Hybrid method for enhancing acoustic leak detection in water distribution systems: Integration of handcrafted features and deep learning approaches

Leak detection of water pipelines is of great significance to ensure water supply safety and conserve water resources. In the context of the prevalence of deep learning which can automatically extract features, deep learning-based acoustic leak detection methods are flourishing. The study compares the deep learning methods and traditional machine learning methods which need human-involved feature engineering, from aspects of detection performance, data requirement, and computational complexity. Furthermore, a hybrid leak detection method that integrates handcrafted features into deep learning networks is proposed. In this study, there are 70 features extracted from acoustic signals, constructing an extreme gradient boosting (XGBoost) classifier. Then three commonly used deep learning classifiers and the hybrid classifier are developed to carry out comparisons. Results show that handcrafted features, especially linear prediction features, still play an important role in acoustic leak detection at the current stage. The XGBoost classifier with key features can outperform deep learning classifiers and achieve acceptable computational complexity. Limited by insufficient data, deep learning classifiers cannot fully exploit their ability to directly extract useful features from model inputs (e.g., acoustic signals' waveforms in the time or frequency domain). The hybrid classifier, requiring neither careful selection of handcrafted features nor complex deep learning network structures, obtains the best detection performance using the same amount of data, showing great promise in practical applications.