PHYSICS-REGULARIZED DEEP LEARNING FOR INTER-FLOOR NOISE PREDICTION IN MULTI-FAMILY RESIDENTIAL BUILDINGS

-

 

Inter-floor noise in multi-family residential buildings represents a persistent structural-acoustic problem, particularly in densely populated urban environments. Excessive impact noise transmission not only contributes to occupant discomfort and health deterioration but also triggers legal disputes and regulatory challenges. Traditional acoustic prediction methods often rely on simplified empirical equations, while purely data-driven deep learning models lack physical interpretability and engineering reliability. Notably, prior studies have reported that 34–71% of predictions from conventional AI models violate fundamental acoustic laws. To address these shortcomings, this study proposes a physics-regularized multi-modal deep learning framework that integrates structural mechanics and acoustic principles into predictive modeling.

Structural-Acoustic Theoretical Foundations

The framework embeds governing physical principles of sound transmission in concrete floor systems directly into the learning process. Core mechanisms such as the Mass Law and the Coincidence Effect, derived from Kirchhoff–Love plate theory, are incorporated as differentiable constraints within the loss function. By enforcing these acoustic laws during optimization, the model ensures that predicted sound pressure levels remain consistent with structural vibration mechanics. This physics-informed strategy enhances interpretability and prevents non-physical prediction behavior commonly observed in black-box neural networks.

Multi-Modal Hybrid CNN–LSTM Architecture

A hybrid convolutional neural network (CNN) and long short-term memory (LSTM) architecture was developed to process multi-modal inputs. Structural parameters—including slab thickness, material properties, and span dimensions—are combined with acoustic spectrogram data representing frequency-domain characteristics of impact noise. The CNN component extracts spatial-frequency features from spectrograms, while the LSTM captures temporal dependencies and sequence behavior. This integration allows the model to simultaneously account for structural configuration and acoustic signal dynamics, improving predictive robustness across diverse building types.

Dataset and Cross-Validation Strategy

The proposed framework was validated using field measurements from 203 Korean multi-family residential buildings. To ensure generalization to unseen structural configurations, a Leave-One-Building-Out (LOBO) cross-validation scheme was adopted. This approach rigorously tests the model’s adaptability by training on all buildings except one and evaluating performance on the excluded structure. Such validation strengthens the reliability of results and demonstrates applicability across varying floor systems and construction conditions.

Prediction Accuracy and Physics Compliance

The physics-regularized model achieved prediction errors of 3.95 dB for light impact noise and 1.15 dB for heavy impact noise, representing approximately 10% improvement over baseline deep learning models. More importantly, physics violation rates were dramatically reduced from 71% to 1.8%, ensuring predictions adhere to structural mechanics and acoustic transmission principles. This substantial reduction confirms that embedding physical constraints effectively bridges the gap between predictive accuracy and engineering validity.

Regulatory Compliance and Engineering Implications

Beyond prediction accuracy, the model demonstrated 89.4% accuracy in forecasting compliance with the 58 dB Korean Building Code standard for inter-floor noise. This capability enables performance-based acoustic evaluation during early design stages, supporting proactive structural optimization and construction quality management. By uniting artificial intelligence with physically grounded constraints, the proposed framework establishes a new benchmark for reliable acoustic prediction in civil engineering, promoting safer, quieter, and regulation-compliant residential infrastructure.

🏗️ Civil Engineering Awards  

👉 Visit our Website: civilengineeringawards.com

#BuildingCode
#KirchhoffLoveTheory
#MassLaw
#CoincidenceEffect
#ConstructionQuality
#CivilEngineeringResearch
#SmartBuildings
#AIinConstruction
#ImpactNoise
#SustainableDesign


Comments

Post a Comment

Popular posts from this blog

Changes in selenium bioavailability in selenium

-
Image
INTRODUCTION This study explores how differing irrigation regimes and organic amendments shape selenium (Se) behaviour in naturally Se‑rich paddy soils. By comparing continuous flooding (CF) with alternating wet‑dry (AWD) cycles and evaluating cotton‑straw biochar (BC) versus sheep manure (SM) at two dosage levels, the work seeks to clarify why Se sometimes remains locked in soil and how it can be mobilised for healthier rice production. WATER‑MANAGEMENT STRATEGIES Switching from CF to AWD proved pivotal: periodic drainage not only elevated root‑surface iron‑plaque formation but also boosted rhizospheric affinity for Se. AWD further hastened the shift from weakly organic‑bound forms toward soluble and exchangeable fractions, creating a more plant‑available Se pool without relying solely on chemical inputs. ORGANIC AMENDMENTS AND RATES Amendment chemistry mattered. A modest 10 g kg⁻¹ SM dose maximised Se bioavailability—especially under AWD—while BC repeatedly suppressed it. Manure’s...
Read more

Extremely Large Telescope Threatened by Energy Project in Chile

-
Image
  The Atacama Desert in northern Chile is renowned for its exceptional conditions for astronomical observation. Its clear, dark skies have attracted multiple world-class observatories. However, a proposed green energy megaproject threatens to disrupt these conditions. This project, backed by AES Andes, aims to construct an extensive energy infrastructure for hydrogen and green ammonia production. Its proximity to the Paranal Observatory, home to the European Southern Observatory’s Extremely Large Telescope, poses risks to astronomical research. Importance of the Atacama Desert for Astronomy The Atacama Desert is considered one of the best locations on Earth for astronomical studies. Its high altitude and minimal light pollution provide ideal conditions. The Paranal Observatory is a key facility in this region, which has become a hub for nearly 40% of the world’s ground-based astronomy. The area is expected to increase this capacity to 60% within the next decade. The Extremely Large...
Read more

Failure mode dependent shear strength of unreinforced concrete brick masonry wall panels

-
  INTRODUCTION This section introduces the purpose and significance of the study, emphasizing the need to evaluate shear strength in unreinforced concrete brick masonry wall panels under diagonal compression. EXPERIMENTAL PROGRAM Details the methodology, including the variables tested—specifically, the bed-joint mortar mixing ratio—and outlines the process of fabricating and testing thirty masonry wall panel specimens. FAILURE MODES IDENTIFICATION Describes the five observed failure modes in the tested panels: diagonal tension, combined failure, bed-joint sliding, toe crushing, and non-diagonal failure, explaining the characteristics of each. SHEAR STRENGTH ANALYSI Presents a detailed discussion on how shear strength varied with each identified failure mode and highlights the dependency of shear performance on the mode of failure. COMPARATIVE EVALUATION WITH EXISTING CODES Compares the experimental results with existing masonry design code provisions, particularly focusing on shear...
Read more