Abstract
Sensors are widely used in inland waterway buoys to monitor their position, but the collected data are often affected by noise, outliers, and irregular sampling intervals. To address these challenges, a standardized data processing framework is proposed. Outliers are identified using a hybrid approach combining interquartile range filtering and Isolation Forest algorithm. Interpolation methods are adaptively selected based on time intervals. For short-term gaps, cubic spline interpolation is applied, otherwise, a method that combines dominant periodicity estimation with physical constraints based on power spectral density (PSD) is proposed. An adaptive unscented Kalman filter (AUKF), integrated with the Singer motion model, are applied for denoising, dynamically adjusting to local noise statistics and capturing acceleration dynamics. Afterwards, a set of time-frequency features are extracted, including centrality, directional dispersion, and wavelet transform-based features. Taking the lower Yangtze River as a case study, representative buoys are selected based on dynamic time warping similarity. The features analysis result show that the movement of buoys is closely related to the dynamics dominated by the semi-diurnal tide, and is also affected by runoff and accidents. The method improves the quality and interpretability of buoy motion data, facilitating more robust monitoring and hydrodynamic analysis.