Predicting the Detectability of Safety Signals in Construction Sounds Using Machine Learning
TL;DR
Whether machine learning can predict warning signal audibility in construction noise is investigated and convolutional neural network models trained on mel-spectrograms will be presented alongside these results as a deep learning comparison.
Abstract
Construction sites produce dynamic, multi-source noise that can mask safety-critical warning signals, contributing to struck-by fatalities. While psychoacoustic models such as Detectsound address masking prediction, they inadequately handle fluctuating noise characteristic of active construction, a gap identified by recent literature. This study investigates whether machine learning can predict warning signal audibility in construction noise. Eighty-one synthetic ISO 7731-compliant warning signals (tonal, multi-tone, and broadband) were mixed with real recorded construction noise at 10 signal-to-noise ratios (−10 to +30 dB), producing 56,700 audio mixtures labeled audible or masked via a two-tier scheme implementing ISO 7731 and ISO 9533 detectability criteria. Thirty-eight acoustic features including spectral, energy, MFCC, and frequency band measures were extracted per mixture. Four tree-based classifiers were evaluated. A Random Forest achieved 98.8% accuracy (F1 = 0.991), confirmed by 10-fold cross-validation (0.991 ± 0.002). The false positive rate, predicting "audible" when a signal is actually masked and the safety-critical error, was 1.19%. RMS energy variability and mid-frequency band power were the most discriminative features, consistent with ISO 9533 evaluation criteria. Convolutional neural network models trained on mel-spectrograms will be presented alongside these results as a deep learning comparison.