Passive Acoustic Monitoring of Underwater Well Leakages with Machine Learning: A Review
Guanlin Zhu, Zechun Deng, Jiaxin Shen, Junchi Yang
TL;DR
This review addresses environmental risks from abandoned subsea wells and evaluates AI-enhanced passive acoustic monitoring (PAM) as a non-invasive, real-time approach to detect leaks in noisy deep-sea conditions. It surveys AI-enabled strategies across signal differentiation, noise suppression, and data interpretation, including semantic front-ends at the ADC and denoising and backhaul enhancements via underwater optical communication (UOC). A key contribution is the proposed hybrid workflow combining non-negative matrix factorisation (NMF) with CNN/GRU/TCN temporal models to classify and quantify leak events, and the framing of semantic communication with an intelligent front-end as formalised by $ \min\limits_{g(\cdot)} H(\widetilde{X})$, subject to $\widetilde{X}=g(X)$ and $I(\widetilde{X};Y)\ge I(X;Y)$. The findings indicate AI-assisted PAM can enable real-time, energy-efficient monitoring and better environmental protection, though challenges like data scarcity, domain shift, and computational limits must be addressed for field deployment.
Abstract
Abandoned oil and gas wells pose significant environmental risks due to the potential leakage of hydrocarbons, brine and chemical pollutants. Detecting such leaks remains extremely challenging due to the weak acoustic emission and high ambient noise in the deep sea. This paper reviews the application of passive sonar systems combined with artificial intelligence (AI) in underwater oil and gas leak detection. The advantages and limitations of traditional monitoring methods, including fibre optic, capacitive and pH sensors, are compared with those of passive sonar systems. Advanced AI methods that enhance signal discrimination, noise suppression and data interpretation capabilities are explored for leak detection. Emerging solutions such as embedded AI analogue-to-digital converters (ADCs), deep learning-based denoising networks and semantically aware underwater optical communication (UOC) frameworks are also discussed to overcome issues such as low signal-to-noise ratio (SNR) and transmission instability. Furthermore, a hybrid approach combining non-negative matrix factorisation (NMF), convolutional neural networks (CNN) and temporal models (GRU, TCN) is proposed to improve the classification and quantification accuracy of leak events. Despite challenges such as data scarcity and environmental change, AI-assisted passive sonar has shown great potential in real-time, energy-efficient and non-invasive underwater monitoring, contributing to sustainable environmental protection and maritime safety management.