Dike conditions monitoring is a challenging task. Algorithms for dike anomaly detection are one of the key components of a dike condition monitoring system. Algorithms for anomaly detection have to detect anomalies in dike behaviour (abnormal behaviour) in an on-line mode based on measurements collected from sensors installed in the dike. A machine-learning-based algorithm presented in this paper is trained on historical data on the normal dike state because data for abnormal dike behaviour is not available and simulation is time-consuming. Detection of abnormal dike behaviour is done by applying a ‘neural clouds’ one-class classification method. The ‘neural clouds’ one-class classifier is used for estimating the nonlinear fuzzy membership function of normal behavior for features from wavelet decomposition. The application of a wavelet transform can detect abnormal dike behaviour hidden in the time-frequency signal properties. Algorithms were tested on real data of a dike located in Boston, United Kingdom.