Boundary-Aware Time-Frequency Sparse Coding for Low-Overhead IoT Time Series Analytics

Distributed IoT water quality monitoring systems are evolving toward greater density, broader coverage, and enhanced real-time performance. However, severe constraints on edge-node bandwidth, power consumption, and computational capability make high-accuracy prediction and reliable anomaly detection extremely challenging under intense noise and non-stationary conditions. Water quality time series are characterized by abrupt changes, complex time–frequency patterns, and dominant noise components. Conventional deep learning models applied to such data tend to over-smooth signals, exhibit limited sensitivity to transient features, and incur excessive communication overhead.To address these limitations, a boundary-aware time–frequency sparse learning codec termed BASM is proposed for resource-constrained environments. Preservation of abrupt boundaries and localized anomaly structures is enhanced through boundary-weighted wavelet sparse reconstruction, while prediction robustness in low-SNR regimes is achieved via sparsity-guided ESN. Additionally, a lightweight adaptive anomaly-triggered reporting mechanism is integrated, enabling substantial reduction in uplink traffic without compromising the integrity of anomaly information. Extensive evaluation on real-world datasets from the Yangtze River basin demonstrates that BASM maintains superior stability and accuracy in low-SNR conditions while significantly lowering both communication and computational costs. The framework establishes a new communication–computation co-design paradigm for efficient edge intelligence in IoT-based water quality monitoring.