Spike Stall Precursor Detection in Axial Compressor System using GNN-RNN Hybrid Architecture

Spike stall events in axial compressor systems lead to performance drops and equipment damage. However, operation at peak efficiency requires running near unstable points. A data-driven early warning system could enable active control to prevent instabilities. In this study, pressure readings from eight 60-blade compressor experiments are utilized to develop a precursor detection method. The data is pre-processed into spatiotemporal matrices and time-evolving graphs. Four neural network architectures are explored for binary classification of impending spikes using prior revolutions: GRU, ConvLSTM, GNN-GRU and GNN-ConvLSTM. The hybrid and non-hybrid models demonstrate 93% - 100% validation accuracy in detecting spikes 30 revolutions prior, representing the best precursor distance. The incorporation of graphical neural networks of-fers comparable performance. The framework establishes deep learning's capability for stall prediction from transient pressure fluctuations. The prospect exists to deploy the algorithms for enhanced efficiency via instability avoidance at near-optimum conditions. Further improvements in accuracy and robustness can be achieved through expansions to the dataset size, model complexity, and evaluation metrics.