Fault Diagnosis and Isolation Prediction for Redundant Relays Based on Discrepancy Analysis

Protective relays are integral to the reliability of any electrical power system, and are fundamental to the decision-making of their protection systems. They support the detection and isolation of problems in the power system, so that the operation of unaffected parts can be maintained. ElectroMechanical Relays (EMRs) are still predominant around the globe in the high and extra high voltage transmission systems. Thus, ensuring the reliability and traceability of relays is of major importance. One way to achieve this is through parallel redundancy by implementing redundant sensor architectures, such as one-out-of-two (1oo2). In this paper, we propose a novel algorithm for the fault prognosis-i.e., detection and failure date prediction – and isolation prediction for redundant 1002 architectures. The algorithm predicts the failure ahead of time and provides an estimated date for the failure event. Our contribution in this work is on the fault isolation prediction, where we infer-ahead of time, before the occurrence of the fault-which relay of the pair will cause the failure. The fault isolation is achieved by means of Machine Learning (ML) based feature extraction and binary classification methods. We apply the algorithm on EMRs based solely on the discrepancy time signals of the opening and closing events of the relays. The algorithm has been tested on data from redundant EMRs from the publicly available SOReDD dataset. While relays are binary switches, our work could potentially not only be applied to other types of binary switches but also to binary sensors as they also produce a binary output signal.