Automatic modulation recognition based on sample-transferable and branch-scalable method for signals in complex multipath channel

Abstract

At present, there are a large number of mature deep learning related studies on automatic modulation recognition (AMR) for signals in the additive white Gaussian noise (AWGN) or fixed multipath channel. However, in actual communication environments, the AMR method is required to have strong generalization ability due to the complexity and variability of multipath channels. Thus, we propose a sample-transferable and branch-scalable method suitable for signals in different multipath channels. According to the generation principle of multipath signals, we first estimate the multipath signals based on the direction of arrival (DOA) estimation algorithm to obtain characteristic parameters such as the number of paths and the direction of arrival. Then we decompose the multipath signals into multi-branch single-path signals using the estimation results. On this basis, we propose a multi-branch neural network trained with signals in the AWGN channel, with the decomposed multi-branch single-path signals serving as inputs. Hence, sample transfer from the training signals in the AWGN channel to the test signals in the multipath channel can be realized, significantly improving the generalization ability of the network. Moreover, we introduce the attention mechanism module to perform feature-level fusion on multi-branch signals, and use multipath signals to obtain additional recognition gain compared to single-path signals. In response to the uncertainty of multipath number in complex multipath channel environments, we propose a branch-scalable dynamic neural network (BSDNN) with novel “dual-branch training, multi-branch recognition”, and realize the recognition of multipath signals with arbitrary path number using the network structure trained with dual-branch signals. The experimental results show that our proposed BSDNN trained with the dual-branch signals in the AWGN channel can successfully transfer to modulation recognition of multipath signals with any number of paths. Furthermore, the method exhibits advantages in terms of lightweight design, with fewer network parameters and training time.