MicroSSM: A Lightweight State-Space Model for Automatic Modulation Classification

Automatic modulation classification (AMC) is a core task in cognitive radio and spectrum monitoring. Mainstream deep-learning-based AMC models typically contain hundreds of thousands to millions of parameters, making them difficult to deploy on resource-constrained edge devices. This work proposes MicroSSM, a lightweight state-space model for automatic modulation classification. MicroSSM decomposes signal processing into three functionally independent stages: a learnable soft-thresholding module performs adaptive denoising at the input stage; a four-layer small-kernel one-dimensional convolutional neural network (1D CNN) extracts local time–frequency features with a minimal parameter count; and a hierarchical state-space dual-channel module (HSMSSD) performs global temporal modeling with linear complexity via a softmax-normalized discrete state selection mechanism. Unlike Mamba-family models, which rely on continuous-time discretization, HSMSSD selects states through this softmax-normalized discrete choice, thereby avoiding the mapping overhead of continuous-time parameterization on short sequences. The capacities of the three modules are jointly regulated by a unified scaling factor, enabling model variants spanning two orders of magnitude in parameter count. On RML2016.10B, the 2.3K-parameter variant achieves 61.63% accuracy, and scaling to 133.5K parameters raises it to 64.13%; consistent results are observed on RML2016.10A, while a 2D grid extension of MicroSSM reaches 87.86% on the long-sequence, multi-class HisarMod2019.1 dataset. Ablation experiments confirm the effectiveness of each module and show that the marginal benefit of model scaling decays sharply: MicroSSM realizes its accuracy gain within the low-parameter window, and the larger variants are not cost-effective for edge deployment. Finally, MicroSSM is deployed on a Raspberry Pi 4B and achieves sub-millisecond per-sample inference latency over a real 433.92 MHz RF link, demonstrating the feasibility of the model on resource-constrained edge devices.