Improved bidirectional echo state network-based time series reconstruction and prediction for structural response

The integrity of the data collected by structural health monitoring systems has a significant impact on structural damage detection and state assessment. The missing or abnormal segments and unacquired future segments can be supplemented through signal reconstruction and prediction models. This paper proposes two novel models toward these two tasks based on bidirectional echo state network, which can exploit both historical and future signal segments to improve accuracies. Adaptive combination coefficient is introduced to control the rate of error accumulation. The effectiveness and robustness of the proposed models are verified by cases of synchronized missing, long-term missing, and boundary effect. A hyperparameter study related to both reservoir and memory is conducted to generate optimal models with maximum processing abilities. An ARIMAX and improved Kalman filter-based preprocessing method is adopted to keep all useful information and provide optimal estimation of the true signal values. The proposed models also show high performance in generating the high-frequency components. The superiority of the proposed models is validated through the datasets measured from Canton Tower, both stationary signals under free vibration and non-stationary signals under earthquake being considered.