A Construction of Hirota Bilinear Neural Network Layer for Deep Learning-Based Classification of Wireless Communication Signals

Deep learning has become an essential tool for analyzing complex patterns in wireless communication signals; however, conventional architectures such as SimpleRNN, LSTM, and Transformer-based networks often show limited capability in capturing highly nonlinear feature interactions and global dependencies simultaneously. To address this limitation, this paper proposes a novel Hirota Bilinear Neural Network (Hirota-BiNN) framework inspired by Hirota’s bilinear formalism from nonlinear mathematical physics. The proposed model introduces a Hirota bilinear layer that explicitly models multiplicative interactions between two parallel feature branches through full bilinear mapping and low-rank bilinear refinement, thereby improving representation capacity while maintaining computational efficiency. The new layer is constructed for deep learning and published at https://pypi.org/project/blinear/ with version 1.0.0. Real-world wireless communication signals from Wi-Fi, LTE, and 5G-NR systems are employed for binary classification tasks. Raw in-phase and quadrature (IQ) signals are transformed into magnitude sequences, segmented into fixed-length samples, and labeled using an energy-based thresholding strategy. For fair evaluation, SimpleRNN, LSTM, Transformer, and the proposed Hirota-BiNN are trained under identical experimental settings. Experimental results demonstrate that conventional recurrent and attention-based models achieve only moderate performance, with testing accuracy around 69.9% and limited precision, recall, and F1-score due to insufficient nonlinear feature modeling. In contrast, the proposed Hirota-BiNN achieves significantly faster convergence, stronger classification stability, and superior predictive performance by explicitly learning nonlinear multiplicative dependencies between signal features. The corrected and reproducible implementation is further released as the publicly available Python package blinear to ensure experimental transparency and practical deployment. These findings confirm that mathematically inspired bilinear neural structures provide a promising direction for intelligent wireless signal analysis and future physics-informed deep learning systems.