Acoustic water pipe leak detection using transformer-based variational autoencoder

Accurate leak detection in water pipes is essential for maintaining infrastructure integrity. However, detecting leaks under noisy real-world conditions remains challenging due to the significant interference caused by dynamic environmental sounds and background noise. This study aims to develop an acoustic approach for detecting water pipe leaks using a free-swimming non-destructive inspection tool. To address the challenge of environmental noise, we explore the feasibility of a transformer-based variational autoencoder trained only on normal signals. First, our approach employs a Swin-Transformer (ST) model combined with a Convolutional Neural Network (CNN) to extract both global and local acoustic features. The extracted features are then projected into a latent space to improve robustness against noise. Finally, reconstruction scores are computed to differentiate leaks. During this process, a kernel density estimation (KDE)-based strategy is used to determine a suitable decision threshold automatically. Experimental results demonstrate that CST-VAE outperforms state-of-the-art methods, achieving a 96.0% accuracy and a 96.2% F1-score. It improves to 3.5% in accuracy and 3.2% in F1-score when compared to the best state-of-the-art performance. The performance indicates its practical applicability and significant potential for automated leak detection in pipeline assessment systems. • This is the first automated approach for detecting water pipe leaks using a free-swimming, non-destructive inspection tool. • A transformer-based autoencoder is proposed to extract both local and global features without requiring labeled training data. • A kernel density estimation-based strategy is employed to automatically determine the leak detection threshold. • The proposed model demonstrates effective leak detection performance on real-world water pipe inspection datasets.