PaperlayerView exact revision on arXiv

Source-linked AI summary

Sub-6 GHz Over-the-Air AMC via Curriculum Fine-Tuned CNN-Transformers

Nurettin Safak, Muhammet Sefa Demirel, Alperen Marasli, Taha Eren Atmaca, Durdu Can Yerdeyatar, Ozgun Ersoy

arXiv:2609.07726v1eess.SPcs.LG

TL;DR

AMC models are often evaluated on synthetic or cabled data, leaving limited evidence about performance on genuine free-space links with path loss and antenna pointing error. This study fine-tunes a hybrid CNN-Transformer from a controlled 915 MHz base dataset to a real 4 GHz link, where matched-distance accuracy stays within 91.8–93.7% across five conditions; the sequential order also limits what can be concluded about misalignment.

  • Problem

    AMC results are commonly based on synthetic or cabled/simulated channels, leaving limited controlled evidence for performance on GHz-range free-space links with path loss and antenna pointing error.

    Source: §I · p. 1

  • Method

    The study trains a hybrid CNN-Transformer on a controlled 915 MHz, all-32-class base dataset and sequentially fine-tunes it at five real 4 GHz distances.

    Source: §I · p. 1, B. Multi-Task Hybrid CNN-Transformer… · p. 2, §IV · p. 3

  • Results

    91.8–93.7% matched-distance test accuracy was reported across all five 4 GHz conditions, while confusion analysis found QPSK/QAM4 and DSSS/BPSK confusions at comparable magnitude in every condition.

    Source: §VI · p. 7

  • Takeaways & Limitations

    The reported curriculum kept the model well-matched to the free-space link across the tested conditions, with the identified confusions suggesting class-taxonomy rather than channel properties.

    Source: §VI · p. 7

    Because partially misaligned Stage 2 checkpoints follow and initialize from Stage 1 checkpoints, the experiment cannot isolate alignment effects from cumulative fine-tuning exposure.

    Source: A. Matched-Distance Accuracy · p. 5

Abstract

from arXiv · show

Automatic modulation classification (AMC) models are frequently trained and validated on synthetic or channel-cabled data, leaving open the question of how they behave once path loss and antenna pointing error are introduced by a genuine free-space link. We report a curriculum fine-tuning study of a hybrid CNN-Transformer AMC model. The general-purpose, all-32-class dataset underlying the model was built entirely at 915 MHz, on a controlled, clock/PPS-synchronized MIMO-expansion-cable link (not spatial-multiplexing transmission); all subsequent free-space, real-hardware experimentation - the sequential fine-tuning curriculum, matched-distance evaluation, and every reported over-the-air accuracy figure - was carried out at 4 GHz, across five directional-antenna distances (25, 35, 50, 70, 75 cm) under fixed TX/RX gain. Three distances used near-ideal antenna alignment (~99%) and two used a deliberately introduced partial misalignment (~85%), fine-tuned last in the curriculum. We report matched-distance test accuracy (91.8-93.7% across all five 4 GHz conditions) and confusion-matrix analysis grounded in RF theory, and report honestly where our sequential fine-tuning order confounds cumulative link adaptation with antenna alignment, rather than overstating what the data can support.

I. INTRODUCTION

The paper addresses the limited evidence on AMC performance over genuine free-space links with path loss and antenna pointing error. It adapts a hybrid CNN-Transformer model from a controlled reference link to a real 4 GHz link and reports the scope and limits of that transition.

I. INTRODUCTION

  • Reported AMC results largely use synthetic or cabled/simulated channels, while controlled transitions to GHz-range free-space links with antenna pointing error remain uncommon.The introduction notes prior evidence of degradation under real over-the-air impairments.Source: §I · p. 1
  • The study trains its 32-class base model at 915 MHz, then fine-tunes it over a real 4 GHz free-space link at five distances, including two with deliberate antenna misalignment.Only the temporal-modeling stage is updated while low-level RF feature extraction remains fixed; matched-distance performance and design limits are reported.Source: §I · p. 1

A. Motivation for a Hybrid CNN-Transformer Design

  • The hybrid design pairs convolutional extraction of local amplitude/phase transitions with Transformer modeling of longer-range temporal dependencies, including FHSS hopping patterns.A convolutional stack has a limited effective receptive field, while a Transformer applied directly to 1024 raw samples is computationally costly and lacks local-signal inductive bias.Source: A. Motivation for a Hybrid… · p. 1
  • Related hybrid convolutional-Transformer AMC designs combine CNN features with attention-based or graph-based long-range modeling, while this study evaluates its design on a physically degraded OTA link.The cited related designs are,.Source: A. Motivation for a Hybrid… · p. 2

B. Multi-Task Hybrid CNN-Transformer Architecture (“Plan A”)

Plan A combines a downsampling 1D-ResNet stem with a Transformer encoder and five task-specific heads. The model jointly predicts modulation and four auxiliary signal properties, with task losses weighted through learnable uncertainty parameters.

B. Multi-Task Hybrid CNN-Transformer Architecture (“Plan A”)

  • The model downsamples raw (2, 1024) I/Q input to 64 time steps with a four-block 1D-ResNet stem, then uses a three-layer Transformer and mean pooling before five task heads.The heads predict 32-class modulation, bandwidth, spread-spectrum type, analog/digital status, and coarse modulation family.Source: B. Multi-Task Hybrid CNN-Transformer… · p. 2
  • The family head predicts coarse modulation-family membership, added to regularize shared features toward distinguishing families behind common within-family confusions.The passage gives examples including adjacent PSK, APSK, QAM, and ASK orders.Source: B. Multi-Task Hybrid CNN-Transformer… · p. 2
  • Learnable task log-variances combine the five losses through homoscedastic uncertainty weighting, automatically down-weighting noisier or harder-to-fit tasks rather than using manual weights.The authors retain the original formulation for simplicity while identifying its limitations as a future refinement direction.Source: B. Multi-Task Hybrid CNN-Transformer… · p. 2

III. EXPERIMENTAL SETUP

The experimental bench uses two USRP N-series radios connected by cable for Stage 0 or operating over the air in later stages. The Plan A architecture shown alongside it has a ResNet-Transformer pipeline and five task-specific heads.

III. EXPERIMENTAL SETUP

  • Two USRP N-series radios face each other, connected by a MIMO expansion cable in Stage 0 or operating purely over the air in Stages 1–2.The cable is used for clock/PPS synchronization rather than multi-antenna transmission.Source: §III · p. 2
  • The setup distinguishes a cable-connected base-training stage from later cable-disconnected over-the-air stages.The cable link is the Stage 0 configuration; Stages 1–2 use free-space transmission.Source: §III · p. 2
  • Plan A maps I/Q input through a four-block 1D-ResNet stem, Transformer encoder, and mean pooling to five task-specific heads.The architecture contains 793,265 parameters.Source: Figure 1 · p. 2

A. Stage 0: 915 MHz Cable-Synchronized Base Training

Stage 0 builds the all-32-class base model on a controlled 915 MHz, clock/PPS-synchronized cable link. A separate longer training run peaked at 94.59% validation accuracy at epoch 34 and showed no further improvement through epoch 95.

A. Stage 0: 915 MHz Cable-Synchronized Base Training

  • The 915 MHz base dataset covered all 32 modulation classes on a clock/PPS-synchronized MIMO expansion cable link that remained single-TX/single-RX.The cable supplied synchronization and did not indicate spatial-multiplexing transmission.Source: A. Stage 0: 915 MHz Cable-Synchronized… · p. 2, Figure 2 · p. 3
  • The dataset comprised approximately 1.85 million training windows, with TX gain swept from 0–10 dB across SNR labels and RX gain held at 5 dB.The base model was trained for 30 epochs.Source: A. Stage 0: 915 MHz Cable-Synchronized… · p. 2
  • The cable-synchronized 915 MHz stage is distinct from the later 4 GHz directional-antenna experiments conducted at five distances.The later setup includes three near-ideal alignments and two partial alignments.Source: Figure 2 · p. 3, Figure 3 · p. 3
  • 94.59% validation accuracy was reached at epoch 34, with no further improvement observed through epoch 95 in the separate longer run.The base model was trained for 30 epochs; the longer run was stopped after epoch 95.Source: A. Stage 0: 915 MHz Cable-Synchronized… · pp. 2–3

B. Stages 1–2: 4 GHz Free-Space Adaptation

The 4 GHz free-space stages use fixed TX/RX gain and directional antennas across five distances. Three distances use near-ideal alignment, while two use deliberately partial alignment.

Fixed-gain free-space link

  • All five formal distances were tested at 4 GHz with TX gain fixed at 31.5 dB and RX gain at 15 dB.Distance and alignment, rather than a TX-power-based SNR ladder, varied across these conditions.Source: B. Stages 1–2: 4 GHz Free-Space… · p. 3

Distance and alignment conditions

  • The formal evaluation covered 25, 35, 50, 70, and 75 cm; a preliminary 105 cm range-finding link was excluded.The formal sweep was bounded at 75 cm to control alignment variables and support reliable performance attribution.Source: B. Stages 1–2: 4 GHz Free-Space… · p. 3
  • At 25, 50, and 75 cm, alignment was approximately 99%; at 35 and 70 cm, it was deliberately reduced to approximately 85%.The partial alignment was intended to emulate pointing or tracking error.Source: B. Stages 1–2: 4 GHz Free-Space… · p. 3

IV. TRAINING STRATEGY: SEQUENTIAL CURRICULUM FINE-TUNING

The model is fine-tuned in sequence from a 915 MHz checkpoint through well-aligned and then partially aligned 4 GHz conditions. The study evaluates each checkpoint at its own distance while noting that distance, alignment, and cumulative fine-tuning exposure are not independently controlled.

Sequential curriculum

  • Fine-tuning proceeds from Stage 0 through 25, 50, and 75 cm in Stage 1, then 35 and 70 cm in Stage 2, with each checkpoint initializing the next.The order moves from well-aligned to partially misaligned conditions.Source: §IV · p. 3, Figure 4 · p. 4

Fine-tuning procedure

  • At each step, the 1D-ResNet stem is frozen while the Transformer encoder and five task heads are updated.Each step uses five epochs of AdamW at η = 10^-5 with cosine annealing.Source: §IV · p. 3

Checkpoint selection and evaluation

  • A checkpoint is retained only if held-out validation accuracy improves over its immediate predecessor, then evaluated on its own distance’s held-out test split.The matched-distance evaluation reports accuracy, macro-F1, and the full 32-class confusion matrix.Source: §IV · pp. 3–4

Study design caveat

  • Distance, alignment condition, and cumulative fine-tuning exposure are not independently controlled variables in this curriculum.Stage 2 follows the three Stage 1 distances in the sequential order.Source: §IV · pp. 3–4

A. Matched-Distance Accuracy

Matched-distance test accuracy stays within a narrow band across the five 4 GHz conditions, including the final partially aligned conditions. The reported results do not isolate the effect of alignment from cumulative fine-tuning order.

Matched-distance accuracy

Accuracy by distance and stage

  • Stage 1 accuracy was 92.23%, 92.86%, and 91.81% at 25, 50, and 75 cm; Stage 2 scored 93.67% at 35 cm and 93.11% at 70 cm.Accuracy did not decrease monotonically with distance, and both Stage 2 values were numerically higher than every Stage 1 value.Source: A. Matched-Distance Accuracy · p. 5

Interpretation caveat

  • The higher Stage 2 values cannot establish that partial misalignment improves or does not harm accuracy, because alignment is confounded with cumulative fine-tuning exposure.The authors describe the overall range as consistent with, but not proof of, robustness to the Stage 2 alignment errors.Source: A. Matched-Distance Accuracy · p. 5

B. Confusion Analysis

The dominant classification errors reflect ambiguities in the modulation labels and signal structure, while real I/Q constellations show how free-space degradation brings QPSK and QAM4 into physical overlap.

B. Confusion Analysis

  • DSSS was also frequently confused with BPSK: 2922 of 6000 windows at 25 cm, consistent with BPSK-chip modulation and the absence of explicit despreading.The passage says both QPSK/QAM4 and DSSS/BPSK confusions appeared at comparable rates across distances, linking them to taxonomy and 1024-sample windows rather than channel conditions.Source: B. Confusion Analysis · p. 5
  • At 25 cm, 4282 of 6000 QPSK windows were predicted as QAM4 versus 1714 correct, and this confusion remained dominant across all five conditions.The passage attributes this to QPSK and 4-QAM sharing a four-point constellation under different modulation-family naming conventions.Source: B. Confusion Analysis · p. 5, Figure 7 · p. 6
  • At 70 cm, the captured QPSK and QAM4 constellations overlap in the I/Q plane, grounding their confusion in physical channel degradation rather than classifier failure.The figure traces broadening and rotation from the 25 cm, ≈99%-aligned link to the 70 cm, ≈85%-aligned link, alongside a tight, separated 915 MHz cable-synchronized baseline.Source: B. Confusion Analysis · pp. 5–6

C. Ablation Studies

Three ablations examine the Transformer encoder, auxiliary-task supervision, and freezing the CNN stem. They show accuracy gains for the hybrid model and multi-task supervision, while stem freezing trades a small accuracy cost for fewer updated parameters.

Transformer contribution

  • 94.60% versus 92.32% best validation modulation accuracy: the full hybrid model exceeded the ResNet-only baseline by 2.28 points.The comparison used the same Stage 0 base-training protocol; Table II summarizes the Transformer contribution.Source: Table II · p. 6, C. Ablation Studies · p. 6

Multi-task vs. single-task supervision

  • 92.74% versus 92.19% overall matched-distance accuracy: multi-task supervision led at four of five distances, while single-task scored 92.20% versus 91.81% at 75 cm.The authors interpret auxiliary tasks as a regularizing signal on the shared representation, while noting the effect was not uniform.Source: C. Ablation Studies · pp. 6–7, Table III · p. 7

Freeze-stem trade-off

  • 93.09% versus 92.74% average matched-distance accuracy: updating the full network scored marginally higher, while stem freezing reduced parameters updated per step by ∼44%.The authors retain stem freezing for constrained onboard hardware, despite its <0.4-point average accuracy cost, and report that it does not improve accuracy in this setting.Source: Table IV · p. 7, C. Ablation Studies · p. 7

VI. CONCLUSION

The study transitions a hybrid CNN-Transformer AMC model from controlled 915 MHz cable-synchronized training to a 4 GHz free-space link. Matched-distance accuracy remains within 91.8–93.7% across five conditions, but the sequential curriculum does not isolate antenna-misalignment effects.

VI. CONCLUSION

  • 91.8–93.7% matched-distance test accuracy was maintained across all five 4 GHz conditions, including two partially misaligned distances fine-tuned last.The base dataset was collected at 915 MHz; the free-space curriculum and evaluation were conducted at 4 GHz.Source: §VI · p. 7
  • QPSK/QAM4 constellation ambiguity and DSSS/BPSK chip-level similarity were consistent error modes across conditions, attributed to class taxonomy rather than channel differences.The reported ablations found a 2.28-point validation-accuracy loss without the Transformer and multi-task supervision outperformed single-task supervision at four of five distances.Source: §VI · p. 7
  • Freezing the CNN stem during fine-tuning did not improve accuracy over full-network updating, but reduced updated parameters per fine-tuning step by ∼44%, retained as a practical trade-off for resource-constrained onboard deployment.Source: §VI · p. 7
  • Because Stage 2 followed Stage 1, the curriculum confounds alignment with cumulative adaptation; the authors propose non-sequential fine-tuning and a distance sweep beyond 75 cm.They therefore make no claim that the experiment isolates antenna-misalignment robustness.Source: §VI · p. 7
Loading 2609.07726v1…