A dual-module deep learning approach for flow field prediction in supersonic isolators from noisy signals

The isolator, a critical component connecting inlet and combustor in supersonic engines, needs precise internal flow field monitoring for safe operation. Accurate wall pressure signal measurement is vital for characterizing complex flow field with separation and shock trains. However, sensor thermal noise and electromagnetic pulse interference reduce the reliability of neural network-based flow field perception models in practical use. To address this, a dual-module model named UT-CL is proposed, integrating U-Net-Transformer (U-Transformer) for denoising and Convolutional Autoencoder-Long Short-Term Memory (CAE-LSTM) for flow field prediction. U-Transformer combines U-Net encoder-decoder with Transformer self-attention to suppress multi-scale noise and preserve high-frequency shock/separation zone features. CAE-LSTM achieves flow field reconstruction via convolutional spatial feature learning and Long Short-Term Memory (LSTM) temporal modeling to capture unsteady spatio-temporal evolution. A multimodal dataset of pressure signals and schlieren images with simulated mixed noise was generated through wind tunnel experiments under varying Mach numbers and backpressures. Compared with baselines, U-Transformer significantly improves signal quality and retains key flow field features. CAE-LSTM-reconstructed flow field, validated by schlieren imaging, outperforms single denoising or traditional models. This approach balances signal-level noise suppression and flow field reconstruction, providing a reusable framework for supersonic flow field prediction in high-noise scenarios to enhance engine monitoring accuracy.