Deep learning for leakage detection in water distribution networks: a systematic review and benchmark

ABSTRACT

Leakage detection in water distribution networks is a critical challenge for water utilities worldwide, with deep learning (DL) emerging as a promising solution. However, the literature lacks standardized benchmarking frameworks, making it difficult to compare claims across studies and assess real-world deployability. This study presents a hybrid systematic review + experimental benchmarking approach to address this gap. A preferred reporting items for systematic reviews and meta-analyses (PRISMA)–guided systematic review of 244 records (2015–2026) from 10 engineering databases identified 71 pure DL studies. The review revealed that convolutional neural network (CNN)-based architectures were most common (35.2%), followed by autoencoders (16.9%) and long short-term memory (LSTM)/recurrent neural network architectures (12.7%). Critical gaps included reliance on simulated data (75%), lack of code or data sharing (59%), and limited reporting of computational metrics, with only 11.3% reporting inference time. The time-transformer (M5) achieved the highest detection accuracy (96.1%) and (AUC) Area Under the Curve-ROC (0.94), but required substantial computational resources. The (GGNN) Gated Graph Neural Network (M4) achieved the highest leak localization success (88.5% within 300 meters). The One-Dimensional Convolutional Neural Network (1D-CNN) deep autoencoder (M2) offered the best practical trade-off: fast inference (3.5 ms), strong noise robustness (9.9% drop at 20 dB signal-to-noise ratio), unsupervised training, and 92.6% accuracy. The lightweight 1D-CNN (M3) was most efficient (12 MB, 1.2 hours) but degraded by 14.5% under noise. The LSTM-autoencoder (M1) achieved 91.3% accuracy and 86.1% recall.