in real-world settings. This paper introduces unsupervised machine learning approaches for fault detection in power transmission lines, eliminating the dependency on labeled data while achieving high detection accuracy. We developed a high-fidelity Cyber-Physical System (CPS) co-simulation testbed utilizing Hardware-in-the-Loop (HIL) simulation, built with HYPERSIM on the OPAL-RT platform and integrated with EXata CPS and SEL-351S relays, to generate comprehensive datasets representing various fault and normal operating conditions in the IEEE test systems. By leveraging advanced clustering and anomaly detection techniques-including Gaussian Mixture Models and Hidden Markov Models-the proposed method effectively identifies patterns in operational data and detects deviations indicative of faults. Experimental results demonstrate that our unsupervised approach achieves detection performance comparable to supervised methods. This work offers a scalable and adaptable solution for power systems where labeled data is limited or unavailable, significantly advancing fault detection methodologies and enhancing grid resilience without the burden of extensive data labeling.