Predicting spatial sound attenuation in buildings with machine learning: Implications for fire alarm placement

Audible and intelligible fire alarms play a critical role in ensuring occupant safety in emergencies. Although various experimental and theoretical methods exist to measure and predict alarm sound levels, the variability and limitations of these methods, especially in complex layouts, remain underexplored. Building on both established fire engineering research and advanced alarm management concepts from process safety, this study evaluates the effectiveness of fire alarm placement within residential units by comparing in situ measurements, calculation-based estimates, and machine learning predictions. The findings show that open doors result in higher sound levels than closed doors, and corridor-based alarms typically fail to meet the recommended 75 dBA threshold needed to awaken sleeping occupants. Moreover, established calculation methods show an average error rate of about 9%, especially in geometrically complex or acoustically variable settings. By contrast, the machine learning model achieves a notably lower error rate at around 2% underscoring its potential to integrate uncertainty factors such as distance, partitions, and acoustic attenuation more effectively than traditional formulas. From a risk management perspective, these results highlight the value of data-driven, risk-based alarm design, aligning with hybrid alarm modeling approaches seen in process industries. The study concludes that installing fire alarms within each dwelling unit, coupled with interconnected sounders in sleeping areas, significantly enhances occupant alertness and system reliability in residential buildings.