Machine learning for industrial acoustic component inspection

This thesis investigates acoustic anomaly detection for industrial quality control, with a primary focus on achieving robust OK (qualified) and NG (not good) classifications within practical manufacturing settings. The research centers on the acoustic characterization of small mechanical and electromechanical components inside specialized test chambers. It directly addresses critical industrial challenges, such as inter-machine variability, non-stationary operating conditions, ambient factory noise, and the prevalence of inconsistent or noisy data labels. To mitigate these issues, a unified data-driven framework is proposed, integrating structured dataset organization, signal preprocessing, time–frequency analysis, and temporal sequence modeling. While several classical machine-learning algorithms are evaluated as baselines, this work introduces a Transformer-based architecture designed to explicitly capture the long-range temporal dependencies inherent in acoustic emission signals. Experimental results demonstrate that the proposed Transformer model consistently yields superior robustness and generalization compared to classical methods, particularly in high-noise environments. System-level reliability is further strengthened through structured hyperparameter design, segment-level weak supervision, and a multi-seed ensemble strategy. Furthermore, an optional dataset-level label auditing mechanism is incorporated to refine training quality. Evaluations performed on both public benchmarks and real-world production recordings show that the proposed approach maintains high performance across diverse machine types and noise levels, while providing significantly clearer decision margins than traditional baseline models.