A Non-annotation Learning for Near-real time Anomaly Detection of Sea Level using Recurrent Neural Network Ensemble-based Model

초록

Real-time sea level data observed from the tidal gauge include not only missing values but also erroneous measurements. The latter classified as abnormal values may be determined by the quality control procedure, which typically applies the 3 (three standard deviation) rule to remove them from the data. However, it is difficult to apply it to sea level records which include extreme values caused by such as unusual weather events, and/or erroneous values existing within the 3 range. To develop a flexible quality control method that overcome these issues, we construct the artificial intelligence model set consists of non-annotated recurrent neural networks and ensemble techniques that do not require pre-labeling of abnormal values. The developed model can diagnose abnormal sea level values recorded within 20 minutes. The verified model well detaches abnormal from the normal values during both normal times and abnormal weather events. In addition, it has been confirmed that abnormal values can be detected even during periods when sea level data is not used for training. Application of the artificial neural network algorithm developed in this study is not limited to coastal sea level records, and hence it can be extended to a detection model of flawed values from various oceanic and atmospheric data.