Distributed measurement of large-scale geotechnical deformation based on a deep learning hybrid model

Abstract To address the challenges of high interference, strong nonlinearity, and poor spatial continuity in large-deformation monitoring of geotechnical structures, this study proposes a hybrid prediction framework that integrates time domain reflectometry (TDR) with deep learning. A CNN-BiLSTM-AdaBoost model is developed, where dynamic time warping (DTW) is used to extract temporal correlation features from TDR waveforms. Spatial features are captured through convolutional layers, temporal dependencies are modeled using BiLSTM, and model generalization is enhanced via AdaBoost ensemble learning. The model achieves mean absolute errors of 2.10 mm for overall deformation ( Y 1 ) and 2.41 mm for deformation localization ( Y 2 ), representing over 40% improvement compared to traditional BP neural networks and Transformer-based models. Cross-validation confirms strong robustness, with R 2 values of 0.957 and 0.937, respectively. Landslide simulation tests under various interference conditions—including noise injection, rainfall, and signal loss—demonstrate the model’s excellent anti-interference performance. Field deployment at a tailings dam in Tibet further validates its practical utility, achieving real-time monitoring with a delay of less than 15 s and cumulative prediction error within 3.5 mm—well below the 250 mm safety threshold. This work offers a high-accuracy and robust monitoring solution for complex geotechnical environments, providing effective technical support for the life-cycle safety management of high-risk slopes and dam infrastructures.