Theory-Assisted Deep Learning Weapon System Combat Effectiveness Prediction

Combat effectiveness prediction is of very importance throughout the entire process of weapon equipment design, production, and actual combat. This paper proposes a theory-assisted deep learning combat effectiveness prediction algorithm that incorporates knowledge from weapon foundation models as auxiliary drivers. The paper provides a brief overview of the general construction form of weapon system foundation models. Based on this foundation, it elaborates on the training method that integrates theoretical knowledge into conventional deep learning models. This process does not require specific knowledge of the underlying physical model, while retaining the advantages of direct and efficient data-driven techniques. Addressing the multidimensional time series prediction involved in combat effectiveness prediction, a GA-CNN-LSTM hybrid prediction model is proposed, which adaptively learns temporal and spatial features of the data. This approach effectively mitigates the problem of poor generalization performance in deep learning large models caused by insufficient training data. The effectiveness and utility of the method are validated through a case study on combat simulation performance evaluation of a certain type of radar equipment. The results demonstrate that the hybrid prediction model improves the R2 score performance metric by approximately 2.9% compared to the original model.