Abstract
Flame flashback is a major concern in hydrogen-fired gas turbines. In order to determine the flashback propensity of a hydrogen burner, several burner design tests at different operating points and fuel blends are performed under engine-relevant conditions at the test facility of Siemens Energy. A camera monitors the flame in the combustion chamber and the occurrence of flame flashback events in the image recordings becomes clearly visible. This anomalous behaviour clearly deviates from normal hydrogen operation. We develop a data-driven approach to detect flame flashback events based on the camera images at 100% hydrogen operation, where all images feature identical characteristics since the pure hydrogen flame is not visible for the camera. Simultaneously, the highest susceptibility to flashback is attained in this regime. We use both facts as well as the good suitability of image data to train a Convolutional Autoencoder (CAE) model to detect anomalies. Here, anomalies correspond to flashback events. Flashback is captured by the CAE using the reconstruction error associated with a dynamic threshold as a measure of anomaly. This newly developed dynamic threshold overcomes the difficulties in the generalization capability of the CAE. Regardless of the test campaign, burner design and camera settings, it reliably identifies flashback events. Along with the CAE, the compressed representation, namely the latent space of the CAE, detects the position of flame flashback events. Our methodology is able to detect flame flashback using only flame images and provides a reliable tool even when unseen data are used.
