关键词: 盾构隧道, 机器学习, 地层识别, 时间序列聚类, 在线学习

Abstract: Current machine learning models for recognizing geological conditions during shield tunneling heavily rely on precise geological data labelling, limiting their applicability in complex geological environments. To address this, we propose a continuous dynamic time warping (CDTW)-based agglomerative hierarchical clustering model (CDTW-Agglomerative), which integrates a linear interpolation framework to overcome DTW’s discretization issues. An online learning mechanism is implemented for dynamic strata recognition. The model’s accuracy and reliability are validated using Xiamen Metro Line 3 data, with generalization tested on Line 6 data. Results show recognition accuracies of 85% and 73% on the two datasets, demonstrating robust generalization. CDTW-Agglomerative outperforms DTW-Agglomerative, SoftDTW-Agglomerative, and CDTW-based models (K-means, K-medoids, Spectral clustering). Notably, it identifies cutterhead stratigraphy without requiring pre-labelled geological data, supporting intelligent decision-making for tunnelling parameters.

Key words: shield tunnel, machine learning, geological condition prediction, time series clustering, online learning