Lightweight U-Net Autoencoders with Semantic Compression for Unsupervised Fault Detection in Water Distribution Systems

Reliable fault detection in water distribution systems (WDS) is critical to maintaining service continuity and mitigating costly infrastructure failures. However, existing solutions often rely on supervised learning and dense sensor deployments, which are impractical in real-world scenarios with scarce labeled data and limited edge resources. To address these challenges, this paper proposes a lightweight U-Net autoencoder framework enhanced with semantic compression for unsupervised fault detection. Semantic compression transforms raw sensor streams into compact statistical and spectral descriptors, thereby reducing computational complexity and communication overhead. The autoencoder leverages separable convolutions in depth and skip connections to efficiently reconstruct normal operating patterns, while faults are identified by thresholding reconstruction errors with data-driven calibration strategies. Experiments on real-world WDS datasets with controlled fault injection (including drift, stuck-at, spike, and noise anomalies) demonstrate that the proposed method outperforms conventional autoencoders and baseline anomaly detectors. Specifically, it achieves higher event-level F1 Scores and recall rates while reducing model size and inference latency, enabling deployment on edge devices. This work presents a novel unsupervised fault-detection pipeline that balances accuracy and efficiency, providing a generalizable paradigm for anomaly detection in resource-constrained Internet of Things (IoT) environments.