Demodulation of BFSK Signal for Super-Low Frequency Wireless Communication Based on Deep Learning

Abstract

Wireless electromagnetic transmission using layered annulus space made up of tubing, casing and packer in production wells is an effective means of overcoming the corrosion, fracture, winding and sealing problems of current cable transmission. Tubing, casing and packer are all good conductors for electromagnetic transmission. So, the high-frequency signal attenuates very fast with transmission due to the skin effect, only the super-low frequency (30 Hz– 300 Hz) band can be selected. The feature difference between frequency 0 and frequency 1 of BFSK modulation signal in super-low frequency segment is very small, and the bandwidth is very narrow too. In addition to low signal-to-noise ratio and channel interference under the conduction of production wells, the frequency characteristics of 0-frequency and 1-frequency signals become fuz- zier. The traditional BFSK demodulation method based on broadband is difficult to demodulate the ultra-low frequency signal. As the strong pattern recognition ability of deep learning neural network, a method of demodulating BFSK signal in ultra-low frequency band based on deep learning is proposed. The convo- lutional neural network is used to design the demodulation model from the modulation signal of super-low frequency segment to the corresponding binary signal. The network model inputs the super-low frequency modulation signal sequence with the size of 20 * 20 and outputs the demodulation signal with the size of 4 * 1. Firstly, 3200 super-low frequency BFSK modulated signals and corresponding original binary signals were obtained through MATLAB simu- lation, in which the carrier frequency of modulation signal element 1 was 100 Hz and the carrier frequency of code element 0 was 200 Hz. The sampling points of each code element in the signal are 100, and the transmission time of 4 code elements is 0.08 s. Then the modulated signal is added with Gaussian white noise to simulate channel interference, among which 2560 signals are used as training data to train the network model, and the other 640 signals are used as test data. The simulation test results show that our method has a stronger anti- interference ability of ultra-low frequency modulation signal than the traditional demodulation method. The next step is to apply the network model to real data to further verify the generalization and practicality of the method.