| dc.description.abstract | Abstract. This study presents and evaluates three independent stall warning approaches based on unsteady casing pressure measurements acquired above the rotor blade tips. The methods are applied to two different axial compressors, each exhibiting distinct stall characteristics and inception mechanisms. The first approach utilizes a cross-correlation technique to detect small perturbations propagating circumferentially just prior to stall onset. The second method employs a convolutional neural network, trained on labeled time-resolved pressure data, to accurately classify operating ranges as stable and unstable. The third approach is based on an autoencoder trained on pressure data exclusively recorded at design flow conditions; deviations from the learned patterns are flagged as anomalies. This unsupervised technique enables the detection of any pre-stall anomalies—including those, not identifiable through visual inspection—that precede conventional stall indicators such as spikes. Importantly, none of the three approaches rely on manually tuned threshold values to trigger stall warnings. As a result, especially the autoencoder remains valid even with changing tip clearances, making this approach highly robust. All methods are validated using independent datasets, and their accuracy is quantified using stall margin-based metrics. Robustness under varying signal-to-noise conditions further demonstrates the applicability of these techniques for reliable stall warning. While the deep learning approaches consistently provide precise warnings at about 5% stall margin, the cross-correlation method only triggers at a similar level if the compressor is free of pre-stall instabilities such as rotating instability. | |
