A machine learning-based approach for broadband sound recognition: A case study on investigating potential fire hazards of grinding machines

Abstract

Fire investigations often rely on visual evidence, which may be compromised by surveillance failures, occlusion, or image loss. Broadband acoustic signals, with strong penetration and temporal resolution, provide complementary clues about equipment conditions, explosions, and structural failures. Yet their non-stationary and overlapping characteristics hinder analysis using conventional methods. This study presents a machine learning-based broadband sound recognition approach, using grinding machines as representative fire hazards. Mel-frequency cepstral coefficients (MFCCs) and spectrogram gray-level co-occurrence matrix (GLCM) features were extracted to capture spectral and texture information, then fused for classification with XGBoost. Bayesian optimization was applied to adapt hyperparameters and improve robustness. The proposed model initially achieved 94.2 % accuracy and 91.5 % recall in multi-condition recognition using default hyperparameters. After applying Bayesian optimization to adapt hyperparameters and improve robustness, the model achieved 96.7 % accuracy and 93.3 % recall, outperforming support vector machines, random forests, and backpropagation neural networks. These results demonstrate the potential of broadband acoustic data to support fire investigations and provide a practical pathway for scene reconstruction and evidence enhancement.