High-Efficiency Modulation Classification With Temporal-Frequency Analysis Based on Multi-channel Filter Bank

Abstract

Efficiency is now a key challenge in automatic modulation classification (AMC), particularly in resource-constrained environments like mobile devices in 6G networks. This paper presents a framework based on the filter-bank channelizer (FBNet) for AMC, which gracefully strikes a balance between accuracy, complexity, and speed. By channelizing signals into simpler sub-band sequences, FBNet captures dependencies in both temporal and frequency dimensions, making modulation features more distinct. Therefore, we use a compact architecture with the custom-designed dilated convolution blocks (DCB) and the adaptive channel aggregation (ACA) module as the backbones. Experimental results demonstrate that, under the same resource constraints, our model consistently outperforms the leading models in accuracy. Further evaluation conducted on the Jetson Nano, an edge platform with computing power similar to a mobile device, reveals a 42.4% advantage in inference speed for our model, underscoring that FBNet holds great potential for deployment in resource-constrained systems.