Many industrial systems rely on some form of “health” estimation to predict and to react suitably to malfunctions. To name a concrete example, in aircraft turboengines, anomalies need to be detected as soon as possible, preferably on the ground, in order to ensure safe and economical flight operations. It is in general a difficult question, however, how to estimate the health of a complex engineering system accurately and efficiently. In particular, over-reporting in a high-stress environment like an airplane cockpit may overload human operators.
The goal of this project is the development, theoretical analysis, and implementation of methods to estimate the health of engineering systems from sensor data in real-time and to make predictions about the health in the future. This will allow better prediction and hence avoidance of failure in complex machinery, and this shall improve the safety of the person and property.
The health of an engineering system is defined as the failure probability distribution over all points in time from the current time onwards. The estimation of this distribution poses many challenges of both theoretical and applied nature. For instance, a gradual change in health is the norm, rather than the exception, and traditional anomaly (changepoint) detection methods are often of little use. Another aspect is the selection of meaningful statistics of the future failure distribution to guide maintenance and inspection efforts in real-world systems. We will investigate idealised theoretical models and, concurrently, embark on an intense exchange of ideas with practitioners and applied statisticians to guide the theoretical investigations. While at least in the beginning the emphasis of the project is predominantly on the theoretical side, we will also implement any new algorithms and apply them to data from industry partners.