Multi-band spectral-temporal deep learning network for structural dynamic data forecasting aided by signal decomposition and bandpass filtering

Abstract

This study proposes a novel online data forecasting strategy for dynamic measurements in structural health monitoring applications, which excels in both time and frequency domain, based on a spectral-temporal deep learning network integrated with signal decomposition and filtering techniques in moving windows. For complex structural dynamic measurements from bridges or buildings, the varying time–frequency features within moving windows are difficult to capture for conventional deep learning networks, e.g., CNN or GNN, resulting in spectral bias or other related issues. To address this challenge, in this study, signal decomposition is adopted to simplify the frequency domain features, i.e., variational mode decomposition (VMD), into multiple intrinsic mode functions (IMFs), each of which has limited bandwidth frequency and regular time domain pattern. The decomposed signals are then fed into a spectral-temporal deep learning network for data forecasting step by step, and the predicted datasets are further finetuned by associated bandpass filters, respectively, to selectively remove the unwanted frequency components to enhance the accuracy. The performance of proposed method is validated on both numerical simulation data and field test data, and the results demonstrate that the predicted datasets match well with ground truth values in both time and frequency domains. Compared to the state-of-the-art solution, the proposed method can achieve much higher forecasting accuracy. In addition, the effects of hyperparameters and VMD are investigated, and the result shows that the introduction of signal decomposition plays a critical role to improve the state-of-the-art solution for data forecasting. The proposed solution can efficiently help to reconstruct dynamic measurements from data anomalies (e.g., missing, bias, and drifts) or to detect structural anomalies (e.g., rare events and damages) in long-term structural health monitoring systems.