Master's Thesis : Data modelling of steam turbine performance

Abstract

Maintaining industrial steam turbines is a very difficult task. It is now possible, using data- driven methods, to detect anomalies quickly to be able to fix them before the anomalies get bigger. In this project, a systematic methodology to detect anomalies on steam turbines is developed. Taking data coming from sensors placed in the turbine we can output the main anomalies as well as the main possible causes of the anomalies. Different algorithms are at the core of the anomaly detection process. One of them, based on Recurrent Neural Network, is a new architecture that can detect big anomalies as well as smaller anomalies. This methodology was developed on a first turbine and then applied to a different turbine proving that this newly developed methodology is systematic and only needs a few days to be adapted to a different case. The ability of this approach to successfully detect steam turbine anomalies and identify the most probable causes of these anomalies has been validated by experts.