Condition monitoring and advanced fault detection techniques for engineering systems

The technological developments are unavoidable paths to ensure the reliability and efficiency of industrial assets, which are requirements for living standard and stable economy both in constant demand. In this context, detecting eventual defect is crucial for ensuring the maximal performance of all the machines within the asset and avoiding their failure and even a complete breakdown. In most of industrial installation nowadays, the strategy is changing from preventive to condition monitoring. The idea consists of scheduling interventions on equipment based on the technical condition of machines, while they are operating. The strategy has tremendous technical and economic advantages. On the other hand, condition monitoring has to be built such that, it gives both the current condition of the machine and an indication on its remaining useful life. The wide community of researchers in condition monitoring concluded that, for signals such as vibrations, they can be modeled as cyclostationary and non-stationary signals, and several mathematical approaches have been developed. Signal processing for a wide class of industrial application is still a subject of serious disagreement both in scientific and industrial community. Indeed, feature extraction and classification are a delicate task, because an error at this step can lead to huge consequences in detection of a potential faulty component and obviously the subsequent failure machine. Another drawback of the signal processing methods currently used in condition monitoring is that they are designed for stationary signals, whereas in real industrial applications, in most of the scenarios, machines operate at varying speed and load, leading in lack of feature extraction performance. The main contributions of this research can inevitably be in methodological development, particularly in the data manipulation in order to improve the results at the detection and diagnosis level, while opening a prognostic window. Applications such as spectral analysis where evaluated, where the development in this thesis suggested an enhancement of the traditional techniques applied to detect electric machines faults. From this enhancement, a particular analysis is done on cascade methods. An advanced spectral analysis is suggested as well where techniques such as spectral negentropy or spectral correlation are used to extract fault information on bearings. In addition, thesis proposed an experimental study using data mining approaches and the use of digital twins for critical industrial processes monitoring. [edited by Author]