An Integrated Natural Language Processing Framework for Automated Seepage Analysis in Construction Engineering: A Deep Learning Approach for Document Processing and Predictive Modeling

Abstract

Seepage analysis is vital in construction engineering yet traditional methods rely heavily on manual extraction of seepage parameters from unstructured geotechnical reports which is time-consuming, error-prone and limits data availability for accurate predictions. Existing approaches often lack automation and integration between document processing and advanced seepage modeling, hindering efficiency and scalability. Our research addresses these gaps by developing a novel framework that combines a hybrid CNN-LSTM-attention-based NLP model to automatically extract seepage-related data from construction documents with a physics-informed neural network for seepage pressure prediction. Validated on the global SoilKsatDB dataset, our system achieved 96.2% extraction accuracy and reduced prediction error by 23.5% compared to traditional methods while processing data 15 times faster than manual techniques. This integrated approach significantly improves both the accuracy and efficiency of seepage analysis, contributing a scalable, intelligent solution that enhances safety and decision-making in critical infrastructure projects.