A Unified Backbone--Expert Framework with Relation-Token and Residual--Classifier Interfaces for Automatic Modulation Recognition

Zhixiang Deng; Houbiao Li; Zongyong Cui

cs.CVAug 15, 2026

A Unified Backbone--Expert Framework with Relation-Token and Residual--Classifier Interfaces for Automatic Modulation Recognition

Authors: Zhixiang Deng, Houbiao Li, Zongyong Cui

Organizations: School of Mathematical Sciences, University of Electronic Science and Technology of China, Chengdu, 611731, China · School of Information and Communication Engineering, University of Electronic Science and Technology of China, Chengdu, 611731, China

Abstract

Automatic modulation recognition (AMR) faces distinct representation bottlenecks under varying observation lengths, where a single model architecture often fails to excel. To address this, we propose a unified backbone-expert framework with a common convolutional state-space backbone and two specialized interfaces. For short sequences, we inject explicit lag-aware complex-plane descriptors as relation tokens before encoding to compensate for information loss. For long sequences, we design a gated multi-scale residual refinement module to correct the feature map, combined with a fixed-averaging classifier collaboration to harness complementary evidence. Our framework achieves overall average accuracies of 67.28 \pm 0.14% on RML2016.10b and 87.19 \pm 0.77% on HisarMod2019 (mean \pm sample standard deviation over three runs), respectively. The framework's efficacy is further validated through three-seed ablations, native-length cross-configuration tests, and controlled window studies, confirming the benefit of expert-interface decoupling over one-size-fits-all architectures.

Figures & tables

Figure 1: Illustration of short and long observation budgets using the same synthetic QPSK signal realization. The 128-sample observation is a contiguous window of the 1024-sample observation. The shorter window contains signal evidence over limited temporal support, whereas the longer window retains evidence distributed over a broader temporal range.

Figure 2: Overall backbone–expert framework. The short and long configurations use the same feature-extraction and state-space topology but independently trained weights. They differ in expert inductive bias and integration interface.

Figure 3: Short-sequence relation-aware token augmentation. Base signal descriptors and multi-lag distance and complex-correlation relations are locally projected into relation tokens, multiplied by a bounded learned scale, and added to the main tokens before state-space encoding.

Figure 4: Long-sequence residual–classifier organization. Multi-scale gated residual refinement corrects the common feature map before patch embedding. The state-space and CNN branches then form separate logits that are combined by fixed averaging.

Dataset	Length	Classes	SNR range	Split (train:validation:test)
RML2016.10b	128	10	$-20{:}2{:}18$ dB	6:2:2
HisarMod2019	1024	26	$-20{:}2{:}18$ dB	8:2:5

Table 1: Dataset and evaluation protocols.

Method	RML2016.10b OAA	HisarMod2019 OAA	RML2016.10b protocol	HisarMod2019 protocol
STF-GCN [ 16 ]	66.04	–	6:2:2	–
IQFormer [ 10 ]	65.65	76.32	6:2:2	8:2:5
CPPCNet [ 11 ]	66.38	83.50	8:2 train/val	8:2 train/val
BioLAMR [ 12 ]	$67.43\pm 0.27$	–	8:1:1	–
MCANet [ 20 ]	66.53	76.58	7:2:1	7:2:1
MAWDN [ 21 ]	–	74.40	–	8:2:5

Table 2: Reported OAA (%), proposed SNR-stratified results, and dataset-specific protocols. Ratios use train:validation:test order unless marked as train/validation or train/test; a dash denotes an unreported result.

Variant	OAA	$\mathrm{SNR}\leq-10$	$\mathrm{SNR}<0$	$\mathrm{SNR}\geq 0$	$\Delta$ OAA
Full short-observation configuration	$67.279\pm 0.135$	$23.725\pm 0.602$	$41.121\pm 0.267$	$93.436\pm 0.033$	–
w/o correlation features	$66.859\pm 1.180$	$21.944\pm 3.164$	$40.267\pm 2.330$	$93.450\pm 0.032$	$-0.420$
w/o distance features	$65.224\pm 0.245$	$17.526\pm 0.067$	$36.968\pm 0.452$	$93.480\pm 0.044$	$-2.054$
Relation gate $=0$	$65.250\pm 0.223$	$17.398\pm 0.069$	$37.054\pm 0.388$	$93.446\pm 0.071$	$-2.029$
Late logit fusion	$65.749\pm 0.163$	$20.089\pm 0.202$	$38.304\pm 0.333$	$93.193\pm 0.060$	$-1.530$

Table 3: Short-configuration ablation on RML2016.10b (%, mean

\pm

sample standard deviation over seeds 2051–2053).

Variant	OAA	$\mathrm{SNR}\leq-10$	$\mathrm{SNR}<0$	$\mathrm{SNR}\geq 0$	$\Delta$ OAA
Equal-weight logit averaging (formal)	$87.186\pm 0.773$	$71.126\pm 1.710$	$75.577\pm 1.484$	$98.794\pm 0.062$	–
w/o residual refinement	$79.893\pm 0.365$	$55.963\pm 0.569$	$61.924\pm 0.547$	$97.862\pm 0.197$	$-7.293$
Learned logit gate	$86.963\pm 0.622$	$70.640\pm 1.316$	$75.131\pm 1.213$	$98.795\pm 0.076$	$-0.223$
Feature concatenation	$85.186\pm 0.258$	$67.209\pm 0.686$	$71.932\pm 0.523$	$98.440\pm 0.017$	$-2.000$
CNN classifier only	$86.013\pm 0.319$	$68.801\pm 0.811$	$73.410\pm 0.670$	$98.617\pm 0.039$	$-1.173$
Sequence classifier only	$76.566\pm 0.119$	$56.938\pm 0.040$	$60.216\pm 0.039$	$92.915\pm 0.215$	$-10.620$

Table 4: Long-configuration expert and interface analysis on HisarMod2019 (%, mean

\pm

sample standard deviation over seeds 2028–2030).

Figure 5: Per-SNR accuracy of representative short- and long-configuration ablations. Curves show the mean over three seeds, shaded bands denote one sample standard deviation, and the gray background marks negative SNRs.

Region	Residual coefficient	Within-run SD	Relative correction norm
All SNRs	$0.642\pm 0.033$	$0.157\pm 0.017$	$1.194\pm 0.096$
$\mathrm{SNR}<0$	$0.678\pm 0.030$	$0.132\pm 0.020$	$1.283\pm 0.094$
$\mathrm{SNR}\geq 0$	$0.606\pm 0.035$	$0.171\pm 0.017$	$1.106\pm 0.097$

Table 5: Residual-interface diagnostics on the HisarMod2019 official test set. The coefficient is bounded above by 0.85. “Within-run SD” measures sample-to-sample coefficient variation within each checkpoint.

Region	Sequence top-1	CNN top-1	Fused top-1	Disagreement	Fusion-only correct	Margin correlation
All SNRs	$60.537\pm 0.718$	$38.590\pm 1.370$	$87.186\pm 0.773$	$84.238\pm 0.537$	$6.524\pm 0.287$	$-0.439\pm 0.015$
$\mathrm{SNR}<0$	$55.175\pm 0.330$	$24.382\pm 0.886$	$75.577\pm 1.484$	$91.468\pm 0.103$	$9.564\pm 0.427$	$-0.324\pm 0.021$
$\mathrm{SNR}\geq 0$	$65.900\pm 1.360$	$52.798\pm 1.858$	$98.794\pm 0.062$	$77.007\pm 0.973$	$3.485\pm 0.189$	$-0.676\pm 0.004$

Table 6: Score-level complementarity of the jointly trained sequence and CNN classifiers on HisarMod2019 (%, except correlation; mean

\pm

sample standard deviation over seeds 2028–2030).

Dataset	Native configuration	Cross-applied configuration	Native advantage	Exploratory 95% interval
RML2016.10b ( $L=128$ )	Short: $67.279\pm 0.135$	Long: $65.326\pm 0.067$	$+1.952$	$[1.726,2.179]$
HisarMod2019 ( $L=1024$ )	Long: $87.186\pm 0.773$	Short: $67.023\pm 1.393$	$+20.163$	$[16.074,24.252]$

Table 7: Native and cross-applied configurations at the original observation length of each dataset (OAA in %, mean

\pm

sample standard deviation). Intervals are computed from three same-seed OAA differences.

Figure 6: Native and cross-applied configuration performance at the original observation length of RML2016.10b and HisarMod2019. Bars and error bars denote mean OAA and one sample standard deviation over three seeds, respectively; markers indicate individual seeds. Numbers above the brackets report the native-configuration advantage in percentage points.

Length	Short	Long	Long–Short
128	$61.783\pm 0.952$	$71.527\pm 0.233$	$9.744\pm 0.719$
256	$65.090\pm 0.646$	$76.970\pm 0.228$	$11.881\pm 0.482$
512	$67.433\pm 0.310$	$81.679\pm 0.527$	$14.247\pm 0.225$
1024	$67.023\pm 1.393$	$87.186\pm 0.773$	$20.163\pm 1.646$

Table 8: Controlled observation-window comparison on HisarMod2019 (OAA in %, mean

\pm

sample standard deviation over seeds 2028–2030).

Short configuration (seeds 2051, 2052, 2053)
Comparison	Seed 1	Seed 2	Seed 3	Mean $\Delta$ OAA
Full $-$ w/o correlation	$-0.135$	$-0.242$	$+1.637$	$+0.420$
Full $-$ w/o distance	$+2.300$	$+1.795$	$+2.067$	$+2.054$
Full $-$ relation gate $=0$	$+2.292$	$+1.795$	$+1.999$	$+2.029$
Full $-$ late fusion	$+1.487$	$+1.557$	$+1.546$	$+1.530$
Long configuration (seeds 2028, 2029, 2030)

Table 9: Same-seed paired OAA differences (percentage points) for the short and long configurations. Positive values favor the corresponding formal configuration.

Configuration	Length	Parameters (M)	MACs (M)	Approx. FLOPs (M)	Latency (ms)	Peak memory (MiB)	OAA (%)
Short-observation configuration	128	0.330	11.13	22.25	2.85	10.51	$67.279\pm 0.135$
Long-observation configuration	1024	0.530	103.08	206.16	2.37	12.44	$87.186\pm 0.773$

Table 10: Active-configuration complexity and batch-1 inference cost. FLOPs use the approximation

1\ \mathrm{MAC}\approx 2

scalar FLOPs.

Automatic Modulation Classification

Automatic modulation classification identifies radio signal modulation schemes from received waveforms, often under noise, fading, and limited labeled data.

See all 18 papers

Geometry-Aware Contrastive Learning for Few-Shot Automatic Modulation Recognition
May 26, 2026Guanqun Zhao, Yitong Liu, Jiaxuan Fang +2Contrastive Learning Self-Supervised Learning
Beijing University of Posts and Telecommunications, Beijing, China
Cross-Validated Cross-Channel Self-Attention and Denoising for Automatic Modulation Classification
Apr 11, 2026Prakash Suman, Yanzhen QuAutomatic Modulation Classification
Computer Science, Engineering, and Technology, Colorado Technical University, 1575 Garden of the Gods Road, Colorado Springs, 80907, CO, USA.
Modulation Consistency-based Contrastive Learning for Self-Supervised Automatic Modulation Classification
May 12, 2026Chenxu Wang, Shuang Wang, Lirong Han +4Contrastive Learning Self-Supervised Learning
School of Artificial Intelligence, Xidian University, Xi’an 710071, China · Unmanned System Research Institute, Northwestern Polytechnical University, Xi’an 710072, China · School of Electronics And Information, Northwestern Polytechnical University, Xi’an 710072, China

May 26, 2026J/K move · Enter open · S save

Geometry-Aware Contrastive Learning for Few-Shot Automatic Modulation Recognition

Guanqun Zhao, Yitong Liu, Jiaxuan Fang, Yufei Mao+1

Beijing University of Posts and Telecommunications, Beijing, China

Contrastive Learning Self-Supervised Learning Automatic Modulation Classification Few-Shot Learning

Open paper AskPDF arXiv

Standard Self-Supervised Learning (SSL) for Automatic Modulation Recognition (AMR) struggles with ineffective isotropic augmentations, spectral instability, and semantic drift. To address these challenges, we propose Dynamic-Consistency Contrastive Learning (DyCo-CL), a geometry-aware framework that couples Virtual Adversarial Augmentation (VAA) with a semantic consistency loss. We provide a theoretical analysis indicating that this strategy acts as an implicit spectral regularizer for the encoder, enabling stable manifold exploration. Complementing this, our Signal-Adaptive Swin Backbone with fixed-window attention improves structural stability by constraining attention locality, while a Hybrid Knowledge Fusion module anchors representations with physical priors. Experiments on RML benchmarks show that DyCo-CL achieves a 6.27% accuracy gain in 1-shot settings over prior methods.

Abstract

Automatic Modulation Classification

Explore similar work

Geometry-Aware Contrastive Learning for Few-Shot Automatic Modulation Recognition

Cross-Validated Cross-Channel Self-Attention and Denoising for Automatic Modulation Classification

Modulation Consistency-based Contrastive Learning for Self-Supervised Automatic Modulation Classification