Practical semi-supervised learning framework for real-time warning of aerodynamic instabilities: Applications from compressors to gas turbine engines

Abstract

This study introduces a semi-supervised learning framework for the aerodynamic instability warning in gas turbine engines, emphasizing effectiveness, generalization, and practicality. The initial preprocessing involves low-pass filtering and downsampling to mitigate noise and high-frequency disruptions in the pressure signal at the compressor outlet. A 5 ms sliding time window then segments the pressure data, followed by the adaptive wavelet synchrosqueezed transform (AWSST) for sample labeling. To address significant dataset imbalance, an anomaly detection approach is adopted, incorporating feature selection with ReliefF and mutual information, a sparse autoencoder with bidirectional gated recurrent units (BiGRU-SAE), and a warning logic based on reconstruction errors and pressure drop amplitudes. The framework's effectiveness and generalization are evaluated across all datasets and validated through real-time warning experiments on a hardware-in-the-loop (HIL) simulation platform. Results show that our method detects instabilities 20 to 45 ms earlier than monitoring the pressure change rate, with a single-step computation time of approximately 3 ms, well within the requirements for real-time processing. This improvement in early detection can significantly enhance engine safety and performance. Notably, our method demonstrates generalizability across different states of the same engine and between different engines, suggesting its potential for developing a widely applicable warning model with limited data.