To address the problems of traditional acoustic methods relying on personnel experience and machine learning methods requiring large amounts of training data with poor interpretability, this paper proposes a bidirectional constant false alarm rate (Bi-CFAR) detection method based on time–frequency images. The method first performs short-time Fourier transform (STFT) on pipeline noise signals to obtain a spectrogram, then processes the spectrogram in blocks to enhance statistical characteristics, and finally employs the variability index (VI)-CFAR algorithm to detect along both time and frequency axes and fuses the detection results. By analyzing the characteristics that leakage noise is continuous along the time axis but discontinuous along the frequency axis, corresponding decision rules are designed. Based on the multiparameter characteristics of the algorithm, multiparameter ablation tests were conducted to determine the optimal parameters corresponding to the highest detection accuracy. Simulation experiments showed that the proposed method achieves 97.67% accuracy and 99.85% recall under ideal conditions. Real data validation results demonstrated that the method maintains 89.27% accuracy and 90.07% recall in complex environments, showing significant advantages over the bandwidth method and matched subspace (MS). Compared with machine learning methods, although the proposed method has slightly reduced detection accuracy, it requires no large training data sets and has stable parameters with interpretability. Furthermore, its algorithmic complexity meets the requirements for deployment on edge devices, providing a practical solution for urban water supply pipeline leak detection.
Zhang et al. (Tue,) studied this question.