Publication: Self-supervised learning frame work and localization using micro air vehicles for water leak detection
Loading...
Date
2024-09-01
Authors
Mohd Yussof, Nurfarah Anisah
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Real-time detection and localization of water leakage are crucial for effective watermanagement in smart buildings. Traditional detection technologies based on static sensors frequently entail significant costs for both installation and operation. Utilizingsmall mobile robots such as Micro Air Vehicles (MAVs) provides a cost-effective and efficient for detecting objects in confined areas. Nevertheless, due to constrained processing power andlimited payload capacities, MAVs canonlydependonlightweight sensors, such as, tiny thermal sensors, which provide low image resolution and hence reduce detection distances typically within 1 𝑚. Therefore, this research presents a Self-Supervised Learning (SSL) framework, where a computer vision algorithm is developed to directly detect water leakage from thermal images, which then is used as supervised output for the training of a deep learning model by using RGB images as input. A pre-trained YOLOv4-tiny model is fine-tuned using 1080 laboratory images. Training test with 50,000 steps and 340 negative images achieves an optimal balance of accuracy, with a detection time of 0.0617 𝑠 and an average precision of 98.97%. In addition, a control strategy that combines the RGB deeplearning modeland the thermal vision algorithm is shown to allow autonomous MAVs for preliminary predictions ofwater leakage from further distances and accurately localize the leakage areas when they get closer. To validate the proposed concept, static detection tests were conducted, followed by flight tests in indoor environments. In static tests, the SSL-trained model extends the detection range from 1 𝑚 to 3 𝑚. In real-world flight tests, two scenarios are conducted: three experiments with varying initial positions and six experiments targeting different leak locations. Both static and flight tests confirm the effectiveness of the control strategy and detection algorithm in localizing water leaks in indoor environments. This research advances the sensory capabilities of MAVs equipped with RGBand thermal cameras and extends their detection range of water leakage, thereby mitigating potential damage in large or complex indoor environments.