Existing tunneling water leakage detection techniques primarily focus on traditional machine learning models, such as decision trees, SVMs,
Existing tunneling water leakage detection techniques primarily focus on traditional machine learning models, such as decision trees, SVMs, and random forests. These models struggle to handle large-scale and complex sensor data. More recent efforts using deep learning models like CNNs and RNNs also face challenges in capturing the diverse relationships among features. This paper introduces HyperClustering, a new framework designed to enhance subway shield tunneling water leakage detection through multimodal deep/shallow feature fusion techniques. The model addresses challenges such as variability in leakage patterns across different environmental conditions and the complexity of sensor data. It combines deep and shallow learning methods to identify high-quality features that capture the unique characteristics of tunneling operations and leakage events. HyperClustering improves upon these methods by constructing a large-scale hypergraph that models complex relationships among leakage features. This hypergraph uncovers high-order interactions that traditional approaches cannot capture. The model also uses a Gaussian Mixture Model (GMM)-based posterior probability to rank candidate patches for water leakage detection. It prioritizes high-risk areas and minimizes false positives. Experiments on datasets from multiple subway tunneling projects show that the model effectively identifies complex feature interactions. It offers adaptive, data-driven optimization for a responsive leakage detection system. This innovative approach sets HyperClustering apart from existing research by leveraging hypergraph modeling and advanced feature fusion. It provides a more robust and accurate solution for real-world tunneling conditions and paves the way for future advancements in multimodal fusion techniques and real-time adaptive learning in leakage detection systems.
