The detection of faulty conditions in steel plants by characterization of the emitted sound is a versatile tool with usually small installation effort.

Furthermore, the question from where the erroneous sound is emitted is an interesting research topic with wide application possibilities, as it directly allows not only the detection but also the immediate localization of potential problems. The advantages of acoustic condition monitoring with localization features range from condition-based maintenance to early diagnosis of unexpected behavior of the plant and thus to an improvement of the economic efficiency.

Causal relation models describing the interdependencies and connections between process variables may also help to perform supervision and quality control in various systems. These will be especially considered in those ones where no quality criteria are measured/supervised over time, i.e. where no direct predictive mappings can be established. Such models, which may be built from high-dimensional and/or transformed system identification cycles might be used for residual signal analysis: significant deviations may provide an indication for upcoming problems either at a late or even at an early stage.

This Ph.D. project focuses on patterns of sound signals in the time-frequency domain in combination with machine and deep learning based classification or regression methods. Appropriate feature extraction, sliding-window and frequency analysis techniques of the audio signals will help to prepare an adequate compact representation of the signals and to reduce the dimensionality of the learning problems and the models.