Pragmatic Decentralized Learning-based Automatic Modulation Classification Method

AMC technology is typically deployed in the receiver to achieve the functionality of identifying and automatically classifying the modulation of received signals.Classic automatic modulation classification methods include likelihood ratio-based methods and feature-based methods.Although the two methods are complete in theory,they have poor universality and robustness.CNN can automatically extract features from data and build classifiers.And hence,DL techniques represented by CNN were applied in AMC methods.However,existing DL-based AMC methods exist three problems.First,traditional CNN models have high complexity and large volume,making it difficult to deploy them on edge devices with limited resources.Second,existing lightweight DL models are unable to meet the requirements of signal classification in scenarios with high performance.Third,existing DL-based AMC technologies employ a centralized learning architecture,which is difficult to protect privacy security and has high training overhead.To address the above three problems,corresponding solutions are proposed.First,a lightweight CNN is designed by using group convolution instead of standard convolution in traditional CNN,removing some fully connected layers,and replacing the flattening layer with global pooling layer.Second,a pragmatic network is designed by considering the classification performance and lightweight of the network,implementing a compact network design based on modulation signal-specific features.Specifically,the phase parameter evaluation module,spatial feature extraction module,and temporal feature extraction module were used to extract the unique phase features,spatial features,and temporal features of signals,combined with multi-channel learning.Third,a decentralized learning framework was used to achieve decentralized model training,where the same global model was trained using local training and model weight sharing methods on edge devices.Experimental results demonstrate that the designed lightweight CNN sacrifices less than 2.5%of the classification performance compared to traditional standard CNN,but the model complexity decreases by over 99%.Compared with the state-of-the-art AMC technologies,the designed pragmatic network achieves the best performance on the public dataset Radio ML.2016 series dataset,while the model complexity is relatively low.Compared with traditional centralized learning,the adopted decentralized learning framework sacrifices less than 4% of the classification performance,achieving decentralized model training.In further conclusion,in the research of DL-based AMC technologies,using group convolution to design CNNs can achieve a lightweight network,which is suitable for scenarios with low classification performance requirements and limited hardware computing resources.Designing neural networks based on signal-specific features can achieve a trade-off between classification performance and lightweight.Decentralized learning can protect edge device privacy and alleviate the pressure on central devices.Deploying pragmatic network in a decentralized learning framework can effectively solve the problems of insufficient of computing power edge device,limited storage space,and high communication costs in actual deployment.