Ayatelawej/Fault-Detection-and-Reliability-Monitering-for-a-Newly-Deployed-Weather-Station-Network-In-Libya

# Weather Station Network Fault Detection and Reliability Monitoring This repository contains the publishable analysis code and evidence for an hour-level fault-detection and reliability-monitoring study of a 26-station personal weather-station network in Libya. The published canonical dataset is frozen through June 2026. The live detection unit is a **station-hour**. The system builds reproducible episode labels, expands them into hour-level fault/not-fault targets, trains a gradient-boosted baseline and a Reliability-Aware Gated Fusion Network (RGFN), and evaluates retrospective event-level mechanism reason codes. Its reliability path adds full/partial outage classification, a causal 0-100 station-health score, transmitting-station health forecasts at 1/3/6/12/24 hours, and a combined 26-station operational scorecard. The final system also includes a read-only Streamlit dashboard that replays the independent July 2026 evaluation. The operational HGB model and all thresholds were selected before July was scored. On 13,565 eligible July station-hours it achieved 0.714 precision, 0.813 recall, 0.761 F1, 0.936 accuracy, 0.977 AUROC, and 0.865 AUPRC without refitting. The separate incident-risk experiments now use continuous clock-hour labels and horizon-purged timestamp partitions. They remain future-work evidence rather than deployable forecasting claims because their held-out precision, recall, and F1 do not jointly meet the project's acceptance criterion. ## Scope and reproducibility This is an **analysis-code-and-evidence repository**, not a clean-clone source-data reconstruction. It begins with the published canonical dataset: ```text data/merged/station_hourly_merged.csv ``` Source retrieval, five-minute normalisation, staging, and canonical-dataset upsert tooling are intentionally excluded. No credentials or source-acquisition capability are included. Public ERA5/Open-Meteo reference fetching is retained. Exact rebuilding of all detection features also …