A beam alignment method based on convolutional long short-term memory network

Abstract

The present invention belongs to the field of wireless communication and deep learning technology, and specifically relates to a beam alignment method based on a convolutional long short-term memory network. The present invention uses spatial correlation to deeply compress the number of detection beams, selects a part of the beams for detection, and then builds an optimal beam inference network based on a convolutional long short-term memory network to extract the time and space information contained in the received signal-to-noise ratio of the detection beam at continuous moments, so as to improve the inference accuracy of the optimal beam. Using the technology proposed in the present invention, the access point does not need to detect all beams, but only needs to detect a small part of fixed beams, so as to accurately infer the optimal beam and complete beam alignment, thereby better balancing the data transmission time and data transmission rate and improving the system throughput.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless communication and deep learning, and specifically relates to a beam alignment method based on a convolutional long short-term memory network. Background Art

[0002] With the rapid growth of demand for mobile communication services, the conflict between limited spectrum resources and capacity requirements has become increasingly apparent. The millimeter wave band has attracted much attention due to its abundant spectrum resources. However, compared with sub-6G technology, the millimeter wave band has serious path loss and air absorption loss. In order to enhance the reception capability of millimeter wave signals, beamforming technology is usually used to concentrate the energy of the transmitted millimeter wave signal into a narrow beam range. Subsequently, beam training is used to align the transmit beam with the receive beam.

[0003] The traditional exhaustive beam search (EBS) method is a beam training method based on a codebook, which uses a set of codewords to define different beams, then detects all beams according to the codewords in the codebook, and defines the beam with the highest received signal-to-noise ratio as the optimal beam. However, the system overhead of the EBS method is proportional to the number of codewords in the codebook, which will reduce the data transmission time of the system and thus limit the average throughput of the system.

[0004] In order to reduce the system overhead of the EBS method and improve the throughput of the millimeter wave communication system, deep neural networks are gradually introduced into the millimeter wave beam training process. Existing beam training technologies can be mainly divided into two categories:

[0005] (1) Beam training technology based on wide beam codebook. Since the number of codewords contained in a wide beam codebook is usually smaller than that in a narrow beam codebook, the overhead of beam scanning is reduced. This technology scans the wide beam to obtain the received signal under the wide beam, and uses the received signal of the wide beam to predict the optimal narrow beam to improve the throughput of the system. In order to improve the inference accuracy of the optimal narrow beam, this technology uses a deep neural network, takes the received wide beam signal as the input of the deep neural network, and takes the corresponding optimal narrow beam as the output, and establishes a mapping relationship between the wide beam received signal and the optimal narrow beam.

[0006] (2) Beam training technology based on narrow beam codebook. This technology reduces the beam scanning overhead by scanning part of the narrow beam. However, scanning part of the narrow beam cannot obtain all the information of the beam space, which is not conducive to the inference of the optimal narrow beam. This technology uses the spatial correlation between the received signal-to-noise ratios of adjacent beams and proposes a method for selecting the narrow beam to be scanned, so that the received signal-to-noise ratio of the beam to be scanned contains as much beam space information as possible. At the same time, a deep convolutional neural network is used to extract the information contained in the received signal-to-noise ratio of the beam to be scanned to improve the inference accuracy of the optimal narrow beam.

[0007] Among the above technologies, the beam training technology based on wide beam codebooks not only requires redesigning the codebook, but also requires scanning all wide beams and narrow beams to construct a data set, which is highly complex. On the one hand, the beam training technology based on narrow beams does not fully explore and utilize the spatial correlation between different beams, and on the other hand, it also ignores the temporal correlation of the received signal-to-noise ratio of the scanned beam in continuous time. Summary of the invention

[0008] In view of the above problems, the present invention aims to further compress the number of scanning beams by utilizing the statistical correlation between narrow beams, and at the same time combine the spatial information and time information contained in the received signal-to-noise ratio of the scanning beam to improve the inference accuracy of the optimal narrow beam.

[0009] The present invention considers a millimeter wave downlink communication link, in which the access point configuration has N ap The antenna array is a uniform planar antenna array with N antennas, and the user equipment is equipped with an omnidirectional antenna. The antennas of the access point are distributed in a rectangular shape, with a total of N x OK, N y Column, that is, N ap =N x N y .

[0010] The present invention adopts the extended Saleh-Valenzuela geometric channel model to represent the sparse millimeter wave channel between the access point and the user equipment. The corresponding channel vector can be expressed as

[0011]

[0012] Where L p represents the number of multipaths, It is the first p The complex gain of the paths, It is the first p The antenna array response vector of the path can be expressed as

[0013]

[0014] in, Indicates the firstp The elevation angle of the path, Indicates the first p The azimuth of the path, λ represents the signal wavelength, and d represents the spacing between antenna elements.

[0015] The present invention uses an analog beamforming method based on a discrete Fourier transform (DFT) codebook. The discrete Fourier transform codebook in the present invention contains M code words. If we choose f i (t) is used to configure the analog beamforming of the access point. Then the received signal at the user at time t can be expressed as

[0016]

[0017] Where P t is the transmission power, s(t) is the transmission signal, is Gaussian white noise.

[0018] The present invention considers that people using mobile devices walk in a certain area, and the blockage of millimeter wave transmission is mainly their bodies. Intuitively speaking, if people walk close to the access point, the link will not be blocked; if people move away from the access point, the link is likely to be blocked. The present invention adopts the blocking model defined in the 3GPP protocol.

[0019] The frame structure used in the present invention is derived from the IEEE 802.11ad / ay protocol, such as Figure 1 As shown in the figure, the beacon interval (BI) is used to represent the time frame of signal transmission, and the millimeter wave channel remains unchanged during the duration of each BI and changes independently between different BIs. Specifically, each BI consists of two parts: the beacon header interval (BHI) phase for signal control and the data transmission interval (DTI) phase for data transmission. In the BHI phase, the access point detects multiple codewords or beams in continuous sector sweep (SSW) subframes for beam training. In each SSW, the access point selects a beam for detection, and the user equipment can obtain the signal-to-noise ratio corresponding to the detection beam. Through multiple detections, the signal-to-noise ratios of different detection beams can be obtained. In the feedback (FB) subframe, the user equipment infers the optimal beam based on the detected signal-to-noise ratio and feeds back the optimal beam number to the access point. After receiving the optimal beam number, the access point returns an acknowledgment (ACK) subframe to the user equipment. Then, in the DTI phase, the access point can use the optimal beam for data transmission.

[0020] Based on the above frame structure, the present invention defines the duration of a BHI as

[0021] T BHI =N p ·T SSW +T FB +T ACK +T IFS ,

[0022] Among them, N p represents the number of detection beams, T SSW represents the duration of each SSW, T FB and T ACK Respectively represent the duration of FB and ACK subframes, T IFS Indicates the duration of all inter-subframe intervals.

[0023] The IEEE 802.11ad protocol uses exhaustive beam detection for beam training, that is, N p =M. After exhaustive detection, the user equipment can obtain the received signal-to-noise ratio of all beams, that is,

[0024] Γ(k)=[γ 1 (k),…,γ i (k),…γ M (k)],

[0025] where γ i (k) is the signal-to-noise ratio of the i-th beam at the k-th BI. According to the received signal-to-noise ratio, the optimal beam can be expressed as

[0026] i opt (k) = arg maxΓ(k),

[0027] The average throughput of the kth BI can be expressed as

[0028]

[0029] Where B represents the system bandwidth.

[0030] According to the above analysis, although exhaustive detection can maximize the transmission rate, it will also reduce the data transmission time. That is, the exhaustive beam detection method cannot balance the data transmission rate and time, and therefore cannot achieve the maximum throughput. The present invention aims to optimize the number and selection method of the detection beams to improve the average throughput.

[0031] Deep spatiotemporal beam training technology design

[0032] 1) Technical ideas of combining time and space information with beam training

[0033] As mentioned above, in order to balance the transmission rate and duration and improve the average throughput, the detection beam needs to be optimized. To this end, the present invention uses the spatial correlation between different beams and the temporal correlation between received detection signals to develop a deep spatiotemporal beam training technology.

[0034] It can be observed that for a user device, the received signal-to-noise ratios of different detection beams are spatially correlated. That is, only part of the beams can be detected to infer the optimal beam. In the present invention, some fixed beams are selected in the entire beam space based on the spatial statistical characteristics between all beams. By detecting these beams, the access point can infer the optimal beam by using the spatial correlation between the signal-to-noise ratio of the detection beam and the optimal beam. In addition, as the user device continues to move in a certain area, the signal-to-noise ratio of the detection beam also has time correlation. That is, the detection beam signal-to-noise ratio in the first few BIs contains some information about the optimal beam in the current BI. In order to accurately infer the optimal beam, the present invention uses a convolutional long short-term memory network (Convolutional Long-short Term Memory, ConvLSTM) to extract the spatiotemporal correlation between the signal-to-noise ratio of the detection beam in the first few and current BIs and the optimal beam of the current BI. Next, the present invention will first introduce the detection beam selection method, and then design an optimal beam inference algorithm based on ConvLSTM.

[0035] 2) Probe beam selection method

[0036] The present invention aims to design partial beams for detection, that is, beams containing similar information or strong spatial correlation should be avoided. The present invention defines spatial mutual information to describe the spatial correlation between two beams. Specifically, the smaller the spatial mutual information between two beams, the richer the amount of spatial information contained therein. In this way, beams with minimum spatial mutual information can be selected for detection, and the received signal-to-noise ratio of these beams can provide the most spatial information, thereby improving the inference accuracy of the optimal beam.

[0037] In order to obtain the spatial mutual information of each beam, it is first necessary to perform exhaustive beam detection in K consecutive BIs to collect the received signal-to-noise ratio corresponding to each beam. It is worth noting that in different BIs, as the user equipment moves, the received signal-to-noise ratio of each beam is different. Subsequently, this technology will test the signal-to-noise ratio set Γ of beam i obtained in different BIs i (i=1,2,…M) is regarded as a random variable, γ i (k)(k=1,2,…,K) is a sample of the random variable.

[0038] According to information theory, Γ i The Shannon entropy can be expressed as:

[0039]

[0040] Γ i and Γ j The joint entropy between can be expressed as:

[0041]

[0042] Among them, p x Represents Γ i The probability density function, p xy Represents Γ i and Γ j The joint probability density function of .

[0043] The present invention uses Γ i and Γ j The discrete histogram distribution between is used to approximate its probability distribution, and then p is obtained x and p xy The approximate estimate of . The histogram width of beam i is:

[0044] W i =max(Γ i )-min(Γ j )

[0045] The histogram is then divided into K equal parts, and the width of each part is:

[0046]

[0047] The present invention defines i =min(Γ i ), p x Approximately equal to γ i (k) The probability of being within the range of the xth histogram, that is,

[0048]

[0049] Where 1(·) is an indicator function, which takes the value 1 when the inequality in the brackets holds, and takes the value 0 otherwise. Similarly, p xy It can be approximated as γ i (k) and γ j (k) are located at Γ i The xth histogram and Γ j The probability of being within the range of the yth histogram is:

[0050]

[0051] The present invention defines Γ i and Γ j The spatial mutual information between is:

[0052] I(Γ i Γ j )=H(Γ i )+H(Γ j )-H(Γ i Γ j ),

[0053] The average spatial mutual information between beam i and all other beams is:

[0054]

[0055] According to the above formula, the average spatial mutual information of all beams can be obtained, that is:

[0056]

[0057] Then, the N with the minimum average spatial mutual information can be selected p A beam is used as a probe beam.

[0058] 3) Design of optimal beam inference algorithm based on ConvLSTM

[0059] Convolutional neural networks have excellent performance in extracting spatial information, but ignore temporal information. Long short-term memory networks can extract temporal information, but cannot fully utilize spatial information. The goal of this technology is to simultaneously utilize the spatial and temporal information contained in the signal-to-noise ratio of the detection beam to improve the accuracy of optimal beam inference. The present invention uses the powerful spatiotemporal information extraction capability of ConvLSTM to model the optimal beam inference problem as a multi-classification task and designs an optimal beam inference algorithm based on ConvLSTM. Its structure is as follows: Figure 2 shown.

[0060] ① Input layer:

[0061] The input of the network is the received signal-to-noise ratio of the detection beam at the current BI and the previous S-1 BI. To represent the input of the network, X 0 Each row of represents the signal-to-noise ratio of the detection beam at a BI, that is:

[0062]

[0063] Where s=1,2,…,S, subscript set is the sequence number set of the detection beam.

[0064] ②ConvLSTM layer:

[0065] After receiving the input data, the present invention uses L layers of ConvLSTM layers from X 0The present invention defines the input and output of the lth ConvLSTM layer (l = 1, 2, ..., L) as X l-1 and H l , then:

[0066] H l =CL l (X l-1 )

[0067] Where CL l (·) is the ConvLSTM function, which represents the calculation process of the lth ConvLSTM layer. It is worth noting that H l It is also the hidden state of the lth ConvLSTM layer, which consists of S extracted feature matrices, corresponding to S BIs. In addition, during the network calculation process, the output of the l-1th layer will be transmitted to the lth layer network as the input of the lth layer.

[0068] The complete feature extraction process can be expressed as:

[0069]

[0070] in Represents the composite operation between ConvLSTM functions, namely:

[0071]

[0072] ③Flatten layer:

[0073] The spatiotemporal information H is extracted through multiple ConvLSTM layers L Afterwards, the present invention selects H L The last hidden state in is taken as the final feature map, that is:

[0074]

[0075] Then c L Input to the flattened layer will c L Flattened into a one-dimensional feature map, that is:

[0076] f 0 =f fla (c L ),

[0077] Among them, f fla (·) is the flattening function.

[0078] ④Fully connected layer:

[0079] After the flattening layer, the flattened feature map will be input into a fully connected layer, and the output is:

[0080] f 1 =wf 0 +b,

[0081] Among them, w and b are the weight matrix and bias matrix respectively.

[0082] ⑤Softmax layer:

[0083] The output of the fully connected layer will be input into the softmax activation layer to get the probability vector:

[0084]

[0085] in Represents the probability that the Sth BI beam i is the optimal beam.

[0086] ⑥Output layer:

[0087] According to the above probability vector, the inferred optimal beam can be expressed as:

[0088]

[0089] ⑦ Training steps of the proposed network:

[0090] As mentioned above, the optimal beam inference problem is modeled as a multi-classification task and is usually trained by minimizing the cross entropy loss function. The loss between the network output and the optimal beam can be expressed as:

[0091]

[0092] Where p(S)=[p 1 (S),p 2 (S),…,p M (S)] is a one-hot vector, where The remaining elements are equal to 0.

[0093] The beneficial effect of the present invention is that the present invention uses spatial correlation to deeply compress the number of detection beams, selects a part of the beams for detection, and then builds an optimal beam inference network based on a convolutional long short-term memory network to extract the time and space information contained in the received signal-to-noise ratio of the detection beams at continuous moments, so as to improve the inference accuracy of the optimal beam. Using the technology proposed in the present invention, the access point does not need to detect all beams, but only needs to detect a small part of fixed beams, and can accurately infer the optimal beam and complete beam alignment, which better balances the data transmission time and data transmission rate and improves the system throughput. BRIEF DESCRIPTION OF THE DRAWINGS

[0094] Figure 1 This is a schematic diagram of the beam training frame structure used in the present invention.

[0095] Figure 2 This is a schematic diagram of the ConvLSTM-based neural network structure proposed in the present invention.

[0096] Figure 3 This is a schematic diagram comparing the optimal beam inference accuracy of the network proposed in the present invention and other commonly used networks.

[0097] Figure 4 Schematic diagram of the comparison of the optimal beam inference accuracy between the present invention and the LABNT technology.

[0098] Figure 5 It is a schematic diagram comparing the throughput of the present invention, the LABNT technology and the EBS technology. DETAILED DESCRIPTION

[0099] The practicality of the present invention is illustrated below with the help of simulation examples and drawings. First, the simulation environment is introduced, and then the simulation results are presented to evaluate the performance of the technology proposed in the present invention. At the same time, the proposed technology is compared with the existing LABNT technology and EBS technology.

[0100] The present invention utilizes the spatial correlation between the received signal-to-noise ratios of different beams and the temporal correlation of the detection beams in continuous time, and proposes a deep spatiotemporal beam training technology based on a convolutional long short-term memory network.

[0101] Consider a millimeter wave downlink with a carrier frequency of 60GHz, a bandwidth of 2.16GHz, and a noise power spectral density of -174dBm / Hz. The access point is 8m from the ground, and its coverage area is regarded as a circular area with a radius of 8m. In this area, a user carries a user device and walks at a speed of 2m / s. The user will randomly change the direction of movement every 1000 BIs. If he walks to the boundary of the area, he will turn in the opposite direction. Based on the motion model and blocking model described above, the present invention collects the received signal-to-noise ratio of all beams on 200,000 consecutive BIs to train the neural network, and uses 20,000 consecutive BIs to test the network performance. The basic input unit of the neural network proposed in the present invention is a time step, which contains the received signal-to-noise ratio data of the detection beam in 16 consecutive BIs.

[0102] Furthermore, in the present invention, T BI ,T SSW ,T FB ,T ACK ,T IFS The durations are set to 10ms, 15μs, 15μs, 15μs, and 92μs respectively. The total number of codewords is 64, and the number of detection beams is 2, 4, 8, 16, and 32 respectively.

[0103] Figure 3The comparison of the optimal beam inference accuracy of different neural networks (including convolutional neural network, long short-term memory network, convolution-long short-term memory network and the proposed ConvLSTM-based network) is shown. In particular, the convolution-long short-term memory network is constructed by directly connecting the convolutional neural network and the long short-term memory network together. As can be seen from the figure, when the number of detection beams N p When N increases from 2 to 16, the ConvLSTM network can achieve the highest inference accuracy. Compared with other networks, the accuracy gain of the ConvLSTM network proposed in this invention is between 13% and 40%. p When N reaches 32, all networks can obtain very high inference accuracy. At this time, the neural network proposed in the present invention still has a small advantage. p The smaller the value, the longer the time available for data transmission in each BI, and the higher the corresponding average throughput. Therefore, we focus on N p Therefore, the ConvLSTM-based neural network proposed in the present invention is superior to the other three networks.

[0104] Figure 4 The optimal beam inference accuracy of the technology proposed in the present invention and the LABNT technology under different numbers of detection beams is compared. It can be noted that for any number of detection beams, the technology proposed in the present invention has a greater advantage over the LABNT technology. In particular, when detecting 4 beams, the inference accuracy of the technology proposed in the present invention reaches 84.87%, while the accuracy of the LABNT technology only reaches 80.82% when detecting 32 beams.

[0105] Figure 5 The average throughput performance of different technologies under different numbers of detection beams was evaluated. As shown in the figure, the average throughput of the EBS technology is independent of the number of detection beams and remains constant, while the average throughput of the LABNT technology increases with the increase in the number of detection beams. Regardless of the value of the number of detection beams, the technology proposed in the present invention is superior to the EBS technology and the LABNT technology. Specifically, when the number of detection beams is 8, the average throughput of the technology proposed in the present invention achieves the best quality, reaching 27.15Gbps, which is 8.82% and 11.87% higher than the EBS technology and the LABNT technology, respectively. This shows that the algorithm can better balance the data transmission rate and duration.

Claims

1. A beam alignment method based on a convolutional long short-term memory network, suitable for a millimeter wave communication system, wherein the system comprises a signal access point and a user equipment, wherein the signal access point is equipped with N x Row, N y Columns, N in total ap =N x N y A planar antenna array of antennas, characterized in that The following steps are involved: S1. Build a dataset: Perform exhaustive beam detection on K consecutive beacon intervals BI to obtain the received signal-to-noise ratio of all beams, from which N p The received signal-to-noise ratio set of the beams with the highest spatial mutual information is taken as the data set, and the index of the optimal beam corresponding to each BI is calculated as the label of the data set; Select N p The method to select the beams with the highest spatial mutual information is to first calculate the average mutual information between each beam and all other beams, that is: In the formula, Γ i and Γ j is the set of received signal-to-noise ratios of beam i and beam j, M represents the total number of beams, i,j = 1, 2, ..., M, I(C i ;C j )=H(Γ i )+H(Γ j )-H(Γ i ;C j ), Where H(Γ i ) and H(Γ j ) is Γ i and Γ j The Shannon entropy, H(Γ i Γ j ) is Γ i and Γ j The joint entropy of Then the average spatial mutual information of all beams is obtained, namely: Select N with the smallest average spatial mutual information p A beam is used as a probe beam; S2. Build an optimal beam inference network based on ConvLSTM, which includes input layer, ConvLSTM layer, flattening layer, fully connected layer, Softmax layer and output layer; The input data of the network is the selected N p The received signal-to-noise ratio of the probe beam in the current BI and the previous S-1 BIs X 0 Each row of represents the signal-to-noise ratio of the detection beam at a BI, that is: Where s=1,2,…,S, subscript set is the sequence number set of the detection beam; After receiving the input data, the ConvLSTM layer extracts the spatiotemporal information from it to obtain the feature map. The process is expressed as: Among them, CL l (·) is the ConvLSTM function, which represents the calculation process of the lth ConvLSTM layer. Represents the composite operation between ConvLSTM functions; The spatiotemporal information H extracted by the ConvLSTM layer L The last hidden state of As the final feature map, it is input into the flattening layer to obtain a one-dimensional feature map: f 0 =f fla (c L ), In the formula, f fla (·) is the flattening function; Then input the flattened feature map into the fully connected layer, that is: f 1 =wf 0 +b, In the formula, w and b are the weight matrix and bias matrix respectively; Finally, the output of the fully connected layer is activated by the softmax function to obtain a probability vector: In the formula represents the probability that the Sth BI beam i is the optimal beam; S3, training the optimal beam inference network based on ConvLSTM: Use the data set constructed in S1 to train the optimal beam inference network built in S2, and use the cross entropy loss function to measure the difference between the actual optimal beam and the inferred optimal beam, that is: Where p(S)=[p1(S),p2(S),…,p M (S)] is a one-hot vector, where The remaining elements are equal to 0. By continuously minimizing the cross entropy in the above formula, a trained intelligent beam alignment network can be obtained. S4, perform beam alignment: The access point selects a detection beam according to the method in S1. Within the first S-1 BIs, the access point performs an exhaustive beam search to obtain the optimal beam and stores the received signal-to-noise ratio of the detection beam. In the Sth BI, only the beam selected in step S1 is detected, and it is input together with the stored received signal-to-noise ratio of the detection beam of the first S-1 BIs into the trained network in S3 to obtain the result of the optimal beam inference, that is: Beam alignment is achieved by the access point using the inferred optimal beam for data transmission.

Images