An Attention-Enhanced YOLOv8 Architecture for Leakage Detection in Building Systems

It has been one of the most critical challenges to maintain a healthy, livable, and sustainable building environment. As part of such efforts, it is desired to efficiently inspect roofs that cover a large portion of building areas and quickly detect anomalies to carry out timely maintenance. However, due to the limitations of traditional inspection technologies, roof leakages caused by material deterioration are hard to detect until the moldy features become visible with certain sizes and shapes during their propagation. This has motivated the use of an artificial intelligence (AI) method to recognize leakage areas from images captured by unmanned aerial vehicles (UAVs). In this work, an attention-enhanced You Only Look Once version 8 (YOLOv8) model is proposed to enable cost-effective monitoring and accurate detection of hidden leakages with the purpose of predictive maintenance through images. The attention-enhanced (AE) module coordinates to features (c2f) was conducted in the backbone to cope with a practical challenge that infiltration features might be extremely tiny and subtle compared with background noise. More connections to the neck were established subsequently by convolution and concatenation operations, generating multistage information fusion from the feature maps in different layers. Finally, a real-world case on University of Arkansas campus was analyzed to validate the model’s performance including accuracy, efficiency, and robustness. It proved to be an effective nondestructive method of promoting AI applications in managing resilient building systems.