A physics-based dimension reduction and modeling method for monitoring data and its application to tunnel engineering

Abstract

The promotion of intelligent control of engineering equipment requires extracting vital information from numerous original features in the monitoring data. Data-driven methods solely rely on data distribution to extract dimensional information, facing difficulties in comprehensively understanding and scrutinizing system behavior. To satisfy the control and optimization requirements of intelligent systems, this paper introduces a physics-based dimension reduction and modeling method. According to scientific principles, original features are transformed into causal variables with physical meaning. Then important causal variables are selected through regularization in machine learning. Finally, a concise prediction model with physical representation is established. An application to tunnel engineering demonstrates that the generated low-dimensional space reveals the essential laws between the control parameter and its main influencing factors (i.e., the geological conditions, operating status, and structural characteristics of the tunneling equipment), which can further provide a better prediction model balancing accuracy and dimensionality.