Research On Intelligent Optimization Methods Of Information Processing For Autonomous Cooperation Of Unmanned Swarms

With the rapid development of key technologies in electronics,communications,artificial intelligence,mechanical manufacturing and other disciplines,unmanned platforms are receiving more and more attention and are increasingly widely used in military and civilian fields.However,with the continuous expansion of the usage scenarios of unmanned platforms,single unmanned platform gradually exposes the problems of low efficiency,poor flexibility and weak robustness.Organizing multiple unmanned platforms to form unmanned swarms has become an important direction of the development of unmanned systems.Effective collaborative interaction between unmanned platforms is the key to the formation of unmanned swarm.To achieve collaborative interaction within unmanned swarm,first of all,we need to solve the data computing and transmission problems,which is the premise to realize the operation of unmanned swarm.Efficient collaborative information processing directly determines the overall efficiency of the swarm.Based on this,focusing on collaborative information processing in unmanned swarms,aiming at optimizing the efficiency of cloud-end fusion based collaborative information processing,this paper studies the important issues of collaborative communication and collaborative computing in unmanned swarms,and breaks through the key technologies of cloud-end fusion based collaborative information processing in unmanned swarms.This paper carries out the research from the following five aspects:Firstly,the cloud-end fusion based collaborative information processing framework for unmanned swarms is proposed,and the key problems to be solved in this paper are clarified.Starting from the real situation and typical applications of unmanned swarms,this paper analyses the system characteristics and general requirements of unmanned swarms,and designs a cloud-end fusion based collaborative information processing framework for unmanned swarms.At the same time,according to the interaction model between the unmanned system and the physical world extracted from the applications of unmanned cluster,the key problems of collaborative information processing optimization methods for unmanned swarms are explored,and the specific application scenario based on battlefield reconnaissance and surveillance is given.The key problems in this specific scenario are concretized,which defines the study scope of this paper.Secondly,an optimization method for data transmission link selection in unmanned swarms is proposed.Aiming at the possible link congestion problem when large-scale unmanned swarm carries out data transmission simultaneously,the link selection problem of data transmission for unmanned swarms is studied.The problem is modeled as a link selection game,and the existence of Nash equilibrium is proved by rigorous theoretical analysis.On this basis,a distributed link selection algorithm based on the fictitious play algorithm is designed.The proposed algorithm does not require the current decisions of other unmanned platforms,and enjoys a light-weight property.Unmanned platforms can independently select data transmission links to generate Nash equilibrium of the link selection game.The experimental results show that more than 70% of unmanned platforms can improve their own utilities of data transmission,and the efficiency of data transmission link is optimize.Thirdly,the optimization method for uploading data with redundancy in complex network environment is proposed.Aiming at the phenomenon that there are a large amount of data with redundancy in the swarm when it performs battlefield reconnaissance and surveillance tasks,and uploading redundant data can not contribute to information utility while waste energy and communication resources,the cooperative data transmission problem of unmanned swarms is studied,and an adaptive distributed optimization method oriented to long-term optimization is proposed.The proposed method is composed of a correlated upload decision and an online distributed scheduling algorithm.Each unmanned platform can independently decide whether to transmit data and which data to transmit according to its own observations.Through rigorous theoretical analysis and complexity pruning,the proposed method can make long-term performance arbitrarily close to the optimal value with acceptable computational complexity.Fourthly,a cloud-end collaborative computing method based on deep learning model segmentation is proposed.In order to realize efficient deep learning model computation on unmanned platforms,this chapter considers that when data transmission link is connected,a cloud-end collaborative computing method is adopted.Through deep neural network segmentation,resource-hungry tasks including model training and complex inference are conducted in the cloud.Unmanned platform only undertakes simple data transformation tasks.In order to eliminate the data security and privacy risks brought by data transmission,a new differentially private perturbation method is designed in this paper.Compared with the traditional perturbation method,the proposed method is more flexible and fits well with the stacking structure of neural networks.At the same time,in order to reduce the impact of local perturbations on the cloud-side inference,a noisy training method is proposed to train cloud-side models.The noisy training method makes the cloud-side model more robust to the noise,and as a result improves the inference performance.The experimental results show that the proposed method can reduce the resource consumption by more than 60% with negligible performance degradation.Finally,a cloud-end collaborative training method for deep learning model based on knowledge distillation is proposed.In order to enable efficient deep learning model calculation on unmanned platforms when data transmission links are disconnected,a method of training a deployable deep learning model is studied.Following the teacher-student scheme,the knowledge embedded in the complex and huge teacher model is extracted and transferred to the student model stage by stage to improve the performance of the compact student model.At the same time,in order to protect the security and privacy of the original training data and the original teacher model,all knowledge transferred to the student model is perturbed to satisfy differential privacy.In addition,a new query sample selection method is designed to significantly reduce the number of queries samples to the teacher model and enhance data security and privacy protection without performance loss.The experimental results show that the compact model trained by the proposed method can obtain 20× compression ratio and 19× speed-up with merely 0.97% accuracy loss.