Monitoring thermoacoustic instability: A comparative analysis of feature-based and end-to-end deep learning approaches

This paper compares two data-driven frameworks for monitoring thermoacoustic instability (TAI) in a gas turbine combustor. A conventional machine learning approach using handcrafted features is contrasted with an end-to-end deep learning method employing a convolutional autoencoder (CNN-AE). Both frameworks generate a continuous stability index to quantify the transition from stable to unstable states. Using experimental data, both indices successfully track the entire dynamic evolution, including intermittent precursors. Critically, the CNN-AE autonomously learns a physically meaningful latent space. Visualizing this space reveals the system’s trajectory, showing a clear transition from a disordered attractor (combustion noise) to well-defined limit cycles (instability) through distinct topological shifts. The study demonstrates that deep representation learning not only automates monitoring but also provides a powerful tool for uncovering the underlying nonlinear dynamics of TAI. • A feature-based ML and a CNN-AE framework are systematically compared for TAI monitoring. Both successfully generate a continuous stability index tracking the full dynamic evolution from stable noise to limit cycle oscillations, including intermittent precursors. • A 1D convolutional autoencoder (CNN-AE) autonomously learns an effective low-dimensional representation from raw pressure signals, achieving monitoring performance comparable to an expert-designed, multi-domain feature set without requiring prior physical knowledge. • The learned latent space provides new physical insights into TAI dynamics. Visualization reveals a clear topological transition from a chaotic attractor (stable state) to well-defined limit cycles (unstable state), demonstrating the model’s ability to capture the underlying nonlinear dynamics of instability onset.