Radio map construction based on BERT for fingerprint-based indoor positioning system

Abstract Due to the heavy workload of RSS collection, instability WLAN signal strength and disappearance signals caused by complex indoor environments, construction radio map for wireless local area network (WLAN) fingerprint-based positioning system is time-consuming laborious. In order rapidly deploy system, bidirectional encoder representation from transformers (BERT) model used fill missing in quickly build map. The imported into BERT form natural language text, filled model. Since number input data cannot exceed 512 words, structure not suitable with large volume. Therefore, we redefine based on original parallel. addition, loss function redefined. Except that each segment has a function, weighted average value all functions defined as total function. experimental results show better than traditional linear interpolation method, compressed sensing algorithm matrix completion filling fingerprint database, probability error within 2 m reaches about 94%.