A load identification and diagnostic framework for aggregate power monitoring

Abstract

Power monitoring solutions have the potential to collect large amounts of data from the operation of electromechanical loads, such as measurements of power, torque, vibration, and acoustic signals. These measurements can act as unique identifiers for the early identification of degrading system performance, providing a rich feature space for fault detection and diagnostics (FDD). However, mainstream machine learning methods may overlook potential features with key physical context in the development of soft faults due to a lack of faulty load data in publicly available datasets. Therefore, a physically informed feature space must be selected and evaluated specifically for FDD applications in which load behaviors evolve over time. This thesis presents both a method for evaluating a potential load disaggregation feature space and a framework for load classification based on adaptive load benchmarks and health tracking.