Acoustic emission detection technology is widely employed for leakage detection in water supply systems. However, this approach heavily relies on extensive field data to develop feature extraction and analysis models. Since field data cannot comprehensively cover all operational conditions—such as variations in pressure, pipe diameter, and leakage size—the limited generalization capability of these models often results in high rates of false negatives and false positives. To address these issues, this study utilizes Large Eddy Simulation (LES) to analyze leakage flow fields, establishing correlations between diverse operating conditions and flow field characteristics, including the areas of negative pressure zones and pressure pulsations. Based on these flow field findings, Computational Aeroacoustics (CAA) is applied to analyze the acoustic radiation field at leakage locations, thus clarifying the sound generation mechanisms of leakage-related acoustic signals, demonstrating strong agreement between simulation results and experimental data. Furthermore, wavelet packet energy ratio, centroid frequency, and frequency entropy are extracted as key feature parameters. A leakage detection model based on Support Vector Machine (SVM) is subsequently developed, achieving an accuracy of 98.6% across a wide range of operating conditions. This research enhances the capability for high-accuracy leakage detection with limited field data, offering valuable technical insights for the development of low-computation and low-hardware-cost leakage detection systems.