Continuous missing original data imputation method for bottom-sitting acoustic ocean wave meter

Abstract

The continuous absence of original data in bottom-sitting acoustic ocean wave observations significantly affects the accuracy of wave parameters. To address the characteristics of multivariate coupling, trend, and transient coexistence in the original data, this study proposes a time series multiple intelligent imputation model. The model employs multi-scale component decomposition based on singular spectrum analysis (SSA) to extract components, uses back propagation neural network fitting (BPNF) to impute the trend data, and integrates gated recurrent unit (GRU) with dynamic time warping (DTW) to impute the transient data. This approach enables the accurate extraction and imputation of two components: large-scale trends and small-scale transients in the original data. It resolves several issues inherent in traditional methods, including excessive smoothing of transient imputation, deviations in trend imputation, and diminished imputation performance as the extent of missing data increases. The proposed method demonstrates superior imputation performance for three types of continuous missing data: 61, 261, and 100 steps. The imputed data exhibits strong distribution consistency with the true data. Compared to the wave eigenvalues before imputation, the errors in wave height and wave period achieved by this method decrease by 91.7% and 93.3%, respectively, outperforming other traditional time series imputation methods.