Wi-Fi Signal Noise Reduction and Multipath Elimination Based on Autoencoder

Abstract

It is known that the signal is noisy and susceptible to multipath interference in indoor positioning, resulting in a significant error in the processing of the signal. The RSS- assisted cross-correlation (RACC) method can reduce noise and eliminate multipath interference to a certain extent, but too environmentally sensitive. Therefore, in this paper, an effective way of using the deep neural network is proposed to address this problem. Accordingly, the performance of the AutoEncoder in signal noise reduction and multipath interference elimination are discussed. To achieve better results, four AutoEncoder models are put forward, fully connection (FC), convolution plus fully connected (C-FC), convolution plus pooling (C-P), inception (ICP), and the performance of these four models are compared when processing signals with different signal to noise ratio (SNR) and multipath interference. The mean square error (MSE) and the time difference of arrival (TDoA) are the standards for evaluating the effect of signal noise reduction and multipath interference removal. Besides simulated data, we also conducted model performance comparisons in terms of ground truth signal. Experimental results show that fully connected layer is essential to automatic signal coding and the model performs better with the appropriate addition of convolution layer when faced with noise and multipath environments. Notably, compared with RACC method, the TDoA of two resultant signals obtained from the model is more accurate, verified by IEEE 802. 11b WLAN.