A CNN-SE-BiLSTM Detection Model for Ultrasonic Wideband Signals Over Intra-Body Fading Channels

Abstract

To enable safe and reliable intra-body communication (IBC) for medical applications, ultrasonic wideband (UsWB) technology employs low-duty-cycle pulses with time-hopping spread spectrum to effectively mitigate multipath and thermal effects. However, transient waveform distortion and multipath delay spread, which are induced by the inhomogeneity of the human body, render UsWB signals difficult to detect. This paper proposes a Squeeze-and-Excitation channel attention mechanism-based hybrid convolutional neural network and bidirectional long short-term memory (CNN-SE-BiLSTM) model for the detection of UsWB signals, in which a one-dimensional convolutional neural network (CNN) is used to capture multi-scale spatial features, the attention mechanism is employed to mitigate the impact of noise through adaptive feature recalibration, and the back-end bidirectional long short-term memory (BiLSTM) network is designed to model the global temporal dependencies of pulses in complex intra-body environments. Extensive Monte Carlo simulations are conducted to compare the bit-error rate (BER) performance of various deep learning-based detection models, including CNN, LSTM, CNN-LSTM, Transformer, and the proposed CNN-SE-BiLSTM, against traditional matched filter and energy detection receivers over intra-body fading channels. Simulation results demonstrate that the CNN-SE-BiLSTM model maintains a Flash size below 1 MB with a peak RAM usage of 280 KB and achieves the lowest BER among the evaluated schemes, offering a promising detection solution for UsWB-based IBC systems.