Thesis Description
During the typical start-up process of the steam turbine, after the nominal speed has been reached, the massfow is continuously increased until the setpoint is reached. Due to the very low ow at the beginning of the start-up process, instationary ow phenomena can occur which can cause elevated blade vibrations [1]. These flow phenomena can also occur in low-load operations. It is important to simulate and assess these instationary flow phenomena and the efects on the blade vibrational behaviour of the blades [2]. For this purpose, both experimental model turbine investigations and transient 3D Computational Fluid Dynamics (CFD) simulations are carried out, which provide a variety of time-dependent data. As a result, the data includes the operating status such as mass ow, pressures, and temperatures, the excitation forces such as pressure changes and frequencies, and vibration data such as amplitudes and modes.
First of all, the existing experimental data will be analyzed. The aim is to investigate whether a powerful method for predicting vibration amplitudes of the blades can be formulated. If necessary, additional data must be generated with the help of CFD simulations to enable an assessment of the Machine Learning (ML) [3] processes for the entire design space. Based on the data analysis, suitable ML algorithms should be identified and tested. Here, for example, Multi-layer Perceptron Regressor (MLP Regressor) [4, 5], Random Forests [6], Decision Tree Regressor (DTR) [7], Support Vector Regressor (SVR) [8], etc. can be used and tested. In particular, ML algorithms are to be tested that are suitable for mapping time series, such as Recurrent Neural Networks (RNN) [9]. The predictive quality of the individual processes should be assessed especially with respect to the entire design space.
In summary, the thesis deals with the following points:
- Analysis of the existing experimental data with respect to:
(a) Coverage of the design space (sampling)
(b) Dimensionality refinement of the design space - Application of suitable ML algorithms:
(a) Various regression models
(b) Deep Learning for mapping time series - Analysis of the generated CFD data by Machine Learning Models:
(a) Examining on how a further improvement of the prediction quality can be achieved by generating data by CFD analysis and correlating with the measured data points
References
[1] Neville Rieger. Progress with the solution of vibration problems of steam turbine blades. 05 2021.
[2] R.S. Mohan, A. Sarkar, and A.S. Sekhar. Vibration analysis of a steam turbine blade. INTERNOISE 2014
– 43rd International Congress on Noise Control Engineering: Improving the World Through Noise Control,
01 2014.
[3] Alfredo Rodriguez, Youness El Hamzaoui, J A Perez, Juan Garcia, Jose Flores-Chan, and A.L. Tejeda. The
use of artificial neural network (ann) for modeling the useful life of the failure assessment in blades of steam
turbines. Engineering Failure Analysis, 35:562{575, 06 2013.
[4] Rishabh Bhardwaj, Navonil Majumder, and Soujanya Poria. Investigating gender bias in bert, 09 2020.
[5] Tejas Subramanya, Davit Harutyunyan, and Roberto Riggio. Machine learning-driven service function chain
placement and scaling in mec-enabled 5g networks. Computer Networks, 166:106980, 11 2019.
[6] Andy Liaw and Matthew Wiener. Classification and regression by randomforest. Forest, 23, 11 2001.
[7] Wei-Yin Loh. Classification and regression trees. Wiley Interdisciplinary Reviews: Data Mining and Knowl-
edge Discovery, 1:14 { 23, 01 2011.
[8] Mariette Awad and Rahul Khanna. Support Vector Regression, pages 67{80. 01 2015.
[9] Alex Sherstinsky. Fundamentals of recurrent neural network (rnn) and long short-term memory (lstm)
network. Physica D: Nonlinear Phenomena, 404:132306, 03 2020.