Source-linked AI summary
Efficient Automatic Modulation Classification for Next-Generation Wireless Networks
To Truong An, Antonios Argyriou, Annisa Anggun Puspitasari, Simon L. Cotton, Byung Moo Lee
TL;DR
6G networks motivate AMC algorithms that combine high classification accuracy with low inference time and computational cost. The paper proposes TDRNN, which adaptively denoises received signals before recurrent modulation classification. TDRNN reports improved accuracy, faster inference, and fewer parameters than comparison methods across RadioML datasets.
Problem
6G-oriented AMC requires high-accuracy modulation identification with low inference time and computational cost.
Method
TDRNN combines an adaptive Threshold Denoiser, whose threshold depends on modulation and noise conditions, with an RNN classifier.
Results
On RadioML 2016.10A, TDRNN improves accuracy over MCLDNN by 4.98%, runs 23 times faster, and requires ten times fewer parameters.
Takeaways & Limitations
TDRNN provides a low-latency AMC design with fewer parameters and competitive or improved accuracy across the evaluated RadioML datasets.
Abstract
from arXiv · showhide
With the imminent development of sixth-generation (6G) networks, there will be a demand for high-accuracy, computationally-efficient, and low-inference time automatic modulation classification (AMC) algorithms. To address this need, we propose a new deep-learning based model for AMC that is called the threshold denoise recurrent neural network (TDRNN). The TDRNN combines an adaptive threshold denoising (TD) algorithm and a recurrent neural network (RNN) that together achieve high accuracy and fast inference. The TD module adaptively reduces the noise level of the received signal, while the RNN module performs the modulation classification on the denoised result. The two subsystems are jointly optimized to reach the optimal architecture. The proposed TDRNN is evaluated for various modulation schemes and signal-to-noise ratios (SNR). The experimental results demonstrate that the TDRNN outperforms existing methods in terms of accuracy, speed, and computational complexity making it an ideal solution for 6G wireless communication systems.
I. INTRODUCTION
6G-oriented AMC requires high accuracy alongside low latency and computational cost. The paper proposes TDRNN, combining adaptive threshold denoising with recurrent classification and evaluating architecture-performance trade-offs.
- Motivation: 6G AMC must support low-latency operation while identifying unknown signal modulations for applications including spectral-efficiency optimization and interference mitigation.The paper frames AMC as important for future 6G systems and their stringent latency requirements.
- Proposed approach: TDRNN combines a threshold denoiser that learns modulation- and SNR-dependent thresholds with an RNN classifier.The denoiser reduces noise before recurrent processing of the signal.
- Design objective: The proposed approach targets URLLC in future networks such as 6G and reports 0.007 ms inference time per modulation example.The contribution also claims reduced computational overhead and energy consumption.
- Results: Nearly 16% average-accuracy improvement is reported from the adaptive Threshold Denoiser.The threshold is learned for each modulation scheme based on SNR.
- Evaluation: Against state-of-the-art AMC methods, TDRNN is reported to use the fewest parameters, achieve the fastest inference, and outperform or match accuracy across the evaluated datasets.The comparison reports higher accuracy on RadioML 2016.10A and comparable performance on RadioML 2018.01A.
C. Paper Organization
The paper reviews deep-learning AMC methods and motivates evaluating accuracy together with parameter count and inference cost. It then presents TDRNN architecture components alongside comparisons with prior lightweight and recurrent approaches.
- Related work: Prior AMC studies use CNN, CLDNN, SCRNN, CGDNet, SCNN, and other architectures to improve classification accuracy, reduce parameters, or accelerate training.The reviewed systems combine convolutional, recurrent, and decentralized-learning components in different ways.
- Related work: SCRNN uses 398K parameters, improves identification accuracy by up to 15%, and reduces training time by 74% versus independent CNN or RNN systems.Its architecture combines CNN feature extraction and dimensionality reduction with LSTM memory.
- Related work: CGDNet uses 119K parameters and is reported to provide faster convergence, greater precision, heightened stability, and up to 37% fewer parameters than previous models.The comparison used identical training sets for the evaluated schemes.
- Design criteria: The related work notes that reducing parameter count can be associated with reduced accuracy, motivating joint consideration of relevant system-performance factors.The paper emphasizes that parameter reduction alone is insufficient for selecting an optimal AMC system.
A. Signal Model
The signal model represents received complex baseband signals under realistic wireless-channel effects and converts them into discrete in-phase and quadrature samples for AMC.
- The received signal model incorporates frequency-selective fading, delay spread, STO, CFO, AWGN, and other real-channel effects.
- The received baseband signal is sampled by an ADC at rate f_s to produce discrete complex samples r[n].
- Each complex sample is decomposed into in-phase I[n] and quadrature Q[n] components, which form the AMC input.
- The TDRNN input vector contains complex samples represented by in-phase and quadrature components over sample length L.
B. Proposed Method
The proposed TDRNN combines adaptive threshold denoising with convolutional and recurrent feature processing before probability-based modulation classification.
- Proposed Method: The TDRNN automatically discovers modulation patterns during training and determines their class during prediction.
- Feature Extraction: The model comprises feature extraction and classification components, with batch normalization applied after the input layer.
- Feature Extraction: The threshold denoiser learns an input-dependent threshold because noise levels vary across modulation types.
- Feature Extraction: Global average pooling estimates a signal average, which is processed by fully connected layers and a Sigmoid function to calculate threshold τ.
- Feature Extraction: The denoiser converts features close to zero into zeros, removing noise-related features according to the learned threshold.
- Feature Extraction: Two ReLU convolutional layers and a GRU with 64 hidden units learn modulation patterns after denoising, with reshaping adapting convolutional output for the GRU.
- Classification: An 11-unit fully connected layer and Softmax produce probabilities for the dataset's 11 modulation types.
2) Classification:
The classification stage uses Adam optimization with categorical cross-entropy to compare one-hot ground-truth vectors against Softmax predictions.
- Classification: Adam is used as the optimizer, while categorical cross-entropy is used as the loss function.
- Classification: The loss compares one-hot ground-truth vectors with predicted vectors across sample types and AMC outputs.
C. Dataset
The evaluation uses RadioML 2016.10A and RadioML 2018.01A, which simulate varied modulation schemes, SNRs, and channel effects.
- Dataset: The study uses RadioML 2016.10A and RadioML 2018.01A with modulation schemes and effects including AWGN, multipath fading, frequency offsets, and timing offsets.
- Dataset: RadioML 2016.10A contains 220K signals spanning 20 SNRs from -20dB to +18dB, with 11 modulation schemes and 128-sample frames.
- Dataset: RML2018.01A contains more than 2.5 million signals across 26 SNRs from -20dB to +30dB, with 24 modulation schemes and 1024-sample frames.
D. Training and Testing
The study trains and evaluates TDRNN primarily on RadioML 2016.10A, using standardized data splits and metrics that capture both classification effectiveness and inference efficiency.
- Dataset: RadioML 2016.10A is selected because it is more challenging and over 30 times smaller than RadioML 2018.01A, reducing training costs.MCLDNN reaches 91% on RadioML 2016.10A and 97% on RadioML 2018.01A.
- Training setup: The data are randomly divided into 60% training, 20% testing, and 20% validation sets.Training uses 500 epochs, a 0.001 learning rate, and batches of 1024, with early stopping and Adam optimization.
- Evaluation metrics: Inference time measures the average time required to classify one modulation example.It is computed from total test-set inference time divided by the number of test examples.
- Evaluation metrics: Average accuracy summarizes classification performance across modulation schemes and SNR levels.The metric aggregates accuracy values over all evaluated modulations and SNR conditions.
IV. EXPERIMENTAL RESULTS AND DISCUSSION
Ablation experiments identify an architecture that balances accuracy and efficiency: 64 GRU hidden units and one GRU layer, together with the Threshold Denoiser.
- GRU hidden-layer width: Approximately 3.5% accuracy improvement occurs when increasing GRU hidden layers from 32 to 64, while gains from 16 to 32 are marginal.The experiment evaluates 16, 32, 64, 128, and 256 hidden-layer configurations within TDRNN.
- GRU hidden-layer width: 0.0067 ms, 0.0071 ms, and 0.0074 ms are the inference times for GRU models with 16, 32, and 64 hidden layers, respectively.The 128- and 256-hidden-layer models require 0.011 ms and 0.017 ms, respectively.
- GRU hidden-layer width: 64 hidden layers balance accuracy with complexity, maintaining relatively few parameters and short inference time.The selected configuration improves accuracy over smaller models without the larger complexity of 128- and 256-layer alternatives.
- Threshold Denoiser: 93% peak accuracy is achieved with the Threshold Denoiser, approximately 11% higher than without it.The denoiser-enhanced model consistently outperforms the version without denoising across SNR conditions.
- Threshold Denoiser: 15.8% average-accuracy improvement raises performance from 54.8% to 63.5% when the Threshold Denoiser is included.The results attribute this improvement to mitigating noise under varying SNR conditions.
- GRU depth: A single GRU layer is sufficient because additional layers increase parameters and inference time while average accuracy remains around 63.3%.The layer-count experiment finds only marginal inference-time increases and no significant average-accuracy gains from deeper GRUs.
B. Comparative Evaluation
Comparative evaluation across two RadioML datasets shows that TDRNN combines strong accuracy with low model complexity and fast inference, although its accuracy is dataset-dependent.
- Accuracy comparison: 93.2% maximum accuracy makes TDRNN the best model on RadioML 2016.10A, ahead of MCLDNN at 90.4% and PET-CGDNN at 89.9%.MCNet and SCNN reach 83.9% and 71.6%, respectively.
- Accuracy comparison: 97% maximum accuracy makes MCLDNN best on RadioML 2018.01A, while TDRNN and PET-CGDNN each reach 96%.The three leading methods have almost indistinguishable accuracy on this dataset.
- Average accuracy: 63.6% average accuracy makes TDRNN best on RadioML 2016.10A, approximately 3-4% above MCLDNN and PET-CGDNN.On RadioML 2018.01A, TDRNN reaches 60.58%, trailing PET-CGDNN and MCLDNN by only 0.5-1%.
- Complexity: TDRNN has approximately two to three times fewer parameters than MCNet, PET-CGDNN, and SCNN, and approximately 10 times fewer than MCLDNN.The comparison evaluates model complexity alongside accuracy and inference time.
- Inference time: 0.007 ms inference time makes TDRNN fastest on RadioML 2016.10A, at three times lower than MCNet and nine times lower than PET-CGDNN.TDRNN is also over 23 times faster than MCLDNN for one modulation example.
C. Discussion
TDRNN combines threshold denoising with recurrent classification to improve AMC accuracy while reducing inference time and model size. Across two RadioML datasets, it achieves strong accuracy and efficiency, supporting broader evaluation criteria for real-time AMC.
- Evaluation perspective: The evaluation considers accuracy, model size, and inference time rather than accuracy alone for real-time signal classification.
- Model contribution: TDRNN combines a Threshold Denoiser with an RNN-based classifier for modulation classification.The TD component removes noise and irrelevant features before recurrent classification.
- Performance: Almost 16% accuracy improvement was achieved compared with models without the Threshold Denoiser.
- Performance: 4.98% higher accuracy, 23 times faster inference, and ten times fewer parameters than MCLDNN were reported on RadioML 2016.10A.
- Performance: 0.007 ms and 0.061 ms inference times were reported across RadioML 2016.10A and RadioML 2018.01A, respectively.