HVAC Fouling Detection with Root-Expansion Tiling-based Image Augmentation for Predictive Maintenance (RETINA-PdM)

Nicholas Strzelczyk, Lucas Hartman, Ethan Pigou, Santiago Gomez-Rosero, Miriam Capretz

Keywords: Computer Vision, Image Augmentation, Predictive Maintenance, Synthetic Data Generation, Fouling, HVAC

TL;DR: We developed a method to generate realistic, synthetic images of HVAC fouling, allowing us to train a highly accurate AI for energy-saving predictive maintenance.

Abstract: Heating, ventilation, and air conditioning (HVAC) systems experience reduced energy efficiency due to fouling on heat exchangers, a problem traditional maintenance struggles to address cost-effectively. While computer vision offers a scalable solution, its effectiveness is limited by the scarcity of labeled data for such rare anomalies. This paper introduces the Root-Expansion Tiling-based Image Augmentation for Predictive Maintenance (RETINA-PdM) algorithm, a method for generating realistic and physics-inspired synthetic data that simulates natural fouling growth. By training convolutional neural networks on data produced by RETINA-PdM, we achieved a predictive F1 score of 0.9642, a significant improvement over previous methods that relied on geometrically simple patterns. This work provides building operators with a highly accurate and cost-efficient tool for predictive maintenance, paving the way for substantial energy savings and optimized building operations.

Submission Number: 26