Hybrid Message Passing Algorithm Based Deep Learning Neural Network for Large-Scale Beamforming Systems

A hybrid deep learning architecture with linear and nonlinear layers, combined with a message passing algorithm, addresses the challenge of adapting to varying SNRs and channel conditions in beamforming systems, enhancing coding gain and reducing overhead in high-frequency wireless communications.

JP7580570B2Active Publication Date: 2024-11-11HUAWEI TECH CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2023504615
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Priority Date
2020-07-24
Filing Date
2021-06-25
Publication Date
2024-11-11
Estimated Expiration
2041-06-25

AI Technical Summary

Technical Problem

Deep learning techniques in wireless systems face challenges in generalizing across different signal-to-noise ratios (SNR) and adapting to time-varying channel conditions, leading to the need for frequent retraining and significant signaling overhead, especially in beamforming systems using high-frequency bands like millimeter waves.

Method used

A hybrid deep learning architecture with linear and nonlinear layers, combined with a message passing algorithm, allows for generalization across varying SNRs and channel conditions, reducing the need for retraining by incorporating power control and utilizing a virtual user concept to optimize information transmission.

Benefits of technology

The system achieves improved coding gain and reduced overhead by tolerating varying channel uncertainties, enabling efficient beamforming in high-frequency wireless systems without the need for frequent retraining and minimizing pilot usage.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007580570000261
    Figure 0007580570000261
  • Figure 0007580570000262
    Figure 0007580570000262
  • Figure 0007580570000263
    Figure 0007580570000263
Patent Text Reader

Abstract

A multi-beam transmission method is provided for transmitting to a receiver having K receive beams using an N-beam transmitter. In the transmitter, a nonlinear encoder implemented by a machine learning block and a linear encoder are trained using gradient descent backpropagation that relies on feedback from the receiver. For each input to be transmitted, the machine learning block is used to process the input and generate N / K sets of L outputs. The linear encoder is used to generate a respective set of K outputs by linearly encoding each set of L outputs to generate a total of K encoded outputs and N / K sets of N encoded outputs. One of the N / K sets of K outputs from each set of K beams. To enable generalization to different SNRs and time-varying channel conditions, the nonlinear layer implemented in the nonlinear encoder is responsible for feature extraction and feature regeneration, and the linear layer implemented in the linear encoder is responsible for further generalization.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] cross reference This application claims priority to U.S. patent application Ser. No. 16 / 937,863, entitled “Hybrid Message-Passing-Algorithm-Based Deep Learning Neural Network for a Massive Beam-Forming System,” filed July 24, 2020, the contents of which are incorporated by reference in their entirety herein.

[0002] Field This application relates to machine learning, and more particularly, to machine learning applied to beamforming systems. [Background technology]

[0003] Deep Learning for Wireless Systems Deep learning is a subset of machine learning that has three or more layers of nonlinear neurons between the input and the output. Deep learning is a branch of artificial intelligence that has made significant progress in its application to highly complex problems. There are several fundamental problems when deep learning techniques are applied to wireless systems. The first problem is the inability to generalize across different signal-to-noise ratios (SNRs). Deep neural networks are highly concerned with information data, i.e., the statistical properties of signals (e.g., expectation and variance, etc.), and random adversariality in wireless systems. Once the SNR changes, the neurons of the deep neural network need to be adjusted to adapt to the different noise, often resulting in complete retraining. In reality, wireless channels, including signal-to-noise ratios and signal-to-interference ratios, are fluctuating and difficult to estimate reliably, so it is impractical to use deep learning neural networks to process time-varying signals.

[0004] Wireless millimeter wave communication 5G proposes to deliver ultra-Gbps levels of throughput using millimeter wave (mmWave) communications. Strictly speaking, 5G mmWave communications may use cm or mm wavelengths.

[0005] Typical mmWave or high frequency communications have different channel conditions and circumstances than those below 6GHz. Propagation losses in mmWave bands are much more severe than those below 6GHz. To address this, a technique called beamforming has been adopted by 5G systems to "focus" the signal energy from several transmitting antennas to one spot or area (on the order of centimeters or mm) to produce a sufficiently high SNR signal within the targeted spot or area. For radio, the signal strength needs to meet the sensitivity level of the receiver.

[0006] However, the higher the frequency of the band used by a wireless system, the greater the propagation loss that the wireless beams will experience. Future wireless systems may adopt bands with frequencies above 100 GHz. The receiver sensitivity of future wireless beams will be significantly constrained by the propagation loss in the air. To compensate for the propagation loss, one option is to increase the transmission power on the wireless beams. On the one hand, the use of higher frequency bands leads to shorter wavelengths and smaller antenna sizes, and on the other hand, the greater propagation loss caused by the higher frequency bands increases the signal receiver sensitivity level on each wireless beam.

[0007] To address these issues, beam-on-chip systems may be used that accommodate many (e.g., thousands) of radio beams simultaneously on one chip die area. If artificial intelligence is used to perform the beamforming in such systems, retraining is generally required when the receiver moves. Retraining may not be able to occur quickly enough to keep up with the changing channel, especially when there are a large number of beams. In addition, the signaling overhead associated with training is very large. Summary of the Invention [Means for solving the problem]

[0008] To achieve generalization to different signal-to-noise ratios, several linear layers are inserted into the autoencoder framework. These linear layers enhance the generalization to different signal-to-noise ratios. In the presented deep learning architecture, the nonlinear encoding layer f(·) acts as a feature extractor that extracts N / K distinct features from the symbols or samples of one image or distribution. The linear layers are N / K phase (w) / bias (b) / power (c) linear encoders. The N / K features can be considered as different N / K viewpoints or observations of the symbol, and the linear layers encode them in parallel. N wireless signals necessarily sum the N / K distinct encoded features in the air, forming a K-rank combiner. The nonlinear decoding layer g(·) is a fully connected combiner that recovers the symbols from the K-dimensional signal following the addition of distortions and noise.

[0009] To allow generalization to different SNRs and time-varying channel conditions, the nonlinear layer is responsible for extracting features and recovering features, while the linear layer is responsible for more generalization. The nonlinear layer itself may have low tolerance to time-varying channel uncertainties such as SNR and attenuation that would normally require a complete retraining of the nonlinear layer as such conditions change. The inclusion of the linear layer has the effect of introducing more tolerance (more generalization) to the overall system, and the linear layer can be easily adjusted when there are changes in channel uncertainties such as SNR and attenuation without requiring retraining of the linear layer.

[0010] In some embodiments, the nonlinear layer is implemented using machine learning. More specifically, in some embodiments, a deep neural network is used for the nonlinear layer. The global solution may approach the optimal solution. In some embodiments, the system is receiver transparent.

[0011] In some embodiments, a message passing algorithm (MPA) or belief propagation algorithm is applied before each transmission to adjust the power control factor.

[0012] For varying channel attenuation, in some embodiments, an additional power control layer is added to allow for taking into account the varying channel attenuation. During initial training, a nonlinear layer and a linear encoding layer (<f(),g(),w,b> ) and power control <c>Tandem training is performed that includes both the message passing algorithm for and the message passing algorithm for . To apply MPA, a virtual user concept is introduced. Useful information about a symbol is directed to a "real" user, and useless information is directed to a "virtual" user. During inference (transmission), before each symbol transmission, the MPA at the transmitter calculates the <c>It works to adjust the other encoding layers (<f(),g(),w,b> ) remains. During training, the attenuation is sampled from the distribution. During actual transmission, the estimated channel coefficients, which have some uncertainty, are used to update the distribution.

[0013] To address channel mismatch and uncertainty, we use attenuation in both the initial training and inference MPA. <h>is inserted into the deep neural network, <h>are samples from the distribution. The nonlinear decoding layer g(·) <h>In practice, if the true decay falls off a bit within the distribution or even outside the distribution, g() can still handle this situation. In some cases, no pilots are needed, or fewer pilots than would normally be needed.

[0014] In some embodiments, the techniques provided are applied to implement a mass beam system.

[0015] In some embodiments, the provided system and method are further adapted to provide coding gain and incremental redundancy hybrid automatic repeat request (IR-HARQ). It is well known in coding theory that encoding one symbol at a coding rate has no coding gain but has power gain. The coding rate only provides power gain in additive white Gaussian noise (AWGN) channels and may provide some MRC gain in selective channels. A system and method are provided for implementing a codec that encodes one symbol at a given coding rate to achieve coding gain.

[0016] Selective attenuation is adversarial: the codec provided learns large-scale attenuation and has built-in logic to tolerate small-scale random attenuation so that no or less pilots are required.

[0017] Incremental redundancy is useful in linear systems. Systems and methods are provided for including IR coding gain through the use of IR transmission with a provided codec.

[0018] Embodiments of the present disclosure will now be described with reference to the accompanying drawings. [Brief description of the drawings]

[0019] [Figure 1] FIG. 1 is a block diagram illustrating multiple beams that compensate for significant propagation losses in the mmWave band. [Figure 2A] FIG. 1 is a block diagram showing a configuration of a beam-on-chip. [Figure 2B] FIG. 1 is a block diagram showing a joint two-beam-on-chip configuration to produce a good quality signal. [Diagram 3] FIG. 2C is a block diagram illustrating an example transmission of one 64QAM symbol by 1000 beams using the beam-on-chip configuration of FIG. 2B. [Figure 4] FIG. 2C is a block diagram illustrating the propagation of an unevenly attenuated beam from the beam-on-chip configuration of FIG. 2B. [Diagram 5] FIG. 2 is a block diagram showing an example configuration of a receiver with K=3 uncorrelated receive dimensions. [Figure 6] FIG. 1 is a block diagram illustrating a general example of a MIMO system for allocating MCS with respect to rank. [Figure 7] FIG. 1 is a block diagram illustrating the ability of transmission of multiple beams to withstand complete blockage of some beams by random obstructions. [Figure 8] FIG. 1 is a block diagram illustrating the ability of transmitting multiple beams to withstand the Doppler effect. [Figure 9A] FIG. 1 is a block diagram illustrating round robin transmission for measuring channel coefficients. [Figure 9B] FIG. 1 is a block diagram illustrating round robin transmission for measuring channel coefficients. [Figure 10] 1 is a plot of measured symbol error rate (SER) versus Es / N0 for 64QAM over AWGN. [Figure 11] 1 is a plot of measured symbol error rate (SER) versus Eb / N0 for 64QAM over AWGN. [Figure 12] 12 is a plot illustrating the expected coding gain for the plot of FIG. 11 when K=3 dimensions are used to transmit the input bits in one 64-QAM symbol. [Figure 13] FIG. 1 is a block diagram illustrating the theoretical limits and effect of coding gain when transmitting input bits in one 64-QAM symbol using K=3 dimensions. [Figure 14] FIG. 1 is a block diagram illustrating the ability of a symbol encoding scheme to withstand selective attenuation. [Figure 15] FIG. 2 is a block diagram illustrating an N-dimensional transmitter and a K-dimensional receiver. [Figure 16] FIG. 16 is a block diagram illustrating an example transmission of 64QAM using the transmitter and receiver of FIG. [Figure 17] FIG. 2 is a block diagram illustrating a high level view of a transceiver. [Figure 18] FIG. 1 is a block diagram illustrating a deep neural network formed by a transceiver with a three-stage encoder. [Figure 19] FIG. 1 is a block diagram illustrating linear and nonlinear layers of a linear / nonlinear hybrid deep neural network. [Figure 20] FIG. 1 is a block diagram illustrating a deep neural network with power control. [Figure 21A] FIG. 13 is a block diagram illustrating the receipt of a "wasteful" combination by a virtual user. [Figure 21B] FIG. 13 is a block diagram illustrating the receipt of a "wasteful" combination by a virtual user. [Figure 22] FIG. 1 is a block diagram showing the iterations of a message passing algorithm (MPA) for training a power factor. [Figure 23A] Backpropagation<f,g,w,b> FIG. 1 is a block diagram showing an example configuration for tandem training in which the MPA adjusts α (power control vector) and the MPA adjusts α. [Figure 23B] Backpropagation<f,g,w,b> FIG. 1 is a block diagram showing an example configuration for tandem training in which the MPA adjusts α (power control vector) and the MPA adjusts α. [Figure 24A] FIG. 13 is a block diagram illustrating transmission where layers trained by backpropagation are frozen and layers by MPA remain adjusted. [Figure 24B] FIG. 13 is a block diagram illustrating transmission where layers trained by backpropagation are frozen and layers by MPA remain adjusted. [Diagram 25] This is a plot of the measured SER performance of encoding one 64QAM signal onto three complex numbers with R=1 / 3. [Figure 26] FIG. 1 is a block diagram illustrating time-varying channel attenuation uncertainty decomposition. [Figure 27A] 1 is a block diagram illustrating a transceiver with random channel attenuation. [Figure 27B] 1 is a block diagram illustrating a transceiver with random channel attenuation. [Figure 28A] FIG. 1 is a block diagram illustrating reception of transmissions by channel attenuation virtual users. [Figure 28B] FIG. 1 is a block diagram illustrating reception of transmissions by channel attenuation virtual users. [Figure 28C] FIG. 1 is a block diagram illustrating reception of transmissions by channel attenuation virtual users. [Figure 29] FIG. 1 is a block diagram illustrating the iterations of MPA during the training phase with channel attenuation. [Figure 30A] FIG. 24 is a block diagram illustrating the tandem training of FIG. 23 with the addition of channel attenuation. [Figure 30B] FIG. 23C is a block diagram illustrating the tandem training of FIGS. 23A and 23B with the addition of channel attenuation. [Figure 30C] FIG. 23C is a block diagram illustrating the tandem training of FIGS. 23A and 23B with the addition of channel attenuation. [Figure 31A] FIG. 1 is a block diagram illustrating transmission with channel attenuation while the MPA continues to operate. [Figure 31B] FIG. 1 is a block diagram illustrating transmission with channel attenuation while the MPA continues to operate. [Figure 31C] FIG. 1 is a block diagram illustrating transmission with channel attenuation while the MPA continues to operate. [Diagram 32] FIG. 1 is a block diagram illustrating an iteration of MPA with time-varying channel attenuation during transmission. [Diagram 33] 11 is a plot showing a simulation of one 64QAM encoded at 1 / 3 rate with attenuated uncertainty. [Diagram 34] FIG. 1 is a block diagram illustrating an exemplary configuration of a beam-on-chip. [Diagram 35] FIG. 1 is a block diagram illustrating a base station transmitting a pre-trained and stored decoder g() neural network to a receiver regarding the location of the terminal. [Figure 36A] FIG. 1 is a block diagram illustrating transmission by an OFDM system to achieve high throughput. [Figure 36B] FIG. 1 is a block diagram illustrating transmission by an OFDM system to achieve high throughput. [Figure 37A] FIG. 1 is a block diagram illustrating training with incremental information retransmission. [Figure 37B] FIG. 1 is a block diagram illustrating training with incremental information retransmission. [Figure 37C] FIG. 1 is a block diagram illustrating training with incremental information retransmission. [Figure 38] 13 is a plot showing a simulation of coding gain (Eb / No) with incremental information retransmission. [Figure 39] FIG. 1 illustrates an example of a network for implementing one or more embodiments of the present disclosure. [Figure 40A] FIG. 1 is a block diagram of an exemplary electronic device. [Figure 40B] FIG. 1 is a block diagram of an exemplary electronic device. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] The applicability of systems using a very large number of antennas is based on the law of large numbers (LLN), and even if one attenuated beam is very weak, hundreds or thousands of attenuated beams can accumulate to a signal energy strong enough to reach the receiver sensitivity level. An example is shown in FIG. 1. In the first example 300, when there is a single beam from the antenna to the terminal, significant propagation loss on the single radio beam results in a very low SNR at the terminal. In the first example 302, when there are multiple N beams from the antenna array to the terminal, even if many of the beams are weak due to propagation loss, they can accumulate to result in an acceptable SNR at the terminal. FIG. 2A shows an example of a multi-beam on-chip, generally designated 400. FIG. 2B shows two of the chips 400 are used in a joint two-beam on-chip system, generally designated 402, to produce an acceptable receiver sensitivity. The two chips 400 may be located in a single base station or in separate cooperating base stations.

[0021] Designing such high density beams-on-chip poses several major design, foundry, and scheduling challenges. First, power density becomes very important since the total energy of each chip is limited by semiconductor technology, so the transmit energy of each beam needs to be reduced in exchange for an increase in the number of radio beams on the chip. Second, due to semiconductor technology limitations, routing (clock trees, power lines) is very difficult, so the system needs to tolerate several time offsets between different beams. Third, scheduling algorithms (power control and time advance) are needed for all beams.

[0022] The capacity of a large-scale beamforming system is determined by the number of beams per radio beam (P beam ) and total signal power (P total ) is equally limited by the transmitted signal power. If the total number of available radio beams is N, then P total =N P beam It is.

[0023] In an AWGN channel, this limit for N equally powered radio beams is the power gain. For example, N=2 provides a 3 dB (=10*log10(3)) boost in received signal power, N=3 provides a 4.77 dB (=10*log10(3)) boost, N=100 provides a 20 dB boost, and N=1000 provides a 30 dB boost. In future mmWave applications, there is interest in high throughput applications, so higher order modulation schemes are likely to be used. For example, a single beam can provide a SNR (P beam ), an N=1000 wireless beamforming system can theoretically reduce this SNR to 25 dB (=-5 dB+30 dB) at the receiver sensitivity level, sufficient for the receiver to decode the 64 QAM symbol. Figure 3 is a block diagram illustrating an example transmission of one 64 QAM symbol by 1000 beams using the beam-on-chip configuration of Figure 2.

[0024] However, the real environment is much more complicated than the ideal evenly attenuated AWGN. Not all beams experience equal channel attenuation (distortion), resulting in different channel uncertainties on different paths. Especially in the mmWave band, non-line-of-sight (NLoS) beams typically experience very high attenuation (i.e., few signals on that path reach the receiver). Even among line-of-sight (LoS) beams, their channel attenuations are usually significantly different from each other, in terms of transmission and arrival angles. Furthermore, not all beams reach their common receiver in a synchronous manner. The delays resulting from beam-on-chip and routing of different radio paths are different for each path (or beam). In non-AWGN channels, it is better to use maximum ratio combining (MRC) gain instead of power gain to indicate this capacity limit according to information theory. Roughly speaking, when scheduling N beams, the system should prioritize “good” beams (less channel hostility means that more transmitted signals are more likely to reach the receiver) over “bad” beams (more channel hostility means that more transmitted signals are less likely to reach the receiver). Figure 4 is a block diagram illustrating the propagation of unevenly attenuated beams from the beam-on-chip configuration of Figure 2B. Clearly, to set up an MRC strategy, the transmitter (here the beamforming base station) needs to estimate the channel conditions on each beam before making proportional resource allocation.

[0025] Since the antenna size is a function of the wavelength, the number K of uncorrelated receive antennas that can be accommodated by the terminal (K is the rank in wireless terminology) is related to the wavelength, and this number can be increased when the wavelength is reduced to the cm or mm level in the millimeter wave band. Therefore, the N-beam base station can group its beams into N / K groups, each having K beams. One group has K uncorrelated beams [beam 1, beam 2, …, beam K] targeting K receive antennas respectively. The N / K groups are added to each other at the receiver. FIG. 5 is a block diagram showing an exemplary configuration of a receiver having K = 3 uncorrelated receive dimensions or K independent channel uses (time, space, code, and any other method by which the receiver can independently utilize the channels).

[0026] In the present disclosure, different receive antennas in a multiple-input multiple-output (MIMO) configuration are used as a specific example of independent (uncorrelated) channel use. However, the provided approach is open for any other independent channel use.

[0027] In existing methods used in some MIMO systems, K independent channel uses are estimated by singular value decomposition (SVD) on the m×m covariance matrix (H·H m×n m<n) of the estimated channel using channel measurements (H H ). The K most important eigenvalues determined and ordered by SVD indicate the importance or weight of the K ranks or components of this channel. Based on the determined quality, the modulation and coding scheme (MCS) is adaptively selected for the rank. For example, if rank #1 is the best, a higher modulation and a higher coding rate are selected for this rank. The performance achieved using the K ranks depends on the granularity of the MCS. In a simplified example of transmitting 6 bits along 3 ranks, if there is one rank after SVD, one 64QAM symbol is transmitted at that rank, If there are two ranks (similar to each other), then two 8QAM symbols are transmitted, one on each rank, If there are three ranks (similar to each other), then three QPSK symbols are transmitted, one across each rank.

[0028] FIG. 6 is a block diagram illustrating a general example of a MIMO system that allocates MCS with respect to rank.

[0029] This system is inefficient for several reasons. First, the quality of the ranks must be reliably estimated and tracked. This requires a high pilot density, resulting in significant pilot overhead. In pilot transmission, the transmitter transmits a reference signal on a percentage of the total resources, and the receiver must measure the reference signal at each receiving port and send feedback such as channel state information (CSI) to the transmitter. Second, a large number of MCSs are defined as candidate MCSs (thousands of MCSs are defined in 5G systems), which means that both the BTS and the UE must store these tables, parameters, and how to fetch them. Third, control signaling must be used to ensure that both the transmitter and the receiver follow the same constellation. Both the transmitter and the receiver must know which MCS is used for which rank. More importantly, if the channel attenuation varies to a certain extent, the quality of the ranks can change very significantly, so that all established constellations (MCSs) become outdated.

[0030] In mmWave, the attenuation on the beam is polarized. More specifically, one beam is generated by one set of antennas. This one set of antennas is configured to direct their energy in one specific direction so that the energy can largely compensate for the propagation loss over that direction to form a strong transmission signal in that direction. In reception, this one set of antennas also has the effect of reducing the reception sensitivity. This action is sometimes called polarization. LoS beams have radio propagation loss attenuation with respect to distance, some NLoS beams will be severely attenuated, and other NLoS beams may be positively enhanced by metallic reflections. Generally speaking, the attenuation model (distribution) is sensitive to its geographic location. According to statistics, an attenuation model with respect to spatial location can be available for a coherent time window called coherent time. However, when high frequency bands are used for signal transmission, small mobility leads to severe Doppler shift and time offset. The coherent window time is small.

[0031] A very large number of beams can be used to protect the coding gain from these hostilities. For example, if some beams are blocked (infinite attenuation, dead beams), the system adjusts the transmission power between the "surviving" and "dead" beams to compensate for the loss. Since the Doppler shift and time offset depend on the angle of the moving orbit, those hostilities can be compensated for by adjusting the transmission power between the "favorable" and "unfavorable" beams.

[0032] Figure 7 is a block diagram illustrating how multiple beam transmission can be used to provide immunity to some beams being completely blocked by random obstructions. Figure 8 is a block diagram illustrating the ability of multiple beam transmission to withstand the Doppler effect.

[0033] Power control and synchronization among N beams becomes paramount in mmWave scenarios. However, it is extremely difficult to achieve. First, the attenuation needs to be known. Although the statistical distribution model of the attenuation for a wide location can be known, the attenuation itself is a random variable. To capture the instantaneous attenuation, the system needs to measure the channel of N / K×K, a total of N paths, which means a large number of downlink reference resources (since the receiver only sees K dimensions resulting from N beams overlapping in the air, one receiving dimension needs to measure N / K beams in a round-robin manner to avoid pilot interference), and a corresponding large amount of uplink reporting resources that serve as overhead. Figures 9A and 9B are block diagrams illustrating round-robin transmission for measuring channel coefficients. At measurement time t illustrated in Figure 9A, only K beams are allowed to transmit their pilots in the downlink. At measurement time t+1 illustrated in Figure 9B, the other K beams are allowed to transmit their pilots in the downlink.

[0034] Other methods are based on the assumption that the downlink and uplink channels are reciprocal. However, in practice there is always some offset between the reciprocities. Second, even if the system had perfect knowledge of the channel attenuation, it would be difficult to calculate the proportional maximal ratio combining (MRC) power offsets of N / K beams in a short time interval.

[0035] Performance gains from dimensional expansion coding gain In modern communication systems, redundancy is widely used to overcome channel attenuation and noise. For example, forward error correction (FEC) is a type of dimensional expansion method that not only inserts some redundant bits but also creates some dependency between the coded bits (of the codeword). In another example, a MIMO encoder is a dimensional expansion method that expands some complex symbols onto multiple spatial dimensions to have coding gain. In general, dimensional expansion results in dimensional gain, but such dimensional gain must include power gain, but not necessarily coding gain.

[0036] The power gain is the equal gain combining (EGC) gain in an AWGN channel. In a uniformly attenuated AWGN channel, if information (e.g., one bit or one QAM symbol) is repeated (or linearly combined) over K dimensions, this necessarily results in a power gain of 10*log10(K) dB. As an example, as shown in Figure 10, the EGC gain of a repeated transmission over three uniformly attenuated dimension gains is equal to 4.77 dB (=10*log10(3)).

[0037] Coding gain exceeds EGC gain. For example, encoding one symbol over three dimensions to achieve a 3 dB gain on top of the ECG gain allows the transmitter to either cut the power in half or double the coverage distance. To demonstrate coding gain, it is better to measure the symbol error rate (SER) in terms of the ratio of energy per bit (Eb) to spectral noise density (No) (Eb / No) instead of the ratio of energy per symbol (Es) to spectral noise density (Es / No). A simple repetition (or any linear combination) scheme leads to zero coding gain as shown in Figure 11.

[0038] To have coding gain, conventional channel encoders (FEC) insert (KM) redundant bits for M information bits or expand N (K>M>1) dimensions to K dimensions. In essence, coding gain is related to the coding rate: M / K. Lower coding rates result in more coding gain. When (M=1) dimensions are expanded to K dimensions to a coding rate of 1 / K, coding theory dictates that there is no coding gain when the input dimension is 1 (M=1) in a uniformly attenuated AWGN channel. It would be advantageous to achieve coding gain for such a system in a uniformly attenuated AWGN channel.

[0039] Unlike the MRC gain, the availability of coding gain in AWGN channels is well understood. If there was coding gain for the 1 → K dimensional extension in Figure 12, it would be possible to realize the gain at the system level in more complex channel conditions.

[0040] Diversity Gain The input bits of a 64 QAM symbol can be transmitted by transmitting three independent QPSK symbols, each carrying three bits, using K=3 dimensions. It is well known that QPSK constellations are optimal in terms of coding distance (e.g., Hamming distance) due to both the constant Euclidean distance between any two adjacent constellation points and the constant magnitude of the constellation points. There are other ways in which 6 bits can be transmitted using K=3 dimensions, and the QPSK curve can serve as a theoretical limit for the potential coding gain of using K=3 dimensions.

[0041] FIG. 13 shows a potential coding gain margin of about 8 dB when transmitting 6 bits using K=3 dimensions compared to transmitting one 64QAM symbol in one dimension. The coding gain margin corresponds to a reduction in transmission power of about 16% or an increase in coverage by about 36 times. It should be noted that the example of transmitting 6 bits in one 64QAM symbol using three dimensions was chosen only to make this example comparable to the QPSK constellation, which has been proven to be optimal in a two-dimensional complex Hilbert space. However, this is not the general case, and it remains unproven and unclear what the theoretical limit is at higher modulation orders (such as 512QAM, 1024QAM, or 4096QAM) with any coding rate of 1 / K. The provided method has been proven to be an optimal solution for QPSK, 16QAM. The provided method can also be applied to higher order modulations. It is speculated that the provided method can approach an optimal solution when applied to higher order modulations. This is in contrast to conventional higher order modulation schemes, which are designed for ease of implementation rather than optimal performance.

[0042] Given a modulator order and any coding rate, it cannot be guaranteed that this leads to log2(modulator order) / K being an integer. In the case of transmitting 6 bits using 3 symbols, this corresponds to a code rate of 1 / 3, log2(64) / 3=2 (→QPSK), whereas in the other case of applying a coding rate of 1 / 3 to the input bits (10) of 1024 QAM symbols, this results in log2(1024) / 3=10 / 3 bits per dimension, which is not an integer. In practice, there can be a variety of numbers K of receiving dimensions. For example, for a terminal equipped with up to four receiving antennas, it is very likely that only three receiving antennas remain uncorrelated. In practice, the number of uncorrelated channels available in a given instance is a function of environmental conditions and cannot be easily controlled in an accurate way.

[0043] It would be advantageous to achieve diversity gain or MRC gain in multipath channel conditions. The K uncorrelated dimensions may be subject to K independent different attenuations. Suppose a transmitter splits the 6 bits of one 64QAM symbol into three sets of 2 bits for transmission in three QPSK symbols, and transmits them through the three uncorrelated dimensions, respectively. If one of the three dimensions is subject to severe negative attenuation (weakened) and the other two dimensions are strengthened, the QPSK symbol on the weakened dimension may be lost and it may be impossible to recover the original 64QAM symbol. In contrast, if one 64QAM symbol is somehow encoded into three dimensions, the receiver may be able to recover this 64QAM symbol to a large extent from the remaining two strengthened dimensions by the wireless channel, as shown in FIG. 14. This is the diversity gain or MRC gain.

[0044] Incremental Redundancy Retransmission In wireless systems, IR-HARQ is widely used together with channel codes to increase reliability (or extend coverage) at the expense of average throughput. Some channel codes, such as turbo codes, convolutional codes, and low-density parity check (LDPC), allow the transmitter to transmit a sub-codeword (encoded block) to the receiver in one transmission interval. If the current channel conditions are good enough, the receiver may be able to recover the entire information block even from this sub-codeword, otherwise the receiver may send feedback to the transmitter for a retransmission. Instead of retransmitting what was previously transmitted (Chase-combining retransmission), the transmitter transmits incrementally redundant parts of the codeword so that the receiver can combine the incrementally redundant parts with previously received parts into a longer codeword for its channel decoder. In this way, additional coding gains beyond the pure power (repetition) gains can be realized.

[0045] IR-HARQ is an opportunistic method that is sensitive to time-varying channel conditions (sometimes the channel usage indicates a positive condition, other times a negative condition). If the first transmission is made in a positive condition, retransmission resources are saved. Furthermore, in some latency-critical and reliability-critical applications, such as Ultra-Reliable Low-Latency Communications (URLLC), the feedback cycle may be too long for the latency requirements, so the transmitter tends to do blind retransmissions instead of waiting for feedback. In both scenarios, the most essential or useful information should be included in the incremental retransmission(s) to obtain coding gain.

[0046] So far, incremental redundancy-based retransmission has only been done for some channel coding schemes, e.g., schemes using specific channel codes such as Turbo, Convolutional, Low-Density Parity Check (LDPC), etc. Other dimensionality expansion-based algorithms such as MIMO, spreading codes, etc. have not reported similar retransmissions (most of them only support Chase combining, i.e., power gain).

[0047] With the development of 5G systems, the number of dimensions of channel usage is increasing, such as frequency (subcarriers, different bands), space (MIMO, dual connectivity), time, and code (PN code, pseudorandom code). A good system should be able to integrate multiple available dimensions to result in an efficient overall system gain in terms of the use of available resources. Among the sources of system gain, coding gain is the most sought after in terms of its impact on system spectral efficiency.

[0048] Nevertheless, one of the reasons why only some channel coding schemes can achieve incremental redundancy (IR) retransmission with some coding gain is due to the orthonormality of these dimensions. Such IR-assisted channel coding schemes benefit from their assumption that the M information bits are independent and identically distributed (IID) and the K coding bits are subject to AWGN noise. In other words, the input is an orthonormal M-dimensional entity and the output is also an orthonormal K-dimensional entity. In fact, efforts are usually made to "maintain" this ideal environment for the channel code by using pilots, channel estimation, and equalization. Unfortunately, this is not the case for most dimension expansion algorithms, and the orthonormality of both the input and output at all times cannot be guaranteed. For example, it cannot be guaranteed that two antennas always remain uncorrelated. In another example, two subcarriers may experience very different channel attenuation in a frequency selective channel due to multipath fading or Doppler effect.

[0049] Faced with these difficulties, most prior art systems resort to Maximal Ratio Combining (MRC) schemes to passively overcome diversity. Although an MRC-based receiver can sometimes achieve MRC gain due to diversity transmission, the gain cannot always be guaranteed.

[0050] Machine Learning It has been proposed to use machine learning techniques to solve the above two problems and create a coding scheme to encode M legacy QAM symbols into K dimensions by power-controlled N / K radio beams. The simplest proposal is to use an autoencoder whose input is M QAM symbols and whose output is the decoded QAM symbols, with a latent layer having K dimensions, where attenuation is modeled by multipliers on this K-dimensional latent layer, and AWGN noise is added on this K-dimensional latent layer. The loss or goal of this autoencoder is to minimize the mean squared error (MSE) between the input and the output.

[0051] One problem with this solution is that the decoding neural network (after the K-dimensional latent layer) attempts to mathematically maximize a likelihood probability that depends on the K-dimensional latent layer. However, this K-dimensional latent layer is subject to varying noise and attenuation. Varying SNR and attenuation change the statistical properties of the K-dimensional latent layer, making it necessary for the decoding neural network to be retrained. In practice, it is impractical to retrain a full deep neural network in situations where the SNR or attenuation are time-varying.

[0052] Another problem is that, to have N beams, the autoencoder architecture includes a layer of N neurons immediately preceding the K neuron layer (N>>K), and the connections between the two layers are multiplied by a given damping factor and summed by random noise. In such a deep neural network, training involves "powering down" most of the beams and giving the very few remaining beams a transmit power P total This tends to "crowd" the beams, which is undesirable because it would require each beam to be designed with a much higher maximum transmit power, resulting in higher costs and larger chip die sizes.

[0053] One of the main reasons why deep learning neural networks are difficult to apply in wireless communication applications is that they are nonlinear functions. Compared with traditional linear functions (most wireless signal processing algorithms are linear or quasi-linear), nonlinear functions have a tendency to overfit, which severely limits their generalization.

[0054] Another problem arises from the presence of varying channel attenuation in wireless systems. Wireless channels are time-varying, frequency-selective, and attenuating. Its presence makes the statistical properties of the received signal (the transmitted signal distorted by the channel) also time-varying and frequency-selective. Neurons in deep neural networks need to be adjusted to adapt to the varying and selective channel attenuation, often resulting in complete retraining.

[0055] EMBODIMENT 1 Model of an N-beam pair K-receiving system. Referring now to Figure 15, a model of the transceiver scenario is shown. An N-dimensional transmitter transmits a Hilbert space vector through a hostile environment.

number

[0056] More specifically, the transmitter samples from manifold A to obtain symbol a using sampling method F 1502 and converts symbol a into a vector using encoder 1506.

number

number

number

[0057] space

number

[0058] The transmitter has N independent or uncorrelated transmit dimensions.

number

number

number

number

number

[0059] There are a total of N connections connecting the N transmit dimensions to the K receive dimensions. Each connection has its own attenuation coefficient

number

number

number

number

number

[0060] Similarly, each connection is assigned a power control factor (real value) normalized to avoid transmit power saturation.

number

number

number

number

number

number

number

[0061] In the kth receiving dimension, N / K connections are summed in the air:

number

number

number

number

[0062] AWGN noise n

number

[0063] Note that when K=1, the model reduces from N independent transmit dimensions to one receive dimension. In this case, there is no longer any coding gain, but there may still be a power gain.

[0064] Next, a specific example of the model of Fig. 15 will be described with reference to Fig. 16. In this example, the Hilbert space

number

number

number

number

number

number

number

[0065] Coding gain problem For the purpose of finding f(·) and g(·), the channel attenuation coefficients are fixed (e.g.,

number

number

number

number

[0066] Using the previous example of FIG. 16, a transmitter wishes to transmit one 64QAM symbol to a receiver. The transmitter knows that the receiver has three independent receive antennas. The transmitter has a very large number of beams available for this transmission. It is known that a repetition or other linear extension from 1 symbol to 3 symbols across N beams results in only power gain (EGC) gain in a uniformly attenuated AWGN channel. It is desirable to find a symbol-level coding scheme that can achieve maximum coding gain in addition to EGC gain in a uniformly attenuated K-dimensional AWGN channel.

[0067] Power Control Issues Another problem concerns setting power control for N beams in the context of varying attenuation and time offsets. The attenuation on one beam (h) is proportional to the static (h s ), dynamic uncertainty (Δh d ), and mismatch uncertainty (Δ′h m ), and h=h s +Δh d +Δ'h m The static component has a strong correlation with the spatial location of the terminal and the characteristics of the surroundings, e.g., buildings, trees, mountains, and any fixed environmental factors. Δh d is mainly due to random events such as mobility and blocking, and h s is the random jitter for Δ'h m is the unavoidable measurement uncertainty, and Δh d The random jitter for d and jitter Δ'h m h s It can be assumed that θ is less important than θ. Dynamic power control for the N beams can be used to compensate for the varying attenuation uncertainties. The power control can be tracked and adapted, but has significant coding gain. The problem model is

number

number

number

[0068] The overall solution addresses both the coding and power control problems and can be divided into subtasks as follows: First task:

number

number

number

number

[0069] For high throughput and low latency, it is important to keep the decoder g(·) the same for time-varying channel attenuation and SNR.

[0070] Continuing with the previous example, the base station wants to transmit one 64QAM symbol to the receiver. The base station knows that the receiver has three receive antennas available to the base station. The base station has a large number of N beams available for this transmission. The channel attenuations across the N beams are not uniform, they are time-varying random variables. The base station may estimate the distribution of the channel attenuations, and the conventional thinking is that this is only possible by transmitting pilots. However, this requires a large number of pilots in systems with large dimensionality. It would be advantageous to avoid transmitting a large number of pilots.

[0071] The third problem is time-varying channel attenuation.

number

number

number

number

[0072] Another problem with the above example is that the targeted receiver needs to be updated, and there is no practical and efficient way to do this: this involves transmitting the new coefficients of g() to the receiver again, which consumes significant radio resources.

[0073] Decomposers and recomposers To investigate the pure coding gain problem, we first consider the uniformly attenuated K-dimensional AWGN channel condition, i.e., the channel attenuation as a constant 1 + 0j.

number

number

[0074] For comparison and practical implementation purposes,

number

number

number

number

[0075] Referring now to FIG. 17, the encoder f(·) and the N beams can be viewed as a black-box encoder 1700 that takes one input symbol a and converts it into a K-element complex vector with average energy 1.

number

number

[0076] In some prior art systems (such as those using MIMO precoding)

number

[0077] Referring now to FIG. 18, the black-box “encoder” of FIG. 17 is split into three stages: a non-linear decomposer 1800, a linear encoder 1802, and a linear (re)synthesizer 1804, based on the following two concepts: Symbol a is considered as an information carrier, not the information itself. In the prior art approach, the transceiver uses all means to maximize the received signal power of symbol a. As a result, the conventional transceiver has a power gain. However, in the system described herein, the focus is instead information-oriented, and the emphasis is on the information carried by symbol a at the receiver side, since the decoder g() can perform the decoding. K element

number

number

[0078] Based on these two concepts, the encoder f(·) derives N / K features from the symbol a.

number

number

number

[0079] The encoder f() converts one symbol a into N / K L-dimensional entities

number

number

number

number

number

number

[0080] An N-beam transmitter can combine N / K K-dimensional complex entities into a single beam by simply adding them together.

number

number

[0081]

number

number

number

number

number

[0082] The whole procedure can be viewed as one complete deep neural network 1900, as illustrated in Figure 19. The input sample a is

number

number

number

number

[0083] f(·) and g(·) are nonlinear neural network layers, and w i and

number

number

[0084] Embodiment 2: Output weighting In this embodiment, various techniques are provided for weighting the output of the linear encoder in order to increase the amount of useful information transmitted to the receiver.

[0085] Referring again to FIG. 18, a set of N / K decomposed features

number

number

number

[0086] In the above embodiment, the convergence is

number

[0087] In this embodiment, an additional power control vector

number

number

number

number

[0088] An example implementation is shown in FIG. 20, which is the same as FIG. 18, except that in FIG. 18, the linear combiner did not apply any weighting so that the combined function is the sum of the inputs, whereas in FIG. 20, a set of weights from the power control vector:

number

[0089] Thus, in FIG. 20, the overall procedure can be summarized by:

number

[0090] The purpose is to limit the training

number

number

[0091] How well the power control works depends on the given

number

number

number

number

number

[0092] In wireless systems, it is better for samples to be widely distributed; the larger the average distance between samples, the better they perform against noise. The distribution of feature importance can vary dramatically from one sample (symbol) to another.

[0093] Third embodiment Virtual user According to another embodiment, the system may actually linearly encode the signal rather than directing the entire output to a single receiver.

number

[0094] Weighted Encoder Output

number

number

number

number

number

number

[0095] The entire system can be summarized as follows:

number

[0096] Figures 21A and 21B are both block diagrams of such a system. Figures 21A and 21B are closely based on Figure 20, but instead of power control block 2004 there is a power control block 2104, which applies weights α i to generate an output directed to the true user, and apply weights β i to generate output directed to the virtual user.

[0097] Note that the outputs to the virtual users are not actually sent. The virtual users are constructed to aid in optimization and training.

[0098] When building such a system, the objective is to find the training constraint α i Using ≦1

number

number

[0099] According to one embodiment of the present disclosure, the training consists of two sub-trainings:

number

number

number

number

number

number

number

number

[0100]

number

number

number

number

number

number

[0101] This is illustrated graphically in Figure 22. The upper part of Figure 22 shows N / K

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

number

[0102] This is a typical messaging passing algorithm. In one iteration, the sender computes (3) for all groups in parallel. Then, each transmission is i and (1-α i ) and then adjust the power for

number

number

number

number

[0103] This message passing algorithm is designed to be embedded in deep neural networks (

number

number

[0104] The purpose is to

number

[0105] For gradient descent backpropagation, [α1,α2,…α N / K ] is a constant. MPA is

number

[0106] Layers trained by backpropagation

number

number

[0107] Referring back to the previous example of encoding one 64QAM symbol into three dimensions in an AWGN channel, in a particular implementation, the transmitter executes the message passing algorithm over 100 64QAM symbols in three iterations. To demonstrate the coding gain, the symbol error rate versus Eb / N0 is determined and plotted in FIG. 25. There is a coding gain of about 8 dB in addition to the power gain. In FIG. 25, the curve is very close to the theoretical limit over a wide range of SNRs with no error floor, but in the example it was trained at a fixed Es / N0 of 10 dB.

[0108] It can be seen that the overall system does not suffer from the generalization problem present in traditional autoencoder approaches: if there are three equal quality ranks for 6 bits, the system produces three QPSK symbols. Without any intervening modifications to the decoder g(), the system can approach the optimal solution over a wide range of SNRs.

[0109] This example shows that, given the number of bits to be transmitted and the number of ranks of equal importance available, this deep neural network architecture allows approaching the theoretical limit. In practice, the number of bits to be transmitted can be 5, 7, or any other value, and does not have to be a power of 2 in particular. Also, the number of ranks can be arbitrary. For example, for 7 bits across 5 ranks, traditional methods have no heuristic limit. The deep neural network approaches the optimal solution.

[0110] Embodiment 4: Multipath channel use Detailed Description of Embodiment 3 Although there is some utility in having coding gain over an evenly attenuated (K equal quality ranks) AWGN channel, when the quality of the ranks differs, multi-rank transmission makes more sense. In the prior art approach, the quality of the K ranks must be estimated and ranked by SVD, and one symbol is separated proportionally into K ways in terms of quality. This mechanism involves significant overhead (pilots, messages and control primitives) and a large set of standardized candidates. This is how some current MIMO systems are designed: Step 1: The transmitter transmits several reference signals on different paths; Step 2: The receiver measures the channel conditions on the different paths and combines them into a channel quality indicator (CQI); Step 3: The receiver sends the CQI back to the transmitter; Step 4, the transmitter performs SVD to find eigenvalues ​​and eigenvectors from the CQI; Step 5, the transmitter ranks the eigenvalues ​​from higher to lower and eliminates some very small eigenvalues. According to the eigenvalues ​​(quality of ranks), it allocates modulation and coding schemes. Step 6, the transmitter sends the MCS and decoding matrix (eigenvector) for each rank to the receiver.

[0111] In mmWave application scenarios, the AWGN channel is an oversimplification and a multipath fading, i.e., time-varying, frequency-selective, fading channel can be used instead, which introduces further complications. In reality, N beams reach the K receive antennas independently and their received signals are mutually added or interfered with. The beams may be correlated in some way. Each beam has a different, frequency-selective, time-varying fading attenuation. The selectivity is due to multiple paths on the beams and is mainly determined by the physical environment. Over a period of wireless communication, the attenuation due to selectivity is static:h s The time-varying attenuation is due to the mobility of the moving terminal as well as nearby moving obstructions. The time-varying attenuation can be measured and estimated by the jitter with respect to the static attenuation: Δh d (t). Furthermore, the time-varying channel attenuation can never be estimated accurately, leading to unavoidable estimation mismatches. Alternatively, the time-varying channel attenuation can become stale after a while, leading to other unavoidable stale mismatches. This mismatch can be considered as Δd m Jitter for (t): Δ'd m (t) may be regarded as

[0112] In an N-to-K scenario (N>>K), pilots (reference signals) across N / K beam groups are summed at the receiver, necessitating a round-robin measurement algorithm across N / K beam groups. Unfortunately, for a given beam group, its two consecutive downlink channel measurements are taken at least every (N / K-1) measurement intervals. Since N can be a number much larger than K, this interval is too long and Δh d It is not possible to account for the variation of (t). Furthermore, this method requires significant overhead in both the downlink (pilot and sounding signals) and uplink (reporting and feedback).

[0113] An alternative approach relies on the reciprocal property of UL / DL in TDD mode, i.e., the base station (transmitter) estimates the channel attenuations on the UL signals transmitted from the receiver and then takes them as the channel attenuations on the next DL (transmitter to receiver). This reciprocal assumption is imperfect due to potential mismatches between UL and DL.

[0114] For mmWave applications, relying on UL / DL reciprocity is preferred since it occupies less overhead than round-robin based approaches. In addition, to support extremely high throughput and ultra-low latency, it is preferred to use a simple decoder g(·) instead of the traditional synchronization, channel estimation, and equalization cascade.

[0115] In some embodiments, the attenuation h(t)=(h s +Δh d (t) + Δh m The approach to address (t) is based on the following: resynthesis

number

number

number

number

[0116] For training purposes, the channel attenuation is split into three parts, as illustrated in FIG. Static (large scale) channel attenuation (h s ) 2600 is mainly due to static time offsets between different beam antennas with respect to the receiver's position and multipath on the beam. The base station (transmitter) typically has good knowledge of the static channel attenuation. Dynamic (medium-scale) channel attenuation (Δh d The delay time Δh(t) 2602 is mainly due to the changing environment (e.g., random passengers) and low mobility (e.g., small movements). The base station uses Δh d Continue to measure (t) and use these measurements for the downlink based on the UL / DL reciprocity assumption. Mismatched (small) channel attenuation (Δh m The (t)) 2604 is mainly due to the mismatch between the channel estimates and the measurements, which can be the mismatch between the UL and DL channels when reciprocity is assumed, or a mismatch that accumulates over time.

[0117] These dampings can be included in the system model as follows:

number

number

number

number

number

[0118]

number

number

number

number

number

number

number

[0119] Next, in the above model,

number

number

number

number

number

[0120] The system of FIG. 20 modified to account for channel attenuation is illustrated in FIGS. 27A and 27B, where the channel attenuation component is introduced at 2700.

[0121] Figures 28A, 28B, and 28C show the system of Figures 21A and 21B including a virtual user, which is defined as an ideal user with no negative or positive attenuation. The systems of Figures 28A, 28B, and 28C show channel attenuation 2800 applied to the output of only the real users. The overall neural network is as follows:

number

number

[0122] During training, the following approach is taken:

number

number

number

number

number

number

number

number

number

[0123] The above example assumes no attenuation for the virtual users, while in other embodiments the same attenuation of the real users is assumed for the virtual users. Simulations show that there is no significant performance difference between the two approaches.

[0124] The training is the same as described above for the AWGN channel.

number

number

number

[0125] After training, f(), g(), w i and

number

number

number

number

number

number

[0126] Figures 30A, 20B, and 30C show how training is split between backpropagation and MPA, as described above for the embodiments of Figures 23A and 23B. As shown in Figures 31A, 31B, and 31C, the gradient descent layer is frozen while the MPA is used for transmission to update the power control coefficients.

[0127] Following training, the following procedure is taken.

number

number

number

number

number

[0128] The behavior of the MPA after the training phase during transmission with attenuation is summarized in FIG.

[0129] Returning to our example of a single 64QAM encoded in three dimensions, in the first simulation,

number

number

number

number

number

number

number

number

[0130] In the second simulation,

number

number

number

number

number

number

number

number

number

[0131] In a randomly attenuated channel, the virtual users (without attenuation) are assigned power control coefficients [α1,α2,…α N / K For each group (containing K beams), MPA has two connections, one for the real users with attenuation and the other for the virtual users without attenuation. The objective of MPA is to move maximum information to the real users and the rest to the virtual users.

[0132] More importantly, in the N-to-K scenario, as mentioned before, it is difficult or impractical to insert pilots in each beam, since all beams are summed at the receiver (the pilots of the beams interfere with each other). With the described technique, no pilots are needed (or much less are needed). All the varying attenuation is handled by the

number

[0133] Embodiment 5: Application to millimeter waves In some embodiments, one or more of the approaches described above are adapted for mmWave applications.

[0134] In some embodiments, the base station includes a beam-on-chip system, where one chip includes many (e.g., hundreds or thousands) of mmWave antennas.

[0135] To address routing issues that can lead to time offsets between antennas, several nearby antenna circuits on a chip can be grouped into a set to generate a beam. Antennas of one beam can be connected to the same clock tree node, so they are chip-wise and physically close and synchronized with each other. To address power density, antennas of one beam can be selected from within the same power island. An example showing a set of K adjacent beams grouped together is shown in FIG. 34. For a given beam group, their clocks are derived from the same clock tree node and they are in the same power island. For one power island, there is a power control tuner that controls all K beams on this island. In FIG. 34, one group is shown at 3400, which has a power control tuner 3402 that controls all K beams and has a clock tree 3404 connected to all beams. The K beams of one group are always scheduled together to point to the same receiver. They are close to each other, but are "synchronized", uniformly and evenly powered, and spatially parallel (no interference). A beam group is the basic scheduling granularity. One beam-on-chip may contain hundreds or thousands of groups of beams. When a large number of beams are implemented on one chip or system-on-chip (SoC), the traditional method of synchronizing all their clocks presents challenges to chip design, especially when designing the routing clock tree. As a result, a large fully synchronous chip will consume a significant portion (more than 30%) of the power for the routing clock tree. Also, the large power consumption inevitably increases the area of ​​the chip for ventilation purposes. Therefore, in chip design, if a large fully synchronous clock tree can be split into several asynchronous sub-clock trees, this will dramatically reduce the power and area and cost. This is because the routing problem can be significantly reduced since the time offset due to chip routing between these groups of beams is tolerated. However, the time offset between the groups varies from chip to chip.They are fixed and can be measured. Static time offsets are:

number

[0136] Multiple beams-on-chips may be used together to arrive at simultaneous N / K beam groups. The time offset between the beams-on-chips is static and is measured and then

number

number

[0137] An example is shown in FIG. 35, showing a base station 3500 and two terminals including a first terminal 3502 located in spot A 3503 and a second terminal 3504 located in spot B 3505 .

[0138] The location of the terminal is known to the base station. Its location, user type, and the K available uncorrelated dimensions,

number

number

[0139] In operation, the receiver begins by transmitting a reference signal on the uplink that is detected / received by the N beams of the base station. The base station then estimates the channel attenuation on each beam.

number

number

number

number

number

number

[0140] In operation, the base station derives N(m hd ,σ hd ) to adjust the distribution of

number

number

[0141] The base station is

number

number

number

number

[0142] Gamma d If (t) exceeds a predefined limit and the user remains in position,

number

number

number

number

number

[0143] Gamma m If (t) exceeds a given limit, then more frequent pilots and / or more pilots are needed. If more frequent and more pilots do not help the situation, then N beams may be insufficient and more beams can be allocated to this receiver. Since the dimensions have changed, new training is needed.

[0144] Although the detailed examples focus on one symbol, in some embodiments OFDM is used and multiple symbols are transmitted in parallel by the OFDM system. Each symbol has its own

number

[0145] Embodiment 6 Incremental Redundancy In this embodiment, an incremental redundancy based retransmission system and method are provided. Retransmission can be a general concept that is not limited to the time domain. "Retransmission" means that incremental information is transmitted from another independent dimension. For example, the first transmission is on three dimensions, and the retransmission is on the fourth dimension. Although multiple retransmissions may be performed, the following example uses one retransmission as an example. In practice, three, four, or more retransmissions may be performed.

[0146] We define the maximum number of dimensions to be K. In the first transmission, the first (K-1) dimensions are used, and in the second transmission (first retransmission), incremental information is sent on the Kth dimension. The deep neural network is almost the same as the one without retransmissions, except that there are two decoders (DNNs): g1() for the first (K-1) dimensions, and g2() for the Kth dimension, which combines the Kth dimension with the previous K-1 dimensions.

[0147] The purpose of the training is to

number

[0148] After training, the receiver has two decoders, g1() and g2(). During the first transmission, the receiver receives a K-1 dimensional signal and inputs it into g1() to estimate the symbol a. During the retransmission, the receiver receives a K dimensional signal and combines it with the previous K-1 dimensions into a K dimensional vector for g2(). Training can be used to control the likelihood that the first transmission will be successfully decoded. By setting the parameter γ higher, more essential information is pushed into the signal processed by g1().

[0149] In a specific simulation example, γ=0.8 for initial transmission, and (1-γ)=0.2 is used for retransmission. The simulation calculation results are shown in FIG.

[0150] The results show a coding gain of about 0.3 dB from incremental information on one dimension.

[0151] 39 illustrates an exemplary communication system 100 in which embodiments of the present disclosure may be implemented. In general, the communication system 100 allows multiple wireless or wired elements to communicate data and other content. The purpose of the communication system 100 may be to provide content (voice, data, video, text) via broadcast, narrowcast, user device to user device, etc. The communication system 100 may operate by sharing resources such as bandwidth.

[0152] In this example, communications system 100 includes electronic devices (EDs) 110a-110c, radio access networks (RANs) 120a-120b, a core network 130, a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160. Although a particular number of these components or elements are shown in Figure 39, communications system 100 may include any reasonable number of these components or elements.

[0153] The EDs 110a-110c are configured to operate, communicate, or both in the communication system 100. For example, the EDs 110a-110c are configured to transmit, receive, or both over wireless or wired communication channels. Each ED 110a-110c represents any suitable end-user device for wireless operation and may include (or be referred to as) a device such as a user equipment / device (UE), a wireless transmit / receive unit (WTRU), a mobile station, a fixed or mobile subscriber unit, a mobile telephone, station (STA), a machine type communication (MTC) device, a personal digital assistant (PDA), a smartphone, a laptop, a computer, a tablet, a wireless sensor, or a home electronic device.

[0154] In Figure 39, the RANs 120a-120b include base stations 170a-170b, respectively. Each base station 170a-170b is configured to wirelessly interface with one or more of the EDs 110a-110c to enable access to any other base stations 170a-170b, the core network 130, the PSTN 140, the Internet 150, and / or other networks 160. For example, the base stations 170a-170b may include (or may be) one or more of several well-known devices, such as a base transceiver station (BTS), a Node B (NodeB), an evolved Node B (eNodeB), a home eNodeB, a gNodeB, a transmission point (TP), a site controller, an access point (AP), or a wireless router. Any of the EDs 110a-110c may alternatively or additionally be configured to interface with, access, or communicate with any other base stations 170a-170b, the Internet 150, the core network 130, the PSTN 140, other networks 160, or any combination thereof. The communication system 100 may include a RAN, such as the RAN 120b, with a corresponding base station 170b accessing the core network 130 via the Internet 150, as shown. Although the detailed embodiments described herein refer to a TP, more generally, any type of base station may be used in any of the embodiments described herein.

[0155] The EDs 110a-110c and the base stations 170a-170b are examples of communication equipment that may be configured to implement some or all of the functions and / or embodiments described herein. In the embodiment shown in FIG. 39, the base station 170a forms part of the RAN 120a and may include other base stations, base station controller(s) (BSC), radio network controller(s) (RNC), relay nodes, elements, and / or devices. Any of the base stations 170a, 170b may be a single element as shown, multiple elements distributed within the corresponding RAN, or otherwise. The base station 170b also forms part of the RAN 120b and may include other base stations, elements, and / or devices. Each base station 170a-170b transmits and / or receives radio signals within a particular geographic region or area, sometimes referred to as a "cell" or "coverage area." The cells may be further divided into cell sectors, and base stations 170a-170b may serve multiple sectors, for example, using multiple transceivers. In some embodiments, picocells or femtocells may be established if the radio access technology supports such cells. In some embodiments, multiple transceivers may be used for each cell, for example, using multiple-input multiple-output (MIMO) techniques. The number of RANs 120a-120b shown is for illustrative purposes only. Any number of RANs may be contemplated in conceiving communication system 100.

[0156] The base stations 170a-170b communicate with one or more of the EDs 110a-110c over one or more air interfaces 190 using wireless communication links, e.g., radio frequency (RF), microwave, infrared, etc. The air interface 190 may utilize any suitable wireless access technology. For example, the communications system 100 may implement one or more channel access methods in the air interface 190, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal FDMA (OFDMA), single carrier FDMA (SC-FDMA), etc.

[0157] The base stations 170a-170b may implement Universal Mobile Telecommunications System (UMTS) Terrestrial Radio Access (UTRA) to establish the air interface 190 using Wideband Code Division Multiple Access (WCDMA). In doing so, the base stations 170a-170b may implement protocols such as HSPA, HSPA+, optionally including HSDPA, HSUPA, or both. Alternatively, the base stations 170a-170b may establish the air interface 190 with Evolved UMTS Terrestrial Radio Access (E-UTRA) using LTE, LTE-A, LTE-B, and / or New Radio (NR). It is contemplated that the communication system 100 may employ multiple channel access capabilities, including those schemes described above. Other wireless technologies for implementing the air interface include IEEE 802.11, IEEE 802.15, IEEE 802.16, CDMA2000, CDMA2000 1X, CDMA2000 EV-DO, IS-2000, IS-95, IS-856, GSM, EDGE, and GERAN. Of course, other multiple access schemes and wireless protocols may be utilized.

[0158] The RANs 120a-120b communicate with the core network 130 to provide various services, such as voice services, data services, and other services, to the EDs 110a-110c. The RANs 120a-120b and / or the core network 130 may communicate directly or indirectly with one or more other RANs (not shown), which may or may not be served directly by the core network 130 and which may or may not use the same radio access technology as the RANs 120a, 120b, or both. The core network 130 may also serve as a gateway access between (i) the RANs 120a-120b and / or the EDs 110a-110c, and (ii) other networks, such as the PSTN 140, the Internet 150, or other networks 160. In addition, some or all of the EDs 110a-110c may include functionality for communicating with various wireless networks over various wireless links using various wireless technologies and / or protocols. Instead of (or in addition to) wireless communication, the EDs may communicate via wired communication channels to a service provider or switch (not shown) and to the Internet 150. The PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). The Internet 150 may include computer networks and / or subnets (intranets) and may incorporate protocols such as IP, TCP, UDP, etc. The EDs 110a-110c may be multi-mode devices capable of operation with multiple wireless access technologies and may incorporate multiple transceivers necessary to support such technologies.

[0159] 40A and 40B show example devices that may implement the methods and teachings of the present disclosure. In particular, 40A shows an example ED 110, and 40B shows an example base station 170. These components can be used in the communication system 100 or in any other suitable system.

[0160] As shown in FIG. 40A, the ED 110 includes at least one processing unit 200. The processing unit 200 implements various processing operations of the ED 110. For example, the processing unit 200 may perform signal coding, data processing, power control, input / output processing, or any other function that enables the ED 110 to operate in the communication system 100. The processing unit 200 may also be configured to implement some or all of the functions and / or embodiments described in more detail above. Each processing unit 200 includes any suitable processing or computing device configured to perform one or more operations. Each processing unit 200 may include, for example, a microprocessor, a microcontroller, a digital signal processor, a field programmable gate array, or an application specific integrated circuit.

[0161] The ED 110 also includes at least one transceiver 202. The transceiver 202 is configured to modulate data or other content for transmission by at least one antenna or network interface controller (NIC) 204. The transceiver 202 is also configured to demodulate data or other content received by at least one antenna 204. Each transceiver 202 includes any suitable structure for generating signals for wireless or wired transmission and / or processing signals received wirelessly or wired. Each antenna 204 includes any suitable structure for transmitting and / or receiving wireless or wired signals. One or more transceivers 202 may be used in the ED 110. One or more antennas 204 may be used in the ED 110. Although illustrated as a single functional unit, the transceiver 202 may also be implemented using at least one transmitter and at least one separate receiver.

[0162] The ED 110 further includes one or more input / output devices 206 or interfaces (such as a wired interface to the Internet 150). The input / output devices 206 enable interaction with a user or other devices in a network. Each input / output device 206 includes any suitable structure for providing information to or receiving information from a user, such as a speaker, microphone, keypad, keyboard, display, touch screen, etc., including network interface communications.

[0163] In addition, the ED 110 includes at least one memory 208. The memory 208 stores instructions and data used, generated, or collected by the ED 110. For example, the memory 208 can store software instructions or modules executed by the processing unit(s) 200 configured to implement some or all of the above-described functions and / or embodiments. Each memory 208 includes any suitable volatile and / or non-volatile storage device(s) and retrieval device(s). Any suitable type of memory may be used, such as random access memory (RAM), read-only memory (ROM), hard disk, optical disk, subscriber identity module (SIM) card, memory stick, secure digital (SD) memory card, etc.

[0164] As shown in FIG. 40B, the base station 170 includes at least one processing unit 250, at least one transmitter 252, at least one receiver 254, one or more antennas 256, at least one memory 258, and one or more input / output devices or interfaces 266. A transceiver, not shown, may be used instead of the transmitter 252 and the receiver 254. A scheduler 253 may be coupled to the processing unit 250. The scheduler 253 may be included in the base station 170 or may operate separately from the base station 170. The processing unit 250 implements various processing operations of the base station 170, such as signal coding, data processing, power control, input / output processing, or any other function. The processing unit 250 may also be configured to implement some or all of the functions and / or embodiments described in more detail above. Each processing unit 250 includes any suitable processing or computing device configured to perform one or more operations. Each processing unit 250 may include, for example, a microprocessor, a microcontroller, a digital signal processor, a field programmable gate array, or an application specific integrated circuit.

[0165] Each transmitter 252 includes any suitable structure for generating signals for wireless or wired transmission to one or more EDs or other devices. Each receiver 254 includes any suitable structure for processing signals received wirelessly or wired from one or more EDs or other devices. Although illustrated as separate components, at least one transmitter 252 and at least one receiver 254 may also be coupled to a transceiver. Each antenna 256 includes any suitable structure for transmitting and / or receiving wireless or wired signals. Although a common antenna 256 is illustrated here as being coupled to both the transmitter 252 and the receiver 254, one or more antennas 256 may also be coupled to the transmitter(s) 252, and one or more separate antennas 256 may also be coupled to the receiver(s) 254. Each memory 258 includes any suitable volatile and / or non-volatile storage device(s) and retrieval device(s), such as those described above in connection with the ED 110. The memory 258 stores instructions and data used, generated, or collected by the base station 170. For example, the memory 258 may store software instructions or modules executed by the processing unit(s) 250 configured to implement some or all of the functions and / or embodiments described above.

[0166] Each input / output device 266 allows for interaction with a user or other devices in the network. Each input / output device 266 includes any suitable structure for providing information to or receiving / providing information from a user, including network interface communications.

[0167] Further details regarding the ED 110 and the base station 170 are known to those skilled in the art, and therefore these details are omitted herein for clarity.

[0168] Numerous modifications and variations of the present disclosure are possible in light of the above teachings, and it is therefore to be understood that, within the scope of the appended claims, the present disclosure may be practiced otherwise than as specifically described herein. [Explanation of symbols]

[0169] 100 Communication Systems 110a Electronic Devices (ED) 110b ED 110c ED 120a Radio Access Network (RAN) 120b RAN 130 Core Network 140 Public Switched Telephone Network (PSTN) 150 Internet 160 other networks 170 base station 170a base station 170b base station 190 Air Interface 200 Processing Units 202 Transceiver 204 Antenna or Network Interface Controller (NIC) 206 Input / Output Devices 208 Memory 250 Processing Units 252 Transmitter 253 Scheduler 254 Receiver 256 Antennas 258 memory 266 Input / Output Devices or Interfaces 400 Beam-on-chip 402 Joint 2-Beam On-Chip System 1500 Manifold A 1502 Sampling Method F 1504 Sample "a" 1506 Encoder 1510 The first K transmission dimensions 1512 Attenuation 1514 Power Control Factor Application 1516 Add Noise 1520 Received Vector a 1550 Decoder 1700 Black Box Encoder 1702 Black Box Decoder 1800 Nonlinear Decomposer 1802 Linear Encoder 1804 Linear (re)synthesizer 1900 Deep Neural Networks 2104 Power Control Block 2600 Static (large scale) channel attenuation 2602 Dynamic (medium scale) channel attenuation 2604 Mismatched (small) channel attenuation 2700 Introducing channel attenuation components 2800 Applying Channel Attenuation 3400 Group 3402 Power Control Tuner 3404 Clock Tree 3500 base station 3502 First Terminal 3503 Spot A 3504 Second Terminal 3505 Spot B 3510 Database< / h> < / h> < / h> < / c> < / c>

Claims

1. 1. A multi-beam transmission method for transmitting using an N-beam transmitter to a receiver having K receive beams, the multi-beam transmission method comprising: training a machine learning block and a linear encoder using gradient descent backpropagation relying on feedback from the receiver, the machine learning block implementing a non-linear encoder; For each input to be sent, processing inputs using the machine learning block to generate a set of N / K outputs; performing linear encoding on each set of outputs using the linear encoder to generate a respective set of K outputs to generate N / K sets each having a total of K encoded outputs and N / K sets each having a total of N encoded outputs; and transmitting respective sets of the N / K sets, each having K outputs from a respective set of K beams.

2. The method of claim 1 , wherein the machine learning block is a deep neural network.

3. The method of claim 1 or 2, wherein the step of training the linear encoder is performed for each input to be transmitted. 【Request 4】 【Number 1】 where "a" is the input to be transmitted, f(a) is the nonlinear encoding operation performed by the nonlinear encoder; [0025] is the output of the non-linear encoding operation, w i and [0030] are the parameters of the linear encoder, [0045] is the set of K recombined entities generated at the output of the N-beam transmitter, [0050] is the noise, [006] is the non-linear decoding operation performed by a machine learning block in the receiver, [0070] is the input to the non-linear decoding operation, 4. The method according to any one of claims 1 to 3.

5. applying a respective weight to each of the N / K sets having K outputs each to generate a weighted set of K outputs. Further comprising: transmitting each of the N / K sets, each having K outputs from each beam, comprises transmitting each of the weighted sets of K outputs; The method further comprises training the weights using a message passing algorithm.

5. The method according to any one of claims 1 to 4. 【Request 6】 【Number 8】 where "a" is the input to be transmitted, f(a) is the nonlinear encoding operation performed by the nonlinear encoder; [0090] is the output of the non-linear encoding operation, w i and [0089] are the parameters of the linear encoder, ##EQU00011## is the set of K recombined entities generated at the output of the N-beam transmitter, ##EQU00012## is the noise, ##EQU00013## are the weights that are trained using the message passing algorithm, ##EQU14## is the non-linear decoding operation performed by a machine learning block in the receiver, ##EQU00015## is the input to the non-linear decoding operation, The method of claim 5.

7. Further, the step of applying a respective weight to each of the N / K sets, each having K outputs, to generate a weighted set of K outputs may further comprise applying a respective weight α i applying Based on the optimization criterion, weight 1-α i using a message passing algorithm to optimize the amount of useful information at the output of the linear encoder contained in the transmitted signal compared to the amount of useful information at the output of the linear encoder contained in other virtual signals calculated using i and determining:

8. During initial training, While training the machine learning block and the linear encoder, the weights α i and fixing the Weight α i treating parameters of the machine learning block and the linear encoder as constants while training following training, using a message passing algorithm to update the weights for each transmission; 8. The method of claim 7, further comprising:

9. During initial training, while training the machine learning block and the linear encoder, the weights α i The step of fixing performing said initial training based on an estimate of channel attenuation=an estimate of static channel attenuation+a probability distribution of dynamic attenuation+a probability distribution of mismatched channel attenuation; and subsequent to training, using a message passing algorithm to update the weights for each transmission based on updated values ​​of channel attenuation.

10. The method of claim 1, further comprising the steps of: driving each group of K antennas from the same clock tree; powering each group of K antennas with a respective uniform power; 10. The method of claim 1, further comprising:

11. storing training results of the machine learning block and the linear encoder for each of a plurality of different portions of a coverage area of ​​the N-beam transmitter, the training results including parameters of a nonlinear encoding operation of the N-beam transmitter and parameters of the linear encoder of the N-beam transmitter and parameters of a nonlinear decoder located within the portions of the coverage area; determining a position of the receiver; Further comprising:

10. The method of claim 9, wherein using the machine learning block and the linear encoder comprises using the stored training results based on the determined position of the receiver.

12. transmitting the parameters of the non-linear decoder for the determined position to the receiver.

12. The method of claim 11, further comprising:

13. determining, based on an analysis of dynamic attenuation compared to static attenuation, that a value of static attenuation is not representative of the portion of the coverage area; updating said value of static damping to a more recent estimate of static damping; retraining the non-linear encoder and the linear encoder using the more recent estimates; if the retraining does not converge, increasing the number of beams and retraining the nonlinear encoder and the linear encoder on the increased number of beams.

13. The method of claim 12, further comprising:

14. Increasing the number and / or frequency of pilots based on an analysis of the mismatch attenuation compared to the static attenuation.

13. The method of claim 12, further comprising:

15. determining additional beams based on further analysis of mismatch attenuation compared to static attenuation after increasing the number and / or frequency of said pilots; Increasing the number of beams; retraining the nonlinear encoder and the linear encoder for the increased number of beams; 15. The method of claim 14, further comprising:

16. Transmitting a respective set of N / K sets, each having K outputs from a respective set of K beams, comprises: transmitting T of the respective sets of K outputs from T beams of the respective set of K beams in a first transmission, where T<K; A method according to any one of claims 1 to 15, comprising: in a second transmission, transmitting the remaining K-T of the respective sets of K outputs from the remaining K-T beams of the respective sets of K beams.

17. 1. An apparatus for transmitting to a receiver having K receive beams, the apparatus comprising: A processor and a memory, The processor and the memory implement a machine learning block having a non-linear encoder and implement a linear encoder; The processor and the memory training the machine learning block and the linear encoder using gradient descent backpropagation relying on feedback from the receiver; For each input to be sent, processing inputs using said machine learning block to generate a set of N / K outputs; configured to linearly encode each set of outputs using the linear encoder to generate a respective set of K outputs, generating N / K sets each having a total of K encoded outputs and N / K sets each having a total of N encoded outputs; The apparatus further comprises an N-beam transmitter for transmitting respective sets of the N / K sets, each having K outputs from a respective set of K beams.

18. apply respective weights to each of the N / K sets, each having K outputs, to generate a weighted set of K outputs. It is further structured as follows: transmitting a respective one of the N / K sets, each having K outputs from a respective beam, includes transmitting the respective weighted sets of K outputs; The apparatus of claim 17 , wherein the apparatus is further configured to train the weights using a message passing algorithm.

19. The i-th set of the N / K sets, each having K outputs, is assigned a respective weight α i Applying Based on the optimization criterion, weight 1-α i using a message passing algorithm to optimize the amount of useful information at the output of the linear encoder contained in the transmitted signal compared to the amount of useful information at the output of the linear encoder contained in other virtual signals calculated using i To determine 20. The apparatus of claim 18, configured to apply a respective weight to each of the N / K sets, each having K outputs, to generate a weighted set of K outputs by:

20. Driving each group of K antennas from the same clock tree; Power each group of K antennas with a uniform power.

20. The apparatus of any one of claims 17 to 19, further configured to:

21. In a first transmission, transmit T of the respective set of K outputs from T beams of the respective set of K beams, where T<K; transmitting the remaining K-T of the respective sets of K outputs from the remaining K-T beams of the respective sets of K beams in a second transmission.

21. The apparatus of claim 17, configured to transmit respective sets of the N / K sets, each having K outputs from a respective set of K beams, by:

22. An apparatus comprising a processor configured to cause the apparatus to perform a method according to any one of claims 1 to 16.

23. A computer program comprising instructions which, when executed by a computer, cause the computer to perform a method according to any one of claims 1 to 16.

Citation Information

Patent Citations

  • Initial synchronization method for spread spectrum communication and its equipment

    JP1997284175A

  • Device for compensating for communication distortion and compensating method

    JP2003078454A

  • Method, base station, system and computer program for massive MIMO communications

    JP2022520901A

  • Deep learning based acceleration for iterative tomographic reconstruction

    US20180197317A1