Large-Span Bridge Strain Reconstruction Based on Bidirectional LSTM and ESN

Abstract

Partly missing and anomalous of the data collected from structural health monitoring (SHM) systems are inevitable due to the failure of sensor and data acquisition equipment, which lead to misjudgment of the target structure state. The data integrity demands for guaranty using reconstruction algorithms before signal processing. Recurrent neural networks (RNN) has been proved effective of reconstruction issue by learning from the historical and future signal segments. The gated RNN represented by long short-term memory (LSTM) networks and reservoir computing represented by echo state networks (ESNs) show superiority on accuracy or training efficiency than standard RNN method. In addition, bidirectional concept can be introduced into these two methods to further improve their reconstruction precision. In this paper, models built by bidirectional LSTM and ESN are used to reconstruct strain data measured by the SHM system of Tsing Kau bridge, during which performances in both time and frequency domains are compared and evaluated. Furthermore, hyperparameters including number of layers, number of hidden units, scale of reservoir, and leaky rate have been optimized to improve the structures of the proposed models.