关键词: 模型可信度评估, 模糊评判框架, 四分支异构神经网络, 损失函数, 分阶段训练策略

Abstract: To address the challenges of flight control modeling—namely, the large number of parameters involved in assessing the reliability of equivalent models, the complexity of comparison dimensions, and the cumbersome evaluation process—an evaluation scheme combining neural networks and a fuzzy evaluation framework was designed. First, the paper introduces the parameter classification and processing workflow for this task, as well as the operational mechanism of the fuzzy evaluation framework. It then details the structural composition and loss function design of a four-branch heterogeneous neural network. Subsequently, this paper uses a specific type of interceptor as an example to demonstrate a complete evaluation process, including feature dimension adjustment and hyperparameter tuning of the neural network. Finally, we compared the evaluation results of the controlled perturbation model between the conventional evaluation approach and the neural network evaluation approach; the overall reliability rankings and reliability scores of the models evaluated by both approaches showed a high degree of consistency. Comparative experiments revealed that introducing the loss in three stages resulted in a 31.5% improvement in performance compared to introducing the entire loss at once, confirming the necessity of a staged training strategy. Furthermore, it was found that as the number of models to be evaluated increased, the acceleration ratio of the neural network evaluation path compared to the conventional error-handling evaluation path showed an overall upward trend, highlighting its efficiency. This study indicates that this evaluation method has significant potential for broader application.

Key words: Model credibility assessment, A fuzzy evaluation framework, A four-branch heterogeneous neural network, Loss function, Phased training strategy