Case C9 - Continuous servomotor monitoring

By monitoring the development in friction forces on the guide vanes in a Francis turbine over time, changes in the condition of the turbine and servo system can be determined. In this case study, changes in servomotor forces are determined using a machine learning method. Four identical large Francis turbines (> 300 MW) from the same plant are studied, called unit A1-A4. The work was carried out by Anders Willersrud from Hymatek Controls, and Master student Asgeir Aasnes [1], [2].

Today the friction forces are typically inspected manually by logging and investigating the differential pressure in the servomotors while performing a "servo indication". During this operation the guide vanes are slowly operated from 0-100-0 % opening. The goal in this case study is to investigate how the manual inspection can be automated by constantly monitoring friction forces during operation, through differential pressure measurements in the hydraulic system.

Friction forces are calculated from pressure measurements during production. As can be seen to the left in the figure, the different servos requires different differential pressure ("Delta force") to change direction, indicating different levels of wear and/or friction in the turbine and mechanical system.

Differential servo forces (Delta force) for all four units during production (left). Support Vector Machines defining the boundary of the data for unit A2 (right).

One Class Support Vector Machine (OC SVM) is used for the analysis, and was found to create a model that accurately defines the boundary for the servo forces, being able to detect changes in the servo forces. The results for unit A2 is seen to the right in the figure. Similar results were found for the three other units, and are therefore left out. The boundary found using OC SVM is plotted as a red line, the support vectors used to create the model are also plotted. All data located inside the boundary is classified as normal, data located on the outside as abnormal.

Any new data will be checked against the boundary found by the machine learning model, where increase in wear will change the friction band for a given opening/load of the turbine. The trained method can then be used to classify new data as normal or degraded, giving a method for continuously detecting wear of the Francis turbine.

[1] A. Ø. Åsnes, "Condition Monitoring of Hydroelectric Power Plants," NTNU, Department of Engineering Cybernetics, Trondheim, master thesis, 2018.
[2] A. Ø. Åsnes and A. Willersrud, "Predictive maintenance and life cycle estimation for hydro power plants with real-time analytics," in Hydro 2018, Gdansk, Poland, 2018.