Acoustic Identification of Water Supply Pipe Leakage Based on Bispectrum Analysis

Abstract

Traditional methods for classification and identification of water supply network leaks rely on manual auditory inspection. Intelligent identification of pipeline leaks based on acoustic signals heavily relies on the extraction of sound signal features. Due to the weak periodicity and “noise-like” characteristics of leak sound signals, power spectrum analysis offers limited value for studying their nonlinear vibrations. Bispectrum features provide non-Gaussian information to the power spectrum, facilitating the detection and characterization of nonlinearity in pipeline systems. The data used to test the hypotheses included acoustic signals of leaks in pipes of various materials (cast iron, PVC, and steel) as well as typical background noise signals at the point of leakage. This study compares bispectrum plots of different types of signals from leak sound signal collection sites, visualizing their unique features. Integrating these insights with a convolutional neural network (CNN) model, the research effectively detects and classifies leak sound signals collected from municipal water networks. Experimental results indicate that the proposed CNN model achieves a test accuracy of 93.08% and an F1 score of 93.43%, with a false positive rate below 10%. These findings demonstrate the model’s robust capabilities in leak identification and misclassification prevention, offering significant savings in detection time and training costs. This study introduces an innovative approach to pipeline leakage detection by combining bispectrum analysis with a convolutional neural network (CNN). Bispectrum analysis captures the nonlinear and phase-coupled characteristics of leakage signals, which are often overlooked by traditional methods relying on manually designed features. The integration with CNN enhances the automatic extraction and emphasis of these complex features, leading to improved detection accuracy and robustness, particularly in noisy and diverse real-world conditions. The dataset, comprising acoustic signals collected from pipelines with varying materials (e.g., cast iron, PVC, steel) and realistic noise environments (e.g., pump stations, mechanical vibrations), ensures the method’s applicability to practical leakage scenarios. This approach demonstrates significant potential for municipal water supply systems and large-scale pipeline networks, achieving a classification accuracy of 93.08%. It offers an efficient, adaptable, and robust solution for detecting leaks in complex operational environments, providing valuable insights for pipeline management and water loss reduction.