Method and apparatus for low complexity beamforming feedback in wireless local area networks
By using an automatic encoder in a wireless communication system to compress and decode the guidance matrix and perform neural network quantization in the training stage, the problems of beamforming feedback calculation complexity and high communication overhead are solved, and the system's efficient throughput and Low overhead.
Patent Information
- Application Number
- CN202410431766.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-11-16
- Filing Date
- 2024-04-11
- Publication Date
- 2024-11-08
AI Technical Summary
The prior art has high computational complexity and communication overhead for beamforming feedback in wireless communication systems, especially in IEEE 802.11 WLAN system, and there is a lack of a method to effectively reduce the communication overhead for training information updates.
Compression and decoding of the guide matrix using an automatic encoder, reduce communication overhead by transmitting the compressed guide matrix between the beamformer and the beamforming receiver, and perform bit width reduction in the training phase to reduce computational complexity. Spend.
The computational complexity and communication overhead of beamforming feedback generation are significantly reduced, and the overall throughput performance of the system is improved, especially when frequent training information updates are performed better than the prior art.
Smart Images

Figure CN118921093A_ABST
Abstract
Description
[0001] This application is based on and claims the benefit of priority of Korean Patent Application No. 10-2023-0059040 filed in the Korean Intellectual Property Office on May 8, 2023, and U.S. Patent Application No. 18 / 511,528 filed in the United States Patent and Trademark Office on November 16, 2023, the disclosures of which are incorporated herein by reference in their entirety. Technical Field
[0002] The disclosure relates to a system and method for reducing computational complexity in beamforming feedback in a wireless communication system, specifically, an Institute of Electrical and Electronics Engineers (IEEE) 802.11 wireless local area network (WLAN) system. Background Art
[0003] Beamforming (abbreviated as “BF”) is one of the key technologies used in wireless communication systems such as IEEE 802.11 WLAN systems (abbreviated as “WLAN”). TM -2021: “Part 11: Wireless LAN Medium Access Control (MAC) and Physical Layer (PHY) Specifications: “Enhancements for High-Efficiency WLANs”, 2021), the IEEE’s Artificial Intelligence and Machine Learning (AIML) Topic Interest Group (TIG) has discussed possible support for artificial intelligence and machine learning in the IEEE 802.11 WLAN standard.
[0004] Beamforming is one of the research topics in AIML TIG. Specifically, one goal of beamforming studied by AIML TIG is to reduce the beamforming feedback overhead under existing standards (e.g., IEEE 802.11 standard), thereby increasing the overall system throughput. In AIML TIG, one technical direction of beamforming uses conventional machine learning algorithms (e.g., K-means). Another technical direction of beamforming involves modern artificial intelligence schemes (which are described, for example, in the following documents: P. Sangdeh, H. Pirayesh, A. Mobiny and H. Zeng, "LB-SciFi: Online Learning-Based Channel Feedback for MU-MIMO in Wireless LANs," 2020 IEEE 28th International Conference on Network Protocols (ICNP), 2020; IEEE 802.11-23 / 0290r1, "Study on AI CSI Compression", 2023).
[0005] According to one document of the prior art (IEEE 802.11-22 / 1563r2, "AI ML Use Case", Sept. 2022 (IEEE 802.11-22 / 1663r2, "AI ML Use Case", September 2022)), data for training an artificial intelligence model, scheme, or operation for beamforming can be obtained from a "traditional" beamforming process. Throughout the disclosure, "traditional" beamforming may refer to a conventional method for performing beamforming in a wireless communication system without using an artificial intelligence or machine learning model, scheme, or operation.
[0006] In addition, prior art documents suggest that training may be performed at a beamformer (e.g., an access point (AP) in a WLAN). In this way, beamforming for WLAN may not introduce additional communication overhead in training data collection. However, the overhead of the beamformer sending the trained database or codebook along with the neural network coefficients to the beamformee may still be required. This may be referred to as "training information update" in the disclosure.
[0007] In the prior art (e.g., M. Nagel, et al., "A White Paper on Neural Network Quantization", Qualcomm AI Research, 2021), "neural network quantization" is a technical field in the field of general AI and is used to reduce memory requirements and computational complexity. However, the prior art does not propose how to use "neural network quantization" to reduce communication overhead. In addition, the prior art does not provide sufficient information on reducing communication overhead for training information updates in beamforming of wireless communication systems (such as IEEE 802.11 WLAN). Summary of the invention
[0008] A method and system for reducing the computational complexity of beamforming in a wireless communication system is provided. For example, an advantage disclosed is that the computational complexity of beamforming feedback generation is significantly reduced.
[0009] According to one disclosed aspect, a computer-implemented method performed by a first electronic device for reducing feedback overhead of beamforming in a wireless communication system includes: sending first data to a second electronic device; sending a data packet to the second electronic device; receiving second data from the second electronic device; extracting a compressed steering matrix from the second data; obtaining an uncompressed steering matrix based on the extracted compressed steering matrix by using a decoder portion of an autoencoder; and sending third data to the second electronic device via a radio signal that is beamformed based on the obtained uncompressed steering matrix.
[0010] According to another aspect of the disclosure, a computer-implemented method performed by a second electronic device for reducing feedback overhead of beamforming in a wireless communication system includes: receiving first data from a first electronic device; receiving a data packet from the first electronic device; estimating a channel between the first electronic device and the second electronic device based on the received data packet; obtaining a steering matrix based on the estimated channel; compressing the obtained steering matrix by using the first data and by using an encoder of an autoencoder, and placing the compressed steering matrix into a field of a data frame; and sending the data frame to the first electronic device.
[0011] According to another aspect of the disclosure, a first electronic device of a wireless communication system includes: a first transceiver; a first memory; a first processor, operatively connected to the first memory and the first transceiver, the first processor being configured to: send first data to a second electronic device; send a data packet to the second electronic device; receive second data from the second electronic device; extract a compressed steering matrix from the second data; obtain an uncompressed steering matrix based on the extracted compressed steering matrix by using a decoder part of an autoencoder; and send third data to the second electronic device via a radio signal beamformed based on the obtained uncompressed steering matrix. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] The above and other aspects, features and advantages of the disclosed specific embodiments will become more apparent from the following description in conjunction with the accompanying drawings, in which:
[0013] Figure 1 shows the overall structure of an Institute of Electrical and Electronics Engineers (IEEE) 802.11 wireless LAN (WLAN) as an example of a wireless communication system;
[0014] Figure 2 A wireless communication system having a beamformer (first electronic device) and a beamforming receiver (second electronic device) according to an embodiment is shown;
[0015] Figure 3 The present invention illustrates a beamforming process in a prior art (eg, a wireless communication system (IEEE 802.11 WLAN));
[0016] Figure 4 The structure of a beamforming (BF) action frame in the prior art (e.g., a wireless communication system (IEEE 802.11 WLAN)) is shown;
[0017] Figure 5A shows the structure of the encoder part of an autoencoder as an example of a neural network module;
[0018] Figure 5B shows the structure of the decoder part of the autoencoder;
[0019] Figure 6 The beamforming stage of the prior art autoencoder is shown;
[0020] Figure 7 shows the structure of an automatic encoder according to a disclosed embodiment;
[0021] Figure 8 shows the training phase of an autoencoder according to an embodiment of the present disclosure;
[0022] Fig. 9shows the beamforming stage of an autoencoder according to the disclosed embodiment;
[0023] Fig. 10A and Fig. 10B An example of software code for implementing the proposed model for autoencoder-based beamforming CSI feedback compression (L=2, N_2=16) is shown;
[0024] Fig.11 A first flow chart illustrating a disclosed embodiment; and
[0025] Fig.12 A second flow chart of the disclosed embodiment is shown. DETAILED DESCRIPTION
[0026] The terms used in the disclosure are provided only for describing specific embodiments and are not intended to limit the scope of other embodiments. Unless the context clearly indicates otherwise, the singular form includes plural indicators. The terms and words (including technical terms or scientific terms) used herein may have the same meaning as those generally understood by those skilled in the art. The terms usually defined in the dictionary may be interpreted as having the same or similar meaning as the contextual meaning of the relevant field. Unless otherwise defined, these terms should not be interpreted as ideal or overly formal meanings. Even if a term is defined in the disclosure, the term should not be interpreted as excluding the disclosed embodiments in a particular case.
[0027] According to one or more embodiments, the electronic device may be one of various types of electronic devices. The electronic device may include, for example, a portable communication device (e.g., a smart phone), a computer device, a portable multimedia device, a portable medical device, a camera, a wearable device, or a household appliance. According to the disclosed embodiments, the electronic device is not limited to the embodiments described above.
[0028] Figure 1 An overall structure of an Institute of Electrical and Electronics Engineers (IEEE) 802.11 wireless LAN (WLAN) as an example of a wireless communication system is shown.
[0029] When describing the disclosed embodiments in detail, the focus will be primarily on IEEE 802.11 WLAN, but the disclosure is also applicable to other communication systems (e.g., cellular communication systems (such as Long Term Evolution (LTE), Advanced LTE (LTE-A), New Radio (NR), Wireless Broadband (WiBro), Global System for Mobile Communications (GSM)) or short-range communication systems (such as Bluetooth and Near Field Communication (NFC))) with "similar technical background" and "with slightly modified channel forms without departing from the scope of the disclosure" as can be determined by a person skilled in the art.
[0030] exist Figure 1 In one embodiment, the wireless communication system 100 may include a first access point AP1, a second access point AP2, a first station STA1, a second station STA2, a third station STA3, and a fourth station STA4. The first access point AP1 and the second access point AP2 may access a network 102, which may include the Internet, an Internet Protocol (IP) network, or any other network.
[0031] The first access point AP1 may provide the first station STA1, the second station STA2, the third station STA3, and the fourth station STA4 with access to the network 102 within the first coverage area 104. Similarly, the second access point AP2 may also provide the third station STA3 and the fourth station STA4 with access to the network 102 within the second coverage area 106. In one embodiment, the first access point AP1 and the second access point AP2 may communicate with at least one of the first station STA1, the second station STA2, the third station STA3, and the fourth station STA4 based on Wireless Fidelity (WiFi) or any other WLAN access technology.
[0032] In one embodiment, access points (e.g., AP1, AP2) may be referred to as routers, gateways, etc. In one embodiment, stations (e.g., STA1, STA2, STA3, and STA4) may be referred to as mobile stations, subscriber stations, terminals, mobile terminals, wireless terminals, user equipment, users, etc. A station may be a mobile device (such as a mobile phone, a laptop computer, a wearable device, etc.) or a fixed device (such as a desktop computer, a smart TV, etc.).
[0033] The access point may allocate at least one resource unit (RU) to at least one station. The access point may send data through at least one allocated resource unit, and at least one station may receive data through at least one allocated resource unit. In 802.11ax (hereinafter referred to as HE), the access point may allocate only a single resource unit to at least one station, and in 802.11be or the next generation IEEE 802.11 standard, the access point may allocate a multiple resource unit (MRU) including two or more resource units to at least one station. For example, the first access point AP1 may allocate a resource unit to at least one of the first station STA1, the second station STA2, the third station STA3, and the fourth station STA4, and may send data through the allocated resource unit.
[0034] In one embodiment, the access point and the station may communicate with each other based on beamforming. Single-user beamforming may improve the transmission or reception performance of a single user, and multi-user beamforming may improve the transmission or reception performance of multiple users by eliminating or reducing interference between multiple users. In one embodiment, the access point may communicate with at least one station based on downlink beamforming. In one embodiment, the access point may communicate with at least one station based on uplink beamforming.
[0035] Throughout the disclosure, access points (e.g., Figure 1 AP1, AP2) may correspond to a "beamformer" or a "beamforming receiver". In one embodiment, a station (e.g., Figure 1 STA1, STA2, STA3 and STA4) may correspond to a “beamformer” or a “beamforming receiver”.
[0036] Figure 2 A wireless communication system 200 according to the disclosed embodiment is shown. In detail, Figure 2 A beamformer 202 (a first electronic device) and a beamforming receiver 204 (a second electronic device) are shown communicating with each other in a wireless communication system 200 . Figure 2 Each of the beamformer 202 (first electronic device) and the beamforming receiver 204 (second electronic device) of the wireless communication system 200 may be any device that communicates in the wireless communication system 200 and may be referred to as a device for wireless communication. In some example embodiments, each of the beamformer 202 (first electronic device) and the beamforming receiver 204 (second electronic device) may be an access point in a WLAN system (e.g., Figure 1 AP1, AP2) or stations (e.g., STA1, STA2, STA3, and STA4).
[0037] exist Figure 2 In the embodiment of the present invention, the beamformer 202 may include a first antenna 205, a first transceiver 206, a first processor 208, and a first memory 210. Similarly, the beamformer 204 may include a second antenna 211, a second transceiver 212, a second processor 214, and a second memory 216.
[0038] In one embodiment, the first antenna 205, the first transceiver 206, the first processor 208 and the first memory 210 may be included in one package or may be included in different packages respectively. In one embodiment, the second antenna 211, the second transceiver 212, the second processor 214 and the second memory 216 may be included in one package or may be included in different packages respectively.
[0039] In one embodiment, the first antenna 205 of the beamformer 202 may receive a signal from the second antenna 211 and provide the received signal to the first transceiver 206, and may transmit the signal provided from the first transceiver 206 to the second antenna 211. Similarly, the second antenna 211 may receive a signal from the first antenna 205 and provide the received signal to the second transceiver 212, and may transmit the signal provided from the second transceiver 212 to the first antenna 205.
[0040] In one embodiment, the first antenna 205 and the second antenna 211 may include multiple antennas for multiple input multiple output (MIMO). In one embodiment, the first antenna 205 and the second antenna 211 may include phased arrays for beamforming.
[0041] In one embodiment, the first transceiver 206 of the beamformer 202 may process a signal received from the beamforming receiver 204 through the first antenna 205, and may provide the processed signal to the first processor 208. In one embodiment, the first transceiver 206 may process the signal provided from the first processor 208 and output the processed signal through the first antenna 205. The second transceiver 212 similarly performs the above functions performed by the first transceiver 206.
[0042] In one embodiment, the first transceiver 206 and the second transceiver 212 may include one or more analog circuits (such as low noise amplifiers, mixers, filters, power amplifiers, oscillators, etc.).
[0043] In one embodiment, the first processor 208 of the beamformer 202 may extract information sent by the beamformed receiver 204 by processing the signal received from the first transceiver 206. For example, the first processor 208 may extract the information by demodulating and / or decoding the signal received from the first transceiver 206.
[0044] In one embodiment, the first processor 208 may generate a signal including information to be transmitted to the beamforming receiver 204 and provide the signal to the first transceiver 206. In one embodiment, the first processor 208 may provide a signal generated by encoding and / or modulating data to be transmitted to the beamforming receiver 204 to the first transceiver 206. The second processor 214 similarly performs the above functions performed by the first processor 208.
[0045] In one embodiment, the first processor 208 and the second processor 214 may include programmable components (such as a central processing unit (CPU), a digital signal processor (DSP), etc.), may include reconfigurable components (such as a field programmable gate array), and may include components that provide fixed functions (such as intellectual property (IP) cores).
[0046] In one embodiment, the first processor 208 may include or have access to a first memory 210 that stores data and / or a series of instructions. In one embodiment, the second processor 214 may include or have access to a second memory 216 that stores data and / or a series of instructions.
[0047] Figure 3 A conventional beamforming process in the prior art (eg, a wireless communication system (IEEE 802.11 WLAN)) is shown.
[0048] In one embodiment, N Tx is the number of antennas at the beamformer, N Rx is the number of antennas at the beamforming receiver. Throughout the disclosure, AP1 may be used as an example of a beamformer, and STA1 may be used as an example of a beamforming receiver. In an optional embodiment, Figure 1 The other devices (AP2, STA2, STA3 and STA4) may be beamformers or beamforming receivers.
[0049] In operation 300, the beamformer receiver 204 receives a null data packet (NDP) from the beamformer 202 for channel sounding.
[0050] In operation 302A, the beamforming receiver 204 estimates the NDPs having size N at all active subcarriers based on the received NDP. Rx ×N Tx Channel H.
[0051] In operation 302B, the beamforming receiver 204 receives the grouping parameters N g Select a plurality of subcarriers, and then decompose the matrix H of the selected subcarriers to obtain a matrix having size N Tx ×N Tx The complex unitary matrix V of . In one embodiment, the decomposition algorithm can be a singular value decomposition (SVD):
[0052] H=U∑V *
[0053] here,(·) * represents the conjugate transpose.
[0054] In operation 302C, the first N c (N c ≤min(N Tx ,N Rx )) column calculates the angle vector (including multiple φ and ψ). Here, N c Indicates the number of spatial streams used for beamforming data transmission.
[0055] In operation 302D, the angle vector is quantized (compressed) and Figure 4 In one embodiment, the "Compressed Beamforming Report" (CBR) field of the BF Action frame is sent. Figure 4 As shown in , a medium access control (MAC) protocol data unit (MMPDU) may include a Category / Action field, a MIMO control field, and a CBR field, wherein the CBR field may include a signal-to-noise ratio field and a compressed angle field.
[0056] As is known in the art, the angle ψ requires N b The angle φ needs to be (N b +2) bits. For each selected subcarrier, the number of bits required for the compression angle in CBR is:
[0057] N tot =N h N b +N h (N b +2)
[0058] Here, N h represents the number of angles φ or ψ for each selected feedback subcarrier, and N h is calculated as:
[0059] N h =[(N r -N c )+(N r -1)]N c / 2
[0060] Here, N r =N Tx Indicates the number of rows in V. CBR( Figure 4 The signal-to-noise ratio (SNR) field in ) can be obtained from Σ, where SNR is N c An indication of the signal quality of each spatial stream.
[0061] In operation 304 , the beamformer 202 receives a BF action frame from the beamforming receiver 204 .
[0062] In operation 306A, the beamformer 202 extracts the CBR received from the beamforming receiver 204 and the matrix V(:, 1:N c ) corresponds to the angle vector.
[0063] In operation 206B, the beamformer 202 constructs a beamformer having size N r ×N c The steering matrix (also called steering matrix) Q is used for beamforming data transmission. Usually, the last row of Q is a real value, indicating that the number of real values required for Q is P = 2N r N c -2. If the number of antennas increases, the number of bits required for the compressed angle in CBR may be very large, so the very large number of bits required would be an undesirable overhead.
[0064] In the prior art (e.g., P. Sangdeh, H. Pirayesh, A. Mobiny and H. Zeng, “LB-SciFi: Online Learning-Based Channel Feedback for MU-MIMO in Wireless LANs,” 2020 IEEE 28th International Conference on Network Protocols (ICNP), 2020), a feedforward neural network (FNN) is used as an autoencoder for the channel feedback process in WLAN.
[0065] Figure 5A An encoder 502 of the autoencoder 500 is shown, and Figure 5B Decoder 504 of autoencoder 500 is shown. Autoencoder 500 is an example of a neural network module that can be used in the disclosure. Other neural network modules (e.g., convolutional neural networks (CNNs) and transformers) can be used in the disclosure instead of autoencoder 500.
[0066] exist Figure 5A In the example, the encoder has (L+1) layers: an input layer, (L-1) hidden layers, and an output layer. The activation function of all neurons in layer l is defined as f l (·), where l∈[0,L]. When l=0,L (i.e., f0 (s 0 )=s 0 and f L (s L )=s l ), the activation function is linear, and for the other l (hidden layers), the activation function is leaky rectified linear unit (ReLU). The number of neurons in the lth layer is defined as N l The input to layer l (before activation) is defined as s l ,s l Size is N l ×1.W l and b l (like Figure 5A ) are the weight coefficients and bias coefficients between the (l-1)th layer and the lth layer respectively. The matrix W l The size is N l ×N l-1 , vector b l The size is N l ×1. Therefore, the following relationship is obtained: s l =W l f l-1 (s l-1 )+b l ,l∈[1,L]
[0067] like Figure 5B The decoder has a similar structure as shown in . The input is And the following relationship is obtained:
[0068]
[0069] Here, the matrix The size is N L-l ×N L-l+1 , and the vector Size is N L-l ×1.
[0070] The number of weight coefficients is the same at the encoder and decoder. However, the number of bias coefficients may be different. For example, the number of bias coefficients in the encoder is and the number of bias coefficients in the decoder is
[0071] In the prior art, during the training phase at the beamformer, the input of the decoder is set to the output of the encoder (i.e. ). Two separate autoencoders are used for φ and ψ respectively, so the encoder input s 0 is φ or ψ. By minimizing s 0The coefficients in the neural network (joint encoder-decoder) are trained by using the difference between the output of the encoder and the decoder. Once trained, the coefficients in the (two separate) encoders are sent to the beamforming receiver (i.e., training information update), which can then use the encoder to compress the angles before putting them into the CBR for feedback transmission. Note that neural networks typically use a "single precision floating point format" (e.g., the IEEE 754 standard format with a total of 32 bits per word) for coefficients to ensure learning performance.
[0072] Figure 6 The beamforming phase in the prior art is shown after the training phase is completed. Figure 6 The bold operations shown in (the conventional beamforming scheme with an autoencoder) are compared with Figure 3 The operation shown in (excluding the traditional beamforming scheme of the autoencoder) is different.
[0073] In operation 600A, the beamforming receiver 204 compresses the angle using an encoder.
[0074] In operation 600B, the beamformer receiver 204 quantizes the compressed angle and places the quantized compressed angle into the CBR. The quantization at the beamformer receiver 204 is to reduce feedback overhead. In one embodiment, both the beamformer receiver 204 and the beamformer 202 know the scale factor used for quantization so that the quantized compressed angle can be represented as an integer.
[0075] In operation 602A, the beamformer 202 extracts the quantized compressed angle from the CBR and dequantizes the compressed angle.
[0076] In operation 602B, the beamformer 202 uses a decoder and decompresses the compressed angles.
[0077] In operation 602C, the beamformer 202 constructs a BF steering matrix Q according to the angles.
[0078] As described below, the disclosure proposes multiple embodiments to improve the beamforming process in the above-mentioned prior art.
[0079] In one embodiment, neural network quantization may be used to reduce communication overhead during training of the autoencoder 500 .
[0080] In one embodiment, the steering matrix is used as input to the autoencoder 500, and thus, only one encoder-decoder pair may be required. Prior art techniques may require two encoder-decoder pairs.
[0081] In one embodiment, information bits may be needed to represent a "floating point" value for beamforming feedback. In the prior art, information bits may be needed to represent an "integer".
[0082] In one embodiment, neural network coefficients may be required in the training information update. In the prior art, both codebooks and neural network coefficients may be required.
[0083] The advantages of the disclosed embodiments are that the computational complexity of beamforming feedback generation is significantly reduced compared to the prior art. In addition, when training information updates are frequently performed, the disclosed embodiments outperform the prior art in terms of effective throughput performance.
[0084] Some disclosed embodiments may require neural network quantization and may require specific quantization parameters (e.g., scale factors) to be sent over the air. Some disclosed embodiments may require feedback for the compressed steering matrix. Prior art techniques may require feedback for two compression angles or for codebook indices.
[0085] Figure 7 , Figure 8 and Fig. 9 A proposed beamforming scheme having an encoder portion 700 and a decoder portion 702 is shown in accordance with the disclosed embodiments.
[0086] Figure 7 The structure of the auto-encoder used in the proposed disclosed beamforming process is shown, and the auto-encoder has an encoder part 700 and a decoder part 702. The encoder part 700 may include Figure 5A The encoder 502 shown in FIG. 1 and the decoder portion 702 may include Figure 5B The decoder 504 shown in FIG. Encoder 502 and decoder 504 may be included in autoencoder 500, but the disclosure is not limited to autoencoder 500. In one embodiment, autoencoder 500 may be replaced by CNN. In one embodiment, autoencoder 500 may be replaced by transformer.
[0087] exist Figure 7 At operation 704, the input to the encoder portion 700 is the steering matrix Q. A single input sample s 0 YesN 0 ×1 vector, N 0 ×1 vector collecting all N sc The selected feedback subcarriers (N 0 =N sc ) is the real or imaginary value of an element in the steering matrix Q of . The information is compressed in the dimension of subcarriers, and P input samples need to be processed by the trained autoencoder during inference.
[0088] At operation 706, the bit width reduction module is executed. In one embodiment, the bit width reduction module is stored in a memory (first memory 210 or second memory 216), which can be executed by a processor (first processor 208 or second processor 214). By executing the bit width reduction module, the bit width of the output of the encoder portion 700 is reduced by compressing the bit width from the single precision floating point format to N Q Bit floating point (N Q <32). In one embodiment, the bit width reduction module may correspond to a combination of software (eg, program, application) and hardware (eg, memory, processor) installed in the beamforming receiver 204 or the beamformer 202.
[0089] In the decoder portion 702, at operation 708, the N Q Bit floating point (N Q <32) to a single precision floating point format. Then, at the decoder 504, the single precision floating point format is used. There is a loss of precision due to the performance of the bit width reduction module 706, however, the feedback overhead can be reduced. In one embodiment, the decoder portion 702 outputs a reconstructed steering matrix 710.
[0090] In the prior art, the scale factor of quantization is "fixed". In contrast, because the output of the encoder 502 (of the encoder portion 700) has a single precision floating point format, the disclosed scale (exponent) is "dynamic". Furthermore, in the prior art, quantization is performed only during the inference phase. In contrast, the bit width reduction module 706 in the proposed encoder portion 700 can perform quantization during both the neural network training phase and the inference phase.
[0091] One of the key drawbacks of beamforming with artificial intelligence / machine learning in the prior art is the introduction of communication overhead during training. Examples of communication overhead are (1) a trained database, (2) a trained codebook, and (3) coefficients of a neural network. For example, these overheads may need to be wirelessly transmitted from a beamformer (e.g., AP1) to a beamformed receiver (e.g., STA1).
[0092] Figure 8 The training phase of the proposed scheme is shown as performed at the beamformer 202. At operation 800, the steering matrix Q (obtained from a conventional beamforming process) is sent to the autoencoder 500 for training. At operation 802, the autoencoder 500 is trained using the obtained steering matrix Q. One feature of the proposed scheme is that "neural network quantization" (804) is performed after the training is completed at operation 802 in order to reduce the communication overhead for sending the coefficients of the encoder portion 700.
[0093] In one embodiment, "neural network quantization" (804) is performed only on the weight coefficients (of the trained autoencoder), and single precision floating point format is maintained for the bias coefficients (806A). l or ) separately performs "neural network quantization" (804). For each weight matrix, first calculate the scaling factor (using N scale bits) (806B), and then quantize the coefficients (806C) to N w The integer (N w ≤32)(806D).
[0094] Note that the scaling factor can be different for different weight matrices, and when N w =32, quantization is not used. Both the scaling factor and the quantized integer weight coefficients need to be sent over the air from the beamformer 202 to the beamformer receiver 204 (operation 808). In this way, storage requirements and communication overhead for the coefficients can be reduced.
[0095] Fig. 9 The beamforming stage of the proposed scheme is shown. The beamforming stage can be Figure 8 At operation 900, the beamforming receiver 204 decomposes the channel (H) to obtain the matrix (V) and then obtains the steering matrix Q. In one embodiment, the beamforming receiver 204 decomposes the channel (H) via a Givens rotation according to the IEEE 802.11 standard.
[0096] At operation 902, the beamforming receiver 204 compresses the steering matrix Q using an encoder and places the compressed steering matrix Q into the CBR. That is, at operation 902, the beamforming receiver 204 uses a plurality of elements (the original single-precision floating-point bias coefficients, the quantized integer form N w The weight coefficient of the bit and N scale When the neural network is quantized (N w <32) is used, the compression (inference) can be expressed as:
[0097] s l =γ l "W l / γ l "f l-1 (s l-1 )+b l ,,l∈[1,L]
[0098] Here, "W l / γl ” is the weight coefficient in quantized integer form, γ l is the matrix W l The scaling factor.
[0099] At operation 904, the beamformer 202 extracts the compressed steering matrix Q from the CBR.
[0100] At operation 906, the beamformer 202 uses the decoder portion 702 of the autoencoder 500 to obtain the uncompressed steering matrix Q. Figure 7 As shown in FIG. 5 , the decoder portion 702 of the autoencoder 500 includes an N-based steering matrix Q for performing compression. Q The decoder 504 then uses the compressed single precision floating point of the steering matrix Q to obtain the uncompressed steering matrix Q.
[0101] At operation 308 , the beamformer 202 uses the obtained steering matrix Q to transmit beamforming data to the beamforming receiver 204 .
[0102] The disclosure also proposes a zero-layer autoencoder, as described below.
[0103] In one embodiment, Figure 7 The scheme proposed in may include the special case when L=0. In this embodiment, encoder 502 (of encoder section 700) is not used for compression, so the steering matrix Q (704) directly enters Figure 7 Therefore, the training phase ( Figure 8 ), and there is no overhead for training information update. The purpose of this special design is to further reduce the overhead in training information update and the complexity of feedback generation, although the feedback overhead may increase.
[0104] The disclosed proposed beamforming scheme performs better than conventional schemes in the prior art. A benefit of the proposed beamforming scheme is a significant reduction in computational complexity. For example, the following [Table 1] shows that the proposed beamforming scheme performs significantly better than the conventional beamforming scheme and the autoencoder scheme in the prior art (P. Sangdeh, H. Pirayesh, A. Mobiny and H. Zeng, "LB-SciFi: Online Learning-Based Channel Feedback for MU-MIMO in Wireless LANs," 2020 IEEE 28th International Conference on Network Protocols (ICNP), Oct. 2020) in at least two performance parameters: "communication overhead in the training phase (KB)" and "number of multiplications per CBR":
[0105] [Table 1]
[0106]
[0107] In addition, another benefit of the proposed beamforming scheme is as mentioned above Figure 7 The accuracy of the described reconstructed steering matrix Q, and hence the effective throughput performance, is much better.
[0108] Fig. 10A and Fig. 10B An example of software code for implementing the proposed model for beamforming channel state information (CSI) feedback compression (L=2, N_2=16) based on an autoencoder is shown. The software code can be run with Python version 3.6 and PyTorch version 1.9.0. For example, the module "def encoder (self, x)" implements the encoder part 700 of the above autoencoder, and the module "def decoder (self, x)" implements the decoder part 702 of the above autoencoder.
[0109] Fig.11 The first flow chart of the disclosed embodiment is shown. Specifically, Fig.11 A computer-implemented operating procedure performed by a first electronic device (beamformer 202) for reducing feedback overhead of beamforming in a wireless communication system is shown. The computer-implemented method includes the following operations.
[0110] At operation 1100, the first electronic device may send first data to the second electronic device (beamforming receiver 204). Operation 1100 may correspond to the above-described operation 808. In one embodiment, the first data may include: (1) a plurality of weight coefficients of the encoder 502 of the autoencoder and (2) a plurality of scale factors of the encoder 502 of the autoencoder.
[0111] At operation 1102, the first electronic device may send a data packet to the second electronic device. Operation 1102 may correspond to the above-described operation 300. In one embodiment, the data packet is a null data packet (NDP) of a wireless communication system.
[0112] At operation 1104, the first electronic device may receive second data from the second electronic device. Operation 1104 may correspond to the above-described operation 304. In one embodiment, the second data may include a beamforming (BF) action frame, and the BF action frame may include a compressed beamforming report (CBR) field.
[0113] At operation 1106, the first electronic device may extract a compressed steering matrix from the second data. Operation 1106 may correspond to operation 904 described above.
[0114] At operation 1108 , the first electronic device may obtain a non-compressed steering matrix by using a decoder part of the autoencoder based on the extracted compressed steering matrix. Operation 1108 may correspond to the above-described operation 906 .
[0115] At operation 1110 , the first electronic device may transmit the third data to the second electronic device via the radio signal beamformed based on the obtained non-compressed steering matrix. Operation 1110 may correspond to the above-described operation 308 .
[0116] Fig.12 A second flow chart of the disclosed embodiment is shown. Specifically, Fig.12 A computer-implemented process for reducing feedback overhead for beamforming in a wireless communication system is shown, performed by a second electronic device. Fig.12 The computer-implemented operation process includes the following operations.
[0117] At operation 1200, the second electronic device may receive first data from the first electronic device. Operation 1200 may correspond to operation 808 described above.
[0118] At operation 1202, the second electronic device may receive a data packet from the first electronic device. Operation 1202 may correspond to operation 300 described above.
[0119] At operation 1204, the second electronic device may estimate a channel between the first electronic device and the second electronic device based on the received data packet. Operation 1204 may correspond to the above-described operation 302A.
[0120] At operation 1206, the second electronic device may obtain a steering matrix based on the estimated channel. Operation 1206 may correspond to operation 900 described above.
[0121] At operation 1208, the second electronic device may compress the obtained steering matrix by using the first data and by using the encoder 502 of the auto encoder, and put the compressed steering matrix into a field (CBR) of the data frame (BF action frame). Operation 1208 may correspond to the above-described operation 902.
[0122] At operation 1210 , the second electronic device may transmit the data frame to the first electronic device. Operation 1210 may correspond to the above-described operation 304 .
[0123] The terms disclosed and used therein are not intended to limit the technical features set forth herein to specific embodiments, and include various changes, equivalents or replacements of the corresponding embodiments. With respect to the description of the accompanying drawings, similar reference numerals may be used to refer to similar or related elements. Unless the relevant context clearly indicates otherwise, the singular form of the noun corresponding to the item may include one or more things. As used herein, each of the phrases such as "A or B", "at least one of A and B", "at least one of A or B", "A, B or C", "at least one of A, B and C" and "at least one of A, B or C" may include any one or all possible combinations of the items listed together in a corresponding phrase. As used herein, terms such as "1st" and "2nd" or "first" and "second" may be used to distinguish only the corresponding component from another component, and do not limit the component in other aspects (e.g., importance or order). It will be understood that if an element (e.g., a first element) is referred to as being "coupled with" another element (e.g., a second element), "coupled to" another element, "connected with" another element, or "connected to" another element, with or without the term "operably" or "communicatively", it means that the element can be coupled to the other element directly (e.g., via a wire), wirelessly, or via a third element.
[0124] As used in conjunction with the disclosure, the term "module" may include units implemented in hardware, software, or firmware, and may be used interchangeably with other terms (e.g., logic, logic block, component, or circuit system). A module may be a single integrated component or its smallest unit or component suitable for performing one or more functions. For example, according to an embodiment, a module may be implemented in the form of an application specific integrated circuit (ASIC).
[0125] One or more embodiments described herein may be implemented as software including one or more instructions stored in a machine-readable storage medium. For example, a processor of a machine may call at least one of the one or more instructions stored in a storage medium and execute it with or without one or more other components under the control of the processor. This allows the machine to be operated to perform at least one function according to at least one instruction called. One or more instructions may include code generated by a compiler or code executable by an interpreter. The machine-readable storage medium may be provided in the form of a non-transitory storage medium. Among them, the term "non-transitory" only means that the storage medium is a tangible device and does not include signals (e.g., electromagnetic waves), but the term does not distinguish between where data is semi-permanently stored in the storage medium and where data is temporarily stored in the storage medium.
[0126] According to an embodiment, the method according to one or more embodiments disclosed may be included and provided in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be released in the form of a machine-readable storage medium (e.g., a compact disc read-only memory (CD-ROM)), or may be downloaded through an application store (e.g., PlayStore). TM ) online (e.g., downloading or uploading), or directly between two user devices (e.g., smart phones) (e.g., downloading or uploading). If published online, at least a portion of the computer program product may be temporarily generated or at least temporarily stored in a machine-readable storage medium (e.g., a memory of a manufacturer's server, a server of an application store, or a relay server).
[0127] According to one or more embodiments, each of the above-mentioned components (e.g., a module or a program) may include a single entity or multiple entities. According to one or more embodiments, one or more of the above-mentioned components may be omitted, or one or more other components may be added. Alternatively or additionally, multiple components (e.g., modules or programs) may be integrated into a single component. In this case, according to one or more embodiments, the integrated component may still perform one or more functions of each of the multiple components in the same or similar manner as performed by the corresponding components in the multiple components before integration. According to one or more embodiments, the operations performed by a module, a program, or another component may be performed sequentially, in parallel, repeatedly, or heuristically, or one or more of the operations may be performed in a different order or omitted, or one or more other operations may be added.
[0128] According to one or more embodiments, in a non-volatile storage medium storing instructions, when executed by at least one processor, the instructions may be configured to cause the at least one processor to perform at least one operation. The at least one operation may include: displaying an application screen of a running application on a display, identifying a data input field included in the application screen, identifying a data type corresponding to the data input field, displaying at least one external electronic device around the electronic device that can provide data corresponding to the identified data type, receiving data corresponding to the identified data type from an external electronic device selected from at least one external electronic device through a communication module, and adding the received data to the data input field.
[0129] The disclosed embodiments described in this specification and the drawings are presented only as specific examples to easily explain the technical content according to the disclosed embodiments and to help understand the disclosed embodiments, and are not intended to limit the scope of the disclosed embodiments. Therefore, the scope of the disclosed one or more embodiments should be interpreted as including all changes or modifications obtained from the technical spirit of the disclosed one or more embodiments in addition to the embodiments disclosed herein.
Claims
1. A computer-implemented method performed by a first electronic device for reducing feedback overhead of beamforming in a wireless communication system, the computer-implemented method include: sending the first data to a second electronic device; sending the data packet to a second electronic device; receiving second data from a second electronic device; extracting a compressed steering matrix from the second data; Based on the extracted compressed steering matrix, an uncompressed steering matrix is obtained by using a decoder part of a neural network module; as well as The third data is transmitted to the second electronic device via the radio signal beamformed based on the obtained non-compressed steering matrix.
2. The computer-implemented method of claim 1, in, In a wireless communication system, a first electronic device is a beamformer and a second electronic device is a beamforming receiver.
3. The computer-implemented method of claim 1, in, The neural network modules correspond to autoencoders, convolutional neural networks, or transformers.
4. The computer-implemented method of claim 1, in, The data packet is a null data packet of the wireless communication system. wherein the second data includes a beamforming action frame, and The beamforming action frame includes a compressed beamforming report field.
5. The computer-implemented method of claim 1, in, The first data includes: a plurality of weight coefficients of an encoder of the neural network module, and a plurality of scaling factors of the encoder of the neural network module.
6. The computer-implemented method according to any one of claims 1 to 5, further comprising: include: Get the steering matrix; Training a neural network module based on the obtained steering matrix, the neural network module comprising an encoder and a decoder; as well as performing neural network quantization on a plurality of weight coefficients of the encoder and a plurality of scale factors of the encoder, The first data includes the multiple weight coefficients of the encoder and the multiple scale factors of the encoder.
7. The computer-implemented method of claim 6, in, The step of obtaining a steering matrix includes obtaining a steering matrix from a conventional beamforming process.
8. The computer-implemented method of claim 6, in, The steps to perform neural network quantization include: Maintains multiple bias coefficients of the encoder in single precision floating point format; quantizing the plurality of scale factors of the encoder; and The plurality of weight coefficients of the encoder are quantized.
9. The computer-implemented method of claim 8, in, The plurality of quantized weight coefficients of the encoder are integers.
10. A computer-implemented method performed by a second electronic device for reducing feedback overhead of beamforming in a wireless communication system, the computer-implemented method include: receiving first data from a first electronic device; receiving a data packet from a first electronic device; estimating a channel between the first electronic device and the second electronic device based on the received data packet; obtaining a steering matrix based on the estimated channel; Compressing the obtained steering matrix by using the first data and by using an encoder of the neural network module, and placing the compressed steering matrix into a field of the data frame; as well as The data frame is sent to the first electronic device.
11. The computer-implemented method of claim 10, in, The first data includes: a plurality of weight coefficients of an encoder of the neural network module, and a plurality of scaling factors of the encoder of the neural network module.
12. The computer-implemented method of claim 10, in, The data packet is a null data packet of the wireless communication system.
13. The computer-implemented method according to any one of claims 10 to 12, further comprising: include: Third data is received from the first electronic device via the radio signal beamformed based on the steering matrix.
14. A first electronic device of a wireless communication system, the first electronic device include: a first transceiver; a first memory; A first processor is operatively connected to the first memory and the first transceiver, and the first processor is configured to: sending the first data to a second electronic device; sending the data packet to a second electronic device; receiving second data from a second electronic device; extracting a compressed steering matrix from the second data; Based on the extracted compressed steering matrix, an uncompressed steering matrix is obtained by using a decoder part of a neural network module; as well as The third data is transmitted to the second electronic device via the radio signal beamformed based on the obtained non-compressed steering matrix.
15. The first electronic device according to claim 14, in, In a wireless communication system, a first electronic device is a beamformer and a second electronic device is a beamforming receiver.
16. The first electronic device according to claim 14, in, The data packet is a null data packet of the wireless communication system.
17. The first electronic device according to claim 14, in, The second data includes a beamforming action frame, and The beamforming action frame includes a compressed beamforming report field.
18. The first electronic device according to claim 14, in, The first data includes: a plurality of weight coefficients of an encoder of the neural network module, and a plurality of scaling factors of the encoder of the neural network module.
19. The first electronic device according to any one of claims 14 to 18, in, The first processor is further configured to: Get the steering matrix; Training a neural network module based on the obtained steering matrix, the neural network module comprising an encoder and a decoder; and performing neural network quantization on a plurality of weight coefficients of the encoder and a plurality of scale factors of the encoder, The first data includes the multiple weight coefficients of the encoder and the multiple scale factors of the encoder.
