Interpretable multitask deep learning framework for robust acoustic leak detection in water pipelines

Non-revenue water from pipeline leaks represents a critical challenge in urban water infrastructure, with detection systems constrained by noise sensitivity and operator trust barriers. We present an interpretable multitask deep learning framework that addresses these limitations through integrated signal enhancement and automatic feature learning from acoustic spectrograms. The multitask CNN architecture jointly optimizes denoising and classification objectives, while contrastive pre-training enables robust performance in data-scarce scenarios. Systematic evaluation on real pipeline data demonstrates substantial improvements over conventional approaches: 98.2% accuracy in optimal conditions and 95.5% accuracy at 0 dB SNR, representing 7.4 percentage points improvement over single-task baselines at matched noise levels. Ablation studies quantify individual component contributions, with signal enhancement providing +7.5% accuracy gain, multitask learning adding +0.7%, and contrastive pre-training yielding +4.2% in few-shot scenarios (10% training data). Critically, Grad-CAM and SHAP analyses reveal learned features align with established hydraulic acoustics—concentrated activation in 0-150 Hz bands corresponding to turbulent flow physics—addressing the black-box barrier to operational deployment. This work establishes a validated pathway toward trustworthy AI-driven infrastructure monitoring with potential for 10-20% non-revenue water reduction in urban systems.