Source-linked AI summary
Robust and Fast Automatic Modulation Classification with CNN under Multipath Fading Channels
Kürşat Tekbıyık, Ali Rıza Ekti, Ali Görçin, Güneş Karabulut Kurt, Cihat Keçeci
TL;DR
AMC classifiers have been difficult to make robust to changing multipath fading, and existing datasets can inadequately represent real-world channel conditions. This paper introduces a diverse controlled dataset and a CNN for classifying modulation families; it reports better performance than CLDNN across the two datasets and shorter training time.
Problem
AMC needs classifiers that remain effective under changing multipath fading and other impairments, while commonly used datasets can lack realistic channel diversity or controlled channel parameters.
Source: §I · p. 1, A. Related Work · p. 2Method
The paper introduces HisarMod2019.1 with varied fading and tap conditions and proposes a CNN to classify modulation families.
Source: B. Contributions · p. 2Results
The CNN performs significantly better than CLDNN on HisarMod2019.1 and is faster and more accurate than CLDNN on RadioML2016.10a; its RadioML maximum accuracy is 90.7%.
Source: §V · p. 4, B. RadioML2016.10a Dataset… · p. 4Takeaways & Limitations
The results demonstrate the proposed CNN on both a diverse multipath dataset and the frequently used RadioML2016.10a dataset.
Source: B. RadioML2016.10a Dataset… · p. 4, §V · p. 4
Abstract
from arXiv · showhide
Automatic modulation classification (AMC) has been studied for more than a quarter of a century; however, it has been difficult to design a classifier that operates successfully under changing multipath fading conditions and other impairments. Recently, deep learning (DL)-based methods are adopted by AMC systems and major improvements are reported. In this paper, a novel convolutional neural network (CNN) classifier model is proposed to classify modulation classes in terms of their families, i.e., types. The proposed classifier is robust against realistic wireless channel impairments and in relation to that when the data sets that are utilized for testing and evaluating the proposed methods are considered, it is seen that RadioML2016.10a is the main dataset utilized for testing and evaluation of the proposed methods. However, the channel effects incorporated in this dataset and some others may lack the appropriate modeling of the real-world conditions since it only considers two distributions for channel models for a single tap configuration. Therefore, in this paper, a more comprehensive dataset, named as HisarMod2019.1, is also introduced, considering real-life applicability. HisarMod2019.1 includes 26 modulation classes passing through the channels with 5 different fading types and several numbers of taps for classification. It is shown that the proposed model performs better than the existing models in terms of both accuracy and training time under more realistic conditions. Even more, surpassed their performance when the RadioML2016.10a dataset is utilized.
I. INTRODUCTION
AMC is used in military and civilian communication systems, but dependable classification under realistic wireless impairments remains a challenge for deep-learning methods.
I. INTRODUCTION
- AMC supports military and civilian applications including electronic warfare, radio surveillance, and spectrum awareness.Source: §I · p. 1
- Deep-learning AMC methods need robustness to wireless impairments to deliver dependable classification under realistic channel conditions.The paper frames this as necessary for efficient and resilient performance.Source: §I · p. 1
A. Related Work
Classical likelihood- and feature-based approaches face computational or robustness limits, while deep-learning AMC research still needs datasets and models suited to realistic channel conditions.
A. Related Work
- Likelihood-based methods maximize correct classification but incur high computational complexity and are vulnerable to model mismatches such as channel estimates and timing offsets.Source: A. Related Work · p. 1
- Feature-based AMC uses statistics and signal transforms, but individual features may not distinguish large signal sets and can degrade under fading or frequency and timing offsets.Cyclostationarity-based features resist mismatches better than other cited features, but remain computationally complex.Source: A. Related Work · p. 1
- CNN, CLDNN, and LSTM are prominent deep-learning architectures studied for AMC, including work on I/Q and FFT inputs, uncertain noise, and online learning [11].Source: A. Related Work · pp. 1–2
- RadioML2016.10a is widely used, but measurement-based RadioML2018.01a lacks channel parameters and laboratory conditions limit its fading and tap diversity.The paper motivates a dataset combining actual channel conditions with controlled parameters.Source: A. Related Work · p. 2
B. Contributions
The paper introduces HisarMod2019.1 for controlled evaluation across diverse channel conditions and proposes a CNN with strong accuracy and training-time performance.
B. Contributions
- HisarMod2019.1 adds a controlled dataset with ideal, static, Rayleigh, Rician, and Nakagami-m channels and varied tap counts.The authors present it as a more challenging resource for observing realistic channel conditions.Source: B. Contributions · p. 2
- The proposed four-convolution, two-dense-layer CNN is reported to outperform available classifiers in accuracy on HisarMod2019.1 and RadioML2016.10a.Source: B. Contributions · p. 2
- The proposed CNN has lower training complexity than available techniques, making its training process relatively short.Source: B. Contributions · p. 2
II. HISARMOD2019.1: A NEW DATASET
HisarMod2019.1 is designed to broaden AMC evaluation through multiple modulation families, fading conditions, and channel-tap configurations.
II. HISARMOD2019.1: A NEW DATASET
- HisarMod2019.1 contains 26 modulation types across analog, FSK, PAM, PSK, and QAM families, with 1,500 signals of 1,024 I/Q samples per type.Source: §II · p. 2
- The dataset totals 780,000 signals across 20 SNR levels from -20 dB to 18 dB, with oversampling rate 2 and a raised-cosine filter roll-off of 0.35.Source: §II · p. 2
- Signals use ideal, static, Rayleigh, Rician (k = 3), or Nakagami-m (m = 2) channels, with 4 or 6 multipath taps selected equally often.The channel types are equally distributed; ideal includes AWGN without fading, while static coefficients remain constant during propagation.Source: §II · p. 2
III. THE PROPOSED CNN MODEL
The proposed CNN uses convolutional and pooling layers for feature extraction, followed by dense layers. Its design narrows the convolutional feature extractor and uses noise during training.
III. THE PROPOSED CNN MODEL
- The proposed CNN has four convolution and pooling layers followed by two dense layers.The architecture is shown in Fig. 1 and its layout is listed for HisarMod2019.1 and RadioML2016.10a in Table II.Source: §III · p. 2, Figure 1 · p. 3, Table II · p. 3
- Across many configurations, the feature extractor narrowed the number of filters in each successive convolutional layer, improving classification results and reducing training time; the first layer used 256 filters.Source: §III · p. 2
- The first dense layer has 128 neurons with ReLU activation, and the following dense layer uses softmax to compute class probabilities.The softmax output covers the 26 modulation types listed across five families in HisarMod2019.1.Source: §III · p. 3, Table I · p. 3
- A noise layer adds noise during training, with its power set according to the desired SNR level.The model uses early stopping when validation loss has converged sufficiently.Source: §III · p. 3
- The proposed CNN has 15 million trainable parameters for HisarMod2019.1, compared with 27 million for CLDNN, and takes one-quarter of LSTM’s time per epoch.Training and testing use four NVIDIA Tesla V100 GPUs in parallel.Source: §III · p. 3
IV. CLASSIFICATION RESULTS
The proposed model is tested on HisarMod2019.1 and RadioML2016.10a; the section introduces the test results.
IV. CLASSIFICATION RESULTS
- The proposed model is tested on both HisarMod2019.1 and RadioML2016.10a.The passage says the test results are provided below but does not report them here.Source: §IV · p. 3
A. HisarMod2019.1 Dataset Classification Results
For HisarMod2019.1, the study evaluates modulation-family classification using a hierarchical approach. The proposed CNN outperforms CLDNN across the reported SNR range, with meaningful results above 2 dB.
A. HisarMod2019.1 Dataset Classification Results
- The evaluation classifies signals first by modulation family, then identifies modulation type within the family subset.The study focuses on family classification, not the order of each modulation type; the dataset is split 8/15, 2/15, and 5/15 for training, validation, and test.Source: A. HisarMod2019.1 Dataset… · pp. 3–4, Figure 2 · p. 4
- The first CNN layer adds noise according to the SNR level, and early stopping is used during training.The authors state that this setup makes the model more robust to overfitting.Source: A. HisarMod2019.1 Dataset… · p. 4
- The proposed CNN exceeds 80% accuracy at 8 dB SNR and 90% at 14 dB and higher, reaching 94% maximum accuracy versus CLDNN’s 85%.CLDNN reaches 80% accuracy at 16 dB and does not achieve 90%; the proposed CNN performs better than both CLDNN and LSTM on this dataset.Source: A. HisarMod2019.1 Dataset… · p. 4
- The model gives meaningful results above 2 dB SNR; at low SNR, it may make random choices between modulation families.Both the proposed CNN and CLDNN have difficulty identifying QAM signals, while LSTM classifies signals as analog regardless of their received type.Source: A. HisarMod2019.1 Dataset… · p. 4
B. RadioML2016.10a Dataset Classification Results
On RadioML2016.10a, the proposed CNN is compared with CLDNN and LSTM on a dataset of synthetic signals with 10 modulation types. The CNN outperforms CLDNN above -2 dB, while LSTM is slightly better overall.
B. RadioML2016.10a Dataset Classification Results
- RadioML2016.10a contains synthetic signals from 10 modulation types and is used to assess the proposed CNN's robustness on a widely used dataset.The dataset is widely used in modulation-classification studies.Source: B. RadioML2016.10a Dataset… · p. 4
- 90.7% maximum accuracy for the proposed CNN, versus 88.5% for CLDNN and 92.3% for LSTM on RadioML2016.10a.The CNN exceeds CLDNN above -2 dB; LSTM has extremely high computational complexity.Source: B. RadioML2016.10a Dataset… · p. 4
- At 6 dB and above, the proposed model achieves very high performance across modulation types except WBFM.At low SNR, it recognizes almost all signals as 8-PSK; at the minimum SNR where accuracy exceeds 50%, results are poor outside 4-PAM.Source: B. RadioML2016.10a Dataset… · p. 4
- The authors report that the proposed model performs with high performance on both RadioML2016.10a and the diverse HisarMod2019.1 dataset.Source: B. RadioML2016.10a Dataset… · p. 4
V. CONCLUDING REMARKS
The conclusion reports that the proposed CNN performs significantly better than CLDNN on HisarMod2019.1 and is faster and more accurate than CLDNN on RadioML2016.10a. It also identifies modulation-order classification and explaining a filter-design observation as future work.
V. CONCLUDING REMARKS
- The proposed CNN performs significantly better than CLDNN on HisarMod2019.1 and is faster and more accurate than CLDNN on RadioML2016.10a.Source: §V · p. 4
- The authors plan to investigate modulation-order classification assuming the modulation family is already identified.Source: §V · p. 4
- The authors also plan to investigate why starting with many filters and reducing their number step by step gives better accuracy.Source: §V · p. 4
- Fig. 3 compares accuracy values for LSTM, CLDNN, and the proposed CNN on HisarMod2019.1 and RadioML2016.10a.Source: Figure 3 · p. 5
- Figs. 4 and 5 show confusion matrices at 0, 6, 12, and 18 dB for the proposed CNN and CLDNN, respectively, on HisarMod2019.1.Source: Figure 4 · p. 5, Figure 5 · p. 5