Research On Readback Verification Of Aviation Radiotelephony Communication Based On Deep Learning

Civil Radiotelephony communication is one of the main ways of communication between ground and air.However,flight accidents caused by Civil Radiotelephony communication errors are very common in actual flight.Among them,the readback error is the main reason.Therefore,achieving readback verification of Aviation Radiotelephony Communication automatically can improve flight safety.In recent years,deep learning has been widely used in NLP and has achieved breakthrough development.So BiLSTM and Attention Mechanism are adopted to construct a readback verification model for Aviation Radiotelephony Communication.Firstly,the original readback corpus of Civil Radiotelephony communication is expanded and preprocessed to obtain the word representation of all the words in the corpus.The word vectors used in this thesis are one-hot and word2vec;Then,the BiLSTM model to achieve the task of readback verification of Civil Radiotelephony communication are built.A parallel BiLSTM networks are used to extract the semantic feature of instructions and readback instructions respectively.The cosine similarity and neural tensor network are used to interactively match the output of BiLSTM at each moment to generate a semantic matching matrix,so that the final matching score is obtained by k-Max pooling and MLP.Finally,BiLSTM-Attention model are proposed,which also uses the BiLSTM network to extract semantic features from instruction and readback instruction,the difference is that the BiLSTM model introduces the idea of interactive matching,semantically interacts the output of the BiLSTM network at each moment.It can obtain more fine-grained matching features between instructions.While the BiLSTM-Attention model adds attention mechanism layer to assign different weight values to the extracted semantic feature vectors,thus improving the model matching effect.In order to verify the effect of the model,the experimental results of word2 vec and one-hot in Chinese corpus of Civil Radiotelephony communication and English corpus of Civil Radiotelephony communication are compared in the thesis.The experimental results show that the two verification models proposed in the thesis are effective in readback verification tasks,and the BiLSTM-Attention model is better.