Background
Guy wires are tensioned cables designed to add stability to the transmission pole. If the wire is not correctly tensioned, it can be a potential risk for the public and the powerline. Being able to find defective guy wires on time helps the client’s customers prioritize their maintenance tasks and prevent further problems in the electric grid.

Solution
The objective of the project was to create a machine learning pipeline to analyze images from power lines and detect lack of tension in guy wires. To solve this problem, the team broke down the problem in different stages and created a multi-model AI solution, in close collaboration with the client. The team developed and assessed different image segmentation models and classification approaches. The solution is a combination of a state-of-the-art segmentation model for identifying wires, an object detection model for selecting the right wire, and an iterative version of a regression model to fit mathematical curves to the point set. Finally, a measure of curvature is calculated and thresholded for detection, indicating if a guy wire is straight or saggy.

Tools/Tech
The training and validation pipelines were developed in Python and PyTorch and data versioning was performed using DVC. Various open-source Python libraries were used for image processing, training, and validation. The solution was deployed in the client’s cloud and is run as a batch job and assesses incoming images on a regular basis.