PL-MIND: Physics-Informed Liquid Neural Networks for Multichannel Identification of NLOS Dynamics in Indoor Positioning

Abstract

Ultra-Wideband (UWB) indoor positioning suffers severe accuracy degradation under Non-Line-of-Sight (NLOS) conditions, where reflected signals introduce positive ranging bias. Existing deep learning approaches treat the Channel Impulse Response (CIR) as a discrete sequence, creating a representational mismatch with the continuous nature of radio propagation. This paper presents Physics-Informed Liquid Neural Networks for Multichannel Identification of NLOS Dynamics (PL-MIND), a three-stage pipeline that models CIR evolution as a continuous-time dynamical system. The dual-circuit liquid encoder adapts its time constant to incoming signal strength, producing a compact 64-dimensional embedding that feeds (i) a binary LOS/NLOS classifier, (ii) a single-bounce identifier, and (iii) a distance regressor. Across a 6,000-sample multichannel dataset spanning ten indoor scenarios, PL-MIND achieves 97% out-of-distribution (OOD) accuracy for Stage 1 classification and 82% for Stage 2 single-bounce identification. This maintains high performance whereas traditional discrete architectures like LSTM, CNN, and BERT suffer significant accuracy drops between 6% and 19%. Furthermore, PL-MIND utilizes only 16,897 parameters, making it 1.14 to 4.33 times more efficient than baselines, and reaches a Mean Absolute Error of 0.2458 m for Stage 3 distance prediction.