Bidirectional Long Short-term Memory Network for Maglev Bridge Acceleration Data Reconstruction

Abstract

Maglev is a developing transportation mode in the recent years, and its safety and stability are required to be ensured by structural health monitoring system. However, the data measured by some sensors may be partially or totally unavailable due to the external electromagnetic interference or sensor failure. While, in some cases, those missing data are important to identify and analyze damage of maglev bridges. Since the structural dynamic characteristics of maglev bridge are complicated, the data reconstruction of those sensors becomes challenging. Therefore, this study proposes a bidirectional long short-term memory (BiLSTM) network to accurately reconstruct the maglev bridge acceleration data. The design of long short-term memory helps to sufficiently exploit the time series data and learn the hidden features of maglev bridge acceleration data between sensors. The bidirectional architecture enables the network to simultaneously learn the time series data from past and future, which is beneficial to extract more hidden features. A dataset collected from an in-site experiment for maglev bridge is used to verify the feasibility of the proposed method. The lost acceleration data from abnormal sensors is predicted by the acceleration data recorded from normally operating sensors. The results shows that the difference between predicted and true acceleration data is at a very low magnitude. Consequently, the proposed method can be applied for the high-performance reconstruction of maglev bridge acceleration data.