Deep learning-based wireless communication synchronization structures

1. A system comprising:

a preprocessor to transform a received wireless signal in the form of a resource grid for a plurality of user equipment devices into a matrix of symbol group features for a narrowband physical random-access channel;

a neural network configured to transform the symbol groups from the preprocessor into predictions of active user equipment devices;

a first twin neural network comprising:

a first neural sub-network to transform the symbol groups from the preprocessor into time-of-arrival predictions for signals from the user equipment devices; and

a second neural sub-network to transform the symbol groups from the preprocessor into carrier frequency offset predictions for signals from the user equipment devices.

2. The system of claim 1 , wherein the preprocessor comprises logic to:

determine averages for each symbol group in the resource grid; and

normalize sequences of the symbol groups based on carrier frequency hopping patterns utilized by the user equipment devices.

3. The system of claim 2 , the preprocessor comprising additional logic to:

determine average received power for each of the sequences of symbol groups; and

concatenate the average received power to the sequences of symbol groups.

4. The system of claim 1 , wherein each neural sub-network comprises a pipeline of convolution layers.

5. The system of claim 4 , wherein the pipeline of convolution layers is configured to determine a feature vector for each subcarrier band of the narrowband physical random-access channel and for each of the symbol groups.

6. The system of claim 4 , wherein the pipeline of convolution layers includes skip connections.

7. The system of claim 4 , wherein the convolution layers comprise kernels configured to operate along a frequency dimension of the symbol groups.

8. The system of claim 1 , wherein the neural network, the first neural sub-network, and the second neural sub-network each comprise a multilayer perceptron.

9. The system of claim 1 , wherein the second neural sub-network comprises a second twin neural network.

10. A system comprising:

a wireless receiver to receive wireless signals; and

a neural network configured to transform the wireless signals into (a) predictions of active user equipment devices that are transmitting carrier frequency hopping signals, (b) time-of-arrival predictions for the signals from the user equipment devices, and (c) carrier frequency offset predictions for the signals from the user equipment devices, the neural network comprising:

a first feed-forward path to transform symbol groups derived from the wireless signals into predictions of active user equipment devices;

a second feed-forward path to transform the symbol groups into time-of-arrival predictions for the signals from the user equipment devices; and

a third feed-forward path to transform the symbol groups into carrier frequency offset predictions for the signals from the user equipment devices.

11. The system of claim 10 , wherein the second and third feed-forward paths each comprise a pipeline of convolution layers.

12. The system of claim 11 , wherein each pipeline of convolution layers is configured to determine a feature vector for each subcarrier band of a narrowband physical random-access channel and for each of the symbol groups.

13. The system of claim 10 , wherein the feed-forward paths each comprise a multilayer perceptron.

14. The system of claim 10 , wherein the second and third feed-forward paths comprise a twin neural network.

15. A wireless base station process comprising:

transforming received wireless signals comprising a resource grid for a plurality of user equipment devices into a matrix of symbol group features for a narrowband physical random-access channel;

operating a first neural network to transform the symbol group features into predictions of user equipment devices active on the base station; and

operating a second neural network to transform the symbol group features into time-of-arrival and carrier frequency offset predictions for signals from the user equipment devices.

16. The process of claim 15 , wherein operating the second neural network comprises:

operating a first neural sub-network of a twin neural network to transform the symbol group features into the time-of-arrival predictions; and

operating a second neural sub-network of the twin neural network to transform the symbol group features into the carrier frequency offset predictions.

17. The process of claim 15 , wherein transforming the received wireless signals into the matrix of symbol group features for the narrowband physical random-access channel comprises:

determining averages for each symbol group in a resource grid for the user equipment devices; and

normalizing sequences of the symbol groups based on carrier frequency hopping patterns utilized by the user equipment devices.

18. The process of claim 17 , wherein transforming the received wireless signals into the matrix of symbol group features for the narrowband physical random-access channel further comprises:

determining average received power for each of the sequences of symbol groups; and

concatenating the average received power to the sequences of symbol groups.

19. The process of claim 15 , wherein operating the second neural network to transform the symbol group features into time-of-arrival and carrier frequency offset predictions for signals from the user equipment devices comprises operating a plurality of independent pipelines of convolution layers to determine feature vectors for each subcarrier band of the narrowband physical random-access channel and for each of the symbol groups.

20. The process of claim 19 , wherein the pipelines of convolution layers include skip connections and kernels configured to operate along a frequency dimension of the symbol groups.