Study on Deformation Forecasting Method for Heritage Buildings Based on AR and Deep Learning Models

The current heritage structural health monitoring practices usually include the high-density monitoring and the low-density continuous monitoring. Aiming at the data at irregular intervals, a deformation forecasting method for heritage buildings based on autoregressive (AR) model and long short-term memory (LSTM) deep learning models is proposed. This new method is used for perceiving, recognizing and predicting the accurate deformation of heritage buildings in complex environment. This method is to use the deformation monitoring data of heritage building at various times as the initial time series. The new time series at regular intervals is generated by monitoring data augmentation based on Monte Carlo simulation method. The signal-to-noise ratio (SNR) of deformation time series are obtained by wavelet threshold denoising method. The method for the order estimation of AR model is improved based on SNR. Four models are studied, which are the proposed AR model and LSTM model with the initial time series at irregular intervals, the proposed AR model and LSTM model with the augmented time series at regular intervals. The prediction effects of four models are compared and analyzed by using the deformation monitoring data of a typical platform structure of Ming-Qing Dynasty. The analyzed results show that, for the time series with finite samples, the proposed AR model with the augmented data at regular intervals is more accuracy than the other three models. This method provides a new idea for deformation forecasting of heritage buildings based on monitoring data.