Deep Learning for Classifying Chaotic Signals Transformed by Advanced Techniques

The rapid evolution of stealthy signals has introduced significant challenges in signal classification, necessitating advanced methodologies for accurate identification and characterization. This study investigates the classification of chaotic signals transformed into Analytic Chaotic Sequences. We utilized a ResNet34 architecture, adapted for one-dimensional signal data, to assess how varying network depths influence classification performance. The dataset comprised sequences with frequency- dependent variations, and the model’s robustness was evaluated under varying noise levels. Results indicate that while the full ResNet34 model maintains high accuracy at elevated signal-to-noise ratios, its performance deteriorates with increased noise. In contrast, models with reduced depths (1, 2, and 3 layers) exhibit improved adaptability and noise resilience. Notably, the 2-layer and 3-layer ResNet34 variants show greater robustness in noisy conditions, suggesting practical benefits for real-world applications. This research highlights the importance of network depth and frequency adaptation in chaotic signal classification, emphasizing that simplified models can provide efficient performance and competitive accuracy, particularly in environments with fluctuating noise levels. Future work will optimize the ResNet34 architecture and expand the dataset to enhance generalization and robustness.