Improving the Accuracy of Location Data in UWB-Based RTLS Using Deep Learning Methods

In Real-Time Location Systems (RTLS) using Ultra-Wideband (UWB) technology, the Decawave DW1000 uses the Two-Way Ranging (TWR) method to obtain the location of a moving object. Multipath propagation occurring under NLOS conditions systematically negatively affects time-leads and distance measurements; this increases the bias (positive bias) and widens the variance, leading to instability in the location data. In this study, an autoencoder-based measurement improvement method proposed for the tag location data obtained using the TWR method. The raw TOF (time of flight) and range measurements obtained from the DW1000 are simultaneously integrated into a low-dimensional latent space with features such as RSSI and CIR-based quality metrics (e.g., first-path amplitude/index, channel energy, pulse width indicators). The denoising/regularized reconstruction process suppresses the jump and bias components in the location data caused by NLOS; thus, the improved measurements can increase the stability and repeatability of location data when used with classical Gauss-Newton location. This approach can trained with a highly dynamic setup (especially using clean LOS records), reducing the burden of relying on field geometry; its modular architecture allows for minimal integration into the existing TWR software chain. Experimental analysis and visualizations were performed on different indoor scenarios (office, corridor, and semi-open space layouts) using MATLAB. This method has been shown to provide a consistent reduction in mean error metrics (MAE/RMSE), a significant improvement in axis bias errors (95th/97th percentile), and location path continuity, while also eliminating erroneous outliers originating from instantaneous NLOS