Detection of abnormal sounds for machine status monitoring

Early detection of faults in machines operating in industrial environments is of critical importance for ensuring operational safety, maintenance planning, and cost optimization. In this context, acoustic signals provide valuable information regarding the operational status of machines and have emerged as a promising source for fault detection. Existing anomalous sound detection (ASD) systems offer a non-contact, cost-effective, and highly applicable solution by utilizing microphones integrated into machines to detect faults via audio signals. However, the performance of these systems is significantly reduced under real-world conditions due to high levels of ambient noise and complex background sounds. This limitation becomes more pronounced when coupled with the noise vulnerability of conventional machine learning-based methods, undermining the consistency and robustness of ASD systems in practical settings. The primary aim of this thesis is to develop a noise-resilient and robust machine learning model for anomaly detection in machine sounds. In this study, a denoising autoencoder (DAE) architecture was adopted to enhance the model's robustness against noise. For this purpose, a baseline autoencoder (AE) model trained with Mel-spectrogram-based features was proposed. The model's performance was evaluated using reconstruction errors along with AUC and pAUC metrics. Regularization techniques were applied to improve generalization capabilities; however, these enhancements provided only limited benefits. Colored noise (white, pink, and brown) represents the primary types of noise commonly encountered in machine environments. Based on the hypothesis that machine sounds may exhibit different noise characteristics, two novel noise injection strategies—hybrid and adaptive—are proposed to integrate colored noise types into the DAE architecture. This approach relies on calculating the power spectral density (PSD) slope for each audio sample and modeling the spectral properties of noise types accordingly. The experiments were conducted on various machine sound recordings from benchmark ASD datasets [13]. Performance comparisons based on AUC and pAUC metrics demonstrate that the integration of hybrid and adaptive noise strategies into the DAE architecture significantly enhances the overall model performance. The proposed DAE-based approach yielded an AUC improvement of 0.9–1.6% in the source domain and a generalization improvement of 1.2–2.3% in the target domain. Specifically, brown noise performed best at moderate-to-high noise levels, while PSD-based adaptive noise demonstrated superior generalization performance under high noise conditions. This thesis focuses on applying denoising autoencoders to machine sound anomaly detection and advances the field with contributions based on different noise types and spectral analyses.