| Availability-Aware Physics-Data Leak and Sensor-Fault Diagnosis for Digital-Twin Decision Support in L-Town Water Distribution Network |
| کد مقاله : 1375-IWWA (R1) |
| نویسندگان |
|
نورالهدی زرنگاریان1، پویان بیات *2 1شرکت مهندسی آب و فاضلاب استان هرمزگان (مادر تخصصی) دانشگاه صنعتی امیرکبیر 2دانشگاه صنعتی امیرکبیر |
| چکیده مقاله |
| Sensor outages can make hydraulic-model discrepancies difficult to attribute to leaks or instrumentation faults. We develop and evaluate a three-class diagnostic framework in simulated L-Town using normalized pressure and flow residuals, sensor-identity-invariant summaries, availability descriptors, and dropout-augmented learning. Development used 600 operating-condition groups; a frozen pipeline was then evaluated on 300 new groups, split equally between familiar and selected unseen targets. Comparators included a residual-threshold rule, an adapted temporal convolutional network, and tabular ablations. At 40% nominal sensor dropout, the proposed ensemble achieved macro-F1 of 0.926 and 0.858 in the familiar and unseen regimes, compared with 0.807 and 0.612 for the closest baselines. Paired improvements were 0.119 (95% group-bootstrap confidence interval: 0.085-0.154) and 0.247 (0.212-0.284), with resolution-limited Holm-adjusted p = 0.0032. Unseen sensor-fault recall remained high, but leak recall for p673 was 0.38. Post hoc analysis associated this failure with a concentrated residual signature resembling a sensor fault. The results demonstrate improved robustness under incomplete sensing in the tested L-Town setting and support dropout-augmented residual learning as an offline diagnostic component for digital-twin decision support. |
| کلیدواژه ها |
| digital-twin decision support, water distribution network, leak diagnosis, sensor fault, missing data, machine learning |
| وضعیت: پذیرفته شده برای ارسال فایل های ارائه پوستر |
