Research On ACARS Packet Parsing Technology Based On Machine Learning

ACARS(Aircraft Communications Addressing and Reporting System)packet parsing is one of the foundations of aviation big data mining,but the flexibility of traditional ACARS packet parsing method based on pattern matching is not high.The old template cannot parse the packet in a new format,and it difficult to make use of the hidden pattern in the whole packet,so the parsing template needs to be added manually frequently according to the characteristics of the new packet.Writing parsing template requires the author to have high professional knowledge in relevant fields,and the reuse rate of written parsing template is not high,which leads to the parsing efficiency of ACARS packet being far lower than the rate of message generation.At the same time,the increasing number of messages put forward higher requirements for its parsing efficiency.For this reason,this dissertation proposes a packet parsing model based on neural network,which can fully learn the pattern hidden in ACARS packets.It can not only improve the success rate of parsing old packet,but also adapt to the parsing of new packets.There are many unique fields such as forwarding time in ACARS packets.If they are directly inputted into the neural network for training,it will lead to some problems,such as too large field dictionary of ACARS packets,slow convergence of the network,and the entries corresponding to the unique fields can not be updated for a long time.Therefore,a <mapping-restore> optimization algorithm is proposed to the conversion of these fields.In order to solve the problem of slow convergence of the randomly initialized neural network,this dissertation uses the word2 vec model to represent the context data such as format,structural information and field order in the ACARS packet into its result vector,and uses it as the Embedding layer of the parsing model to realize the knowledge transfer of the ACARS packet,so as to improve the training starting point of the parsing model.Finally,the experimental results show that the parsing accuracy of the parsing model designed in this paper can reach 0.97,which is 0.12 higher than 0.85 based on pattern matching.The addition of transfer learning can shorten the convergence time of the parsing model by about 1/6.At the same time,the accuracy of message parsing can be increased to0.98.After accelerating the operation of the neural network by using GPU,the convergence time of the neural network is only about 1 hand 4 of the CPU method.