Remaining Useful Lifetime for New and Used Technical Systems under Non-Stationary Conditions

Condition-based maintenance and predictive maintenance are increasingly applied in the industry due to their ability of ensuring an optimum utilization of the monitored system. These maintenance strategies allow for diagnosing and predicting the health states of the system under stationary operating conditions. However, technical systems mostly operate under non-stationary conditions, e.g. a wind turbine affected by different loads and speeds due to stochastic wind excitation. Non-stationary conditions lead to changed sensor data and thereby mask alterations caused by either faults or degradation of the system. Therefore, condition monitoring methods need to be extended and adapted for systems operating under non-stationary conditions.The proposed project aims to develop methods for remaining useful lifetime prediction for systems operated under non-stationary conditions. Therefore, classical data-driven and model-based approaches from engineering are combined with approaches from the field of artificial intelligence. By a hybrid combination of clustering and classification with knowledge-based approaches, operating conditions are categorized and failure modes are identified. Based on uncertainty quantification and analyzed relationships between the operating conditions, the sensor data und the degradation evolution, suitable features for enabling the prediction of the remaining useful lifetime are developed and evaluated. Embedding non-stationary future operating conditions is realized by the use of different machine learning methods such as learning on data streams. These methods enable incremental learning and adaption to changes like variation of operating conditions. Moreover, hybrid methods are developed to allow a prediction of the remaining useful lifetime for used systems that are retrofitted with suitable sensors but lack sensor data of their past operation.For validation of the methods for remaining useful lifetime predictions, three application examples are chosen which have been selected from various thematic fields. To generate data for the first example, a suitable ball bearing test rig needs to be developed and constructed. The test rig should allow varying operating conditions regarding speed and bearing load. Run-to-failure data is acquired by different sensors such as acceleration sensors, temperature sensors, and strain gauges. For the second example, a laboratory experiment based on piezoelectric transducers is also implemented, whose failure is characterized by cracks and should be monitored. The third example is based on simulated data of a turbofan engine whose degradation under six conditions has been detected by various sensors.