Target RCS Prediction Based on Residual Structure and Data Augmentation

The existing network structures used for radar cross-sectional area (RCS) prediction include mainstream network structures such as BPNN and LSTM neural network. These networks are capable of learning the RCS characteristics of objects and making predictions on the RCS of objects. However, due to the limited RCS data features, increasing the network depth to learn more features may result in model degradation. Therefore, this paper proposes adding a residual connection module to the BPNN to address the issue of network degradation. Moreover, in practical applications, obtaining RCS data is challenging, and the available data is limited. Training the network solely on existing data can lead to poor robustness and overfitting. To tackle this problem, this paper improves the dataset by augmenting the original data with multiple sets of Gaussian noise data for data augmentation. When using the BP neural network for prediction, the mean absolute error (MAE) is 2.84. After incorporating the residual connection module, the MAE predicted by the model decreases to 2.21. Furthermore, after data enhancement based on the network structure, the model achieves a further decrease in MAE to 1.83. Experimental results demonstrate that the BPNN augmented with deep residual connections improves the prediction accuracy of the target RCS. Additionally, the application of Gaussian data augmentation further enhances the accuracy of the RCS prediction model.