Research On Intelligent Identification Method Of Wireless Signal Modulation Mode Based On Knowledge Graph

Modulation mode identification has a very wide application demand in radio communication,and this topic introduces knowledge graph technology to modulation mode identification to improve the performance of identification.Knowledge graph is a kind of semantic network to reveal the relationship between entities,which is a popular direction in the field of semantic database and data mining.The research of using knowledge graph to achieve modulation mode recognition is just beginning,and this topic tries to find a feasible and effective solution in this field.To improve recognition performance and perform subsequent inference and prediction,a knowledge graph construction framework applicable to modulation mode recognition is proposed and implemented in this topic,and intelligent recognition of signal modulation modes is realized based on the framework.The framework includes three modules:a multimodal entity relationship joint extraction module,an entity alignment module based on representation learning,and a knowledge inference module based on hierarchical reinforcement learning of Option.The main work and contributions of this topic are summarized as follows.First,the joint entity-relationship extraction module for multimodality is divided into two main parts:entity extraction of multimodal data(simulated signal data,signal images and text)and end-to-end relationship extraction.The multimodal data entity extraction part is designed with a multi-headed attention mechanism based on weighted combined similarity,which is added as an Attention layer in the Bert-BiLSTM-CRF model and conditional GAN to achieve further extraction of key features in text and Image Caption.In the named entity recognition of the signal graph,firstly,the signal graph generates word vectors that can imply the semantics of the described text,and then the word vectors are input to the BiLSTM layer for text entity extraction,and then the named entity recognition of the signal graph is finally realized by the Attention layer and CRF layer,which realizes the two modal data represented by a unified vector pattern.In the relationship extraction module,the unified representation of the two modal data is input into PCNN to realize the extraction of relationships.Second,in this study,the entity alignment module based on representation learning is designed and implemented to achieve knowledge alignment,which utilizes the Embedding module and Alignment module of the open source OpenEA-Tutorial framework for extended programming.In the Embedding module,the MTransE is improved by a weighted combination of two techniques,axis calibration and linear transformation,to enhance both the distance factor and the spatial structure factor,which together effectively achieve the Embedding task.The improved extensions in the Alignment module focus on similarity calculation,using Jaccard similarity and cosine distance to construct the combined similarity,considering both the co-occurrence degree and semantic similarity of the words.Third,the knowledge inference module of hierarchical reinforcement learning based on Option consists of a historical information layer,which acquires the action space of this model,and a decision layer,which implements the inference task.The decision layer corresponds each Option to a modulation,which establishes both the connection between the Option discovery problem and the learning problem,and determines the number of Options.The action space of the decision layer is generated by identifying the task learning according to the modulation method,so that the transfer of states unfolds only in a limited action space,avoiding unnecessary state transfer of the model.The Option-Critic model is optimized by a restrictive definition of the Option termination function βω.To verify the feasibility and effectiveness of the model,this study conducts experimental validation for the three modules of the model and quantitatively evaluates the experimental result data to prove that the design is feasible and effective.