An integrated AR-CNN-LSTM method for missing data repair in rockfill dam deformation

Abstract

Deformation is a critical indicator of dam structure integrity. Rockfill dams are exposed to harsh operating conditions and diverse environmental factors, making deformation monitoring instruments prone to failure from mechanical damage and environmental impacts. These failures create significant data gaps, undermining the accuracy of monitoring models and compromising structural safety assessments. However, current models for repairing missing deformation data often overlook the underlying physical mechanisms, struggle to capture complex spatiotemporal correlations and suffer from poor training stability, limiting their ability for high-precision data repair. Therefore, this article proposes an integrated autoregressive-convolutional neural network-long short-term memory method for repairing missing deformation data. It addresses the instability of traditional unidimensional approaches in high-missing-rate scenarios by enabling collaborative learning of spatiotemporal deformation patterns and complex nonlinear features. The approach combines a CNN to extract spatial features, a LSTM network to capture temporal dynamics, and an AR integrated moving average residual module to enhance the method’s response to sudden events. This spatiotemporal framework effectively addresses data repair challenges in both discrete and continuous missing data scenarios. An engineering case study of the PB rockfill dam demonstrates that the proposed method achieves high-precision repair of complex missing data patterns, reducing the mean error by 57.3–81.7% and the mean absolute percentage error to below 23%. This approach provides a novel solution for missing data repair in dam deformation monitoring, with significant theoretical and practical implications for ensuring the accuracy and reliability for long-term structural safety assessments.