Missing data imputation model for dam health monitoring based on mode decomposition and deep learning

Abstract

Dam health monitoring is an important method for quantitative evaluation of dam safety. After long-term operation, there have missing data in dam monitoring data series inevitably due to the sensor damage or monitoring system failure problem which seriously affects the correctness of dam safety evaluation. The imputation accuracy of missing value is affected by data decomposition, reconstruction, and prediction methods. Therefore, in view of the high-precision imputation model of missing data in dam health monitoring, this paper proposes a data-driven fusion imputation model based on novel mode decomposition and deep learning method. First, the fusion imputation model is constructed based on extreme-point symmetric mode decomposition (ESMD), permutation entropy (PE), and bidirectional gate recurrent unit neural network (BiGRU). The ESMD-PE data preprocessing module can decompose the original data into a series of stable subsequences which can be input into the advanced deep learning BiGRU model to improve the interpolation accuracy. Then, the types of dam missing data and interpolation steps are studied. The engineering example illustrates that the root mean square error of the proposed model is decreased by 55.32% on average compared with four classical imputation models. The ESMD-PE–BiGRU fusion model can effectively simulate the inherent law of dam monitoring data and predict the missing data, which provides complete monitoring data for dam safety analysis.