| description abstract | Effectively detecting leaks is critical to improving leakage control management. Acoustic detection is one of the main leakage detection methods, and has been widely used in water utilities. Nevertheless, the effectiveness of this method is unsatisfactory in cases with various types of noise. To tackle this problem, this work proposes a leakage spectrogram to represent the features of leakage signals, and developed a time–frequency convolutional neural network (TFCNN) model to identify leakage signals. The performance of the TFCNN model was compared with other classification models (i.e., decision tree, support vector machine, multilayer perceptron, random forest, and extreme gradient boosting) under different signal-to-noise ratio (SNR) conditions. The results showed that the proposed method improves the accuracy and stability of leakage detection. Compared with other classification models, the TFCNN model had the best performance, and its mean accuracy reached 98% under different SNR conditions. Even in the case of −10 dB SNR, the mean detection accuracy reached 90%. In practice, the mean detection accuracy reached 99% for different time–frequency resolutions. The transfer learning–based TFCNN model is a promising method for leakage detection in cases in which there are insufficient data sets. | |
