Source-linked AI summary
An Efficient Deep Learning Model for Automatic Modulation Recognition Based on Parameter Estimation and Transformation
Fuxin Zhang, Chunbo Luo, Jialang Xu, Yang Luo
TL;DR
DL-AMR needs models that balance recognition accuracy with parameter count and computational cost, particularly for resource-limited devices. The paper proposes PET-CGDNN, which combines phase parameter estimation and transformation with CNN-GRU feature extraction, then applies pruning for further compression. The reported model retains high accuracy while substantially reducing model size and maintaining competitive timing.
Problem
Existing DL-AMR models often prioritize recognition accuracy at the expense of model size and complexity, while lightweight models may not meet accuracy requirements for resource-limited devices.
Method
PET-CGDNN estimates and transforms phase-related parameters before extracting spatial and temporal signal features with CNN and GRU layers; pruning progressively removes low-magnitude weights.
Results
With fewer parameters and competitive training and test times, the model maintains high recognition accuracy; after pruning, fewer than 15K parameters can retain accuracy above 90%.
Takeaways & Limitations
The proposed approach supports compact AMR models for resource-limited systems and demonstrates further compression even when the unpruned model is already small.
Abstract
from arXiv · showhide
Automatic modulation recognition (AMR) is a promising technology for intelligent communication receivers to detect signal modulation schemes. Recently, the emerging deep learning (DL) research has facilitated high-performance DL-AMR approaches. However, most DL-AMR models only focus on recognition accuracy, leading to huge model sizes and high computational complexity, while some lightweight and low-complexity models struggle to meet the accuracy requirements. This letter proposes an efficient DL-AMR model based on phase parameter estimation and transformation, with convolutional neural network (CNN) and gated recurrent unit (GRU) as the feature extraction layers, which can achieve high recognition accuracy equivalent to the existing state-of-the-art models but reduces more than a third of the volume of their parameters. Meanwhile, our model is more competitive in training time and test time than the benchmark models with similar recognition accuracy. Moreover, we further propose to compress our model by pruning, which maintains the recognition accuracy higher than 90% while has less than 1/8 of the number of parameters comparing with state-of-the-art models.
I. INTRODUCTION
Existing DL-AMR models face a trade-off between recognition accuracy and resource demands. PET-CGDNN addresses this trade-off through phase-aware processing and lightweight architecture design, with pruning for more constrained deployments.
- Motivation: High-accuracy DL-AMR models rarely consider model size and complexity, while lightweight models may struggle to achieve high accuracy.The gap is especially relevant for IoT devices with limited computing and energy resources.
- Proposed approach: PET-CGDNN processes signals through phase parameter estimation and transformation before CNN-GRU feature extraction and classification.The design is inspired by radio transformer networks, CNNs, and GRUs.
- Proposed approach: The proposed model targets state-of-the-art recognition accuracy with fewer parameters and competitive training and test times.The paper reports that its accuracy is equivalent to state-of-the-art models while using much fewer parameters.
- Design contributions: Parameter estimator and transformer modules are combined with reduced CNN kernel sizes and feature maps to improve recognition accuracy while efficiently using spatial-temporal features.The modules are intended to reduce adverse effects associated with phase.
- Design contributions: Pruning further compresses the lightweight model by five times for scenarios with extremely limited resources.The stated goal is to preserve high recognition accuracy while reducing model size.
A. Signal Model
The received baseband signal is modeled as a modulated signal affected by channel gain, frequency and phase offsets, and additive Gaussian noise. The receiver stores the resulting samples in I/Q form for processing and modulation recognition.
- Signal equation: The received sample y[l] consists of a channel-affected modulated signal plus complex additive Gaussian noise.The model is defined for l = 1, ..., L symbols.
- Signal equation: A[l] denotes channel gain, ω frequency offset, ϕ phase offset, y[l] the l-th received value, and L the number of symbols in a sample.These parameters characterize channel and hardware-related distortions in the received signal.
- I/Q representation: The received signal is stored as an I/Q vector containing the real and imaginary parts of all L samples.This representation facilitates data processing and modulation recognition.
B. The Proposed DL Model
PET-CGDNN combines phase-aware parameter estimation and inverse transformation with CNN-GRU feature extraction. The architecture uses these stages to process distorted I/Q signals and classify modulation schemes.
- Architecture: PET-CGDNN comprises a parameter estimator, parameter transformer, and hybrid neural network.The hybrid network performs feature extraction and classification.
- Phase processing: Channel and hardware effects can distort I/Q data through temporal shifting, linear mixing or rotation, and spinning.The parameter modules extract phase-offset information and perform a phase parameter transformation.
- Parameter estimation: The estimator maps each 2 × L I/Q sample, with L = 128 or 1024, to one estimated phase parameter using Flatten and Dense layers.The Dense layer uses a Linear activation to produce a continuous unbounded phase estimate.
- Parameter transformation: The customized transformer takes the input y and estimated phase ˆϕ to perform a parametric inverse transformation and output ˆy.The output is the transformed signal passed to subsequent processing.
- Feature extraction and classification: CNN layers extract spatial features, a 128-unit GRU extracts temporal features, and a Dense layer classifies the modulation classes.The CNN layers use 75 filters with a 2 × 8 kernel and 25 filters with a 1 × 5 kernel.
C. Model Pruning Method
The pruning method progressively removes low-magnitude weights during fine-tuning to reduce redundancy and model size while targeting high recognition accuracy.
- Motivation: Pruning is applied because CNN-based structures can contain redundancy that increases model size and computational cost.The method aims to preserve high recognition accuracy during compression.
- Pruning procedure: Binary masks are added to Dense, CNN, and GRU weights, and the smallest weights are zeroed progressively until the target sparsity is reached.Sparsity increases gradually during fine-tuning according to the pruning schedule.
III. DATASETS AND IMPLEMENTATION DETAILS
The experiments use three RadioML datasets with different input dimensions, modulation counts, and dataset sizes, using standardized data splits and benchmark implementations.
- Experiments use RML2016.10a, RML2016.10b, and RML2018.01a with input dimensions 2 × 128, 2 × 128, and 2 × 1024, respectively.
- The datasets contain 220,000, 1,200,000, and over 2.5 million signals, spanning 11, 10, and 24 modulation schemes, respectively.Only half of RML2018.01a was randomly selected because of hardware limitations.
- Each dataset is randomly divided by class into training, validation, and test sets at a 6:2:2 ratio.Training uses categorical cross-entropy and Adam, with validation-loss-based learning-rate reduction and early stopping.
- Benchmark comparisons include IC-AMCNET, MCNET, LSTM2, GRU2, and MCLDNN, implemented using a GeForce GTX 1080Ti GPU and Keras with TensorFlow.
A. Model Performance Measurement
Performance is evaluated across three datasets using accuracy, parameter count, training time, and test time, including vanilla and pruned-model comparisons.
- The evaluation compares parameter count, training time, test time, highest accuracy across SNRs, and average accuracy across SNRs on three datasets.Input and output layers are adjusted to each dataset, so parameter counts vary; Table I reports unpruned models.
- PET-CGDNN has the fewest parameters among the benchmark models without pruning.
- PET-CGDNN has lower time cost than GRU2, LSTM2, and MCLDNN, while remaining comparable with IC-AMCNET and MCNET at higher accuracy.
- PET-CGDNN parameters remain relatively stable as input dimension increases, whereas IC-AMCNET requires 6.8 times more parameters for RML2018.01a.PET-CGDNN has the shortest training time on RML2016.10b.
- PET-CGDNN maintains stable and consistent recognition performance across all three datasets against benchmark high-accuracy models.Figure 2 reports recognition accuracy against SNR for the three datasets, while its fourth panel compares PET-CGDNN with PET-CGDNN-Part 3.
- With sparsity of 0.8, the pruned model has fewer than 15K parameters while maintaining accuracy above 90%.On RML2016.10a and RML2016.10b, accuracy remains stable up to sparsity 0.8, when model size is one-fifth of the original; accuracy decreases slightly on RML2018.01a.
C. Effectiveness Analysis of PET-CGDNN
Ablation and visualization analyses examine the contribution of parameter estimation and transformation to PET-CGDNN’s recognition performance.
- PET-CGDNN-Part 3 omits the parameter estimation and transformation modules, enabling an ablation comparison with the complete PET-CGDNN model.
- Above 0 dB SNR, PET-CGDNN-Part 3 does not reach PET-CGDNN’s equivalent recognition accuracy.The complete model shows overall better average and best recognition accuracy in the reported comparison.
- At +10 dB SNR, Part 2 transforms the signals into tighter I/Q-plane clusters than the original inputs.The visualization focuses on constellation distribution features rather than point values, and coordinate origins differ between images.
- The tighter clusters benefit the classification module and lead to improved overall recognition accuracy compared with PET-CGDNN-Part 3.
V. CONCLUSION
The conclusion presents PET-CGDNN as a lightweight, low-complexity AMR model with high accuracy and stable performance, further compressed through pruning.
- PET-CGDNN combines phase-offset estimation and compensation with lightweight, low-complexity design and high recognition accuracy.
- The model exhibits good stability across different datasets.
- Network pruning further reduces model size even though the unpruned model already has relatively few parameters.
- The authors identify potential applications in massive machine-type communications and ultra-reliable and low-latency communications.This is presented as a potential future application scope.