Enhancing SHM data reconstruction with MA-CNN-BiLSTM: A deep learning approach for missing acceleration responses

Abstract

Complete, accurate, and continuous dynamic response data are essential for ensuring real-time structural safety in structural health monitoring (SHM) systems. However, sensor failures often result in missing dynamic response data, substantially compromising SHM performance. To address this challenge, this paper proposes a CNN-BiLSTM network (MA-CNN-BiLSTM) that integrates a multi-head attention mechanism to efficiently reconstruct long-term missing acceleration response data in SHM. This approach leverages the complementary strengths of CNN and BiLSTM in time series prediction to deeply analyze intricate spatiotemporal correlations among sensors. Furthermore, the multi-head attention mechanism accurately captures key patterns and detects abnormal fluctuations in acceleration responses. The proposed method was validated using a numerical study on a two-dimensional truss structure and experimental data from the Liujiaxia Bridge. Reconstruction results were benchmarked against those of a traditional LSTM network and a BiLSTM network (MA-BiLSTM) integrating a multi-head attention mechanism. Results show that the proposed method surpasses the other two networks, particularly in reconstructing high-frequency components of the response data. This leads to substantial improvements in reconstruction accuracy across both time and frequency domains. The reconstructed acceleration responses were subsequently applied in modal analysis, effectively identifying the structure's natural frequencies and vibration modes. The study further investigated the impact of sensor quantity and layout on reconstruction performance, providing valuable insights for practical applications.