Integrating advanced frequency-domain signal processing with machine learning for accurate leak detection in subsurface CO2 storage

Abstract

Ensuring the integrity of geological CO₂ storage is critical for the long-term success of carbon capture and storage (CCS) technologies. The detection and localisation of potential leakage events rapidly and accurately remains a key challenge, particularly under constraints of limited monitoring data. This study presents a proof-of-concept framework that integrates advanced frequency-domain signal processing with machine learning to address this challenge using only pressure data from a monitoring well in a CO₂ storage site. Here, pressure signals are translated into the frequency domain using Fast Fourier Transform (FFT) in order to extract physically meaningful features that are highly sensitive to leakage phenomena. These features capture subtle variations in signal behaviour that are often missed in time-domain analysis. A two-stage machine learning pipeline is also developed, involving a classification stage to distinguish leak versus no-leak conditions, followed by leak localisation in a regression stage.

The results showed that in the leak detection stage, ensemble and probabilistic classifiers, particularly Naive Bayes (test accuracy = 0.9873, F1 = 0.9788) and Random Forest (test accuracy = 0.9823, F1 = 0.9016), outperformed linear models by a substantial margin. In the localisation stage, the K-Nearest Neighbours Regressor (test R² = 0.9899, MAE ≈ 6.8 m) and Gradient Boosting Regressor (test R² = 0.9790, MAE ≈ 9.5 m) achieved the highest spatial prediction accuracy. Additionally, the findings demonstrate that frequency-domain feature engineering substantially enhances leak-detection sensitivity and spatial inference accuracy compared to time-domain methods. The proposed framework is computationally efficient, requiring only sparse pressure data, and can be integrated into real-time monitoring systems.