CNN-LSTM Hybrid Architecture for Over-the-Air Automatic Modulation Classification Using SDR

arXiv:2511.21040 · cs.LG, cs.CV · Submitted 2025-11-27 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "CNN-LSTM Hybrid Architecture for Over-the-Air Automatic Modulation Classification Using SDR".

Jane: The paper was written by Dinanath Padhya, Krishna Acharya, Bipul Kumar Dahal and Dinesh Baniya Kshatri from Department of Electronics and Computer Engineering, Thapathali Campus, Institute of Engineering, Tribhuvan University.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Jane: We also have Lu with us today — senior AI researcher at Tsinghua.

Tom: We also have Meng with us today — lead engineer at a mysterious AI startup.

Jane: We also have Lalam with us today — the in-house Large Language Model.

Tom: Alright, let's get started.

Summary of Results: Tom: We’ve talked about the structure, so let's look at the summary of results in "CNN-LSTM Hybrid Architecture for Over-the-Air Automatic Modulation Classification Using SDR." Jane, what are the key performance numbers they achieved, and how does that high reliability translate into actual benefit?

Jane: The authors found that by combining these two components, their optimized model achieved a very high level of performance. Specifically, the results show an accuracy of ninety-three point four eight percent on the test set after tuning its internal parameters.

Lu: That accuracy is impressive, but the truly fascinating detail is how consistently they maintained it across an entire range of noise conditions—a Signal-to-Noise Ratio from zero to twenty dB.

Meng: From a practical standpoint, that range of zero to twenty decibels is extremely challenging for robust systems. It means the system can handle both very clean signals and heavily distorted ones without any failure.

Lalam: This stability under extreme noise levels suggests a fundamental shift in how we design communication hardware; we're moving away from assuming perfect conditions toward accepting real-world chaos as the norm.

Tom: So, it’s not just about peak performance when things are pristine, but reliable performance across the entire operational spectrum?

Jane: Exactly. The fact that they kept that ninety-three point four eight percent accuracy even when the signal was heavily distorted is a massive indicator of robustness for any real-world application we can imagine.

Lu: This suggests that separating features within each window, combined with temporal context, allows them to distinguish between modulation types even if their characteristics are blurred by noise.

Meng: The consistency in precision and recall also tells us they aren’t just guessing; the model is highly confident when it makes a classification.

Lalam: By achieving such high reliability across the entire SNR range, we are paving the way for self-healing or adaptive networks that can operate consistently regardless of external interference.

Tom: It's clear they didn't just achieve a snapshot of success; they managed to make it sustainable and dependable in a noisy environment.

Jane: And this consistency is visible across all their metrics, like the F1 score, which also remained very high at ninety-three point four five percent.

Lu: The success of this hybrid approach shows that we can capture both the local structure and the long-term sequence without losing either gaining accuracy.

Meng: This data is vital because it means that if they are deploying this solution, the reliability will be constant, which is what operators need most when managing critical infrastructure.

Lalam: This reliable performance ensures that we are building AI tools capable of handling complex operational realities for society's benefit.

Improvements Over Existing Methods: Tom: We’ve seen the general concept and the strong results, but what specific improvements does this paper claim over existing methods in "CNN-LSTM Hybrid Architecture for Over-the-Air Automatic Modulation Classification Using SDR"? Jane, how is this approach fundamentally better than older models?

Jane: The core advantage is that it's moving past simple models that only look at raw data in the time domain. They are intelligently processing the signal by combining spatial feature extraction with temporal modeling.

Lu: It's about creating an adaptive flow; instead of just stacking layers, they designed a synergy where each component optimizes the input for the next layer in a very sophisticated, overlapping manner.

Meng: If I understand this correctly, older sequential models often struggle with feature loss or difficulty maintaining sequence integrity when trying to capture both patterns and time.

Lalam: This refinement is pushing us toward true cognitive radio capabilities, moving past simple detection and into understanding the intent of a transmission itself.

Tom: So, it’s not just about higher accuracy; it’s about making the entire process more robust and comprehensive than what was previously possible?

Jane: Exactly. By explicitly designing the handoff from CNN to LSTM, they are managing which parts of the signal—the local shape or the long-term trend—are most important for classification at any given moment.

Lu: The specific benefit I see is in noise resilience; previous methods might fail if a single burst of noise distorted a key feature, but this layered approach can filter out that transient distortion using memory.

Meng: For deployment, this suggests the system requires less complex pre-processing than older models, which is a major advantage for real-world implementation.

Lalam: This architectural advancement means we are creating systems that don't just process data but actually understand the context of managing our increasingly crowded wireless spectrum.

Tom: It sounds like they are building a detection system that understands the "why" behind the signal, not just its physical appearance.

Jane: That’s a great way to put it; they are capturing both local details and long-term trends simultaneously using this hybrid method.

Lu: The sequential nature of the signal demands this dual perspective, and this hybrid architecture delivers that necessary depth of understanding.

Meng: It's exciting because it provides a clear path for practical, real-world deployment without relying on idealized test conditions that often fail in the field.

Lalam: This work demonstrates how AI can help us manage complexity by leveraging the the full potential of both spatial and temporal data structures to improve societal systems.

Conclusion: Tom: We've seen just how effective this "CNN-LSTM Hybrid Architecture for Over-the-Air Automatic Modulation Classification Using SDR" is, and it's clear that its ability handle noisy real-world signals makes a huge difference. Jane, what’s the final takeaway regarding the practical application of this ninety-three point four eight percent accuracy?

Jane: It’s incredibly reassuring to see such high reliability across that entire zero to twenty dB SNR range, because this covers almost every chaotic environment we encounter in the field today.

Lu: I think this model is capable of so much more than just classification; it opens a path toward truly intelligent radio where the network can adapt its behavior based on what it's actually hearing and seeing.

Meng: From a deployment standpoint, achieving ninety-three point four eight percent accuracy while using an SDR setup feels like a very practical and achievable milestone for building robust spectrum monitoring tools right now.

Lalam: My vision is that this could lead to a shift where wireless devices don't just connect, but intelligently negotiate how they communicate based on the precise characteristics of their environment, improving overall efficiency.

Tom: That’s a powerful idea, Lu; moving beyond just seeing signals to making intelligent decisions about future potential connections.

Jane: And it confirms that when we use AI to handle both local spatial patterns and long-term temporal flow, we get a level of robustness that simulation simply cannot match.

Lu: Exactly, Jane; the sequential nature of the signal requires both the short-term context and that deep temporal memory, which is exactly what this hybrid design provides.

Meng: It’s exciting because it shows a clear path for real-world deployment without relying on idealized test conditions, so I think we are seeing a very viable product here.

Lalam: This work demonstrates how AI can help us manage the complexities of our crowded radio spectrum in ways that benefit society as a whole.

Tom: We are definitely feeling very optimistic about the future of wireless communication with this architecture, and that leaves us with so much to think about for our listeners today.

Conclusion: Tom: So, we've spent time dissecting this research, and it's clear that the ability to reliably identify communication signals under real-world noise conditions is a massive leap forward for any system monitoring the radio spectrum.

Jane: It’s so encouraging to see that ninety-three point four eight percent accuracy achieved consistently across all those challenging signal-to-noise ratios, because simply put, it means this technology works when it matters most in real life.

Lu: I think the biggest impact is that we are moving past simple detection; we’re now building systems capable of truly understanding the operational context of the entire communication environment.

Meng: That reliability combined with a practical SDR implementation suggests that this isn't just a lab curiosity, but a deployable solution for building robust monitoring tools today.

Lalam: This work shows us how AI can help us manage the complexities of our crowded wireless world in ways that actively benefit society as a whole.

Tom: It’s not just about getting the numbers right, though; it's about making the entire system dependable, which is a huge win for anyone working with critical infrastructure.

Jane: And this consistency across all metrics, like the F1 score and the ROC curves, shows that we are achieving a level of quality that simulation simply cannot match.

Lu: The sequential nature of the signal demands this dual perspective—the spatial detail plus temporal memory—and this hybrid architecture delivers exactly that depth.

Meng: It’s impressive how they designed a system to handle both structural analysis and real-time tracking without needing perfect conditions, making it a very practical tool.

Lalam: This research demonstrates how we can leverage AI to see through noise and perceive the true nature of any complex communication environment for society.

Tom: We’re feeling incredibly optimistic about the future of wireless communication with this CNN-LSTM Hybrid Architecture for Over-the-Air Automatic Modulation Classification Using SDR, which is truly a monumental piece of work.

Jane: It sets such a high bar, and I think it makes us all very excited to see what the next breakthrough in automated signal processing will be.

Dinanath Padhya, Krishna Acharya, Bipul Kumar Dahal, Dinesh Baniya Kshatri

Department of Electronics and Computer Engineering, Thapathali Campus, Institute of Engineering, Tribhuvan University

cs.LG, cs.CV

Submitted: 2025-11-27

Updated: 2026-08-25

Importance score: 88/100

The gist: This paper proposes a hybrid CNN-LSTM architecture for Automatic Modulation Classification (AMC) integrated with a Software Defined Radio (SDR) platform to identify modulation schemes without prior

Key concepts

CNN-LSTM Hybrid Architecture: This model combines two components: CNN for spatial feature extraction (local structure) and LSTM for temporal modeling (long-term sequence). This synergy allows the system to capture both local details and long-term trends simultaneously, providing a deeper understanding of the signal.
Modulation Classification: This is a technique where signals are classified automatically based on their characteristics as they travel through the air. The hybrid architecture helps distinguish between different types of modulation even if noise blurs or distorts their physical features.
Signal-to-Noise Ratio (SNR) Robustness: The system's ability to maintain high accuracy across a range of noise conditions, specifically from zero to twenty dB. This indicates that the model performs reliably whether the signal is very clean or heavily distorted by real-world interference.

Terminology

Accuracy — 33× in this paper · explained in 1 episode
Automatic Modulation Classification — 22× in this paper
F1 Score — 11× in this paper · explained in 3 episodes
AUC — 9× in this paper · explained in 2 episodes
Feature Extraction — 6× in this paper · explained in 2 episodes
Robustness — 6× in this paper · explained in 5 episodes
Convolutional Neural Network — 5× in this paper
Software Defined Radio — 5× in this paper
Generalization — 4× in this paper · explained in 13 episodes
Long Short-Term Memory — 4× in this paper
Attention Mechanism — 3× in this paper · explained in 2 episodes
Consistency — 3× in this paper · explained in 3 episodes
Core — 3× in this paper · explained in 2 episodes
Convolutional Neural Networks (CNNs) — 2× in this paper · explained in 2 episodes
Global Average Pooling — 2× in this paper
Hybrid Models — 2× in this paper · explained in 2 episodes
Internet of Things — 2× in this paper
Receiver Operating Characteristic — 2× in this paper
Transfer Learning — 2× in this paper · explained in 10 episodes
Transformer — 2× in this paper · explained in 2 episodes
Adam — 1× in this paper · explained in 2 episodes
Adaptability — 1× in this paper · explained in 1 episode
Adaptation — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Domain Adaptation — 1× in this paper · explained in 5 episodes
Frequency Modulation — 1× in this paper
Generative Adversarial Networks (GANs) — 1× in this paper · explained in 3 episodes
Gradient Flow — 1× in this paper · explained in 2 episodes
Hybrid Approach — 1× in this paper · explained in 4 episodes
Multimodal Fusion — 1× in this paper · explained in 3 episodes
Semi-supervised Learning — 1× in this paper · explained in 4 episodes
Verification — 1× in this paper · explained in 2 episodes

Summary

This paper proposes a hybrid CNN-LSTM architecture for Automatic Modulation Classification (AMC) integrated with a Software Defined Radio (SDR) platform to identify modulation schemes without prior knowledge. This capability is essential for advancing cognitive radio, spectrum monitoring, and intelligent communication networks, particularly as the expansion of IoT and smart cities increases radio frequency spectrum congestion.

The Problem and Proposed Solution

Traditional AMC methods, such as likelihood-based or feature-based approaches, often struggle with real-world signal variations like noise, fading, and Doppler shifts. Existing deep learning research has addressed this using CNNs for spatial features or LSTMs for temporal dependencies, but architectures utilizing them separately often struggle to capture both spatial and temporal dependencies simultaneously.

To address these gaps, the authors propose a novel hybrid architecture that leverages:

> Convolutional Neural Networks (CNNs)

> Long Short-Term Memory (LSTM) networks

The system is uniquely designed to move beyond simulation by implementing a custom SDR testbed to evaluate the model’s performance on over-the-air (OTA) signals, ensuring practical, real-world applicability.

Methodology and Data Processing

The system utilizes a hybrid dataset combining the RadioML2018.01a dataset with custom-generated signals from GNU Radio, featuring Signal-to-Noise Ratios (SNRs) ranging from 0 to 30 dB. The signal acquisition process involves:

> Generation of modulated signals in GNU Radio

> I/Q signal reception using RTL-SDR hardware

> Transformation into amplitude, phase, and I/Q domains

Each signal instance is segmented into eight overlapping windows of 224 samples. These are reshaped into a unified 224 × 224 × 3 tensor, which resembles an RGB image where each channel encodes a distinct physical property (amplitude, phase, and raw I/Q). This tensor serves as the input to the CNN backbone.

Model Architecture

The classification model integrates a modified AlexNet for spatial feature extraction with an LSTM for temporal modeling. The CNN component uses five convolutional layers with ReLU activations and Global Average Pooling (GAP) to produce a 256-dimensional feature vector. This is followed by:

> An LSTM module

> A fully connected classification head consisting of two dense layers (512 and 256 neurons)

The authors performed comparative experiments by replacing the LSTM’s flattening mechanism with a single-headed temporal attention layer. However, this modification resulted in degraded performance, suggesting that preserving the full temporal sequence is more beneficial for differentiating spectrally similar modulation schemes.

Experimental Results and Performance

The optimized model achieved an accuracy of 93.48%, 93.53% precision, 93.48% recall, and an F1 score of 93.45%. The ROC analysis confirmed the model's outstanding ability in identifying modulation classes, with AUC values approaching 1.00 for most classes, such as BPSK and QPSK. While the model is robust across a wide SNR range (0–20 dB), the authors note that accuracy drops to approximately 80% in live, real-time environments due to operational complexities like multipath fading and carrier frequency offsets. Despite this, the results validate the effectiveness of the hybrid architecture for adaptive spectrum management. Drawing from their findings, they conclude that preserving the complete temporal feature map via flattening outperforms attention-based compression for this task.

Improvements for AI systems

Based on the technical findings and architectural comparisons in this paper, I propose the following specific improvements to existing Automatic Modulation Classification (AMC) systems:

Implements a Flattened Temporal State mechanism instead of Attention-based compression for sequential modeling.
Integrates a multi-channel 3D tensor input representing Amplitude, Phase, and raw I/Q values simultaneously (mimicking RGB structure).
Utilizes an AlexNet-based spatial feature extractor coupled with an LSTM layer trained on hybrid datasets (synthetic + real-world OTA signals).

The improved AI system will be able to:

Accurately differentiate between spectrally similar analog modulation schemes (specifically AM-DSB-SC vs. AM-SSB-SC) that standard attention mechanisms fail to separate due to information bottlenecks.
Maintain high classification stability and precision across a wide Signal-to-Noise Ratio (SNR) range (0 dB to 20 dB), avoiding the performance degradation typical of models optimized only for high SNR environments.
Provide robust, real-time modulation identification in live RF environments by effectively capturing both local spatial patterns in I/Q samples and long-term temporal dependencies across signal windows.