Colloquia

Samenvatting

This master's thesis details the development of a machine learning based anomaly detection system for misaligned guide vanes in a Francis (hydropower) turbine. A LSTM autoencoder model was developed and optimised using experimental data that was collected from a working physical model Francis turbine that was constructed as part of this research. This master's thesis was undertaken in cooperation with Gugler Water Turbines GmbH in Feldkirchen an der Donau, Austria, with the goal of creating an optimised LSTM autoencoder which can be used to monitor real turbines. The approach of creating an experimental turbine and using it to collect training and testing data for the model was used because there was no failure data available from real turbines, which is a common problem for the application of AI to hydrpower equipment due to the reliable nature of the equipment.

The model turbine was predominantly constructed from 3D printed parts with the exception of the turbine shaft, and the electronic components. The turbine was controlled using an Arduino Uno microcontroller, which also served as a data logger for a two pressure sensors, a position sensor, an RPM sensor, and a voltage and current sensor. Using this experimental model, data was collected for both steady state operating conditions and dynamic operating conditions (wherin the turbine regulator was opened and closed repeatitively) with and without a single misaligned guide vane.

The experimental data was used to create and then optimise the LSTM autoencoder model which was designed to detect the misaligned guide vanes. The optimisation process was conducted using the Optuna optimisation library in python, followed by model trials using different input sequence lengths and input variable combinations. The model was able to accurately detect a misaligned guide vane during steady state operation, with slightly decreased accuracy during dynamic operation. When the optimised model structure was used for a model trained using data from an actual turbine without anomalies, no anomalies were detected except for in the period immediately following changes in the regulator position induced by the operator.