Low resolution OFDM receivers via deep learning
View Patent ↗Various embodiments provide for deep learning-based architectures and design methodologies for an orthogonal frequency division multiplexing (OFDM) receiver under the constraint of one-bit complex quantization. Single bit quantization greatly reduces complexity and power consumption in the receivers, but makes accurate channel estimation and data detection difficult. This is particularly true for OFDM waveforms, which have high peak-to average (signal power) ratio in the time domain and fragile subcarrier orthogonality in the frequency domain. The severe distortion for one-bit quantization typically results in an error floor even at moderately low signal-to-noise-ratio (SNR) such as 5 dB. For channel estimation (using pilots), various embodiments use novel generative supervised deep neural networks (DNNs) that can be trained with a reasonable number of pilots. After channel estimation, a neural network-based receiver specifically, an autoencoder jointly learns a precoder and decoder for data symbol detection.
1. A communication system receiver using low-resolution quantization to reduce power consumption, the communication system receiver comprising:
a low resolution analog to digital convertor that produces a quantized complex valued received signal;
a first neural network to assist in channel estimation, wherein the first neural network uses the quantized complex valued received signal as part of labeled data that trains the first neural network to learn a channel;
a second neural network to enhance quality of a channel estimator which is made by the first neural network, wherein the second neural network is a special unsupervised untrained generative deep neural network; and
a third neural network for data detection to decode transmitted data, wherein a detector is trained offline irrespective of a current channel estimate to be able to detect data symbols that were encoded by a neural network based precoder at a transmitter whose weights depend on the current channel estimate.
2. The communication system receiver of claim 1 , wherein the low resolution analog to digital converter is a one-bit analog to digital converter.
3. The communication system receiver of claim 1 , wherein the communication system receiver feeds back an output of the first neural network to the transmitter to be used in the neural network based precoder at the transmitter.
4. The communication system receiver of claim 1 , wherein a decoder is trained without explicitly considering the channel and an orthogonal frequency-division multiplexing (OFDM) modulator.
5. The communication system receiver of claim 1 , wherein the channel estimator is trained using labeled data produced via pilot symbols and corresponding one-bit quantized observations.
6. The communication system receiver of claim 1 , wherein the precoder is learned after a decoder is trained while taking into account the channel and an orthogonal frequency-division multiplexing (OFDM) modulator.
7. The communication system receiver of claim 1 , wherein:
precoded symbols are multiplied with a learned precoder matrix; and
the precoded symbols are normalized to ensure an average transmission power constraint.
8. The communication system receiver of claim 1 , wherein the channel is a frequency selective fading channel with multiple subcarriers.
9. A neural network-based receiver using low-resolution quantization to reduce power consumption, the neural network-based receiver comprising:
an antenna to receive a wireless signal from a transmitter using a number of subcarriers in a frequency selective fading channel;
an analog to digital converter connected to the antenna to quantize the wireless signal;
a deep learning-based channel estimator that utilizes an unsupervised untrained generative deep learning technique to improve channel estimation quality; and
an autoencoder that learns a precoder and a decoder to detect transmitted data symbols.
10. The neural network-based receiver of claim 9 , wherein parameters learned for the precoder are transmitted to a transceiver before transmission of the wireless signal.
11. The neural network-based receiver of claim 9 , wherein the decoder is trained offline and irrespective of the frequency selective fading channel.
12. The neural network-based receiver of claim 9 , wherein the precoder is trained online in conjunction with a trained decoder and while taking into account the frequency selective fading channel.
13. The neural network-based receiver of claim 9 , wherein the analog to digital converter has a resolution of six bits or less per in-phase (I) and quadrature (Q) channel.
14. A method for operating a orthogonal frequency division-multiplexing communication system receiver, the method comprising:
performing channel estimation using a first neural network and a second neural network,
wherein the first neural network uses a set of pilot symbols as an input layer and identifies parameters for multiple hidden layers to produce an output,
wherein the parameters are identified by minimizing an error between a function of the output and a labeled data set, and
wherein the second neural network does not require training and combines a special untrained generative deep neural network with a least squares (LS) estimator; and
performing data detection using a third neural network.
15. The method of claim 14 , wherein the first neural network is trained using labeled data produced via pilot symbols and corresponding one-bit quantized observations.
16. The method of claim 14 , wherein performing channel estimation includes performing channel estimation using a generative supervised deep neural network that was trained with a number of pilot symbols.
17. The method of claim 14 , further comprising:
generating a learned precoder by training a precoder after the channel estimation; and
notifying a transmitter of the learned precoder before data transmission begins.
18. The method of claim 17 , wherein the learned precoder is generated with stochastic computing-based hardware.
19. The method of claim 14 , wherein the first neural network is a generative model that generates random inputs from a learned distribution.
20. The method of claim 14 , wherein the first neural network and the third neural network are separately trained.