Algorithm And Application Of Modulated Signal Classification Based On Edge Computing

Automatic identification of digital modulation signals refers to a technology that automatically identifies the type of signal modulation by analyzing the amplitude,frequency,and phase characteristics of the signal after receiving the digital modulation signal.At present,the automatic identification technology of digital modulation signals is mainly divided into two categories.One is to calculate the high-level statistical characteristics of the signal and set the appropriate thresholds to determine the type of signal modulation.The other is based on feature extraction.The idea is to extract the features of the signal by some methods to obtain the higher-order features of the modulation.Then,the classifier is used to complete the classification process,such as the deep learning,which essentially acts as the role of a feature extraction and classifier.In the research of this article,through the analysis of the current mainstream network models such as AlexNet,GoogLeNet,ResNet and other models,combined with digital modulation signal classification tasks to make some improvements,such as in AlexNet By replacing the maximum pooling layer with the Long-short Term Memory module,it can avoid the irreversible damage caused by the pooling layer,provide a better classification features for the classifier,increase the performance and robust of the classification process.It can also to enable edge device of classification models.In this paper,the network pruning method is regarded as a pameter combination optimization problem that has the least impact on the loss function.In order to simplify the calculation,the Taylor expansion at the clipping parameters is used to approximate the change of the loss function.Actually,The Taylor expansion is the sum of the product of the gradient value and the corresponding activation value at the clipping parameter.In terms of parameter importance evaluation,this article uses an evaluation method based on the importance of the channel,that means,the minimum granularity of the clipping unit is the feature channel.In this paper,the fixed parameters cutting method is improved,and the concept of accuracy redundancy ratio is proposed,and the pruning method is dynamically adjusted by using the accuracy redundancy ratio.This has a great effect on pruning for the purpose of preserving model performance.It fully balances the relationship of the accuracy and the complexity of the model.It can lead to approximate the global optimal point of the compression model,which has certain significance for the actual edge computing.At the end of this paper,we compared the forward inference time and parameters of the model between and after the pruning process.The results show that the model can greatly reduce the calculation time with acceptable accuracy loss.