Water Leak Detection and Classification With Complex-Valued Neural Networks and Sensor Fusion

Water scarcity is an escalating global concern, with pipeline leaks and breaks accounting for up to 30% of urban water loss. While many recent studies have explored leak detection using intelligent systems and acoustic sensing, major challenges remain, particularly in achieving robust performance across sensor types, handling the complex nature of signal information, and extending detection to more informative leak classification tasks. In this study, we address these limitations by proposing a novel, application-driven method that integrates complex-valued Convolutional Neural Networks (CNNs) with a multi-sensor fusion strategy involving hydrophones and accelerometers. Unlike prior works that rely solely on signal magnitude, our method leverages both magnitude and phase information via the Short Time Fourier Transform (STFT), significantly enhancing feature expressiveness. The extracted features are fused and passed to a Multi Layer Perceptron (MLP) to perform leak detection or classification, the latter being a task rarely explored in the literature. We validate our method on two distinct real-world datasets, achieving a balanced accuracy of 99% in both detection and classification tasks. These results demonstrate our model’s strong generalization and competitive advantage over existing state-of-the-art approaches. Our contribution lies not only in the novel use of complex CNNs and sensor fusion, but also in advancing practical, data-driven leak classification, paving the way for more reliable and informative water infrastructure monitoring systems.