Development and performance evaluation of prognostic approaches for technical systems

As the need to improve reliability, availability and safety of technical systems increases, a large number of proactive maintenance strategies have been proposed. Of greater interest is the development of prognostic and health management strategies where maintenance is scheduled based on the current and the predicted future health state of a technical system. In addition, prognostic information can be used to control the reliability of intelligent mechatronic systems to ensure their mission objective is achieved. Therefore methodologies for estimating these current and future health states reliably and accurately are imperative. With the advancement in sensor technology, majority of the present day technical systems are installed with a network of sensors for condition or performance monitoring. This has led to the increased application of machine learning algorithms in condition monitoring. Depending on the sensor data available, different approaches for utilizing the data with machine learning algorithms can be applied. However, a guide for selecting the appropriate approach for a given system is either lacking or has not been explored extensively. Therefore, this work aims at providing a guide for selecting suitable approaches and machine learning algorithms for a given system depending on the available sensor data. Five approaches for prognostics and an ensemble of the best performing approaches are presented. Since the performance of machine learning algorithms is highly dependent on the input features, methods for feature extraction and selection are also presented. The approaches are evaluated and validated with run-to-failure condition monitoring data of actual systems.