<p> A reliable uninterrupted electricity supply is essential to a functioning society and a growing economy. Traditionally, low voltage distribution networks (LVDN) are less well monitored, despite being more extensive. The reliability, cost, and compatibility of sensors are the primary constraints to implementing a large-scale monitoring system at LVDN. Faults at this level have the potential to damage public property and cause fatal accidents, but their detection is complicated by the characteristics of some faults. Therefore, this thesis’s overall goal is to evaluate the feasibility of non-contact giant magneto resistive (GMR) sensors for LVDN overhead line monitoring. Two applications were evaluated: 1) non-contact current sensing and 2) fault detection and classification. A review of low and high impedance faults, their characteristics, and the existing detection methods indicates that fault detection schemes and characterization have mainly focused on the high voltage levels. In this thesis, a novel 400-volt (3φ-neutral) physical test facility is developed with proper personnel safety and system security following the industry guidelines, to generate fault and system event data for analysis. Cross-coupled alternating current (AC) magnetic fields from the overhead line arrangement were captured via a GMR sensor. Three typical off the shelf single-axis GMR sensors were used to develop a 3-axis GMR sensor head to capture the spatial magnetic fields. The sensors were individually calibrated with respect to one single true calibrated sensor. This thesis describes the procedure and for achieving the correct alignment of the sensor output with the AC phase direction. A method for fault detection and classification using GMR sensor measurements is also presented in this thesis. This was developed using deep learning algorithms. Two supervised deep learning algorithms were evaluated to classify the fault and system event signatures from the magnetic field measurements. The analysis shows a hybrid model designed with a convolution neural network and gate recurrent unit performed best. This followed by a proposed decision framework is shown to detect and classify faults and normal system events with high security and reliability. In addition to the detection of faults, this thesis also confirms the feasibility of using non-contact calibrated GMR sensors for overhead line current sensing. The analysis shows a minimum of two vertically placed 2D sensor heads are required to calculate the individual phase current data from magnetic field measurements. Compared to measured data, this calculation was found to be more than 90% accurate. In summary, this thesis presents the differentiating characteristics of faults and system events observed in experiments in a purpose-built 400-volt network simulation facility. Non-contact GMR sensors are shown to be an alternative to traditional current sensors for LVDN current sensing, and for fault detection and classification via deep learning (DL)-based methods. </p>
