Exploring Encoder-Decoder Transformer Structure for Signal Classification

Abstract

Automatic modulation classification (AMC) is one of the fundamental technologies in adaptive communication systems, supporting various tasks such as spectrum surveillance and cognitive radio. Recently, transformer-based architectures for AMC have been explored due to their strong sequence modeling capability. However, existing approaches have primarily relied on encoder-only architectures with limited focus on the decoder and thus leaving half of the transformer framework underutilized in AMC. To adress this, in this paper, we explore how the decoder can be exploited to fully utilize the transformer for AMC by comparing an encoder-only architecture with a full encoder–decoder architecture, where learnable vector parameters are injected as decoder inputs. To this end, we conduct simulations on various types of signals in a noisy channel. Simulation results show that incorporating the proposed encoder-decoder architecture can yield consistent performance improvements over the encoder-only counterpart, highlighting the potential of decoder-assisted designs for transformer-based AMC.