Channel State Identification in Complex Indoor Environments With ST-CNN and Transfer Learning
Zhiguo Sun, Kaixuan Wang, Rongchen Sun, Zengmao Chen
Abstract
Accurate channel state identification can improve the accuracy of ultrawideband (UWB) indoor positioning. In complex indoor environments, existing channel state identification methods suffer from high costs. Meanwhile, these methods pay little attention to the different localization errors caused by weak -non-line-of-sight (WLOS) and non-line-of-sight (NLOS). In this letter, a channel state identification method that can identify line-of-sight (LOS), WLOS, NLOS is proposed for use in a single indoor scenario based on the Stockwell transform and convolutional neural networks (ST-CNN). Moreover, combined with transfer learning (TL), ST-CNN can reduce the cost of identification in complex indoor environments. Simulations based on open source datasets show that ST-CNN can achieves F2 scores of 97.17% and 97.14% in NLOS and WLOS. For complex indoor environments consisting of multiple indoor scenarios, compared with the ST-CNN, the ST-CNN+TL method can reduce the training time by 62% and the training data by 40%.