Comparative analysis of machine learning approaches for leak detection in pipelines using hydrophone acoustic sensing

• Comparative study of feature-based and deep learning methods for leak detection. • Hydrophone sensing evaluated under varying pipe geometries and sensor placements. • Proposed a Feature-Informed CNN integrating domain features with raw signal learning. • Feature-Informed CNN achieved outperformed standard CNN and feature-based models. • Findings support robust, real-time leak detection for complex pipeline infrastructures. Reliable leak detection in pipeline systems is essential for minimizing water losses, protecting infrastructure, and ensuring the safety of distribution networks. This study investigates hydrophone-based acoustic sensing combined with machine learning (ML) methods for detecting and localizing leaks under varying pipeline geometries and sensor placements. A range of ML models was systematically compared, including conventional feature-based algorithms and deep learning approaches such as a standard Convolutional Neural Network (CNN) and an enhanced Feature-Informed CNN (FI-CNN). The key contribution of this work lies in the definition of statistical features—derived from fluid–structure interaction dynamics—and their integration with raw acoustic signals within the unified FI-CNN architecture. Controlled laboratory experiments were conducted on straight and U-shaped pipelines using multiple hydrophone configurations. Experimental results show that the Random Forest (RF) achieved the highest accuracy (81.2%) among feature-based models, the standard CNN reached 82.9%, and the FI-CNN achieved 90.4% accuracy with minimal performance variation across all configurations. These findings demonstrate that incorporating statistical features within deep learning architectures enhances robustness to pipeline geometry and sensor placement, offering a scalable foundation for real-time, data-driven leak detection in complex water distribution networks.