Acceleration response reconstruction of a television tower via deep learning with multi-scale feature fusion

Abstract

Accurate reconstruction of structural response signals in structural health monitoring (SHM) remains challenging under complex operating conditions involving communication interruptions, sensor malfunctions, and environmental variability. To overcome these challenges, this study proposes a multi-scale temporal feature fusion method for multi-source SHM data in complex engineering environments. First, the proposed method employs a dilated temporal convolutional network (TCN) as a front-end feature extractor to capture multi-scale temporal dependencies in the historical environmental-variable sequences. The extracted representations are then fed into a bidirectional LSTM (BiLSTM), whose forward and backward branches encode the same historical feature window in chronological and reverse order, respectively. To further emphasize informative segments of the sequence, an attention mechanism is introduced to assign adaptive weights across time steps, enabling the method to emphasize temporal patterns most relevant to acceleration reconstruction. In addition, a time-domain and frequency-domain evaluation framework is employed to assess the consistency between the reconstructed and measured signals. Experiments are conducted using measured acceleration responses from a representative television tower, with three missing-data scenarios constructed to evaluate the effectiveness and robustness of the proposed method under real-world conditions. In the mixed missing-data scenario, the proposed method achieves a coefficient of determination R 2 of 0.9656 and a SpecError of 0.1146. Compared with four baseline models, including the gated recurrent unit (GRU), Transformer, graph neural network (GNN), and arithmetic optimization algorithm-based temporal convolutional network (AOA-TCN), the proposed method achieves superior overall reconstruction performance, with the most pronounced improvements under incomplete or missing measurements. The results demonstrate that the proposed method is effective for high-accuracy structural response reconstruction under complex missing-data scenarios in SHM.