Prediction of Construction-Induced Ground Vibrations Using Field Measurements and Bidirectional Gated Recurrent Unit Neural Network

This paper proposes a sequential bidirectional gated recurrent unit (BGRU) model to predict construction-induced ground vibrations. The ground vibration time histories for twelve real construction projects in Toronto, Canada, are collected and used to develop the BGRU model. A single time-step method is used to predict the vibrations, and the time window is swept continuously over the whole training data. In addition to the BGRU method, and for comparison, two other methods, autoregressive integrated moving average (ARIMA) and random forest (RF), are used to predict the ground vibrations. The results show that the BGRU method performs much better than ARIMA and RF methods in forecasting construction-induced ground vibrations. The BGRU method captures the construction-induced and background vibrations very well, and this method remains accurate when the training data includes both background and construction vibrations. Therefore, this method can be used to predict ground vibrations in real projects where there is always a potential for missing some parts of the ground vibration data due to the malfunction of the vibration recording units.