Signal analysis and anomaly detection for flood early warning systems

Abstract

We describe the detection methods and the results of anomalous conditions in dikes (earthen dams/levees) based on a simultaneous processing of several data streams originating from sensors installed in these dikes. Applied methods are especially valuable in cases where lack of information or computational resources prohibit computing the state of the dike with finite element and other mathematical models. The data-driven methods are part of the artificial intelligence (AI) component of the 'Urbanflood' early warning system. This AI component includes pre-processing (e.g., gap filling and measurements synchronization procedures) of data streams, feature extraction and anomaly detection by one-side (also known as one-class) classification methods. Our approach has been successfully validated during a non-destructive piping experiment at the Zeeland dike (The Netherlands).