Real-time diagnostics of complex mechanical systems

Real-time damage diagnostics of complex mechanical systems, such as aircraft engines or gas turbines in power plants, is one of the greatest current challenges in maintenance industry to control cost and time. In the project we focus on a combination of compressive sensing and machine learning in order to extract critical damage indicators from the complex information of key vibration sensors. A wavelet basis will allow for addressing nonlinear behavior, tracking frequency changes in time. The retrieved signal characteristics are then evaluated against small, partial mechanical models, data sets representing characteristic damage patterns from previous cases, data from healthy systems, and expert assessments. Uncertainty will be considered in the partial mechanical models, in the data, as well as in the expert assessments. Hence, the aspired damage identification will be developed as a statistical test assessing the distance between distribution functions, prospectively with the Bhattacharyya distance expanded to epistemic uncertainties from expert assessments. A self-generating mechanism for consecutive performance improvement of the algorithm will be implemented through a machine learning approach in form of Bayesian compressive sensing. In this regard, the general and fundamental algorithms can achieve optimal performance in a broad range of application areas.

Doctoral Researcher: Jan Grashorn

Scientific Advisors: Michael Beer, Ludovic Chamoin