RSLoc: An Accuracy Indoor Localization System Based on KAN Convolution Network

With the rapid growth of mobile communication technologies and smart devices, demand for location-based services has surged. Fingerprint-Based positioning using wireless signals is now a leading method for indoor localization. However, wireless signals from access points (APs) fluctuate due to time-varying noise, causing fingerprint signals sampled at the same location to differ over time, sometimes significantly. This variability not only hinders accurate fingerprint map construction but also disrupts precise matching between the target point’s fingerprint and the map, leading to inconsistent positioning results. To address this, we propose RSLoc, an indoor localization system based on a KAN convolutional network. First, APs with weak or erratic signals are excluded, and a refined fingerprint map is generated. We then develop a deep network model, FFEM, based on KAN convolution. The model is trained using a sample library derived from fingerprint data, optimized through entropy regularization and pruning. The FFEM model generates a mapping of the target point, calculates its similarity to the fingerprint map, and selects appropriate neighboring points for effective localization. Experiments using different mobile devices show that RSLoc improves average positioning accuracy by 33%-44% compared to methods like LORI and FML. Moreover, removing invalid APs and constructing the sample library reduce errors, enhancing accuracy by over 48% and 29%, respectively.