Date of Award

January 2018

Document Type


Degree Name

Master of Science (MS)


Mechanical Engineering

First Advisor

Jeremiah Neubert


This work presents a novel system utilizing previously developed convolutional neural network (CNN) architectures to aid in automating maintenance inspections of the dead-end body component (DEBC) from high-tension power lines. To maximize resolution of inspection images gathered via unmanned aerial systems (UAS), two different CNNs were developed. One to detect and crop the DEBC from an image. The second to classify the likelihood the component in question contains a crack. The DEBC detection CNN utilized a Python implementation of Faster R-CNN fine-tuned for three classes via 270 inspection photos collected during UAS inspection, alongside 111 images from provided simulated imagery. The data was augmented to develop 2,707 training images. The detection was tested with 111 UAS inspections images. The resulting CNN was capable of 97.8% accuracy in detecting and cropping DEBC welds. To train the classification CNN if the DEBC weld region cropped from the DEBC detection CNN was cracked, 1,149 manually cropped images from both the simulated images, as well images collected of components previously replaced both inside and outside a warehouse, were augmented to provide a training set of 4,632 images. The crack detection network was developed using the VGG16 model implemented with the Caffe framework. Training and testing of the crack detection CNNs performance was accomplished using a random 5-fold cross validation strategy resulting in an overall 98.8% accuracy. Testing the combined object detection and crack classification networks on the same 5-fold cross validation test images resulted in an average accuracy of 73.79%.