Machine learning for channel estimation
Systems and methods are disclosed for performing training using superimposed pilot subcarriers to determine training data. The training includes starting with a training duration (T) equal to a number of antennas (M) and running a Convolutional Neural Network (CNN) model using training samples to determine if a testing variance meets a predefined threshold. When the testing variance meets a predefined threshold, then reducing T by one half and repeating the running Convolutional Neural Network (CNN) model until the testing variance fails to meet the predefined threshold. When the testing variance fails to meet the predefined threshold, then multiplying T by two and using the new value of T as the new training duration to be used. Generating a run-time model based on the training data, updating the run-time model with new feedback data received from a User Equipment (UE), producing a DL channel estimation from the run-time model; and producing an optimal precoding matrix from the DL channel estimation.
1. A method of machine learning for channel estimation, comprising:
performing training using superimposed pilot subcarriers to determine training data; the performing training comprising:
starting with a training duration (T) equal to a number of antennas (M),
running a Convolutional Neural Network (CNN) model using training samples to determine if a testing variance meets a predefined threshold,
when the testing variance meets a predefined threshold, then reducing T by one half and repeating the running Convolutional Neural Network (CNN) model using training samples and reducing the T by one half until the testing variance fails to meet the predefined threshold, and
when the testing variance fails to meet the predefined threshold, then multiplying T by two and using the new value of T as the new training duration to be used;
generating a run-time model based on the training data;
updating the run-time model with new feedback data received from a User Equipment (UP;
producing a downlink channel estimation from the run-time model; and
producing a precoding matrix from the downlink channel estimation.
2. The method of claim 1 further comprising:
allowing a predetermined number of paths to include both of an original static nature of the path and a new dynamic path; and
learning by the model based on the dynamic path data.
3. The method of claim 1 further comprising using a different number of antennas to be used in a different configuration of a radio transceiver.
4. The method of claim 1 wherein the channel estimation is per-slot, per-subcarrier, or per-User Equipment (UE).
5. The method of claim 1 wherein the predefined threshold comprises a means-square error against the fully-known channel.
6. The method of claim 5 wherein the means-square error against the fully-known channel is less than 0.5 percent of the known channel.
7. The method of claim 1 wherein the model is recalibrated offline.
8. The method of claim 1 wherein the model is recalibrated periodically.
9. The method of claim 1 wherein the channels are 2G, 3G, 4G, 5G, or Wi-Fi channels.
10. A non-transitory computer-readable medium containing instructions for providing machine learning for channel estimation which, when executed, cause a network system to perform steps comprising:
performing training using superimposed pilot subcarriers to determine training data; the performing training comprising:
starting with a training duration (T) equal to a number of antennas (M);
running a Convolutional Neural Network (CNN) model using training samples to determine if a testing variance meets a predefined threshold;
when the testing variance meets a predefined threshold, then reducing T by one half and repeating the running Convolutional Neural Network (CNN) model using training samples and reducing the T by one half until the testing variance fails to meet the predefined threshold; and
when the testing variance fails to meet the predefined threshold, then multiplying T by two and using the new value of T as the new training duration to be used;
generating a run-time model based on the training data;
updating the run-time model with new feedback data received from a User Equipment (UE);
producing a DL channel estimation from the run-time model; and
producing a precoding matrix from the DL channel estimation.
11. The non-transitory computer-readable medium of claim 10 further comprising instructions for allowing a predetermined number of paths to include both of an original static nature of the path and a new dynamic path; and
learning by the model based on the dynamic path data.
12. The non-transitory computer-readable medium of claim 10 further comprising instructions for using a different number of antennas.
13. The non-transitory computer-readable medium of claim 10 further comprising instructions wherein the channel estimation is per-slot, per-subcarrier, or per-User Equipment (UE).
14. The non-transitory computer-readable medium of claim 10 further comprising instructions wherein the predefined threshold comprises a means-square error against the fully-known channel.
15. The non-transitory computer-readable medium of claim 10 further comprising instructions wherein the model is recalibrated offline and wherein the model is recalibrated periodically.
16. The non-transitory computer-readable medium of claim 10 further comprising instructions wherein the channels are 2G, 3G, 4G, 5G, or Wi-Fi channels.