A coarse-to-fine modulation classification scheme utilizing Transformers and Liquid Neural Networks for Smart Grids
- 1Power Dispatching and Control Center, China Southern Power Grid (CSG), Guangzhou, Guangdong, China
- 2College of Electronics and Information Technology, South China University of Technology, Guangzhou, Guangdong, China
Abstract
Dependable data transmission is vital for modern Smart Grid operations, facilitating essential services like demand-side management and Advanced Metering Infrastructure (AMI). Within these networks, Automatic Modulation Recognition (AMR) acts as a crucial technique for blind signal processing and system security. However, the highly volatile nature of Power Line Communication (PLC) channels severely restricts AMR performance. Traditional statistical algorithms generally fail under severe noise conditions (low SNR), whereas standard single-stage deep learning architectures lack transparency and struggle to distinguish highly similar modulation schemes. To overcome these limitations, we introduce an innovative two-tier recognition architecture that integrates statistical feature engineering with cutting-edge neural networks. The preliminary stage executes coarse classification by feeding Higher-Order Cumulants (HOCs) into a Transformer module, effectively categorizing signals into general modulation families (e.g., PSK or QAM). Here, the Transformer’s self-attention mechanism excels at identifying complex correlations among statistical parameters. Subsequently, based on the preliminary result, a specialized Liquid Neural Network (LNN) is deployed for fine-grained analysis. By processing 3D constellation matrices, these specialized LNN models pinpoint the exact modulation order. Their inherent continuous-time properties provide improved adaptability to the temporal fluctuations and heavy noise typical of PLC environments, as evidenced by our low-SNR evaluations. Experimental evaluations confirm that the introduced hierarchical strategy yields markedly higher accuracy rates than traditional algorithms and unified deep learning baseline models, especially in high-noise scenarios. Ultimately, this structured design delivers a highly transparent and robust solution for AMR in challenging communication setups.
Related
- Coarse-graining metabolic networks via feature learning reveals cross-species growth predictors
- Computational alternative to cardiac biopsy in multiple myeloma: utilizing the shape of routine laboratory data for the classification of AL cardiac amyloidosis
- Co2ODE: Conditionally Coarse-Grained Model Discovery for Rewiring in Biochemical Networks
- Vision Transformers versus convolutional neural networks for fine-grained orchid genus identification in a species-rich, data-poor flora: a controlled benchmark on the Orchidaceae of New Guinea
- Transfer Learning-Driven Convolutional Neural Networks for Classification of Plant Species from Leaf Vein Patterns
BibTeX
@article{hu2026coarsetofine,
title = {{A coarse-to-fine modulation classification scheme utilizing Transformers and Liquid Neural Networks for Smart Grids}},
author = {FeiFei Hu and Yu Huang and Xubin Lin and Guanhuan Huang},
journal = {PLOS One},
year = {2026},
month = oct,
doi = {10.1371/journal.pone.0359526},
url = {https://doi.org/10.1371/journal.pone.0359526},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - A coarse-to-fine modulation classification scheme utilizing Transformers and Liquid Neural Networks for Smart Grids
AU - FeiFei Hu
AU - Yu Huang
AU - Xubin Lin
AU - Guanhuan Huang
PY - 2026
DA - 2026/10/26
JO - PLOS One
DO - 10.1371/journal.pone.0359526
UR - https://doi.org/10.1371/journal.pone.0359526
AB - Dependable data transmission is vital for modern Smart Grid operations, facilitating essential services like demand-side management and Advanced Metering Infrastructure (AMI). Within these networks, Automatic Modulation Recognition (AMR) acts as a crucial technique for blind signal processing and system security. However, the highly volatile nature of Power Line Communication (PLC) channels severely restricts AMR performance. Traditional statistical algorithms generally fail under severe noise conditions (low SNR), whereas standard single-stage deep learning architectures lack transparency and struggle to distinguish highly similar modulation schemes. To overcome these limitations, we introduce an innovative two-tier recognition architecture that integrates statistical feature engineering with cutting-edge neural networks. The preliminary stage executes coarse classification by feeding Higher-Order Cumulants (HOCs) into a Transformer module, effectively categorizing signals into general modulation families (e.g., PSK or QAM). Here, the Transformer’s self-attention mechanism excels at identifying complex correlations among statistical parameters. Subsequently, based on the preliminary result, a specialized Liquid Neural Network (LNN) is deployed for fine-grained analysis. By processing 3D constellation matrices, these specialized LNN models pinpoint the exact modulation order. Their inherent continuous-time properties provide improved adaptability to the temporal fluctuations and heavy noise typical of PLC environments, as evidenced by our low-SNR evaluations. Experimental evaluations confirm that the introduced hierarchical strategy yields markedly higher accuracy rates than traditional algorithms and unified deep learning baseline models, especially in high-noise scenarios. Ultimately, this structured design delivers a highly transparent and robust solution for AMR in challenging communication setups.
SN - 1932-6203
ER -