Abstract:

In sea environment, unmanned system relies on edge artificial intelligence (AI) model collaboration to implement data collection and edge processing tasks. However, facing with problems such as poor communication links, limited communication bandwidth, and sensitive to interferences, this paper firstly proposes a model collaborative training method, federated mutual distillation, to reduce the bandwidth requirements for model training data, from the perspective of AI model collaboration among unmanned aerial vehicle. Secondly, from the perspective of data de-redundancy transmission, a data differential dynamic compression method is proposed to reduce the frequency of data transmission. Simulation results show that the performance of the model, trained with the federated mutual distillation training method, is better than that of the benchmarks, and a cost of communication bandwidth is reduced compared to the centralized training models. The proposed data differential dynamic compression method can greatly reduce the sending length and frequency of communication messages, and adapt to the bandwidth bottleneck in the weak communication connection environment.

Key words: weak communication link, edge model collaboration, federated mutual distillation, data compression