AI-powered NUN-SEDFN framework for addressing sparse data challenges in geotechnical parameter prediction

Abstract

Accurate and comprehensive geological parameter acquisition is a persistent challenge in engineering geological mapping, particularly in construction environments with complex conditions where conventional drilling is impractical. Addressing the issue of sparse drilling data, this study introduces a novel prediction framework combining Non-Uniform Normalization (NUN) and a Spectral-Enhanced Deep Fusion Network (SEDFN). The proposed framework enhances the ability to predict geotechnical characteristic parameters critical for construction and infrastructure management. Specifically, the NUN-SEDFN framework transforms sparse textual drilling data into high-resolution geotechnical parameter maps by leveraging advanced AI techniques for data processing and prediction. The characterization stage employs NUN to ensure robust mapping between geotechnical data and image representations, addressing challenges in integrating large-span geological feature parameters. The prediction stage uses Frequency-Domain High Preservation Fast Fourier Convolution (FHP-FFC) and a Modified Super Resolution Convolutional Neural Network (mSRCNN) to learn and reconstruct high- and low-frequency geotechnical features, achieving over 80% prediction accuracy. This method enhances the reliability of geological mapping, offering significant potential for optimizing resource allocation, cost reduction, and safety in engineering and construction tasks. Furthermore, it demonstrates how AI can address data scarcity and improve decision-making in construction environments, aligning with current industry needs and technological trends.