<![CDATA[The quality of air temperature data from Automatic Weather Stations (AWS) is crucial for meteorological analysis, climatology, and early warning systems. However, flatline anomalies, a condition where sensor values tend to remain constant over a period of time, can degrade data quality and are often not optimally detected by conventional rule-based quality control (QC) methods. Previous research is also limited in specifically examining flatline detection, with most studies focusing on general anomalies and not integrating deep learning approaches with operational quality control systems. This study proposes a data-driven approach using a Long Short-Term Memory Autoencoder (LSTM-AE) combined with Level-1 QC. The novelty of this study lies in the use of a normal-only training scheme, anomaly threshold determination based on the reconstruction error distribution, and post-detection diagnosis to identify flatline characteristics. The methods include QC filtering, sliding window formation, model training, threshold determination, and anomaly detection. The results show stable model performance with an anomaly threshold value of 0.01177 (MSE). Of the 985,730 data windows, approximately 0.578% were detected as anomalies, indicating that flatline occurrences are relatively small but still significant to data quality. Most anomalies are short-lived and discontinuous, indicating localized sensor noise. This study demonstrates that LSTM-AE is effective as an adaptive flatline detection method and has the potential to be implemented as an automated QC module in AWS systems to improve data reliability.]]>
Copyrights © 2026
