Using Synthetic Data to Train a Visual Condition Monitoring System for Leak Detection
Ralf Gitzel, Arzam Muzaffar Kotriwala, Tobias Fechner, Martin Schiefer · 2021
Machine Learning algorithms excel at tasks such as classification but require a large amount of balanced training data. In this paper we use synthetic image data to train a Convolutional Neural Network to recognize industrial non-pipeline leaks. Using the open source raytracer Blender, a set of images is rendered. We discuss the key steps for obtaining the data, explain our network architecture, and discuss the performance on real-world data. The results using purely synthetic training data are quite encouraging with an F1 score of 0.95 on real test data.