Physics-regularized deep learning framework for floor impact sound prediction in residential buildings: A civil engineering approach

초록

Inter-floor noise in multi-family residential buildings constitutes a critical structural-acoustic challenge, causing health deterioration and legal disputes in densely populated urban areas. While deep learning offers promising predictive capabilities, purely data-driven models lack the physical coherence required for reliable civil engineering applications—with 34–71% of predictions violating fundamental acoustic laws. This study develops a physics-regularized multi-modal deep learning framework integrating structural parameters and acoustic spectrograms through a hybrid CNN-LSTM architecture. Fundamental principles governing sound transmission in concrete structures, including the Mass Law and Coincidence Effect derived from Kirchhoff-Love plate theory, are embedded as differentiable constraints within the loss function. The framework leverages field measurements from 203 Korean residential buildings with Leave-One-Building-Out cross-validation ensuring generalization to unseen building configurations. The proposed model achieves prediction errors of 3.95 dB for light impact and 1.15 dB for heavy impact noise, representing 10% improvement over baseline models. Physics violation rates are reduced from 71% to 1.8%, ensuring predictions respect structural mechanics principles. The framework demonstrates 89.4% accuracy in predicting regulatory compliance with the 58 dB Korean Building Code standard, enabling reliable performance-based acoustic evaluation during design phases. This work establishes a new benchmark for physically coherent acoustic prediction in building structures, bridging the gap between data-driven artificial intelligence and the engineering reliability required for construction quality management and proactive structural design optimization. © 2026 The Authors.

키워드