Machine Learning Model for Detecting Leakage in Water Distribution Networks Through Road Surface Leakage Noise Analysis: Feature Extraction Using Fourier Transform and MFCC

Research on machine learning and neural network classification models for detecting leaks or anomalies in water distribution pipelines has been actively conducted. Detecting leaks using “leakage noise” transmitted to the road surface is traditionally one of the most commonly used methods by human inspectors. However, the characteristics of these road surface leakage noises vary significantly over time across various environmental conditions, posing a challenge in finding a model that maintains consistent classification performance despite these variations. This study focuses on the premise that data preprocessing methods have a greater impact on improving classification performance than model selection and hyperparameter tuning. To effectively extract features from highly variable leakage noise, Fourier transform and Mel-frequency cepstral coefficients were used. Considering the possibility of redundant information, a tree-based model, which is less sensitive to multicollinearity, was employed to evaluate the classification performance of the leakage noise. Through this approach, we aim to propose a data preprocessing method that provides stable classification performance despite the variability in leakage noise, thereby contributing to the development of robust machine learning models.