Automatic Modulation Open-Set Recognition Based on Random Convolutional Prototype Network-MultiOCSVM

Automatic Modulation Recognition (AMR) is a crucial technology in signal recognition for detecting target signals in radio communications. The accuracy of AMR has been improved by using machine and deep learning methods. However, these methods often have high computational complexity and lower accuracy in open set recognition. This paper proposes a modulation recognition approach based on the Random Convolution Prototype Network-MultiOCSVM to address the challenges mentioned. The method extracts modulation signal features effectively by establishing a random convolution prototype network using various random convolution kernels. It then performs prototype learning on the retrieved features within a fully connected neural network for dimension-reduced features and category prototype extraction. A Multiple One Classifier-Support Vector Machine (MultiOCSVM) is constructed to fit the dimension-reduced feature boundaries and establish a decision boundary. The outer and inner bounds represent unknown and known signals respectively, and the known signals are classified by combining category prototypes. The experimental results indicate that this approach is more efficient and accurate in open-set automatic modulation recognition than existing algorithms for automatic modulation recognition.