Deep learning for weak signal acquisition: correlation peak pixel absolute position using deep CNNs

Abstract

Listen

Translate article

In a fading environment, satellite navigation signal is so weak that a receiver hardly has a high detection probability using traditional acquisition algorithm. In the acquisition process, a receiver obtains two-dimensions correlation image so as to search doppler frequency shift and pseudo random phase. This paper considers two-dimensions correlation image features and builds a deep learning classification and absolute position networks framework with a Peak Classification Convolutional Neural Networks (PC-CNNs) and a Peak Position Convolutional Neural Networks (PP-CNNs). A maximum correlation peak is a pixel that has an absolute two-dimensions coordinate in a correlation image. After classifying whether there is a correlation peak or not in PC-CNNs, the convolutional layers of the PP-CNNs can encode absolute position of correlation peak features. The PP-CNNs reveals the absolute position information with supervised learning, then the PP-CNNs can obtain code phase and doppler frequency shift with a high probability for a weak satellite navigation signal correlation image. This paper also introduces a regional coordinate encoding method for absolute position with CNNs and generalizes training data set with different signal power. We also compare the accuracy of different number of convolutional layers and different training data set. According to simulation experiments, the results indicate that the deep CNNs has a high probability to acquire weak satellite navigation signal.