Automatic Modulation Classification And Software Implementation Based On Signal Image Transform Depth Feature

Automatic modulation classification is the key technology of the non-cooperative communication.The main implementation methods can be divided into methods that based on the likelihood ratio and the feature extraction.However,there are shortcomings in the practical application.First,Modulation classification methods based on likelihood ratio are often computationally complex and not suitable for practical applications.While,the machine learning algorithms based on feature extraction rely on manual design and expert experience.Second,deep learning algorithms based on feature extraction can be further divided into methods based on original IQ signals and images transformation.In the former method,the waveform data is easily damaged by interference and noise,which makes the waveform regularity lost.So,it is not conducive to deep network learning any more,and its robustness and anti-noise performance are poor.Third,among the deep learning algorithms based on feature extraction,there is a method based on constellation diagrams.The original constellation point pixels are often sparse and scattered,and the feature expression ability is limited.Especially under the influence of noise,the visual effect of the constellation points is poor.In the meanwhile,the contrast between constellation points and background is not obvious enough.Therefore,the visual effect needs to be enhanced.Fourth,most modulation classification networks are not specially designed for the characteristics of the input signals,but are stacking network layer structures simpley.In view of the above problems,this paper starts with two image conversion methods(constellation diagram and eye diagram)for signal quality evaluation in communication.Then,we design a deep network to realize the classification and identification of the modulated signals.Finally,we transplant the algorithm into the application software to achieve engineering implementation.The main research contents and results are as follows:(1)An enhanced constellation diagram(ECDs)algorithm is proposed based on the signal image conversion method of constellation diagram,which can enhance the visual effect of the constellation diagram.It makes the constellation cluster more prominent in the visual contrast effect,and enhances its representation ability.Further,in order to extract the contour,size,edge and other multi-level features of the ECDs,a Convolution-FP-SE Network(CFSN)is designed to realize the multi-level features extraction and fusion of the ECDs.It improves the deep feature representation capabilities.We conduct experiments on the 23 types of modulation data sets with different frequency offset and signal-to-noise ratio combinations and the effectiveness of the enhanced constellation method is verified by comparing the effects before and after enhancement on the various networks.The comparison with other metheds verifies the effectiveness and robustness of the network and the Overall Accuracy of the CFSN achieves 98.17% on the dataset with no frequency bias and high signal-to-noise ratio.(2)The third-generation Spiking Neural Network(SNN)is introduced for modulation classification based on eye diagram signal image conversion method.The spiking neural network is modeled based on the characteristic mechanism of biological neurons and uses the RC charging and discharging structure.So,it is more interpretable and can be analyzed by circuits and mathematical tools.The SNN performs pulse encoding on the input eye diagram image and uses the Spatio-Temporal Backpropagation(STBP)algorithm for backpropagation training.The Overall Accuracy of the proposed method achieves 100% at 25 d B on a 10-category eye-diagram datasets with a signal-to-noise ratio in the range of [0:5:25]d B.At the same time compares favorably with.Besides,our method is superior to other eyediagram-based methods and the classification network models in each signal-to-noise ratio,which verifies the effectiveness and robustness of the SNN network.(3)We develop the intelligent modulation classification interface based on the PyQt programming architecture and MVC design pattern.PyQt has the characteristics of diverse library functions,convenient design tools,high design efficiency and cross-platform.At the same time,Model-View-Controller(MVC)has the characteristics of reduced degree of association between modules,high programming efficiency and code reuse.The advantages of the two are combined to achieve efficient and high-quality development and design.The designed intelligent modulation classification interface mainly includes a total of eight parts such as menu,setting area and various display areas,etc.It supports multiple input modes including the single file input,the batch file input and the test set file input.Also,it supports multiple input sources including constellation input,enhanced constellation diagram input and eye diagram input with.It matches the corresponding input with the corresponding identification steps later.Besides,the functions of the signal analysis,the folder monitoring,the continuous identification and other functions are expanded,which is more in line with the actual needs of the project,and further enhances the practicability of the software.Finally,we realize the construction of intelligent modulation classification software with complete functions,friendly interface,high efficiency and practicality.