Abstract
IncepSeqNet is a novel deep learning architecture designed to classify 5G signal sequences by rethinking the "imagification" of 1D data. By utilizing a "Multi-Shape Augmentation" technique that reshapes complex signal data into multiple rectangular formats, the model achieves superior classification accuracy in noisy environments and significantly faster training times than standard CNN architectures like ResNet and InceptionV4.
Introduction
Deep learning, particularly Convolutional Neural Networks (CNNs), has revolutionized image recognition. Naturally, researchers have sought to apply these tools to wireless communications, where signals are often represented as sequences of complex numbers. However, a fundamental "incongruity" exists: standard CNNs and their associated data augmentation tricks—such as rotating or flipping an image—are designed for spatial data, not the temporal and phase-sensitive nature of radio signals.
When 1D signal data is simply "imagified" (flattened into a 2D grid) to fit into a CNN, traditional augmentation often leads to performance degradation because it destroys the underlying physical characteristics of the signal. IncepSeqNet addresses this gap by treating the shape of the data as a form of augmentation itself, specifically for Zadoff-Chu sequences used in 5G systems.
Methodology: Multi-Shape Augmentation
The core innovation of IncepSeqNet is the Multi-Shape Augmentation technique. Instead of selecting a single fixed 2D grid size for signal data, the model's "Imagification Layer" transforms a sequence of 144 complex numbers into five distinct rectangular shapes ($R_0$ through $R_4$).
This approach allows the model to capture various correlations between signal elements that might be overlooked in a single fixed grid. The architecture feeds these five shapes into parallel convolution layers, utilizing adaptive filter sizes ($2 \times 2$ and $3 \times 3$) to extract features. By stacking the In-phase (I) and Quadrature-phase (Q) components together, the model performs joint operations that better capture the relationship between the real and imaginary parts of the signal.
Experimental Setup
The study evaluated IncepSeqNet using a specialized dataset of DeModulation Reference Signal (DMRS) sequences. These signals consist of Zadoff-Chu sequences with eight different indices, collected from practical commercial 5G mobile base stations. To test robustness, the data was evaluated across eight different Signal-to-Noise Ratio (SNR) levels, representing the noisy environments typical of real-world wireless communication.
Key Results
IncepSeqNet was benchmarked against established architectures, including GoogleNet, ResNet18, and InceptionV4. The results demonstrate significant improvements in both classification accuracy and computational efficiency.
Accuracy in Noisy Environments
IncepSeqNet outperformed all baseline models across the entire SNR spectrum. The performance gain is most pronounced in low-SNR conditions, where the signal is weakest. At the lowest tested SNR (-4.11 dB), IncepSeqNet showed a 5% percentage point improvement over the advanced InceptionV4 model.
Training Efficiency
Despite processing multiple shapes in parallel, IncepSeqNet is remarkably efficient. The Inception-style parallel processing leads to faster convergence and lower training times compared to deeper, more rigid architectures. The model achieved an average training time of 228.9 seconds, which is approximately 1.3x to 3.87x faster than the baseline models (e.g., ResNet18 required 886.2 seconds).
Discussion and Future Work
The success of IncepSeqNet suggests that reshaping input tensors can act as a powerful augmentation technique for non-image data. By providing the network with multiple "views" of the same sequence, the authors enable the extraction of a richer feature set without the destructive effects of traditional image-based rotation or cropping.
Future research will focus on further optimizing the model for deployment on edge devices and base station hardware. The authors aim to refine the multi-shape technique to explore non-rectangular shapes and more complex stacking methods for I/Q components to further enhance feature diversity.
Summary updated