PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zhixiang Deng, Houbiao Li, Zongyong Cui

arXiv:2608.15160v1cs.CVcs.AI

TL;DR

Automatic modulation recognition faces different representation bottlenecks across short and long observation regimes, challenging one generic expert form. This paper pairs a common convolutional state-space backbone with regime-specific interfaces and achieves 67.28 ± 0.14% OAA on RML2016.10b and 87.19 ± 0.77% on HisarMod2019.

  • Problem

    Different observation regimes create distinct representation bottlenecks, motivating evaluation of whether one generic expert can address both effectively.

  • Method

    A common convolutional state-space backbone is paired with short-sequence relation-token augmentation and long-sequence residual refinement with decision-level classifier collaboration.

  • Results

    The framework achieves overall average accuracies of 67.28 ± 0.14% on RML2016.10b and 87.19 ± 0.77% on HisarMod2019.

  • Takeaways & Limitations

    Three-seed ablations, cross-configuration tests, and controlled windows support expert-interface decoupling across the evaluated benchmark conditions.

  • Takeaways & Limitations

    The controlled window study uses nested windows from fixed recordings, so it does not isolate a causal rule based on sequence length alone.

Abstract

from arXiv · show

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.

1. Introduction

The paper proposes a unified convolutional state-space backbone with two observation-regime-specific expert interfaces. Relation-aware token augmentation addresses short-sequence local-relation bottlenecks, while residual refinement and decision collaboration address long-sequence feature distortion and complementary evidence.

  • Motivation: Different observation budgets create distinct representation bottlenecks: short sequences limit recovery of informative local relations, whereas long sequences require feature refinement and complementary cues.These challenges motivate separate expert designs for short- and long-observation benchmarks.
  • Short-observation configuration: For RML2016.10b, lag-aware distance and complex-correlation descriptors provide explicit local complex-plane relations through early token augmentation before contextual encoding.The sequence encoder jointly contextualizes these relation tokens with learned backbone features.
  • Long-observation configuration: For HisarMod2019, multi-scale residual refinement corrects long-receptive-field feature maps, while CNN–state-space decision collaboration combines complementary evidence.The framework is evaluated with three-seed ablations, SNR-stratified analysis, native-length cross-configuration experiments, and controlled windows.
  • Framework: The unified framework uses a common convolutional state-space topology to support two expert inductive biases and their corresponding integration interfaces.The shared backbone is paired with complementary expert configurations rather than a single one-size-fits-all design.

2. Related work

Prior AMR research spans learned convolutional and recurrent representations, attention, graph, and state-space models, while exploiting complex-signal structure, explicit relations, long-sequence processing, and interface-specific fusion. The present framework instead assigns experts and interfaces to benchmark-specific representation bottlenecks within one backbone template.

  • Learned representations: Deep AMR learns representations directly from signal samples using convolutional, recurrent, and convolutional–recurrent architectures, alongside broader reviews of architectures, representations, datasets, complexity, and deployment.These developments replace handcrafted decision statistics with learned features.
  • Contextual modeling: Attention, graph, and state-space methods improve contextual modeling through nonlocal dependencies, cross-scale token fusion, adaptive graph connectivity, and combined convolution–Mamba–Transformer designs.Examples include IQFormer, MST, STF-GCN, DWMTN, ConvMamba, and LM-GDMAF.
  • Complex-signal structure: Complex-aware methods preserve I/Q coupling through complex convolution, shared-parameter dual streams, or explicit relations derived from multidomain correlations and multimodal representations.The proposed short path instead uses local lagged distance and complex correlation at the token interface.
  • Long-observation modeling: Long-observation AMR methods combine denoising, wavelet decomposition, residual decision flows, multi-scale convolutions, recurrence, hierarchical attention, and Mamba-inspired attention.These approaches address restoration, receptive-field growth, and temporal-scale diversity through different hybrid designs.
  • Fusion and specialization: Existing fusion mechanisms place specialization at token embedding, gated multimodal fusion, prediction heads, learned SNR gates, or multi-resolution sequence processing.The proposed framework instead assigns a corresponding expert and interface to each benchmark-condition representation bottleneck within one backbone template.

3. Proposed method

The proposed method uses a common backbone with independently specialized short- and long-observation experts, differing in inductive bias and integration interface. The short path injects lag-aware relation tokens, while the long path performs multi-scale residual refinement and classifier collaboration.

  • Unified backbone–expert organization: A shared backbone template separates configuration-specific expert processing for short and long observations while preserving the same topology and feature dimensions.The configurations are independently trained and executed with different learned weights.
  • Short-observation expert: The relation channels are invariant to common phase rotation, although the complete network also receives raw I/Q samples.The distance and correlation channels provide explicit local complex-plane relations before representation learning.
  • Short-observation expert: Three-seed ablations show that removing the relation branch or distance channels mainly harms low-SNR recognition, while early token integration outperforms the corresponding late-logit interface.High-SNR accuracy changes little when these relation components are removed.
  • Long-observation expert: The long configuration applies three depthwise temporal branches with effective receptive fields of 15, 61, and 125 samples before patch embedding.The branches use kernel–dilation pairs (15, 1), (31, 2), and (63, 2), then concatenate and fuse their outputs.
  • Long-observation expert: The long interface averages separately formed sequence-classifier and CNN-expert logits, retaining complementary class evidence without sample-dependent fusion.Feature concatenation and learned logit gating are evaluated as alternative interfaces.

4. Experimental setup

Experiments evaluate short- and long-observation configurations on RML2016.10b and HisarMod2019 using matched convolutional state-space backbones, dataset-specific protocols, and three-seed reporting. Cross-configuration and controlled-window studies further test whether expert-interface benefits depend on observation length.

  • Datasets and evaluation protocols: Short and long configurations are evaluated on RML2016.10b and HisarMod2019, respectively, retaining each dataset’s native observation length and reported split protocol.RML2016.10b contains 10 simulated-I/Q modulation classes, while HisarMod2019 contains 26 classes generated under multiple fading conditions.
  • Implementation: Both configurations use a parameter-matched convolutional state-space backbone with width 96, two Mamba blocks of state dimension 24, and grouped patch embedding with kernel 16 and stride 8.The short configuration adds distance and complex-correlation relation channels, whereas the long configuration uses residual refinement.
  • Metrics and reporting: Evaluation reports overall average accuracy, mean accuracy for SNR < 0, mean accuracy for SNR ≥0, and descriptive SNR ≤−10 accuracy as mean ± sample standard deviation over three seeds.Same-seed OAA differences are treated as descriptive stability checks rather than formal population-level significance tests.
  • Cross-configuration evaluation: A 2 × 2 native-length cross-configuration evaluation compares each dataset’s designated configuration with the cross-applied alternative without discarding signal samples.Both short- and long-observation configurations are evaluated under each dataset’s native input length and original test protocol.
  • Controlled window study: Within HisarMod2019, nested center windows of 128, 256, 512, and 1024 samples assess how independently trained short and long configurations use increasing temporal support.No padding, resampling, or cross-record concatenation changes sample identity, labels, SNR, membership, or the train/validation split.

5. Results … 5.3. Long-sequence expert and interface analysis

The proposed framework achieves competitive protocol-matched AMR performance and benefits from decoupled short- and long-sequence expert interfaces. Ablations show relation descriptors primarily improve difficult short-sequence conditions, while residual refinement and classifier collaboration strengthen long-sequence recognition, especially under negative SNR.

  • 5.1. Comparison with literature-reported methods: These literature comparisons support competitive state-of-the-art performance under protocol-matched settings, not a protocol-independent ranking across published results.Competing values come from original publications, whereas proposed results average three predefined seeds.
  • 5.1. Comparison with literature-reported methods: Under matched protocols, the proposed short configuration achieves the best listed RML2016.10b OAA, while the long configuration leads HisarMod2019 methods reporting the same protocol.The HisarMod2019 result exceeds DAE-CNN-BiLSTM by 0.26 percentage points.
  • 5.2. Short-sequence ablation: Removing distance features causes a similar OAA loss of 2.054 points, indicating distance accounts for most of the measured relation-branch gain.The zero-gate control removes the complete relation-token update while preserving the backbone and main sequence classifier.
  • 5.2. Short-sequence ablation: The full short configuration improves OAA by 2.029 percentage points and SNR < 0 accuracy by 4.067 points over the zero-gate variant.SNR ≥0 accuracies differ by only 0.010 points.
  • 5.3. Long-sequence expert and interface analysis: Removing residual refinement lowers OAA by 7.293 points and SNR < 0 accuracy by 13.653 points despite retaining equal-weight decision collaboration.This confirms the importance of residual refinement in the long configuration.
  • 5.3. Long-sequence expert and interface analysis: Equal-weight averaging and learned gating achieve comparable accuracy, and both outperform feature concatenation and separately optimized single-branch controls.Because learned gating adds a sample-dependent fusion module without repeatable gain, fixed averaging is adopted.
  • 5.3. Long-sequence expert and interface analysis: Residual refinement provides its largest long-configuration advantage in negative-SNR conditions, while sequence-classifier collaboration supplies a smaller gain with the same concentration.Short-configuration relation-aware token augmentation and distance descriptors likewise gain mainly below 0 dB.

5.4. Long-configuration mechanism diagnostics · 5.5. Native-length cross-configuration comparison

Long-configuration diagnostics show that residual refinement is sample-dependent and stronger in difficult low-SNR conditions, while averaged classifier logits provide complementary evidence. Native-length comparisons favor the configuration matched to each dataset, with substantial dataset-dependent advantages.

  • 5.4. Long-configuration mechanism diagnostics: Residual coefficient averages and relative correction norms increase below 0 dB, indicating stronger residual intervention in the difficult SNR region.The coefficient is bounded, and reported statistics are mean ± sample standard deviation across three seed-level statistics.
  • 5.4. Long-configuration mechanism diagnostics: Nonzero within-run coefficient variation shows that the residual gate performs sample-dependent adjustment rather than acting as a fixed global multiplier.The coefficient distributions overlap across SNR regions, so the gate should not be interpreted as an explicit SNR estimator.
  • 5.4. Long-configuration mechanism diagnostics: The sequence and CNN branches exhibit high prediction disagreement and negative true-class-margin correlation, forming complementary score residuals.Margin correlation is computed between the branches’ true-class margins within the same jointly trained checkpoint.
  • 5.4. Long-configuration mechanism diagnostics: Averaged logits recover correct decisions when neither branch is individually top-1 correct, with the largest fusion-only recovery rate below 0 dB.The formal model supervises only the averaged logits, so its extracted branch accuracies are not equivalent to separately optimized single-branch controls.
  • 5.5. Native-length cross-configuration comparison: The native configuration wins all six same-seed comparisons at each dataset’s original observation length.Table 7 reports native and cross-applied configurations using OAA means and sample standard deviations over three seeds.
  • 5.5. Native-length cross-configuration comparison: 1.952 points is the short-configuration mean advantage over the cross-applied long configuration on RML2016.10b.This advantage is measured at the dataset’s original observation length.
  • 5.5. Native-length cross-configuration comparison: 20.163 points is the long-configuration mean advantage over the cross-applied short configuration on HisarMod2019.The larger advantage further supports matching configuration choice to dataset observation length.

5.6. Controlled observation-window comparison · 5.7. Paired stability analysis · 5.8. Accuracy–complexity trade-off

Controlled window studies show that the long configuration benefits more from additional temporal support, while paired analyses confirm consistent directional gains across matched seeds. The long path also uses substantially more computation but can achieve slightly lower measured GPU latency.

  • 5.6. Controlled observation-window comparison: 15.659 points: the long configuration improves from 128 to 1024 samples, versus 5.240 points for the short configuration.The short configuration saturates after 512 samples.
  • 5.6. Controlled observation-window comparison: The Long–Short advantage increases monotonically with window length for every seed.The comparison uses nested windows from the same HisarMod2019 observations, with differences computed seed by seed before aggregation.
  • 5.7. Paired stability analysis: The paired OAA differences expose run-to-run directions and magnitudes as a reproducibility-oriented stability check.Positive values favor the corresponding formal configuration, rather than establishing formal statistical significance.
  • 5.7. Paired stability analysis: Across three matched seeds, the short full configuration consistently outperforms the zero-gate, w/o-distance, and late-fusion variants.The paired analysis is descriptive and is not presented as a definitive significance test.
  • 5.7. Paired stability analysis: Across three matched seeds, the long full configuration consistently outperforms the w/o-residual, feature-concatenation, sequence-only, and both-experts-off variants.The two native-length cross-configuration comparisons retain the same direction for every seed.
  • 5.8. Accuracy–complexity trade-off: 9.3× more MACs: the long configuration requires approximately this increase because of its longer feature sequence and residual branches.FLOPs use the approximation 1 MAC ≈2 scalar FLOPs.
  • 5.8. Accuracy–complexity trade-off: Slightly lower measured batch-1 latency: the long configuration achieves this in the reported GPU setting despite its higher analytical MAC count.The measurements use float32 batch-1 forward passes on the specified RTX 4070 SUPER; hardware latency is not proportional to analytical MAC count.

6. Discussion

The discussion supports decoupling expert information and classifier interfaces according to observation length. It also identifies controlled-window and cross-paper limitations that constrain interpretation.

  • Short-configuration ablations: Explicit lagged distance provides the dominant measured gain in the short relation branch, while correlation remains complementary and pre-encoding insertion outperforms output decision combination.These ablations distinguish both the information contributed and where it enters the classifier.
  • Long-configuration ablations: Removing multi-scale residual refinement causes a clear loss below 0 dB, while the CNN-only control is stronger than the sequence-only control and the complete model exceeds both.Residual refinement corrects the feature map before tokenization, whereas CNN and sequence classifiers collaborate at the decision level.
  • Long-configuration ablations: The larger residual coefficient at negative SNRs and negatively correlated classifier margins support adaptive feature correction and score-level collaboration without a learned fusion gate.The long design separates feature correction from decision collaboration within jointly trained formal checkpoints.
  • Cross-configuration results: Across all seeds, the short configuration wins on RML2016.10b and the long configuration wins on HisarMod2019, while controlled windows show a stable within-source trend.The results support relation-token augmentation for limited observations and residual–classifier collaboration with extended support.
  • Limitations: Dataset differences confound direct comparisons, although controlled HisarMod2019 windows hold sample identity and labels fixed while varying retained temporal support.The windows derive from fixed 1024-sample recordings rather than independent acquisitions at several durations; reported literature results also limit strict cross-paper ranking.

7. Conclusion

The work introduces a unified backbone–expert framework for AMR representation bottlenecks across benchmark conditions. It pairs a common convolutional state-space backbone with specialized short- and long-configuration interfaces, achieving strong reported benchmark accuracies.

  • Framework: The framework addresses AMR representation bottlenecks under different benchmark conditions with a unified backbone–expert design.Its architecture uses a common backbone and complementary expert designs.
  • Framework: The common architecture uses a convolutional state-space backbone topology paired with two complementary expert designs and integration interfaces.The conclusion describes the backbone as shared across configurations.
  • Short configuration: The short configuration adds lag-aware distance and complex-correlation tokens before sequence encoding.These tokens are introduced as the short-configuration interface.
  • Long configuration: The long configuration combines long-receptive-field residual refinement with sequence–CNN decision collaboration.This interface is designed for the long configuration.
  • Benchmark results: 67.28 ± 0.14% OAA was achieved on RML2016.10b, while 87.19 ± 0.77% was achieved on HisarMod2019.The reported values are overall average accuracies for the two benchmarks.

CRediT authorship contribution statement

The paper assigns primary technical development and manuscript drafting to Zhixiang Deng, while Houbiao Li led conceptualization and supervision and Zongyong Cui contributed validation and review.

  • Zhixiang Deng handled methodology, software, investigation, validation, formal analysis, visualization, and writing the original draft.
  • Houbiao Li contributed conceptualization, supervision, and writing–review and editing.
  • Zongyong Cui contributed validation and writing–review and editing.

Data availability

The study uses publicly accessible RML2016.10b and HisarMod2019 datasets, while withholding dataset redistribution and promising code and experiment configurations upon publication.

  • Data availability: RML2016.10b and HisarMod2019 are available through the Kaggle and IEEE DataPort dataset pages, respectively.The authors do not redistribute the datasets.
  • Data availability: Code and experiment configurations will be made publicly available upon publication.
Loading 2608.15160v1…