CTLE: A Hybrid CNN-Transformer-LSTM Equalizer with Multi-Head Attention for Low-BER Signal Recovery in Multipath Fading Channel

Traditional channel equalizers, such as linear FIR filters and decision feedback equalizers (DFE), face limitations including noise amplification at frequency nulls, error propagation from symbol decisions, and exponential complexity in maximum likelihood sequence estimation. Although deep learning-based approaches like convolutional recurrent neural networks (CRNN) improve performance through spatiotemporal feature fusion, their effectiveness diminishes at high SNR conditions. This paper proposes a CNN-TransformerLSTM equalizer (CTLE) that synergistically combines convolutional operations for local feature extraction, LSTM layers for temporal dependency modeling, and multi-head attention mechanisms for global channel correlation learning. Under SUI-3 multipath scenarios with 5 Hz Doppler shift and 0.9 μs delay spread, experiments demonstrate that CTLE reduces bit error rate (BER) by an order of magnitude compared to DFE across 2-20 dB SNR. The model achieves stable convergence within 22 training epochs using AdamW optimizer (learning rate 0.001, weight decay 1e-4), processing 250,000 QPSK symbols with 100 Mbps bit rate. This architecture effectively addresses the trade-off between computational complexity and anti-noise capability in timevarying channels, providing a viable solution for real-time equalization in 5G/B5G multipath environments.