Lightweight Adaptive Residual Feedforward Neural Network for Enhanced Pseudolite Indoor Positioning

To address the challenges of modeling pseudosatellite signal propagation in complex indoor environments and the tendency of deep learning models to overfit, this paper proposes a Lightweight Adaptive Residual Feedforward Neural Network (LARFNN). The model employs a cross-layer feature fusion architecture, integrating raw signal features with deep abstract features via residual connections to mitigate feature degradation in deep networks. It incorporates a tripleoptimization strategy combining Xavier initialization, the Adam optimizer, and a dynamic early stopping mechanism to accelerate convergence and reduce the risk of overfitting on small datasets. Experimental results show that within a $\mathbf{1 0 m} \times \mathbf{1 0 m}$ test area, LARFNN achieves an average positioning accuracy of 0.73 meters using a commercial smartphone, improving by $51.7 \%$ over the random forest method. The $\mathbf{R}^{\mathbf{2}}$ values for $\mathbf{X}$ and $\mathbf{Y}$ coordinate predictions reach 0.954 and 0.959, respectively, offering a highprecision, low-cost positioning solution for scenarios such as large shopping malls and intelligent warehouses.