This project was carried out for an internship as part of the Master Mechanical Engineering. The project considered plant-wide oscillation detection and root cause diagnosis, based upon data-driven methods as described in the papers Thornhill et al.  and  and in the book Choudhury et al. . The ISC report and corresponding oscillation detection and diagnosis software of Guan and van der Molen  acted as a starting point for this project.
A critical view on the software of Guan and van der Molen  revealed some important implementation mistakes, especially in the diagnosis part. This also explained the fact that the diagnosis results did not make sense in the report of Guan and van der Molen .
The software was improved to a certain extent. Three data sources were used as input for the improved software:
¿ Data sets of the Hunterston nuclear power plant, provided to ISC by EDF Energy;
¿ Data sets of a Pilkington windshield factory;
¿ Data sets generated with a two-boiler steam raising system Simulink model.
Oscillation detection on the Hunterston data set "EDF Data: R3 2 hours" resulted in the same largest plant-wide oscillation group as was the case for Guan and van der Molen . However, the calculated nonlinearity indexes were completely different. Whether the new software results, i.e. the root cause(s) indicated by the diagnosis software, for the EDF data were more reliable than the results in report , could not be stated with certainty since the actual root cause(s) were unknown to the authors.
The Pilkington data turned out to be unsuitable for the software due to its inconsistent sample time used for data measurement.
The results with the new software on data sets generated by the two-boiler steam system Simulink model were very promising. Adding backlash at a certain valve in the system to reflect nonlinear stick and slip behaviour of the valve, did introduce a plant-wide oscillation. Besides detecting the plant-wide oscillation group, the software successfully indicated the correct root cause of the plant-wide oscillation.
Unfortunately, some issues still occurred with the current detection and diagnosis software. For the detection part, the frequency-domain filter, used for multiple oscillation detection in one signal, did not work optimally yet. Improving this filter would be a main task when continuing this research. For the diagnosis part, the uncertainty of the calculated nonlinearity index was higher than desired. Probably the most important issue consisted of the strict requirements for the data set used as input in order to get reliable software results, including a constant sample time and an absence of (significant) set-point changes.
Concluding, generally the software was improved. Especially the application on the steam system simulation data gave satisfactory results. Unfortunately the software was not proven for real plant data yet (e.g. EDF data). In order to accomplish this goal a data set with fewer set-point changes would be needed. Furthermore cooperation with an engineer of EDF with more in-depth process understanding of the Hunterston nuclear power plant would be needed, in order to analyse the software results and check the root cause diagnosis