The existing model relies too much on the pressure data, which not only increases the workload and cost of experiment and measurement, but also makes the model extremely sensitive to the quality and accuracy of the data. In addition, it is often difficult for existing models to capture and retain the characteristics of complex wave system structure in an all-round way, and it is unable to effectively capture the time evolution characteristics of the unsteady flow field, resulting in the inaccurate identification of the leading edge position of the shock string in isolator. In this paper, the flow field prediction model was established to extract the complex features of the flow field and enhanceed the details of the wave system features, so as to realize the high-precision prediction of the flow field.

A NNCDFE (neural network model based on combined detail feature enhancement) for isolator flow field reconstruction was proposed, comprising a wave system structure feature extraction network and a detail feature enhancement network. The model achieved high-fidelity reconstruction through a composite architecture design. The wave system structure feature extraction network employed an encoder-decoder framework, where the encoder utilized transposed convolutional layers to progressively reconstruct high-resolution feature maps from compressed input data, establishing the primary shock wave system characteristics. The decoder further processed these features through convolutional layers and refined the spatial distribution via fully connected layers. To address the loss of subtle flow details in multi-layer convolutions, a residual network with skip connections was implemented in the detail enhancement module, enabling multi-scale feature refinement by fusing shallow and deep layer outputs. Ultimately, the synthesized output delivered reconstructed flow field images with enhanced shock wave system integrity and resolved boundary layer vortices, demonstrating superior capability in preserving transient flow features compared to conventional methods.

Experimental validation on the isolator flow field dataset demonstrated the superior performance of the proposed NNCDFE method. Compared with four benchmark approaches, NNCDFE achieved average metrics of 24.661 dB peak signal-to-noise ratio, 0.886 structural similarity index measure, and 0.857 correlation coefficient, indicating significantly enhanced reconstruction fidelity. Furthermore, the model attained a 0.71% average relative error in STLE (shock train leading edge) localization against numerical simulation references, validating its precision in resolving wave system details. Notably, NNCDFE maintained robust performance under progressively sparse pressure input conditions, proving its capability for high-accuracy density gradient field prediction from limited pressure measurements.

The NNCDFE model proposed in this study effectively compensates for the flow field details overlooked by multi-layer convolutional neural networks during feature extraction through its combined neural network framework. By enhancing the details of the reconstructed flow field, it improves the detection accuracy of the STLE position. Simultaneously, the NNCDFE model retains density gradient field prediction capability under sparse pressure data conditions, demonstrating strong robustness. This provides data support for subsequent scramjet STLE position control.