Enhancing the Structural Health Monitoring (SHM) through data reconstruction: Integrating 1D convolutional neural networks (1DCNN) with bidirectional long short-term memory networks (Bi-LSTM)

Abstract

Time series data plays an important role in structural health monitoring (SHM), but is often compromised by many factors including sensor failure, transmission errors, and adverse weather conditions. These issues render data incomplete, potentially leading to incorrect structural assessments. Although many studies have attempted to address data loss, reconstructing time series data for SHM remains challenging due to several factors: (1) Time series data may exhibit complex trends and fluctuations over time, making accurate reconstruction difficult; (2) Extensive data loss complicates understanding the underlying trends and relationships between data points; (3) The ìnluance of random or unpredictable factors often necessitates statistical models for replication. This research introduces a novel approach that combines a one-dimensional convolutional neural network (1DCNN) with a Bidirectional Long Short-Term Memory (Bi-LSTM) network to reconstruct missing sensor data in SHM. The proposed method leverages the strengths of two deep learning networks: the robust feature extraction capabilities of 1DCNN and the enhanced temporal processing power of Bi-LSTM, which analyses time series data from past and future contexts. The effectiveness of this hybrid model is validated through two distinct projects involving a continuous 3-span steel truss bridge and a cable-stayed bridge. Results demonstrate that combining 1DCNN and Bi-LSTM effectively reconstructs data and outperforms traditional models based on 1DCNN, LSTM, or Bi-LSTM alone, offering significantly improved accuracy.