Learning the hard-to-learn: Active learning for imbalanced datasets in data-centric tunnel engineering

Abstract

Data-driven methodologies hold the potential to revolutionize mechanized tunneling by informing decision-making. However, imbalanced data categories are prevalent within the realm of mechanized tunneling. Data-driven models often struggle to accurately predict data from minor sample categories, which, despite their scarcity, are crucial in practice. To address this bottleneck, this study introduces a committee-based active learning strategy to tackle the imbalanced sample identification problem. This strategy involves constructing multiple committee models to quantify the uncertainty in surrogate predictions. These specimens exhibiting high forecast uncertainty will be resampled with replacement from the validation data to iteratively augment the training data, thereby enhancing the model’s capability to handle imbalanced datasets. This study validates the strategy by applying it to two real tunneling engineering scenarios. The first case is to predict the grade of surrounding rock and the other one involves forecasting muck clogging in mechanised tunneling. The results demonstrate that the proposed strategy significantly improves the prediction accuracy of minority categories. Further, the study reveals that both the “entropy method” and “weight method” can effectively quantify the learning difficulty of samples. The same strategy can be equally applied to other fields of engineering and science.