Research on predicting flow fields based on discrete monitoring data

This study addresses the issues of high cost and low efficiency in the analysis of flow fields around underwater moving bodies. It proposes a rapid prediction method for flow field parameters based on sparse pressure monitoring points and deep learning technology. By constructing a U-Net neural network model that integrates a multi-path encoding structure, an end-to-end mapping from limited monitoring data to full-field velocity and pressure distributions is achieved. The study systematically explores the influence of different nose geometries and monitoring point configurations on prediction accuracy. The results indicate that a 36-point medium-density monitoring configuration exhibits a favorable balance between information capture and generalization ability, while configurations with fewer points can also meet practical needs through reasonable distribution design. The model maintains good structural similarity (average SSIM >0.92) even outside the training data distribution, though there is some degradation in local amplitude accuracy (PSNR). This method preserves the prediction accuracy of large-scale flow structures while significantly improving the efficiency of flow field analysis, with the inference time per working condition reduced to within 10 s. It provides an effective technical approach for multi-condition performance analysis and shape optimization of underwater moving bodies.