Experimental Artefacts for HARDIO: A Health-Aware Reliable Detection-Imputation Orchestrator for Operational Data Quality in Instrumented Water Infrastructure

Online sensor telemetry feeds the digital-twin, forecasting, control, and maintenance tools of instrumented water infrastructure. Probes accumulate drift, flatlines, volatility bursts, and outages, yet detection, imputation, and sensor degradation are usually evaluated separately and under inflated synthetic metrics. This work proposes HARDIO, a Health-Aware Reliable Detection-Imputation Orchestrator, an operational data-quality framework that pairs a duration-weighted sensor decay scorecard with a per-variable, chained detector-imputer benchmark. The assignment is selected by a four-pillar composite score (operational 35%, health coverage 30%, reconstruction 20%, calibration 15%) that scores the signal an operator actually receives after detection and imputation. HARDIO processes sensor installations independently, enabling parallel execution. HARDIO is validated on the Cheles wastewater treatment plant (WWTP), a real 1,177 population-equivalent (PE) plant, over 432 days, 19 channels, and eight sensor installations. No single detector-imputer pair is best for every channel: detection is distributed across SAITS, KNN Residual, LSTM-AE, and RuleBased, and chained imputation across LSTM-AE and KNN. Relative R² exceeds 0.90 on 12 of the 16 channels where it is interpretable. On the held-out final 20% of the record (12 channels in five installations), HARDIO reaches a mean chained 1-NRMSE of 0.976, close to the per-channel oracle (0.979), and significantly outperforms SAITS-only (p=0.003). Differences from LSTM-AE-only and KNN-Residual+KNN are not significant, and a single pair selected on validation data reaches 0.978, so HARDIO's benefit lies in traceable, health-aware orchestration rather than higher accuracy. The released artefact exposes the model map, confidence labels, margins, and production masks for decision support.