Abstract
Unmanned Aerial Vehicles (UAVs), also known as drones, have seen increasingly widespread use in recent years with advancing technology in a variety of fields, and present an option to improve the capability and accessibility of remote data collection. One area that may benefit from this capacity is the maintenance of electrical equipment, particularly in circumstances in which the equipment owned by electrical utilities covers wide areas, and more conventional sensor options may be expensive to implement at scale. However, the cost and accessibility of UAVs and their sensor payloads can also limit the use of this technology, and options for specific sensor capabilities may be limited to certain UAV platforms by weight, power usage, and compatibility. This dissertation will investigate an approach to implementing a custom, relatively low-cost camera system with accessible components, and compare the performance of machine learning based image classification to integrate the separate components for the example application of the detection of electrical arcing. Specifically, this research will examine a low-cost camera system composed of three industrial cameras, which collect light in different parts of the electromagnetic spectrum: visual, ultraviolet, and infrared. Each of these wavelength bands contains information that could assist in the detection of an electrical arc, with light emitted from the arc in the visual and ultraviolet spectra, and infrared radiation from the arc increasing the radiant heat of nearby metal components which in turn emit infrared radiation. This dissertation will use this prototype camera system to generate a custom dataset of electrical arcing examples in varying ambient outdoor conditions, including bright sunlight and low light, and use this data to train separate and ensemble machine learning based image classification models. A comparison of different model parameters and ensemble approaches will be used to address the strengths and limitations of the cameras and the approach to arc detection. The goal of this research is to provide insights into the use of machine learning for custom low-cost sensor systems, what strengths it can demonstrate, and what potential weaknesses a developer or user of such a system may benefit from addressing.
