Evaluation of Bidirectional Cross-Scale Alignment Networks for Short-Window Signal Classification under Severe Noise and Receiver Distortion

• Mohamed Trabelsi

  National School of Engineers of Gabes, Department of Electrical Engineering and Industrial Informatics, Omar Ibn El Khattab Street 6029, Gabes, Tunisia

  Author

• Sami Ben Salem

  Higher Institute of Applied Sciences and Technology of Sousse, Electrical Engineering Department, Cite Erriadh 4023, Sousse, Tunisia

  Author

• Youssef Khelifi

  National Engineering School of Monastir, Department of Electrical Engineering and Microelectronics, Avenue Ibn El Jazzar 5019, Monastir, Tunisia

  Author

Signal classifiers are increasingly deployed in settings where decisions must be made from short and degraded waveform snippets rather than from long and carefully synchronized observations. Background work in wireless sensing has therefore shifted from purely handcrafted descriptors toward trainable representations that can absorb distortion, timing uncertainty, and class overlap. Yet the practical difficulty is not only to build a deeper encoder. It is also to prevent the decision system from discarding brief discriminative waveform traits while it learns broader invariances. This paper presents an empirical study of a bidirectional cross-scale alignment network for signal classification from raw in-phase and quadrature samples. The proposed model maintains a fine-resolution processing route and a coarse-context processing route, then exchanges information between them through reciprocal alignment bridges before classification. The study evaluates the method on an eleven-class complex-baseband benchmark across a wide signal-to-noise range, with special attention to short-window classification, repeated-seed stability, off-diagonal confusion patterns, certainty fidelity, and behavior under receiver mismatch. The primary task uses cropped observation windows shorter than the original benchmark span, so that the classifier must operate under deliberate information scarcity. Relative to matched single-stream and late-concatenation baselines, the proposed network improves global accuracy and macro-averaged recall, with the clearest gains appearing in the intermediate low-SNR region where fragile local evidence still exists but is easily suppressed by conventional deep encoders. Additional tests under carrier offset, IQ imbalance, clipping, and impulsive corruption show that the model degrades more smoothly than comparison systems. The results suggest that reciprocal cross-scale interaction is a useful empirical design principle when short-window signal classification must remain stable under noise and hardware-related perturbation.

• PDF

Copyright (c) 2024 International Journal of Science, Engineering and Applied Technologies

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.