一种基于迁移学习的多任务制导算法

A Multitask Guidance Algorithm Based on Transfer Learning

针对典型的飞行器制导任务,利用深度学习算法可以有效地拟合导弹飞行状态与制导指令之间的函数关系。然而当制导任务发生变化时,其二者之间的映射关系也会随之改变,从而导致在当前环境下预训练好的模型无法直接作用于新环境,重新训练制导模型需要大量的弹道数据和巨额的时间成本。为解决上述问题,基于迁移学习的思想引入域对抗神经网络,提出基于迁移学习的多任务制导算法。以1个含有大量标签数据的源域任务辅助2个含有极少量标签数据的目标域任务进行迁移学习,从而克服预训练与在线控制之间的环境差异。使用特征提取器和域判别器提取出对任务环境不敏感的关键特征,使神经网络学习到各个任务所共享的底层信息;为提高预测精度,分别设计针对不同任务的偏置加速度预测器。数值仿真结果表明:基于迁移学习的多任务制导算法实现了导弹在不同任务中的加速度指令预测。

For typical aircraft guidance missions,the deep learning algorithm can be used to effectively fit the functional relationship between missile flight state and guidance command.However,when the guidance mission changes,the mapping relationship between them will also change.As a result,a pre-trained model in the current environment cannot directly act on a new environment,and retraining the guidance model requires a large amount of ballistic data and a huge amount of time cost.In order to solve the above problems,a domain adversarial neural network is introduced based on the idea of transfer learning,and a multitask guidance algorithm based on transfer learning is proposed.One task in the source domain containing a large amount of tag data is used to assist two tasks in the target domain containing a small amount of tag data for transfer learning,so as to overcome the environmental difference between pre-training and online control.The key features that are not sensitive to the task environment are extracted by using feature extractor and domain discriminator so that the neural network learn the underlying information shared by each task.In order to improve the prediction accuracy,the bias acceleration predictors for different tasks are designed,respectively.The simulated results show that the multitask guidance algorithm based on transfer learning can predict the acceleration instruction of a missile in different missions.