Back to results
US 12,732,398 B2Granted
Method and system for acquiring channel images
OwnerKeke Zuinventor
Life of this patent
Estimating the term…
Abstract
Disclosed herein is a method and a system for acquiring channel images, relating to the field of cross-integration of artificial intelligence neural network and wireless communication system. Based on data-driven artificial intelligence neural network channel estimation, the disclosure obtains a sufficient number of channel images for training the neural network. It overcomes the limitation that the traditional acquisition of channel images dependent on the types of deployed antennas and geometric dimensions, so that the artificial intelligence neural network can be effectively used for channel estimation of wireless communication systems in practice.
Claims10
A method for acquiring channel images, comprising the following steps:
- step 1:
- off-line training;
- in an off-line training stage, gathering signals at time t (t≥1) to obtain initial channel signals H=[H1, H2, . . . , HT]∈CMXSXT of a three-dimensional tensor and received signals HLS=[HLS,1, HLS,2, . . . , HLS,T]∈CMXSXT after passing through channels, wherein the initial channel signals H and the received signals HLS are used as real signals for a neural network training and input signals to be estimated, respectively, where M represents a number of antennas of a receiving device, S is a number of transmitted data streams, and a time dimension T represents a time observation length (T≥1);
- step 2, dimension reconstruction;
- performing dimension reconstruction on the signals H and HLS of the three-dimensional tensor, fixing the time dimension T, transforming each MXS matrix of the signals H and HLS into a vector of MSX1, and arranging the vector to obtain matrices of real signals H′ and received signals
- H
- L
-
- S
- ′
- with a size of MSXT based on the time dimension;
- step 3, sliding window selection;
- employing a sliding window with a size of PXK, with a step size of p, to slide on the time dimensions of the real signals H′ and the received signals
- H
- L
-
- S
- ′
- ,
- respectively, und selecting a channel image with the size of PXK, where P≤MS, K is a time dimension and K≤T;
- step 4, real part and imaginary part separation;
- performing a real part and an imaginary part separation on the selected channel image of PXK to obtain a channel image of PXKX2;
- step 5, neural network training;
- using the obtained channel image of PXKX2 to train a neural network having a plurality of layers, obtaining a weight coefficient for each layer, and deploying the trained neural network to the receiving device;
- step 6, online inference;
- in an online inference stage, lasting for time K to obtain initial received signals from time 1 to the time K;
- and
- step 7:
- the received signal at time K+1 constitutes an observation signal consisting of the received signals from time 2 to the time K+1, and repeating the above steps at subsequent times.
26 more elements
A system for acquiring channel images, comprising:
- a memory;
- and a processor including a plurality of modules by executing one or more programs stored in the memory, wherein:
- an off-line training module, which is used for collecting signals at time t (t≥1) in an off-line training stage to obtain initial channel signals H=[H1, H2, . . . , HT]∈CMXSXT of a three-dimensional tensor and received signals HLS=[HLS,1, HLS,2, . . . , HLS,T]∈CMXSXT after passing through channels, wherein the initial channel signals H and the received signals HLS are used as real signals for a neural network training and input signals to be estimated, respectively, where M represents a number of antennas of a receiving device, S is a number of transmitted data streams, and a time dimension T represents a time observation length (T≥1);
- a dimension reconstruction module, which is used for performing dimension reconstruction on the signals H and HLS of the three-dimensional tensor, fixing the time dimension T, transforming each MXS matrix of the signals H and HLS into a vector of MSX1, and arranging the vector to obtain matrices of real signals H′ and received signals
- H
- L
-
- S
- ′
- with a size or MSXT based on the time dimension;
- a sliding window selection module, which is used for employing a sliding window with a size of PXK, with a step size of p, to slide on the time dimensions of the real signals H′ and the received signals
- H
- L
-
- S
- ′
- ,
- respectively, and selecting a channel image with the size of PXK, where P≤MS, K is a time dimension and K≤T;
- a real part and imaginary part separation module, which is used for performing real part and imaginary part separation on the selected channel image of PXK to obtain a channel image of PXKX2;
- a neural network training module, which is used for training the neural network having a plurality of layers by using the obtained channel image of PXKX2 to obtain a weight coefficient for each layer, and deploying the trained neural network to the receiving device;
- and
- an online inference module, which is used in an online inference stage for the system lasting for time K to obtain initial received signals from time 1 to the time K.
19 more elements
Dependencies are inferred from the claim text.