A hybrid deep learning neural network based on message passing algorithm for large-scale beamforming systems

Abstract

A multi-beam transmission method using an N-beam transmitter to transmit to a receiver with K receive beams. In the transmitter, a linear encoder and a nonlinear encoder implemented by a machine learning block are trained using gradient descent backpropagation dependent on feedback from the receiver. For each input to be transmitted, the machine learning block processes the input to generate N / K groups of L outputs. Each group of L outputs is linearly encoded using the linear encoder to generate groups of K outputs, thereby generating N / K groups of K encoded outputs and N encoded outputs in total. One of the N / K groups of K outputs is transmitted from each group of K beams. To allow generalization to different signal-to-noise ratios (SNRs) and to accommodate time-varying channel conditions, the nonlinear layers implemented in the nonlinear encoder are responsible for extracting and regenerating features, while the linear layers implemented in the linear encoder are responsible for stronger generalization.

Description

[0001] Cross-references

[0002] This application claims priority to U.S. patent application Ser. No. 16 / 937,863, filed on July 24, 2020, entitled “A Hybrid Deep Learning Neural Network Based on Message Passing Algorithm for Large Beamforming Systems,” the contents of which are incorporated herein by reference. Technical Field

[0003] The present application relates to machine learning, and more particularly, to machine learning applied to beamforming systems. Background Art

[0004] Deep Learning for Wireless Systems

[0005] Deep learning is a subset of machine learning in which there are more than two layers of nonlinear neurons between the input and output. Deep learning is a branch of artificial intelligence that has made significant progress in its application to very complex problems.

[0006] When deep learning techniques are applied to wireless systems, there are several fundamental problems. The first is the inability to generalize to different signal-to-noise ratios (SNRs). Deep neural networks are closely related to the statistical properties (such as expectation and variance) and random adversarial properties of the information data (i.e., the signal) in wireless systems. Once the SNR changes, the neurons in the deep neural network need to be adjusted to adapt to the different noise, which often requires complete retraining. In practice, because wireless channels, including signal-to-noise ratio and signal-to-interference ratio, are variable and difficult to reliably estimate, it is impractical to use deep learning neural networks to process time-varying signals.

[0007] Wireless mmWAVE communications

[0008] 5G proposes to use millimeter wave (mmWAVE) communications to provide throughput exceeding Gbps. Strictly speaking, 5G mmWAVE communications can use centimeter or millimeter wavelengths.

[0009] Typical mmWave (mmWave) or high-frequency communications have different channel conditions and geography than sub-6 GHz. Path loss in mmWave bands is significantly more severe than in sub-6 GHz. To address this, 5G systems employ a technique called beamforming, which uses multiple transmit antennas to "focus" signal energy onto a single point or area (at the centimeter or millimeter level). This generates a sufficiently high signal-to-noise ratio (SNR) signal within the target point or area. In wireless communication, the signal strength must meet the receiver's sensitivity level.

[0010] However, the higher the frequency band used by wireless systems, the greater the path loss experienced by the radio beam. Future wireless systems may utilize frequency bands exceeding 100 GHz. The receive sensitivity of future radio beams will be significantly limited by the path loss over the air. To compensate for path loss, one option is to increase the transmit power within the radio beam. On the one hand, using a higher frequency band results in a shorter wavelength and smaller antenna size; on the other hand, the greater path loss resulting from a higher frequency band increases the signal receive sensitivity level within each radio beam.

[0011] To address these issues, on-chip beamforming systems can be used, which simultaneously accommodate many (e.g., thousands) of radio beams within a single chip die area. When using artificial intelligence for beamforming in such systems, retraining is often required when the receiver moves. Retraining may not be fast enough to keep up with the changing channel, especially with a large number of beams. Furthermore, the signaling overhead associated with training is significant. Summary of the Invention

[0012] To generalize across different signal-to-noise ratios (SNRs), linear layers are inserted into the autoencoder framework. These linear layers enhance generalization across different SNRs. In the proposed deep learning architecture, the nonlinear encoding layer f(·) acts as a feature extractor, extracting N / K distinct features from an image or a distribution of symbols or samples. The linear layer is an N / K phase (w), bias (b), and power (c) linear encoder. These N / K features can be viewed as N / K different perspectives or observations of a symbol, and the linear layer encodes these N / K features in parallel. N radio signals naturally superimpose these N / K distinct encoded features over the air, forming a K-level encoder. The nonlinear decoding layer g(·) is a fully connected synthesizer that regenerates the symbol from the K-dimensional signal after distortion and noise superposition.

[0013] To allow generalization to different SNRs and time-varying channel conditions, the nonlinear layer is responsible for extracting features and regenerating them, while the linear layer is responsible for stronger generalization. The nonlinear layer itself can have low tolerance to time-varying channel uncertainties such as SNR and attenuation, which usually require a complete retraining of the nonlinear layer when such conditions change. Including the linear layer has the following effects on the entire system: it introduces additional tolerance (stronger generalization) and allows the linear layer to easily adjust to changes in channel uncertainties such as SNR or attenuation without having to retrain the linear layer.

[0014] 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 overall solution can be close to the optimal solution. In some embodiments, the system is receiver transparent.

[0015] In some embodiments, a message passing algorithm (MPA) or a belief propagation algorithm is applied before each transmission to adjust the power control factor.

[0016] For varying channel attenuation: In some embodiments, adding an additional power control layer allows varying channel attenuation to be taken into account. During initial training, a tandem training is performed, which includes performing a power control on each layer (including the nonlinear layer and the linear coding layer).<f(),g(),w,b> )) and back propagation for power control <c>The message passing algorithm is used to apply MPA. The concept of virtual users is introduced. Useful information about the symbol is sent to the "real" user, while useless information is directed to the "virtual" user. During inference (transmission), before each symbol transmission, the MPA at the transmitter is run to adjust the current input symbol and the current channel attenuation. <c>, but other encoding layers (<f(),g(),w,b> ) remains unchanged. During training, attenuations are sampled from the distribution. During actual transmission, the distribution is updated using the estimated channel coefficients with some uncertainty.

[0017] In order to solve the channel mismatch and uncertainty, the attenuation is set in both the initial training and inference MPA. <h>Inserted into this deep neural network, <h>are samples from each distribution. The nonlinear decoding layer g(·) has the function of learning these <h>In fact, if a real attenuation falls within the distribution, or even falls a little outside the distribution, g() can still handle this case. In some cases no pilot is needed, or alternatively, fewer pilots are needed than in other cases.

[0018] In some embodiments, the provided methods are applied to implement large-scale beamforming systems.

[0019] In some embodiments, the provided systems and methods are also adapted to provide coding gain and incremental redundancy hybrid automatic repeat request (IR-HARQ). It is well known that encoding a symbol using a code rate in coding principles yields power gain, not coding gain. In additive white Gaussian noise (AWGN) channels, this code rate only yields power gain, while in selective channels, some MRC gain may result. A system and method are provided to implement a codec that encodes a symbol at a given code rate to achieve coding gain.

[0020] Selective fading is hostile. The provided codec has built-in logic to learn large-scale fading and tolerate small-scale random fading, thus eliminating the need for pilots or requiring fewer pilots.

[0021] Incremental redundancy is effective for linear systems. A system and method are provided for utilizing IR transmission through the provided codec to include IR coding gain. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Embodiments of the present invention will be described below with reference to the accompanying drawings.

[0023] Figure 1 is a block diagram depicting a large number of beams to compensate for significant path loss in the mmWAVE band;

[0024] Figure 2A is a block diagram illustrating an on-chip beam configuration;

[0025] Figure 2B is a block diagram illustrating a joint 2-on-chip beam configuration for transmitting high-quality signals;

[0026] Figure 3 Describes the use of Figure 2B Block diagram of an example transmission of a 64QAM symbol using 1000 beams using the on-chip beam configuration.

[0027] Figure 4 It depicts the Figure 2B Block diagram of the propagation of unequal attenuation beams of an on-chip beam configuration;

[0028] Figure 5 is a block diagram illustrating an example configuration of a receiver with K=3 uncorrelated receive dimensions;

[0029] Figure 6 is a block diagram illustrating a general example of a MIMO system in which MCSs are assigned according to rank;

[0030] Figure 7 is a block diagram depicting the ability to transmit a large number of beams to allow random obstructions to completely block part of the beam;

[0031] Figure 8 is a block diagram depicting the ability to transmit a large number of beams to accommodate the Doppler effect;

[0032] Figure 9A is a block diagram depicting a cyclic transmission for measuring channel coefficients;

[0033] Figure 9B is a block diagram depicting a cyclic transmission for measuring channel coefficients;

[0034] Figure 10 This is a graph measuring the symbol error rate (SER) of 64QAM over AWGN versus Es / N0;

[0035] Figure 11 This is a graph measuring the symbol error rate (SER) of 64QAM over AWGN versus Eb / N0;

[0036] Figure 12 It depicts the case when K=3 dimensions are used to transmit the input bits of a 64QAM symbol. Figure 11 a graph of the expected coding gain in a graph of ;

[0037] Figure 13 is a block diagram depicting the theoretical bounds and effects of coding gain when transmitting the input bits of a 64QAM symbol using K = 3 dimensions;

[0038] Figure 14 is a block diagram depicting the ability of a symbol coding scheme to tolerate selective fading;

[0039] Figure 15 is a block diagram depicting an N-dimensional transmitter and a K-dimensional receiver;

[0040] Figure 16 Describes the use of Figure 15 Block diagram of an example transmission of a 64QAM symbol by a transmitter and receiver;

[0041] Figure 17 is a block diagram showing a high-level view of a transceiver;

[0042] Figure 18 is a block diagram depicting a deep neural network formed by a transceiver including a three-stage encoder;

[0043] Figure 19 is a block diagram depicting the linear and nonlinear layers of a linear / nonlinear hybrid deep neural network;

[0044] Figure 20 is a block diagram depicting a deep neural network with power control;

[0045] Figure 21A is a block diagram depicting a virtual user receiving a "useless" combination;

[0046] Figure 21B is a block diagram depicting a virtual user receiving a "useless" combination;

[0047] Figure 22 is a block diagram illustrating an iterative message passing algorithm (MPA) to train a power factor;

[0048] Figure 23A is a block diagram showing an example configuration for tandem training, where backpropagation adjusts<f,g,w,b> And MPA adjusts α (power control vector);

[0049] Figure 23B is a block diagram showing an example configuration for tandem training, where backpropagation adjusts<f,g,w,b> And MPA adjusts α (power control vector);

[0050] Figure 24A is a block diagram depicting transfer, where layers trained via backpropagation are frozen, while layers trained via MPA remain adjusted;

[0051] Figure 24B is a block diagram depicting transfer, where layers trained via backpropagation are frozen, while layers trained via MPA remain adjusted;

[0052] Figure 25 is a graph measuring the SER performance of encoding a 64QAM symbol on three complex numbers with R = 1 / 3;

[0053] Figure 26 is a block diagram depicting the uncertainty decomposition of time-varying channel attenuation;

[0054] Figure 27A is a block diagram depicting a transceiver with random channel fading;

[0055] Figure 27B is a block diagram depicting a transceiver with random channel fading;

[0056] Figure 28A is a block diagram depicting a virtual user receiving transmission with channel attenuation;

[0057] Figure 28B is a block diagram depicting a virtual user receiving transmission with channel attenuation;

[0058] Figure 28C is a block diagram depicting a virtual user receiving transmission with channel attenuation;

[0059] Figure 29 is a block diagram illustrating MPA iterations during a training phase with channel fading;

[0060] Figure 30A Increased channel attenuation Figure 23A and Figure 23B Block diagram of tandem training;

[0061] Figure 30B Increased channel attenuation Figure 23A and Figure 23B Block diagram of tandem training;

[0062] Figure 30C Increased channel attenuation Figure 23A and Figure 23B Block diagram of tandem training;

[0063] Figure 31A is a block diagram depicting transmission with channel fading, where the MPA operates continuously;

[0064] Figure 31B is a block diagram depicting transmission with channel fading, where the MPA operates continuously;

[0065] Figure 31C is a block diagram depicting transmission with channel fading, where the MPA operates continuously;

[0066] Figure 32 is a block diagram illustrating MPA iteration with time-varying channel attenuation during transmission;

[0067] Figure 33 is a simulation diagram showing a 64QAM symbol (coded at 1 / 3 rate) after fading with uncertainty;

[0068] Figure 34 is a block diagram depicting an example configuration of an on-chip beamformation;

[0069] Figure 35 It is a block diagram depicting a base station sending a pre-trained and stored decoder g() neural network to a receiver according to the terminal's location;

[0070] Figure 36A is a block diagram depicting transmission by an OFDM system to achieve high throughput;

[0071] Figure 36B is a block diagram depicting transmission by an OFDM system to achieve high throughput;

[0072] Figure 37A is a block diagram depicting training with incremental information retransmission;

[0073] Figure 37B is a block diagram depicting training with incremental information retransmission;

[0074] Figure 37C is a block diagram depicting training with incremental information retransmission;

[0075] Figure 38 is a simulation diagram showing the coding gain (Eb / N0) with incremental information retransmission;

[0076] Figure 39 An example of a network for implementing one or more embodiments of the present disclosure is shown;

[0077] Figure 40A is a block diagram of an example electronic device; and

[0078] Figure 40B is a block diagram of an example electronic device. DETAILED DESCRIPTION

[0079] The applicability of systems using a large number of antennas is based on the law of large numbers (LLN): even if one attenuated beam is very weak, hundreds or thousands of such beams can accumulate to a signal energy strong enough to reach the receiver's sensitivity level. Figure 1 In the first example 300, when there is a single beam from the antenna to the terminal, the significant path loss on the single radio beam results in a very low SNR at the terminal. In the first example 302, when there are a large number (N) of beams from the antenna array to the terminal, even if many beams are weak due to path loss, they can be accumulated to achieve an acceptable SNR at the terminal. Figure 2A An example of multiple on-chip beams, generally indicated at 400, is shown. Figure 2B Two chips 400 are shown for combining two on-chip beamforming systems (generally indicated at 402) to provide acceptable receive sensitivity. The two chips 400 can be located in a single base station or in separate cooperating base stations.

[0080] Designing such high-density on-chip beams presents several major design, manufacturing, and scheduling challenges. First, because the total energy of the chip is limited by semiconductor technology, power density becomes critical, necessitating a reduction in the transmitted energy per beam to accommodate more radio beams on the chip. Second, due to the limitations of semiconductor technology, routing (clock trees, power lines) is difficult, and the system must tolerate some timing skew between different beams. Third, scheduling algorithms (power control and timing advance) are required for all beams.

[0081] The capacity of a large-scale beamforming system is determined by the equal transmission signal power (P beam ) and the total signal power (P total ) limit. If the total number of effective radio beams is N, then P total =N·P beam .

[0082] In an AWGN channel, the power gain of N equal-power radio beams is limited. For example, in terms of received signal power, N = 2 increases by 3dB (= 10*log10(3)); N = 3 increases by 4.77dB (= 10*log10(3)); N = 100 increases by 20dB; and N = 1000 increases by 30dB. In future mmWAVE applications, high-throughput applications are attracting attention, so higher-order modulation schemes are likely to be used. For example, although the SNR (P beam ) is approximately –5dB, but theoretically, at this receiver sensitivity level, a radio beamforming system with N = 1000 can reduce the SNR to 25dB (= –5dB + 30dB), which is sufficient for the receiver to decode 64QAM symbols. Figure 3 Describes the use of Figure 2B Block diagram of the on-chip beam configuration for an example transmission of one 64QAM symbol via 1000 beams.

[0083] However, the real environment is much more complex than the ideal uniformly attenuated AWGN. Not all beams experience the same channel attenuation (distortion), resulting in varying channel uncertainties along different paths. In particular, in the mmWAVE band, non-line-of-sight (NLoS) beams often experience very high attenuation (i.e., signals along these paths rarely reach the receiver). Even between line-of-sight (LoS) beams, channel attenuation can vary significantly in terms of transmission angle and arrival angle. Furthermore, not all beams arrive at their common receiver synchronously. The latency introduced by on-chip beam routing and different radio paths varies across paths (or beams). In non-AWGN channels, it is best to use maximum ratio combining (MRC) gain rather than power gain to indicate this capacity bound based on the information principle. Roughly speaking, when scheduling N beams, the system should favor "good" beams (less channel hostility means the transmitted signal is more likely to reach the receiver) over "bad" beams (more channel hostility means the transmitted signal is less likely to reach the receiver). Figure 4 It depicts the Figure 2B Block diagram of the propagation of unequal attenuation beams for an on-chip beam configuration. Obviously, in order to establish an MRC strategy, the transmitter (here, a beamforming base station) needs to estimate the channel conditions on each beam before performing proportional resource allocation.

[0084] Because antenna size is determined by wavelength, the number K (K is "level" in wireless terminology) of uncorrelated receive antennas a terminal can accommodate is wavelength-dependent and increases as wavelength decreases to the centimeter or millimeter scale in mmWAVE bands. Therefore, an N-beam base station can group its beams into N / K groups, each with K beams. A group consists of K uncorrelated beams [Beam-1, Beam-2, ..., Beam-K], each aimed at K receive antennas. These N / K groups are superimposed at the receiver. Figure 5 is a block diagram illustrating an example configuration of a receiver with K=3 uncorrelated reception dimensions or K independent channel usages (which may be time, space, coding, or any other method by which the receiver independently utilizes the channel).

[0085] In this disclosure, different receive antennas in a Multiple Input Multiple Output (MIMO) configuration are used as a specific example of independent (uncorrelated) channel usage. However, the provided approach is valid for any other independent channel usage.

[0086] In the existing methods adopted by some MIMO systems, channel measurement (H m×n , m<n) and the m×m covariance matrix of the estimated channel (H·H H ) is used to estimate K independent channel usages. The K most important eigenvalues ​​sorted and determined by SVD indicate the importance or weight of the K levels or components of the channel. Based on the determined quality, a modulation and coding scheme (MCS) is adaptively selected for that level. For example, if level #1 is the best, a higher modulation and a higher coding rate are selected for that level. The performance achieved using these K levels depends on the granularity of the MCS. For a simplified example of sending 6 bits along 3 levels:

[0087] - If there is a level after SVD, send one 64QAM symbol on that level;

[0088] - If there are two levels (similar to each other), two 8QAM symbols are sent, one for each level;

[0089] - If there are three levels (similar to each other), three QPSK symbols are sent, one for each level.

[0090] Figure 6 is a block diagram illustrating a general example of a MIMO system in which MCSs are assigned according to rank.

[0091] This system is inefficient for a variety of reasons. First, the quality of the tiers must be reliably estimated and tracked. This requires a high pilot density, which results in significant pilot overhead. For pilot transmission, the transmitter sends reference signals on a certain percentage of the total resources, and the receiver should measure these reference signals on each receive port and send feedback to the transmitter, such as channel state information (CSI). 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 methods to obtain these tables and parameters. Third, control signaling must be used to ensure that both the transmitter and the receiver follow the same convention. Both the transmitter and the receiver must know which MCS is being used at which tier. More importantly, if the channel attenuation changes to some extent, the quality of the tiers can change significantly, making each established convention (MCS) obsolete.

[0092] In mmWAVE, the attenuation on the beam is polarized. More specifically, a beam is created by a group of antennas. This group of antennas is used to direct its energy in a specific direction, allowing it to significantly compensate for path loss in that direction, resulting in a strong transmission signal in that direction. For reception, this group of antennas also has the effect of reducing receive sensitivity. This operation is sometimes referred to as polarization. In terms of distance, LoS beams experience attenuation due to air path loss; some NLoS beams experience significant attenuation; and other NLoS beams can be positively enhanced by reflections from metal. Generally speaking, the attenuation pattern (distribution) is sensitive to its geographic location. Statistically, the attenuation pattern based on spatial location becomes usable within a coherence time window (called the coherence time). However, when high-frequency bands are used for signal transmission, the low mobility can lead to significant Doppler frequency and time offsets. The coherence window time is very small.

[0093] A large number of beams can be used to mitigate these challenges for coding gain. For example, if some beams are blocked (infinite attenuation, null beams), the system adjusts the transmit power between the "surviving" and "null" beams to compensate for the loss. Because Doppler frequency and time offset depend on the angle of the motion trajectory, their impact can be compensated by adjusting the transmit power between the "favorable" and "unfavorable" beams.

[0094] Figure 7 A block diagram depicting how a large number of beams can be transmitted to allow random obstructions to completely block some beams. Figure 8 is a block diagram depicting the ability to transmit a large number of beams to accommodate the Doppler effect.

[0095] In mmWAVE scenarios, power control and synchronization between N beams are critical. However, achieving this is extremely difficult. First, the attenuation needs to be known. Although the statistical distribution model of attenuation based on open space locations can be obtained, attenuation itself is a random variable. To capture instantaneous attenuation, the system needs to measure N / K×K channels, a total of N paths, which means a large number of downlink reference resources (because the receiver only sees K dimensions generated by the N beams overlapping in the air, so one receiving dimension needs to measure N / K beams in a cyclic manner to avoid pilot interference) and a corresponding large number of uplink reporting resources for overhead. Figure 9A and 9B is a block diagram depicting a cyclic transmission to measure the channel coefficients. Figure 9A At the measurement time t shown, only K beams are allowed to transmit their pilots on the downlink. Figure 9B At the measurement time t+1 shown, only the other K beams are allowed to transmit their pilots on the downlink.

[0096] Another approach assumes reciprocity between the downlink and uplink channels. However, in reality, this reciprocity always has a certain offset. Furthermore, even if the system fully understands the channel attenuation, calculating the maximum ratio combining (MRC) power offset for N / K beams within a short time interval remains challenging.

[0097] Performance gains with dimension expansion

[0098] Coding Gain

[0099] In modern communication systems, redundancy is widely used to overcome channel attenuation and noise. For example, forward error correction (FEC) is a dimensionality expansion method that not only injects some redundant bits but also creates certain dependencies between coded bits (codewords). In another example, a MIMO encoder is a dimensionality expansion method that expands several complex symbols into multiple spatial dimensions to achieve coding gain. In general, although dimensionality expansion brings dimensionality gain, this dimensionality gain must include power gain but does not necessarily produce any coding gain.

[0100] The power gain is the equal gain combining (EGC) gain in the AWGN channel. In an AWGN channel with equal attenuation, if a piece of information (e.g., a bit or a QAM symbol) is repeated (or linearly combined) in K dimensions, this naturally leads to a power gain of 10*log10(K)dB. As an example, the EGC gain of repeated transmission in three equal attenuation dimensions is equal to 4.77dB (=10*log10(3)), as shown in Figure 10 shown.

[0101] Coding gain exceeds the EGC gain. For example, encoding a symbol in 3 dimensions to achieve a 3dB gain in addition to the ECG gain allows the transmitter to halve the power or double the range. To illustrate 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) rather than the ratio of energy per symbol (Es) to spectral noise density (Es / No). Simple repetition (or any linear combining) scheme results in Figure 11 Zero coding gain is shown.

[0102] To achieve coding gain, traditional channel encoders (FEC) inject (K–M) redundant bits for each M information bit, or expand N dimensions (K>M>1) to K dimensions. Essentially, coding gain is related to the coding rate: M / K. The lower the coding rate, the greater the coding gain. When (M=1) dimensions are expanded to K dimensions, resulting in a coding rate of 1 / K, according to coding principles, if the input dimension is 1 (M=1) in a uniformly attenuated AWGN channel, there is no coding gain. For such systems, achieving coding gain in a uniformly attenuated AWGN channel is advantageous.

[0103] Different from the MRC gain, the availability of coding gain in AWGN channel is easy to understand. Figure 12 If there is a 1->K dimensional extended coding gain in the CNN, system-level advantages can be achieved under more complex channel conditions.

[0104] Diversity gain

[0105] The input bits of a 64QAM symbol can be transmitted using K = 3 dimensions by sending three independent QPSK symbols, each carrying 3 bits. It is well known that QPSK constellations are optimal with respect to coding distance (e.g., Hamming distance) due to the constant Euclidean distance between any two adjacent constellation points and the constant amplitude of the constellation points. There are other methods for transmitting 6 bits using K = 3 dimensions, and the QPSK curve can serve as a theoretical bound on the potential coding gain using K = 3 dimensions.

[0106] Figure 13 A potential coding gain of approximately 8 dB is shown when sending 6 bits using K = 3 dimensions, compared to sending a 64QAM symbol in one dimension. This coding gain is equivalent to reducing the transmit power by approximately 16%, or increasing the coverage by approximately 36 times. It should be noted that the example of sending 6 bits of a 64QAM symbol using three dimensions was chosen simply to make it comparable to the QPSK constellation that has been proven to be optimal in 2D complex Hilbert space. However, this is not the general case: for higher modulation orders (such as 512QAM, 1024QAM, or 4096QAM) with arbitrary code rates of 1 / K, the theoretical bounds remain unproven and unknown. The proposed method has been shown to be optimal for QPSK and 16QAM. The proposed method is also applicable to higher-order modulations. It is speculated that the proposed method can approach the optimal solution when applied to higher-order modulations. This contrasts with traditional high-order modulation schemes, which are not designed for optimal performance but for ease of implementation.

[0107] There is no guarantee that for a given modulator order and an arbitrary coding rate, log2(modulator order) / K will be an integer. In the case where 6 bits are sent using 3 symbols, this is equivalent to a coding rate of 1 / 3, and log2(64) / 3=2 (-> QPSK). In the other case where a coding rate of 1 / 3 is applied to the input bits (10) of a 1024QAM symbol, this results in log2(1024) / 3=10 / 3 bits per dimension, which is not an integer. In practice, there may be a different number of receive dimensions K. For example, for a terminal with a maximum of 4 receive antennas, it is likely that only 3 receive antennas remain uncorrelated. In practice, the number of uncorrelated channels available in a given instance is determined by the environmental conditions and is not easy to control in an accurate manner.

[0108] It is advantageous to achieve diversity gain or MRC gain in multipath channel conditions. K uncorrelated dimensions may be subject to K independent and different attenuations. Assume that the transmitter divides the 6 bits of a 64QAM symbol into 3 groups of 2 bits for transmission on 3 QPSK symbols, and sends them through 3 uncorrelated dimensions respectively. If one of the three dimensions undergoes severe negative attenuation (weakened) and the other two dimensions are enhanced, the QPSK symbol on the weakened dimension may be lost, resulting in the inability to recover the original 64QAM symbol. In contrast, if a 64QAM symbol is encoded into 3 dimensions in some way, the receiver can recover the 64QAM symbol from the remaining two enhanced dimensions through the radio channel, such as Figure 14 This is the diversity gain or MRC gain.

[0109] Incremental Redundancy Retransmission

[0110] In wireless systems, IR-HARQ is widely used in conjunction with channel coding to improve 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) codes, allow the transmitter to send a partial codeword (coded block) to the receiver during a transmission interval. If the current channel conditions are good enough, the receiver may even be able to recover the entire information block from this partial codeword; otherwise, the receiver can send feedback to the transmitter for retransmission. Instead of retransmitting previously sent content (tracking and combining retransmissions), the transmitter sends an incremental redundant portion of the codeword, allowing the receiver to combine this incremental redundant portion with the previously received portion into a longer codeword for its channel decoder. This allows for additional coding gain beyond the pure power (repetition) gain.

[0111] IR-HARQ is an opportunistic method that is sensitive to time-varying channel conditions (sometimes channel usage exhibits positive conditions and sometimes negative conditions). If the first transmission occurs under positive conditions, retransmission resources are saved. Furthermore, in some delay-critical and reliability-critical applications, such as Ultra-Reliable Low Latency Communication (URLLC), the transmitter prefers blind retransmissions rather than waiting for feedback, as the feedback cycle may be too long for their latency requirements. In both scenarios, the most essential or useful information should be included in the incremental retransmission(s) to generate coding gain.

[0112] So far, incremental redundancy-based retransmission has only been performed with some channel coding schemes, such as those using specific channel coding schemes such as Turbo codes, convolutional codes, and low-density parity-check (LDPC) codes. Similar retransmission has not been reported for other algorithms based on dimensionality expansion, such as MIMO and spreading codes (most of which only support tracking combining, i.e., power gain).

[0113] With the development of 5G systems, the number of channel dimensions, such as frequency (subcarriers, different frequency bands), space (MIMO, dual connectivity), time, and coding (PN codes, pseudorandom codes), is also increasing. A good system should be able to integrate multiple available dimensions to provide overall system gain that efficiently uses available resources. Among the sources of system gain, coding gain is the most popular in terms of its impact on system spectral efficiency.

[0114] However, due to the orthogonality of these dimensions, only some channel coding schemes can achieve incremental redundancy (IR) retransmission with some coding gain. Such channel coding schemes that support IR benefit from the assumption that the M information bits are independent and identically distributed (IID) and the K coded bits are subject to AWGN noise. In other words, the input is an orthogonal M-dimensional entity and the output is also an orthogonal K-dimensional entity. In practice, efforts are usually made to "maintain" this ideal environment for channel coding through the use of pilots, channel estimation, and equalization. Unfortunately, this is not the case for most dimensionality expansion algorithms, and the orthogonality of the input and output cannot always be guaranteed. For example, it cannot be guaranteed that two antennas will always remain uncorrelated. In another example, due to multipath fading or the Doppler effect, two subcarriers can experience very different channel attenuation in a frequency selective channel.

[0115] Faced with these difficulties, most prior art systems adopt Maximum Ratio Combining (MRC) schemes to passively overcome diversity. Although in some cases, an MRC-based receiver can achieve MRC gain due to diversity transmission, this gain cannot always be guaranteed.

[0116] Machine Learning

[0117] It is proposed to use machine learning techniques to address the two aforementioned issues and create a coding scheme that encodes M traditional QAM symbols into K dimensions using N / K power-controlled radio beams. The most straightforward proposal is to use an autoencoder, where the input is M QAM symbols and the output is the decoded QAM symbols. There is a latent layer with K dimensions, attenuation is modeled via multipliers on this K-dimensional latent layer, and AWGN noise is added to this K-dimensional latent layer. The loss or objective of this autoencoder is to minimize the mean square error (MSE) between the input and output.

[0118] One problem with this solution is that the decoding neural network (after the K-dimensional latent layer) will try to mathematically maximize the likelihood probability that depends on the K-dimensional latent layer. However, the K-dimensional latent layer is subject to varying noise and attenuation. As the SNR and attenuation change, the statistical properties of the K-dimensional latent layer change, requiring the decoding neural network to be retrained. In practice, it is impractical to retrain the entire deep neural network when the SNR or attenuation changes over time.

[0119] Another problem is that in order to have N beams, the autoencoder architecture includes an N-neuron layer just before a K-neuron layer (N>>K), and the connection between the two layers is multiplied by a given attenuation coefficient and random noise is added. In such a deep neural network, training tends to "power off" most of the beams and reduce the transmit power P total This results in a "focus" on the very few remaining beams. This is not a desirable result because it would require each beam to be designed with a much higher maximum transmit power, which results in higher cost and larger chip die size.

[0120] One of the main reasons deep learning neural networks are difficult to apply to wireless telecommunications applications is that they are nonlinear functions. Compared to traditional linear functions (most wireless signal processing algorithms are linear or quasi-linear), nonlinear functions tend to overfit and their generalization is severely limited.

[0121] Another problem arises from the variable channel attenuation in wireless systems. Radio channels are time-varying, frequency-selective, and fading. Their presence causes the statistical properties of the received signal (the transmitted signal distorted by the channel) to also be time-varying and frequency-selective. The neurons in the deep neural network need to adapt to the variable and selective channel attenuation, often requiring complete retraining.

[0122] Example 1

[0123] Model of N-beam to K receiving system

[0124] Now refer to Figure 15 , which is a model of the transceiver scenario. The N-dimensional transmitter sends a specific sample "a" 1504. This sample is obtained from the Hilbert space through the hostile environment using the sampling method F 1502. A sample "a" is sampled from a manifold A (probability distribution) 1500 on a signal path, and a K-dimensional receiver including a decoder 1550 estimates the sample "a" from the distorted and noisy received signal. The transmitter encodes the sample using an encoder 1506 to generate K×N / K outputs (i.e., N / K K-dimensional vectors), where N is the dimensionality of the transmitter and K is the dimensionality of the receiver (i.e., these N / K K-dimensional vectors accumulate into one K-dimensional vector for the receiver). In a specific example, the N transmit dimensions are achieved by N transmit beams from N transmit antenna elements, and the K receive dimensions are achieved by K receive antenna elements.

[0125] More specifically, the transmitter samples from the manifold A using sampling method F 1502 to obtain symbol a, and encodes the symbol a into a vector [s1, s2, ..., s N ]: where f(·) is the transmitter encoder unknown to the receiver. The vector [s1,s2,…,s N ] is sent on N transmission paths, which naturally superimpose in the air and generate a receive vector at the receiver The receiver decodes the received signal using decoder g(·) 1550 to generate:

[0126] Including space The topology of the manifold A and the sampling method F are specified and standardized, and are known to both the transmitter and the receiver.

[0127] The transmitter has N independent or uncorrelated emission dimensions Each transmit dimension sends a sine wave represented by a complex value. The N components of are grouped into N / K groups: Each group has K components: And align the K uncorrelated receiving dimensions of the receiver. Without loss of generality, the transmitting dimension s i,1 Indicates the connection from the i-th transmitting group to the first receiving port; s i,2 represents the connection from the ith transmitting group to the second receiving port, and so on. For example, Figure 15 1510 is shown transmitting on the first set of K transmission dimensions.

[0128] There are N connections connecting the N transmit dimensions to the K receive dimensions. Each connection has its own attenuation coefficient Each attenuation coefficient is represented by a complex value (I, Q), and its amplitude (I 2 +Q 2 ) represents power attenuation, and the angle (inner product<I,Q> ) represents the phase offset, that is, the time offset. After the grouping, can be divided into N / K groups, where each group has K complex elements: and Attenuation is generally Figure 15 Indicated at 1512.

[0129] Similarly, each link has a normalized power control coefficient (real value) To avoid saturation of the transmission power. After the grouping, can be divided into N / K groups, where each group has K real elements: and To avoid excessive power concentration, limit: This implies that the total power is less than N / K, i.e., The maximum total power of a group is N / K to ensure that there is no damage due to excessive power concentration. Figure 15 The output of power control 1514 is what is sent over the air.

[0130] In the kth receiving dimension, these N / K connections are superimposed in the air: For the receiver, it receives the K-element complex vector Each of these transmits a sine wave represented by a complex (I, Q) number: in is element-wise multiplication.

[0131] AWGN noise n is added to r k ′=r k +n k The addition of noise Figure 15 1516 places indicated.

[0132] Note that if K = 1, the model simplifies to N independent transmit dimensions to 1 receive dimension. In this case, there is no longer a coding gain, but there may still be a power gain.

[0133] Now refer to Figure 16 To describe Figure 15 A specific example of a model of . In this example, the Hilbert space is a two-dimensional space with real and imaginary parts. The 64QAM constellation is a standardized manifold A, which defines Gray coding F is a method for sampling this manifold, ensuring that each constellation point is sampled with equal probability (i.e. 1 / 64). When sampling, the 64QAM symbol a = (aI, aQ) (because it is sampled from a two-dimensional space Sampling) by Encoded N = 768. The transmitter in this example has 768 independent beams. The receiver has 3 (K = 3) independent receive antennas, so the 768 beams are grouped into 256 groups (each group has 3 beams). A group of 3 beams is parallel and aimed at 3 receive antenna ports respectively. Since there are 256 groups, the significant path loss of a single group can be compensated (256->SNR of the receive antenna is improved by 24dB). As a result, the receiver has And use the decoder g() to estimate the symbols: This example is used in the following description, but it should be understood that the provided systems and methods are valid for any Hilbert space, manifold, and sampling method.

[0134] Coding gain problem

[0135] For the purpose of finding f(·) and g(·), the channel attenuation coefficient is fixed (e.g., set and the power control is fixed (e.g. by setting is fixed to 1). The problem is simplified to: Given a, A and F, and N and K, find f(·) and g(·) to provide the maximum coding gain.

[0136] use Figure 16 In the aforementioned example, a transmitter wishes to send a 64QAM symbol to a receiver. The transmitter knows that the receiver has three independent receive antennas. The transmitter has a large number of beams available for this transmission. It is known that repetition or other linear expansion from one symbol to three symbols across N beams will produce a gain that does not exceed the power gain (EGC) in a uniformly attenuated AWGN channel. It is desirable to find a symbol-level coding scheme that achieves maximum coding gain in addition to the EGC gain in a uniformly attenuated K-dimensional AWGN channel.

[0137] Power control issues

[0138] Another problem concerns setting power control for N beams in the presence of varying attenuation and time offset. The attenuation on a beam (h) consists of three components: static uncertainty (h s ), dynamic uncertainty (Δh d ) and mismatch uncertainty (Δ'h m ):h=h s +Δh d +Δ'h m This static component has a strong correlation with the spatial location of the terminal and the surrounding environmental features (such as buildings, trees, mountains, and some fixed environmental factors). d Mainly due to random events such as mobility and blocking, is the s Random jitter on Δ'h m is the unavoidable measurement uncertainty, which is Δh d It can be assumed that the jitter Δh d and Δ'h m Not as good as h s Important. Dynamic power control of N beams can be used to compensate for the varying attenuation uncertainty. Power control can be tracked and adjusted, but significant coding gain is required. The problem model is: Right now ( is element-wise multiplication).

[0139] The overall solution addresses both the coding and power control issues and can be divided into the following subtasks:

[0140] The first task: Given a, A, F, N, K and static decay Find the codecs f(·) and g(·) to provide maximum Coding gain, and find the power control [c1,c2,…,c N ] to compensate for static attenuation.

[0141] The second task: given a time-varying attenuation Dynamically adjust power control to compensate for the dynamic attenuation without changing g(·).

[0142] It is very important to keep the same decoder g(·) against time-varying channel fading and SNR to achieve high throughput and low latency.

[0143] Continuing with the previous example, a base station wishes to transmit a 64QAM symbol to a receiver. The base station knows that the receiver has three receive antennas, and it can utilize these antennas. The base station has a large number (N) of beams available for this transmission. The channel attenuation across the N beams is not equal and is a time-varying random variable. The base station can estimate the distribution of the channel attenuation; conventional wisdom suggests that this can only be achieved by transmitting pilots. However, in a system with large dimensionality, this requires a large number of pilots. It is advantageous to avoid the transmission of a large number of pilots.

[0144] The third question is how to and the given A, F, N, K and static decay Get the codec f(·) and g(·) and power control While machine learning techniques can be used, time-consuming back-propagation algorithms are impractical for real-time communications in mmWAVE applications.

[0145] Another problem with the above example is that the target receiver needs to be updated, and there is no practical and efficient way to implement this update. This involves sending the new g() coefficients to the receiver again, which consumes a lot of radio resources.

[0146] Decomposer and Recomposer

[0147] In order to study the pure coding gain problem, we first consider the equal attenuation K-dimensional AWGN channel condition, that is, the channel attenuation As a constant 1+0j. AWGN noise is added to r k ′:

[0148] For comparison and practical implementation purposes, r k 'Normalized to: to ensure The average energy of is 1 / K, so The average energy is 1 (=1 / K*K).

[0149] Reference Figure 17 , the encoder f(·) and the N beams can be viewed as a Black Box encoder 1700; the Black Box encoder inputs a symbol a and outputs a K-element complex vector with an average energy of 1 After being combined with white noise n according to Es / N0, Output to decoder g(·). From this point of view, a symbol a is encoded into Kv k Following the above detailed example, one 64QAM symbol can be encoded into three complex symbols. A Black Box decoder 1702 is also shown.

[0150] In some prior art systems (such as those using MIMO precoding), where w is the K×1 encoder matrix. Under AWGN channel conditions, this linear combination does not provide coding gain, but does provide a power (EGC) gain (10*log10(K)dB). Coding gain occurs when a set of symbols is encoded as a set of 2K complex values. This is a channel coding scheme. However, the focus here is on encoding a symbol across multiple dimensions.

[0151] Reference Figure 18 , based on the following two concepts, Figure 17 The Black Box "encoder" is divided into three stages: a non-linear decomposer 1800, a linear encoder 1802, and a linear (re)encoder 1804:

[0152] The symbol a is considered an information carrier rather than the information itself. In prior art approaches, the transceiver uses all available means to maximize the received signal power of the symbol a. Consequently, conventional transceivers exhibit power gain. However, in the system described here, the focus is instead information-oriented, with emphasis placed on the information carried by the symbol a at the receiver, which the decoder g() is able to decode.

[0153] The K element is considered as a K-dimensional atomic entity rather than a divisible vector consisting of K complex elements. In the prior art system, the decoder is Recover or derive a redundancy reducer for the symbol a.

[0154] Based on these two concepts, the encoder f(·) extracts N / K features from the symbol a Each feature is an L-dimensional complex entity: The function f(·) is a nonlinear function implemented by a machine learning block. In a specific example, the machine learning block is a deep neural network with a specific kernel and nonlinear activations such as ReLU, Sigmoid, etc. Given different (Hilbert space), A (manifold), F (sampling method), N (number of uncorrelated transmit dimensions), and K (number of uncorrelated receive dimensions), the machine learning block f(·), such as a deep neural network, is valid for a single implementation.

[0155] The encoder f() decomposes a symbol a into N / K L-dimensional entities Afterwards, each entity is encoded independently (in parallel) by its own linear encoding:

[0156]

[0157] Among them, w i is an L×K complex matrix, and is a 1×K complex matrix. It is a K-dimensional complex entity: In total, there are N / K K-dimensional complex entities:

[0158] The N-beam transmitter re-edits the N / K K-dimensional complex entities into one by simply summing

[0159]

[0160] In the normalized and after adding noise, a single Input to the decoder g(·) to estimate g(·) is a machine learning block, such as a deep neural network, with a specific kernel and nonlinear activations such as ReLU, Sigmoid, etc. (Hilbert space), A (manifold), F (sampling method), N (number of uncorrelated transmit dimensions), and K (number of uncorrelated receive dimensions), the machine learning block g(·) is valid for each implementation. Figure 18 In the AWGN channel, the overall process is:

[0161]

[0162] The overall process can be viewed as a complete deep neural network 1900, as shown in Figure 19 As shown. The input sample a is given according to (Hilbert space), A (manifold), F (sampling method), N (number of uncorrelated emission dimensions) and K (number of uncorrelated reception dimensions). The encoder f(·) and decoder g(·) are the nonlinear layers of the deep neural network. i and is the linear layer of the deep neural network. The training loss or reward is to minimize the variance (MSE): The neural network is differentiable. In this embodiment, gradient descent back propagation is used to adjust the f(·) layer, g(·) layer, w i Layer and The layer is close to the MSE optimal point.

[0163] Since the f(·) layer and the g(·) layer are nonlinear neural network layers and w i and This is a linear layer. This deep neural network is a linear / nonlinear hybrid neural network architecture, which is different from the traditional nonlinear deep neural network. i and ) is used to assist f(·) and g(·) to make them more tolerant to varying channel conditions and noise levels.

[0164] Example 2: Weighting the output

[0165] In this embodiment, various methods of weighting the output of the linear encoder are provided with the goal of increasing the amount of useful information sent to the receiver.

[0166] Reference again Figure 18 , the N / K features in this group decomposition In, not every are all equally important. Some features are more important than others, in the sense that they carry more important information than others so that the receiver g() can recover the information sent. The importance distribution depends on the given (Hilbert space), A (manifold), F (sampling method), N (number of uncorrelated transmit dimensions), and K (number of uncorrelated receive dimensions). It is difficult to distinguish which features are more important.

[0167] In the above embodiment, the convergence is This implies that the features are merged in an equal or equally important manner.In another embodiment, the modification system can place less emphasis on these less important features and place more emphasis on more important features.

[0168] For this embodiment, an additional power control vector is introduced In a particular embodiment, this vector is simplified to include one power control factor (real value) for each feature: When the power control vector has a large value for a given feature, it will emphasize the feature relatively more, and when the power control vector has a small value for a given feature, it will emphasize the feature relatively less. Convergence then becomes and

[0169] Example implementations such as Figure 20 As shown, it is Figure 18 Same, except in Figure 18 In the linear editor, no weighting is applied so that the edited function is the sum of the inputs, whereas in Figure 20 According to the power control vector A set of weights is applied to the output of linear encoder 1802.

[0170] Therefore, in Figure 20 The entire process can be summarized as follows:

[0171]

[0172] The goal is to constrain to determine

[0173] How well power control is performed depends primarily on the given (Hilbert space), A (manifold) and F (sampling method). If the samples are concentrated in the space an area, then Relatively static and easy to train. If the samples are scattered in space A "wider" area, then It is dynamic, which makes training difficult or takes longer.

[0174] In wireless systems, it is better if the samples are widely dispersed; the larger the average distance between samples, the better the samples are resistant to noise. The distribution of feature importance can vary significantly between samples (symbols).

[0175] Example 3: Virtual User

[0176] According to another embodiment, the system is further modified to implement virtual users so that, in effect, instead of directing the entire output to a single receiver, the linear coding is split between the real and virtual users by using different power control vectors for the real and virtual users. By appropriately setting the two power control vectors, the system is trained to direct more important features (also referred to herein as “useful” information) to real users, while directing less important features (also referred to herein as “useless” information) to virtual users.

[0177] Weighted encoder output is sent to the real user, and the weighted encoder output is sent to the "virtual" user. The two users share power: α i +β i =c i .like carries more important information about the symbol a, then α i >β i Otherwise α i <β i The original power control vector in this case can be set to Every one There are two connections: one to the real factor α with power control i real user, another connected to a real user with power control factor 1-α i Using the original power control vector setting, it satisfies

[0178] The entire system can be summarized as follows:

[0179]

[0180] Figure 21A and Figure 21B is a block diagram of such a system. Figure 21A and Figure 21B Closely based on Figure 20 , but the power control block 2104 replaces the power control block 2004. The power control block 2104 sets the weight α i is applied to the output of the linear encoder 1802 to generate an output directed to the real user, and the weight β i The output of linear encoder 1802 is applied to generate an output directed to the virtual user.

[0181] Note that the output is not actually sent to the virtual user. This virtual user is constructed to assist in optimization and training.

[0182] After constructing such a system, the goal is to train the system to constrain α by training i ≤1 to determine If traditional gradient descent back propagation is used, then Not reflected in training.

[0183] According to an embodiment of the present disclosure, the training consists of two sub-trainings:

[0184] against Gradient descent back propagation;

[0185] Targeting α i Message passing algorithm. In order to satisfy α i +β i =1, use the following softmax function:

[0186]

[0187] in, yes and The metric distance between yes and The distance between them.

[0188] To minimize and The average metric distance between: This distance is defined as and The inner product between: The following is an example of a message passing algorithm (MPA) that can be used to determine weights:

[0189] initialization:

[0190]

[0191] For iteration = 1:r

[0192]

[0193] End of iteration. / /

[0194] The iteration is as Figure 22 shown. Figure 22 The top shows the number of Node to two r nodes (real and virtual). There are two inputs on the node and Each Node Computing That is, the power weight, and then send to the real user and send To this virtual user. Note that all N / K Nodes all perform the same work. However, because of their Different, its weight α i Also different. Figure 22 The bottom of the figure shows the number of nodes from two r nodes (real and virtual) to N / K The real r node pair comes from the N / K All nodes Sum and Send back all N / K Node. The virtual r node pair comes from the N / K All nodes Perform the summation, where and will Send back all N / K Node. Now, all N / K All nodes are updated and The whole process is then repeated for multiple iterations.

[0195] This is a typical message passing algorithm. In one iteration, (3) is calculated in parallel for all groups on the sending side. Then, each transmission is updated according to the updated α i and (1-α i ) to adjust the power, and then update and In the next iteration, the updated and will be used to update all α via (3) i After r iterations, the obtained α i is the power control factor.

[0196] This message passing algorithm is differentiable (because ), so it can be embedded in deep

[0197] In neural network:

[0198]

[0199] The goal is to determine

[0200] For gradient descent backpropagation, [α1,α2,…α N / K ] is a constant. MPA training [α1,α2,…α N / K ], processing f, g, w at the same time i , is a constant. Figure 23A and Figure 23B A tandem training approach is described, showing backpropagation training and MPA training.

[0201] Layers trained via backpropagation For the power control layer ([α1,α2,…α N / K ]) assumes fixed / frozen values, and the layers ([α1,α2,…α N / K ]) For layers trained via backpropagation Assume a fixed value. In some embodiments, when transmitting, the power control layer keeps adjusting continuously, while the layers trained by backpropagation are frozen, e.g. Figure 24A and Figure 24B shown.

[0202] Referring again to the previously introduced example of encoding a 64QAM symbol into three dimensions in an AWGN channel, in a specific implementation, the transmitter runs the message passing algorithm on 100 64QAM symbols in three iterations. To illustrate the coding gain, the symbol error rate versus Eb / N0 is determined and plotted in Figure 25 There is a coding gain of about 8dB above the power gain. Although it is trained with a fixed Es / N0 of 10dB in this example, Figure 25 The curve in is very close to the theoretical boundary over a wide SNR range without any error plateau.

[0203] As can be seen, the entire system does not suffer from the generalization issues found in conventional autoencoder methods. If three equal quality levels exist for 6 bits, the system will generate three QPSK symbols. Without intervening and changing the decoder g(), the system can achieve near-optimal results over a wide range of SNRs.

[0204] The above examples demonstrate that, given the number of bits to be transmitted and the number of available equally important levels, this deep neural network architecture allows one to approach theoretical bounds. In practice, the number of bits to be transmitted can be 5, 7, or other values, and in particular, need not be a power of 2. Furthermore, the number of levels can be arbitrary. For example, when 7 bits span 5 levels, the traditional heuristic bounds do not exist. The deep neural network will approach the optimal solution.

[0205] Example 4: Multipath Channel Application

[0206] Detailed description of Example 3

[0207] While coding gain over an AWGN channel with equal attenuation (K equal quality levels) has some utility, multi-level transmission makes more sense when the levels have different qualities. In prior art methods, the quality of the K levels must be estimated and ranked using SVD; a symbol is divided into K levels proportionally in terms of quality. This mechanism involves a large amount of overhead (pilots, control messages, and primitives) and a large set of standardized candidates. The following are some current designs for MIMO systems: Step 1: The transmitter sends 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 runs SVD to find eigenvalues ​​and eigenvectors from the CQI; Step 5: The transmitter ranks the eigenvalues ​​from high to low and removes some very small eigenvalues. Based on the eigenvalues ​​(quality of the level), the transmitter assigns a modulation and coding scheme. Step 6: The transmitter sends the MCS and decoding matrix (eigenvector) for each level to the receiver.

[0208] In mmWAVE applications, the AWGN channel is oversimplified; instead, a multipath fading channel, i.e., a time-varying and frequency-selective fading channel, can be used, which introduces additional complexity. In practice, N beams arrive at K receiving antennas independently, and the received signals of these beams are summed or interfered together. These beams may be somewhat correlated. Each beam has a different, frequency-selective, and time-varying fading attenuation. The selectivity is due to the multiple paths on the beam, which is mainly determined by the physical environment. The attenuation caused by the selectivity can be considered static over a certain period of wireless communication: h s Time-varying attenuation is caused by mobility, not only the moving terminal, but also nearby moving obstacles. Time-varying attenuation can be viewed as jitter on static attenuation: Δh d (t), which can be measured and estimated. In addition, the time-varying channel attenuation is never accurately estimated, resulting in an inevitable estimation mismatch. Or the time-varying channel attenuation will become outdated after a period of time, resulting in another inevitable outdated mismatch. This mismatch can be regarded as Δd m Jitter on (t): Δ'd m (t).

[0209] In the N to K scenario (N>>K), the pilots (reference signals) on the N / K groups of beams will be superimposed at the receiver, making a cyclic measurement algorithm on the N / K groups of beams necessary. Unfortunately, for a given set of beams, two consecutive downlink channel measurements occur at least every (N / K–1) measurement interval. Because N can be much larger than K, this interval is too long to explain Δh d Furthermore, this approach requires significant overhead both in the downlink (pilot and sounding signals) and uplink (reporting and feedback).

[0210] An alternative approach relies on UL / DL reciprocity in TDD mode: the base station (transmitter) estimates the channel attenuation on the UL signal sent by the receiver and then considers it as the channel attenuation on the next DL (transmitter to receiver). This reciprocity assumption has drawbacks due to potential mismatches between UL and DL.

[0211] In mmWAVE applications, reliance on UL / DL reciprocity is preferred as it consumes less overhead than loop-based approaches. Furthermore, to support very high throughput and ultra-low latency, a simple decoder g(·) is preferred instead of extensive traditional synchronization, channel estimation, and equalization.

[0212] In some embodiments, the attenuation h(t)=(h s +Δh d (t)+Δh m The method (t)) is based on:

[0213] -As a result of re-editing Its inherent editing logic is conducive to the transceiver's tolerance of attenuation h(t) = (h s +Δh d (t)+Δh m (t)).

[0214] -Since machine learning is a (non-linear) fitting function, each layer Trained and fixed by a few channel fading samples. To a (limited) extent, they can handle channel fading that has never been encountered before.

[0215] -After training the fixed layers ( is a trained and fixed layer), the transmitter continuously runs MPA to target Update the power control vector [α1,α2,…α N / K ]. Using h s +Δh d (t) This knowledge can help the transceiver adapt to time-varying channel attenuation.

[0216] For training purposes, the channel attenuation is divided into three parts, such as Figure 26 As shown:

[0217] Static (large-scale) channel attenuation (h s )2600 is mainly due to the static time offset between different beam antennas and multipath on the beam according to the receiver location. The base station (transmitter) usually has good static channel attenuation knowledge.

[0218] - Dynamic (medium-scale) channel attenuation (Δh d (t)) 2602 is mainly due to the changing environment (e.g., random passers-by) and low mobility (e.g., small movement). The base station continuously measures Δh on the uplink. d (t), and use these measurements for the downlink based on the UL / DL reciprocity assumption.

[0219] -Mismatch (small scale) channel attenuation (Δh m (t)) 2604 is mainly due to the mismatch between the channel estimation and the channel measurement. It can be the mismatch between the UL and DL channels when reciprocity is assumed, or the mismatch accumulates over time.

[0220] These attenuations can be included in the system model as follows:

[0221]

[0222] In this formula:

[0223] is static and always known to the transmitter;

[0224] is time-varying but known to the transmitter;

[0225] It is time-varying, but theoretically unknown to the transmitter;

[0226] is element-wise complex multiplication.

[0227] Although and are time-varying, but their distribution can be known or estimated using some statistics. During the training phase, and The value is not immediately available, but An estimate of may be available. To address this, numerical assumptions are made for training purposes. In the following example, two known Gaussian distributions for N / K beams are used, namely and These distributions are described in further detail below.

[0228] Then, in the above model, Replace with and Replace with

[0229]

[0230] Figure 27A and Figure 27B Depicts a modified version of the Figure 20 system, where the channel attenuation component is introduced at 2700.

[0231] Figure 28A 、 Figure 28B and Figure 28C Shown Figure 21A and Figure 21B A system comprising a virtual user, wherein the virtual user is defined as an ideal user with no attenuation, whether negative or positive attenuation. Figure 28A 、 Figure 28B and Figure 28C The system shows the channel attenuation 2800 applied only to the output of the real user. The entire neural network is:

[0232]

[0233] During the training phase,

[0234] During training, the following methods were adopted:

[0235]

[0236] For iteration = 1:r

[0237]

[0238] and

[0239] End of iteration. / /

[0240] The above example assumes that the virtual user has no attenuation; in another embodiment, it is assumed that the virtual user has the same attenuation as the real user. According to the simulation results, there is no significant difference in the performance of the two methods.

[0241] The training is the same as the training for AWGN channel mentioned above. The goal is Gradient descent backpropagation is fixed [α1, α2, … α N / K ] while training MPA is fixed f, g, w i , While training [α1,α2,…α N / K ]. Figure 29 MPA during the training phase with decay was summarized.

[0242] After training, f(), g(), w are frozen i and Layer. The transmitter will use them to encode sample a into The transmitter will continue to use MPA to generate [α1, α2, … α N / K ]. In each transmission interval t, the transmitter receives the channel attenuation Δh d (t) For mismatch, MPA continuously uses Updated During the transmission phase,

[0243] Figure 30A 、 Figure 30B and Figure 30C shows how to divide the training between backpropagation and MPA, as previously described for Figure 23A and Figure 23B As described in the embodiment of Figure 31A 、 Figure 31B and Figure 31C As shown, the gradient descent layer is frozen while the MPA is used for transmission to update the power control coefficients.

[0244] After training, the following methods were taken:

[0245]

[0246] For iteration = 1:r

[0247]

[0248] The iteration ends.

[0249] Figure 32 The MPA operation after the training phase and during transmission with attenuation is summarized.

[0250] Returning to the example of encoding a 64QAM symbol into 3 dimensions, in the first simulation only and During the training phase, the decay is in From the known distribution N(m hd ,σ hd ) i,k During transmission, MPA uses the attenuation in is a time-varying random variable, but is known to the transmitter. The receiver, using only the decoder g(·), neither knows I don't know

[0251] In the second simulation, we considered and During the training phase, the decay is in From the known distribution N(m hd ,σ hd ) i，k A sample sampled from , and From the known distribution N(m hm ,σ hm ) ik During transmission, MPA uses the attenuation in is a time-varying random variable but is known to the transmitter, and It is from The updated distribution N(m hm ,σ hm ) i,k A sample sampled from . The simulation results are as follows Figure 33 shown.

[0252] In a random fading channel, the virtual user (no fading) generates the power control factor [α1, α2, … α N / K ] plays an important role. For each group (including K beams), MPA has two connections: one connection for real users with attenuation and the other connection for virtual users without attenuation. The goal of MPA is to move the maximum amount of information to real users and the rest to virtual users.

[0253] More importantly, in the N to K scenario, as mentioned before, it is difficult or impractical to insert a pilot on each beam because all beams are superimposed at the receiver (the pilots of these beams interfere with each other). With the described method, no pilot is needed (or much fewer pilots are needed). For the case where g(·) is used as the decoder to estimate All varying attenuation is transparent to the receiver, eliminating the need for synchronization, channel estimation, and equalization, enabling very high throughput and very low latency.

[0254] Example 5: Application to mmWAVE

[0255] In some embodiments, one or more of the above methods are applicable to mmWAVE applications.

[0256] In some embodiments, the base station includes an on-chip beamforming system. One chip includes a large number (e.g., hundreds or thousands) of mmWAVE antennas.

[0257] To solve the routing problem that may cause time offset between antennas, several adjacent antenna circuits on the chip can be coordinated into a group to generate a beam. Since the antennas of a beam can be connected to the same clock tree node, the antennas of this beam are closer to each other in chip intelligence and physics and synchronized with each other. To solve the power density problem, the antennas of a beam can be selected from the same power island. For example, Figure 34 As shown in Figure 1, groups of K adjacent beams are grouped together. For a given group of beams, their clocks come from the same clock tree node and they are located in the same power island. For a power island, there is a power control regulator to control all K beams on the island. Figure 34 In the diagram, a group is indicated at 3400, which has a power control regulator 3402 that controls all K beams and a clock tree 3404 connected to all beams. A group of K beams is always scheduled together, directed toward the same receiver. Although they are located in close proximity, they are "synchronized," evenly and equally powered, and spatially parallel (non-interfering). A group of beams is the basic scheduling granularity. An on-chip beamset can contain hundreds or thousands of beam groups. If a large number of beams are implemented on a chip or system-on-chip (SoC), traditional methods of synchronizing the clocks of all beams pose challenges to chip design, especially when designing a routed clock tree. As a result, a large, fully synchronized chip consumes a significant portion of its power (over 30%) due to the routed clock tree. Furthermore, this significant power consumption inevitably increases the chip area dedicated to ventilation. Therefore, for chip design, if a large, fully synchronized clock tree can be divided into several asynchronous sub-clock trees, it can significantly reduce power, area, and cost. This is because the time offset caused by chip routing between these beam groups is allowed, which can significantly alleviate routing issues. However, the time offsets between groups vary from chip to chip. They are fixed and can be measured. Static time offsets can be included in the of static decay.

[0258] Multiple on-chip beams can be used together to achieve N / K synchronized groups of beams. The time offset between on-chip beams is static and measurable, which can then be taken into account to be incorporated into the The static attenuation of the multiple on-chip beams can be installed separately and connected via wired connections. Given the location of the terminal, the air time delay can be determined and can be taken into account to be incorporated into the static attenuation.

[0259] Examples include Figure 35 As shown, a base station 3500 and two terminals are shown, including a first terminal 3502 located at point A 3503 and a second terminal 3504 located at point B 3505 .

[0260] The base station knows the location of the terminal. Given its location, user type and the available K uncorrelated dimensions, we can obtain (static delay and attenuation) and N(m hd ,σ hd ) and N(m hm ,σ hm In some embodiments, these are stored in the memory or database of the base station; or, these are predicted by, for example, a deep neural network. In some embodiments, It can be trained based on location, user type and available K irrelevant dimensions and stored in memory. For example, Figure 35 As shown, database 3510 stores pre-trained g() for each point within the base station's coverage area. In this case, the base station does not need to retrain g(), but instead simply sends the coefficients and architecture of g() to the terminal. If the terminal moves to a different point, the coefficients and architecture need to be updated. Ideally, the coverage area associated with each point is large enough that updates do not need to be frequent.

[0261] In operation, the receiver starts by sending a reference signal on the uplink, which will be detected / received by the base station's N beams. The base station then estimates the channel attenuation on each beam. The base station applies the UL / DL reciprocity assumption to form The base station uses MPA to target the current attenuation ( and or ) Calculate the power of each group of K beams [α1(t),α2(t),…α N / K (t)]. By updating [α1(t),α2(t),…α N / K (t)], the transmitter runs f(), w i and The power is adjusted to the granularity of the beam group and finally the signal is sent over the air in the downlink. The receiver receives the signals from K uncorrelated antennas and inputs them into g(). The terminal then continuously sends the reference signal on the uplink so that the base station will have the reference signal for the next iteration. The updated value of .

[0262] In operation, the base station stores N(m) can be adjusted by traditional statistical methods or methods based on deep neural networks. hd ,σ hd ) distribution. Base station utilization N(m) can be adjusted by traditional statistical methods or methods based on deep neural networks. hm ,σ hm ) distribution.

[0263] The base station calculates and and to monitor the relevance of and and

[0264] If γ d (t) exceeds the predefined limit and the user stays at that location, then may not be an adequate representation of the location, so Should be the latest However, after the update Will trigger the In some embodiments, the concept of transfer learning is utilized: this involves training based on the current neurons rather than an undefined set of neurons. If the retraining converges, the base station sends the new g() to the receiver. If the retraining does not converge, the location may be a "blind spot," meaning that N beams may not be sufficient and more beams should be used to communicate with the receiver in that part of the coverage area. In this case, the base station can allocate more beams to that area. Because the overall system dimensions have changed, new training is required. These "blind spot" situations may occur during the startup of a new base station.

[0265] If γ m (t) exceeds predefined limits, then more frequencies and / or more pilots are needed. If more frequencies and more pilots do not help, then N beams may not be enough, and more beams can be assigned to the receiver. Because the dimensionality has changed, new training is required.

[0266] Although this detailed example focuses 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 In some embodiments, an encoder is used that encodes more than one symbol together to have more coding gain. The same process is used when two QAM symbols are treated as one "symbol". Examples of OFDM-based systems are as follows: Figure 36A and Figure 36B shown.

[0267] Example 6: Incremental Redundancy

[0268] This embodiment provides a retransmission system and method based on incremental redundancy. Retransmission is a general concept and is not limited to the time domain. "Retransmission" means that incremental information is sent from another independent dimension. For example, the first transmission is performed along the three dimensions, while the retransmission is performed along the fourth dimension. Although multiple retransmissions are possible, the following example uses a single retransmission. In practice, three, four, or even more retransmissions are possible.

[0269] The maximum number of dimensions is defined as K. In the first transmission, the first (K–1) dimensions are used; in the second transmission (the first retransmission), incremental information is sent on the Kth dimension. The deep neural network is almost the same as the deep neural network without retransmission, except that there are two decoders (DNN): g1() for the first (K–1) dimensions and g2() for the Kth dimension, which merges the Kth dimension with the previous (K–1) dimensions.

[0270] The training goal is to achieve Where γ is the weight factor used at the receiver to weight the transmission, and (1 – γ) is the weight factor used for the second transmission. For a high value of γ, such as 0.9, it is expected that the first transmission will be successfully decoded most of the time. If there are more than two transmissions, their weights should sum to 1.

[0271] After training, the receiver will have two decoders, g1() and g2(). During the first transmission, the receiver will receive the K-1-dimensional signal and input it to g1() to estimate the symbol a. During the retransmission, the receiver will receive the K-th dimension of the signal and combine it with the previous K-1 dimensions into a K-dimensional vector for g2(). Training can be used to control the probability of successful decoding on the first transmission. By setting the parameter γ higher, more important information is pushed into the signal processed by g1().

[0272] In a specific simulation example, γ = 0.8 is used for initial transmission and (1 – γ) = 0.2 is used for retransmission. Figure 38 shown.

[0273] The results show that the coding gain obtained from the incremental information in one dimension is about 0.3dB.

[0274] Figure 39 An example communication system 100 is shown in which embodiments of the present disclosure may be implemented. Generally speaking, the communication system 100 enables multiple wireless or wired elements to transmit 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.

[0275] In this example, the communication system 100 includes electronic devices (EDs) 110a to 110c, radio access networks (RANs) 120a and 120b, a core network 130, a public switched telephone network (PSTN) 140, the Internet 150, and other networks 160. Figure 39 A certain number of these components or elements are shown, but any number of these components or elements may be included in the communication system 100 .

[0276] The EDs 110a to 110c are configured to operate and / or communicate in the communication system 100. For example, the EDs 110a to 110c are configured to transmit and / or receive via a wireless or wired communication channel. The EDs 110a to 110c represent any suitable end-user device configured for wireless operation and may include (or may be referred to as): user equipment (UE / user device), wireless transmit / receive unit (WTRU), mobile station, fixed or mobile subscriber unit, cellular phone, station (STA), machine type communication (MTC) device, personal digital assistant (PDA), smartphone, laptop, computer, tablet, wireless sensor, or consumer electronic device.

[0277] exist Figure 39 In the embodiment, RANs 120a and 120b include base stations 170a and 170b, respectively. Base stations 170a and 170b are configured to establish wireless connections with one or more of EDs 110a to 110c to enable access to any other base stations 170a and 170b, the core network 130, the PSTN 140, the Internet 150, and / or other networks 160. For example, base stations 170a and 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 NodeB (eNodeB), a home eNodeB, a gNodeB, a transmission point (TP), a site controller, an access point (AP), or a wireless router. Any ED 110a to 110c may alternatively or additionally be configured to connect to, access, or communicate with any other base station 170a and 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 RAN 120b, where the corresponding base station 170b accesses the core network 130 via the Internet 150, as shown. The detailed embodiments described herein refer to TPs, but more generally, any type of base station can be used with any of the embodiments described herein.

[0278] EDs 110a to 110c and base stations 170a and 170b are examples of communication devices that may be used to implement some or all of the functionality and / or embodiments described herein. Figure 39 In the illustrated embodiment, base station 170a forms part of the RAN 120a, which may include other base stations, base station controller(s) (BSCs), radio network controller(s) (RNCs), relay nodes, components, and / or devices. Any base station 170a, 170b may be a standalone component, as shown, or may be multiple components distributed across the corresponding RAN, and so on. Similarly, base station 170b is part of the RAN 120b, which may include other base stations, components, and / or devices. Each of base stations 170a and 170b transmits and / or receives wireless signals within a specific geographic area or region (sometimes referred to as a "cell" or "coverage area"). Cells may be further divided into cell sectors, and base stations 170a and 170b may, for example, employ multiple transceivers to provide service to multiple sectors. In some embodiments, there may be established pico or femto cells, which are supported by the radio access technology. In some embodiments, multiple transceivers may be used for each cell using multiple-input multiple-output (MIMO) technology, etc. The number of RANs 120a and 120b shown is merely exemplary, and any number of RANs may be considered when designing the communication system 100.

[0279] Base stations 170a and 170b communicate with one or more of EDs 110a through 110c via one or more air interfaces 190 using wireless communication links, such as radio frequency (RF), microwave, infrared, or the like. Air interfaces 190 may utilize any suitable wireless access technology. For example, communication system 100 may implement one or more channel access methods, such as code division multiple access (CDMA), time division multiple access (TDMA), frequency division multiple access (FDMA), orthogonal frequency division multiple access (OFDMA), or single carrier frequency division multiple access (SC-FDMA), in air interfaces 190.

[0280] Base stations 170a and 170b can implement Universal Mobile Telecommunication System (UMTS) terrestrial radio access (UTRA) to establish an air interface 190 using wideband CDMA (WCDMA). In this case, base stations 170a and 170b can implement protocols such as HSPA, HSPA+, etc., where HSPA+ optionally includes HSDPA and / or HSUPA. In addition, base stations 170a and 170b can establish an air interface 190 using LTE, LTE-A, LTE-B, and / or New Radio (NR) with Evolved UMTS Radio Access (E-UTRA). It is contemplated that communication system 100 may utilize multi-channel access capabilities, including such schemes as described above. Other wireless technologies used to implement the air interface include IEEE 802.11, 802.15, 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 can be used.

[0281] The RANs 120a and 120b communicate with the core network 130 to provide various services, such as voice, data, and other services, to the EDs 110a through 110c. The RANs 120a and 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 directly served by the core network 130 and may or may not employ the same radio access technology as the RANs 120a and / or RANs 120b. The core network 130 may also serve as a gateway access between (i) the RANs 120a and 120b and / or the EDs 110a through 110c, and (ii) other networks, such as the PSTN 140, the Internet 150, and other networks 160. In addition, some or all of the EDs 110a through 110c may include functionality to communicate with different wireless networks over different wireless links using different radio technologies and / or protocols. EDs may communicate with a service provider or switch (not shown) and with the Internet 150 via wired communication channels, rather than wirelessly (or in addition to wirelessly). PSTN 140 may include a circuit-switched telephone network for providing plain old telephone service (POTS). Internet 150 may include computer networks and / or subnets (intranets) and may include protocols such as IP, TCP, and UDP. EDs 110a through 110c may be multimode devices capable of operating in accordance with a variety of wireless access technologies and may include multiple transceivers necessary to support these technologies.

[0282] Figure 40A and Figure 40B An example device that can implement the methods and concepts provided by the present disclosure is shown. Specifically, Figure 40A An exemplary ED 110 is shown, Figure 40B An exemplary base station 170 is shown. These components may be used in the communication system 100 or any other suitable system.

[0283] like Figure 40A As shown, ED 110 includes at least one processing unit 200. Processing unit 200 implements various processing operations of ED 110. For example, processing unit 200 may perform signal encoding, data processing, power control, input / output processing, or any other function that enables ED 110 to operate within communication system 100. Processing unit 200 may also be used to implement some or all of the functions and / or embodiments detailed above. Each processing unit 200 includes any suitable processing or computing device for performing one or more operations. Each processing unit 200 may include a microprocessor, a microcontroller, a digital signal processor, a field programmable gate array, an application-specific integrated circuit, or the like.

[0284] ED 110 also includes at least one transceiver 202. Transceiver 202 is used to modulate data or other content for transmission by at least one antenna or network interface controller (NIC) 204. Transceiver 202 is also used to demodulate data or other content received via at least one antenna 204. Each transceiver 202 includes any suitable structure for generating signals for wireless or wired transmission and / or for processing signals received via wireless or wired means. Each antenna 204 includes any suitable structure for transmitting and / or receiving wireless or wired signals. One or more transceivers 202 can be used in ED 110. One or more antennas 204 can be used in ED 110. Although transceiver 202 is shown as a separate functional unit, it can also be implemented using at least one transmitter and at least one separate receiver.

[0285] ED 110 also includes one or more input / output devices 206 or interfaces (e.g., a wired interface connected to the Internet 150). Input / output devices 206 can interact with users or other devices on the 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, or touch screen, including network interface communications.

[0286] In addition, ED 110 includes at least one memory 208. Memory 208 stores instructions and data used, generated, or collected by ED 110. For example, memory 208 may store software instructions or modules for implementing some or all of the functionality and / or embodiments described above and executed by one or more processing units 200. Each memory 208 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices. Any suitable type of memory may be used, such as random access memory (RAM), read-only memory (ROM), a hard disk, an optical disk, a subscriber identity module (SIM) card, a memory stick, a secure digital (SD) memory card, etc.

[0287] like Figure 40B As shown, 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. Transceivers (not shown) may be used in place of transmitter 252 and receiver 254. Scheduler 253 may be coupled to processing unit 250. Scheduler 253 may be included within base station 170 or may operate separately from base station 170. Processing unit 250 implements various processing operations for base station 170, such as signal encoding, data processing, power control, input / output processing, or any other functions. Processing unit 250 may also be used to implement some or all of the functions and / or embodiments described in detail above. Each processing unit 250 includes any suitable processing or computing device for performing one or more operations. Each processing unit 250 may include a microprocessor, a microcontroller, a digital signal processor, a field programmable gate array, or an application-specific integrated circuit, among others.

[0288] 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 at least one transmitter 252 and at least one receiver 254 are shown as separate components, they may be combined into a transceiver. Each antenna 256 includes any suitable structure for transmitting and / or receiving wireless or wired signals. Although a shared antenna 256 is shown coupled to both transmitter 252 and receiver 254, one or more antennas 256 may be coupled to one or more transmitters 252, and one or more separate antennas 256 may be coupled to one or more receivers 254. Each memory 258 includes any suitable one or more volatile and / or non-volatile storage and retrieval devices, such as those described above in connection with ED 110. Memory 258 stores instructions and data used, generated, or collected by base station 170. For example, memory 258 may store software instructions or modules for implementing some or all of the functionality and / or embodiments described above and executed by one or more processing units 250.

[0289] Each input / output device 266 can interact with users or other devices in the network. Each input / output device 266 includes any suitable structure for providing information to a user or receiving information from a user / providing information from a user, including network interface communications.

[0290] Additional details regarding ED 110 and base station 170 are known to those skilled in the art. Therefore, for the sake of clarity, these details are omitted here.

[0291] Many modifications and variations of the present invention are possible in light of the above teachings. It is therefore to be understood that, within the scope of the appended claims, the present invention may be practiced otherwise than as specifically described herein.< / h> < / h> < / h> < / c> < / c>

Claims

1. A multi-beam transmission method using an N-beam transmitter to transmit to a receiver having K receive beams, characterized in that The method comprises: training a machine learning block and a linear encoder using gradient descent backpropagation dependent on feedback from the receiver, wherein the machine learning block implements a nonlinear encoder; For each input to be sent: Processing the input using the machine learning blocks to generate N / K sets of L-dimensional outputs; and Performing linear encoding on each group of L dimensional outputs using the linear encoder to generate groups of K dimensional outputs, thereby generating N / K groups of K dimensional encoded outputs and a total of N encoded outputs; and Each of the N / K groups of K outputs is transmitted from each group of K beams.

2. The method according to claim 1, characterized in that The machine learning block is a deep neural network.

3. The method according to any one of claims 1 to 2, characterized in that The training of the linear encoder is performed for each input to be sent.

4. The method according to any one of claims 1 to 2, characterized in that: Where "a" is the input to be sent; f(a) is the nonlinear encoding operation performed by the nonlinear encoder; represents the N / K sets of feature vectors generated by the machine learning block; represents the i-th group of feature vectors generated by the machine learning block; w i and are the parameters of the linear encoder; is a set of K re-edited entities generated at the output of said N-beam transmitter; Used for normalization; It’s noise; is a non-linear decoding operation performed by a machine learning block in the receiver; is the receiver's estimate of symbol a via decoder g(); and is the input to the nonlinear decoding operation.

5. The method according to any one of claims 1 to 2, characterized in that The method further comprises: applying a respective weight to each of the N / K groups of K outputs to generate a weighted group of K outputs; wherein transmitting each of the N / K groups of K outputs from each beam comprises transmitting the weighted group of K outputs; and The method also includes training the weights using a message passing algorithm.

6. The method according to claim 5, characterized in that: Where "a" is the input to be sent; f(a) is the nonlinear encoding operation performed by the nonlinear encoder; represents the N / K sets of feature vectors generated by the machine learning block; represents the i-th group of feature vectors generated by the machine learning block; w i and are the parameters of the linear encoder; is a set of K re-edited entities generated at the output of said N-beam transmitter; Used for normalization; It’s noise; are the weights trained by the message passing algorithm, c1, c2, ... c N / K represents the power control factor; is a non-linear decoding operation performed by a machine learning block in the receiver; is the receiver's estimate of symbol a via decoder g(); and is the input to the nonlinear decoding operation.

7. The method according to claim 5, characterized in that The applying a respective weight to each of the N / K groups of K outputs to generate a weighted group of K outputs includes applying a corresponding weight α to the i-th group of the N / K groups of K outputs. i ; and Use the message passing algorithm to determine α i , so that the contrasts are included in the i The amount of useful information in the output of the linear encoder in the calculated further virtual signal optimizes the amount of useful information in the output of the linear encoder included in the transmitted signal based on an optimization criterion.

8. The method according to claim 7, characterized in that The method further comprises: During initial training: Fix the weight α while training the machine learning block and the linear encoder i ;and In training the weight α i while treating the parameters of the machine learning block and the linear encoder as constants; and After training, a message passing algorithm is used to update the weights for each transmission.

9. The method according to claim 8, characterized in that During the initial training, the weight α is fixed while training the machine learning block and the linear encoder. i include: Performing the initial training based on the estimation of channel attenuation = the estimation of static channel attenuation + the probability distribution of dynamic attenuation + the probability distribution of mismatched channel attenuation; After training, the weights are updated for each transmission using a message passing algorithm, including updating values ​​based on channel attenuation.

10. The method according to any one of claims 1 to 2, 6 to 9, characterized in that The method further comprises: driving each group of K antennas from its own common clock tree; and Each group of K antennas is powered with their own uniform power.

11. The method according to claim 9, characterized in that Also includes: 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 comprising parameters of a nonlinear encoding operation of the transmitter and parameters of the linear encoder of the transmitter, and parameters of a nonlinear decoder located within the portion of the coverage area; and determining a location of the receiver; Wherein using the machine learning block and the linear encoder comprises using the stored training results based on the determined receiver position.

12. The method according to claim 11, characterized in that The method further comprises: Parameters of the non-linear decoder are sent to the receiver for the determined receiver location.

13. The method according to claim 12, characterized in that The method further comprises: Based on an analysis of the dynamic attenuation compared to the static attenuation, it is determined that the value of the static attenuation is not representative of the portion of the coverage area: Updating the value of the static attenuation to the latest estimate of the static attenuation; retraining the nonlinear encoder and the linear encoder using the latest estimate; and If the retraining does not converge, the number of beams is increased, and the nonlinear encoder and the linear encoder are retrained for more beams.

14. The method according to claim 12, characterized in that The method further comprises: Based on analysis of mismatched attenuation compared to static attenuation, the number and / or frequency of pilots are increased.

15. The method according to claim 14, characterized in that The method further comprises: After increasing the number and / or frequency of the pilots, more beams are determined based on further analysis of mismatch attenuation compared to static attenuation: Increasing the number of beams; and The nonlinear encoder and the linear encoder are retrained for more beams.

16. The method according to any one of claims 1 to 2, 6 to 9, 11 to 15, characterized in that The transmitting each of the N / K groups of K outputs from each group of K beams comprises: In a first transmission, T outputs of the respective groups of K outputs are transmitted from T beams of the respective groups of K beams, where T < K; In a second transmission, the remaining K - T outputs of the respective groups of K outputs are sent from the remaining K - T beams of the respective groups of K beams.

17. An apparatus for transmitting to a receiver having K receive beams, characterized in that The device comprises: processor and memory; wherein the processor and the memory implement a linear encoder and implement a machine learning block via a nonlinear encoder; The processor and the memory are configured to: training the machine learning block and the linear encoder using gradient descent backpropagation dependent on feedback from the receiver; For each input to be sent: Processing the input using the machine learning blocks to generate N / K sets of L-dimensional outputs; and Performing linear encoding on each group of L dimensional outputs using the linear encoder to generate groups of K dimensional outputs, thereby generating N / K groups of K dimensional encoded outputs and a total of N encoded outputs; and The apparatus further includes an N-beam transmitter for transmitting respective groups of the N / K groups of K outputs from respective groups of K beams.

18. The device according to claim 17, characterized in that Also used for: applying a respective weight to each of the N / K groups of K outputs to generate a weighted group of K outputs; wherein transmitting each of the N / K groups of K outputs from each beam comprises transmitting the weighted group of K outputs; and The apparatus is further configured to train the weights using a message passing algorithm.

19. The device according to claim 18, characterized in that Applying a respective weight to each of the N / K groups of K outputs to generate a weighted group of K outputs by: Apply the corresponding weight α to the i-th group of the N / K groups of K outputs i ; and Use the message passing algorithm to determine α i , so that the contrasts are included in the i The amount of useful information in the output of the linear encoder in the calculated further virtual signal optimizes the amount of useful information in the output of the linear encoder included in the transmitted signal based on an optimization criterion.

20. The device according to any one of claims 17 to 19, characterized in that Also used for: driving each group of K antennas from its own common clock tree; and Each group of K antennas is powered with their own uniform power.

21. The device according to any one of claims 17 to 19, characterized in that For transmitting each of the N / K groups of K outputs from each group of K beams by: In a first transmission, T outputs of the respective groups of K outputs are transmitted from T beams of the respective groups of K beams, where T < K; In a second transmission, the remaining K - T outputs of the respective groups of K outputs are sent from the remaining K - T beams of the respective groups of K beams.

Images