A novel identification technology and real-time classification forecasting model based on hybrid machine learning methods in mixed weathered mudstone-sand-pebble formation

Abstract

Geological challenges in tunnel construction have consistently played a pivotal role in influencing project progress and safety. Accurate tunnel formation data holds the potential to assist in effectively managing various issues encountered during shield tunneling operations. This paper introduces a machine learning methodology for real-time tunnel geological prediction, based on the shield machine tunnelling parameters, as applied to the construction project of Chengdu Metro Line No.18 III phase. This method serves to enable timely formation identification and swift classification forecasting while tunneling progresses. Firstly, a new data pre-processing framework for real-time geological prediction is proposed. The 18 shield driving parameters in the daily report of the shield machine were selected as input features, which reduced the data by 2 orders of magnitude while retaining the geological characteristics of the data. Subsequently, leveraging the Dung Beetle Optimizer (DBO) and K-means algorithm, formation identification is carried out and validated against borehole data, enabling the acquisition of shield excavation data integrated with geological labels. Finally, 9 machine learning classification methods are used to classify and predict the data with geological label information, which proves that the tree-based classifier has strong interpretability for stratigraphic information and summarizes three boundary recognition modes of shield traversing different strata. The results show that: (1) the DBO+K-means algorithm has a lower clustering error rate and can successfully identify all 7 strata; (2) Considering the training time, RF is the optimal algorithm for this project due to its brief training time of only 3.808 s, coupled with high predictive performance. The research outcomes outlined in this paper offer a promising methodology for identifying stratigraphic boundaries during shield operation.