Long short-term memory and Kalman filter with attention mechanism as approach for covariance shift problem in water leakage

Abstract

Urban water systems continue to face a major problem with water leakage, which results in substantial waste, shortages, damage to infrastructure, and monetary losses. While deep learning models have been effective in locating and identifying leaks, overfitting may result from their complexity over several training epochs. By including an attention mechanism, prominent features are given priority, improving model performance without compromising simplicity. Furthermore, layer normalization reduces problems in long short-term memory networks such as exploding gradients. Notable F1-scores are achieved by the proposed approach, demonstrating strong performance in both leak detection and localization tasks. Performance analysis under three different conditions for leak detection task such as source adaptation, target adaptation and adversarial simulation have shown an increase with scores of 91.59, 86.25 and 82.51 yielding 8.2%, 8.7% and 6.8% of improvement in F1-score, respectively. Similarly, performance analysis under three different conditions for leak localization task such as source adaptation, target adaptation and adversarial simulation has shown an increase with scores of 89.86, 84.39 and 80.77, yielding 7.4%, 8.5% and 8.6% of improvement in F1-score, respectively. Also, analysis using Wasserstein distance indicates reduced covariate shift through significant increase in accuracy (around 6.5%–9.5%, respectively), which is essential for adapting to varying water demand scenarios. The effectiveness of the proposed approach in urban water management is underscored by these results, emphasizing its potential for enhancing resource conservation and infrastructure sustainability.