Rotating machines system identification by means of big data analysis

In the field of mechanics, design values such as the modal parameters are essential: they are governing system response to vibration sources. To avoid problems such as the transmission of undesired vibrations and the generation of tonalities, we propose an experimental approach to gain insights in machine dynamics and verify simulation predictions. Particularly for non-linear systems, where there is dependency on operation conditions, detailed long-time-duration experiments can be of added value. With this work, we want to propose an innovative approach for long-term operational modal analysis (OMA) testing of machines operating in the field or in a laboratory environment. The goal, is to give the designers information that allows to take decisions on how to improve the subsequent design variants based on what can be learnt in real operating conditions, i.e. with the machine running through every possible scenario that it could meet during its operating life. This is limited for current testing procedures, aimed to catch only specific operating conditions that can be predicted. At this purpose, OMA represents a powerful approach. The scope of this research is to make OMA suitable for processing a continuous stream of data coming from the operating machines. To do so, state-of-the-art vibration processing techniques are combined with methodologies coming from big data analysis and machine learning disciplines. First, the state-of-the-art algorithm for applying OMA both on stand still and rotating machines are improved to eliminate user-algorithm interactions and make modal parameter estimation completely automatic. Subsequently, the procedure is embedded in an automatic processing context using machine learning algorithms to pre-process the signals and post-process the results.

Firstly, when a raw signal is available, the quality of the acquired data is checked and the signal is classified in function of the operating condition of the machine. This will allow to use the appropriate OMA approach to get optimal estimates. After that, when the estimates are available, a tracking procedure is used to track the evolution of the modal behavior of the machine along different data sets, i.e.: along longer periods.