Classically, it is assumed that the underlying operating conditions of a technical system are known a priori and stationary, i.e. constant or periodic, during the lifecycle of the system. This assumption simplifies the implementation of condition monitoring methods in various aspects. For example, when diagnosing bearing fault/failure, i.e. the estimation of the current health state and the identification of the failed bearing component, bearing fault frequencies are calculated or examined in relation to the (constant) speed or frequency. During prognosis, the current trend can then be propagated in the future using appropriate methods.
However, many technical systems are operated under non-stationary, such as stochastic or discrete conditions, which are necessarily not known a priori. In a production line, where it might be required to manufacture different components at different operating conditions, the necessary force for press fitting as an example, might assume discrete states due to the different material properties of the components. This non-stationary character is reflected in the recorded measurement data, i.e. changes due to friction and degradation are masked by varying operating conditions, which poses new challenges and difficulties. The subject of current research is how to tackle challenges and problems posed while developing condition monitoring algorithms for technical systems operating under non-stationary conditions.