| With the increasing importance of wireless communication in the development of information technology,electromagnetic information security is particularly crucial in the context of increasingly complex electromagnetic environment applications.Radio monitoring is one of the important means to ensure the safety of electromagnetic information.In response to the urgent needs in the field of automatic detection and classification of illegal electromagnetic signals,it is of great significance to quickly search and discover from massive signal data based on target signals,extract electromagnetic signal features,and complete signal detection and modulation pattern recognition.The extensive application of deep learning methods in the field of artificial intelligence has pointed out a new research direction for signal detection and modulation recognition.Currently,the iterative upgrade of the hardware platform of computing devices utilizes multithreaded parallel computing,while leveraging network model lightweight technology to accelerate signal processing and analysis,and achieve the requirements of deploying models to embedded edge devices.In this thesis,a signal fast detection is realized based on parallel computing and convolutional autoencoder.Then,a lightweight model based on feature fusion and knowledge distillation is used to complete the modulation recognition task of detected signals.Finally,a system application of signal detection and recognition is built on edge devices.Aiming at the detection task of large amount of spectrum data in practical application scenarios,this thesis designs a signal fast detection model by using neural network and CPU parallel architecture.Firstly,a one-dimensional convolutional autoencoder is used to extract key features from power spectrum data.Then,detection is achieved through a classifier.Compared with traditional fully connected networks,the proposed method improves the detection accuracy.To further accelerate the process of signal detection,this thesis optimizes the detection algorithm based on multi-threaded parallel computing,and obtains the relationship between the number of threads and acceleration time through experiments,achieving good acceleration effect.In order to identify the modulation type of detected signals on edge devices,the recognition performance of CLDNN network model based on the in-phase and quadrature component is studied in this thesis.However,the recognition accuracy of the signals is poor.Next,from the perspective of signal input characteristics,this thesis proposes a modulation recognition method based on feature fusion.The in-phase and quadrature components,phase information,and power spectrum data of the signal are used as model inputs,and feature extraction is performed on the signal in the time and frequency domains,followed by feature fusion,and then output through a classification network.The experiment proves that the model has a higher accuracy.Finally,in order to meet the requirements of model lightweight,this thesis compresses and optimizes the model based on knowledge distillation methods,using feature fusion models as a teacher network to train the student network,making its performance as close to the teacher network as possible.Through experimental verification,the student network has a performance that is close to the accuracy of the teacher network,and the algorithm execution efficiency has also been improved.Finally,the signal detection and recognition system has been developed based on edge devices.In combination with the specific system software requirements of the national major instrument research and development project-Development and Application of Illegal Electromagnetic Signal Monitoring and Classification Technology,a testing system is built,and the trained signal detection and modulation recognition models are added to the system software,and the software interface display and other functions are also completed.From the test results,the system software achieves good performance of detection and identification. |
