Open Access

Marcel Achzet

• Ultrasound-based condition monitoring (CM) offers a non-invasive, sensitive means of tracking technical states in industrial processes. However, challenges such as signal complexity, sensor limitations, and performance degradation over time hinder its broader application. This dissertation addresses these challenges by introducing three advanced techniques: application-specific feature adaptation, robust feature selection, and data reconstruction for unmonitored areas. The first method focuses on tailoring feature parameters—such as window size and frequency range—to enhance the predictive accuracy of process variables. The second develops a selection strategy aimed at long-term model stability, identifying features resilient to environmental and operational drift. The third approach addresses physical sensor limitations by proposing data reconstruction methods to infer conditions in unsensed regions. These techniques are validated on diverse experimental setups including aUltrasound-based condition monitoring (CM) offers a non-invasive, sensitive means of tracking technical states in industrial processes. However, challenges such as signal complexity, sensor limitations, and performance degradation over time hinder its broader application. This dissertation addresses these challenges by introducing three advanced techniques: application-specific feature adaptation, robust feature selection, and data reconstruction for unmonitored areas. The first method focuses on tailoring feature parameters—such as window size and frequency range—to enhance the predictive accuracy of process variables. The second develops a selection strategy aimed at long-term model stability, identifying features resilient to environmental and operational drift. The third approach addresses physical sensor limitations by proposing data reconstruction methods to infer conditions in unsensed regions. These techniques are validated on diverse experimental setups including a gearbox test bench, rheological measurements, and resin transfer molding processes. Results show that the proposed methods significantly improve the accuracy, robustness, and applicability of ultrasound CM, reducing the need for frequent recalibration and enabling broader industrial adoption.…