Robuuste akoestische monitoring van roterende machines

The robust condition monitoring of rotating machinery using non-contact acoustic sensors is a critical objective for modern industry, yet its practical deployment remains challenging. Industrial acoustic environments are characterized by complex soundscapes, strong background interference, and limited availability of fault data for training diagnostic models. In addition, fault-related acoustic signatures are often subtle and highly specific, further complicating reliable detection. This PhD research addresses these challenges by proposing a suite of advanced methodologies that together enhance the robustness, applicability, and practicality of acoustic monitoring systems. The contributions are structured around three central themes: robust source separation, data-efficient learning under fault data scarcity, and the exploitation of physics-informed signal characteristics. First, to improve robustness against acoustic interference and background noise, this PhD research develops and progressively refines a deep-learning-based framework for unambiguous acoustic source separation. Conventional blind source separation (BSS) techniques suffer from a fundamental ambiguity, resulting in so-called "virtual" rather than physically meaningful sources. To address this limitation, the proposed TUNE-BSS framework learns a transformation from virtual components to physical source maps. This framework is further enhanced by replacing an adversarial loss with a spatial entropy regularization term, which serves as a physically motivated prior and improves both accuracy and zero-shot generalization. Importantly, this approach is not limited to inverse acoustic imaging methods but is also, for the first time, successfully adapted through transfer learning to improve the outputs of conventional beamforming techniques, representing a significant methodological advance for industrial applications. Second, this PhD research tackles the critical challenge of fault data scarcity, which severely restricts the deployment of data-driven diagnostic models in real-world settings. A novel simulation-driven partial domain adaptation strategy is introduced, in which a physics-based bearing model is used to generate synthetic fault data. These simulated fault representations are aligned with real-world healthy data through contrastive learning, enabling the training of an effective fault classifier in the realistic scenario where only healthy data from the physical asset are available. This approach substantially improves the feasibility of applying deep learning techniques to industrial fault diagnosis. Finally, this PhD research demonstrates that further robustness can be achieved by explicitly exploiting the physical characteristics of bearing fault signatures. A cost-effective acoustic monitoring method based on a sparse microphone array is presented, which applies beamforming to the low-frequency envelope of acoustic signals to localize amplitude-modulated fault signatures using minimal hardware resources. Taken together, the contributions of this PhD research provide a comprehensive methodological toolkit for robust acoustic monitoring of rotating machinery. All proposed approaches are rigorously validated using experimental and simulated data, demonstrating significant progress toward making acoustic-based condition monitoring a reliable and practical solution for industrial maintenance.