A millimeter-wave MIMO codebook search method based on deep learning

Abstract

The present invention discloses a millimeter wave MIMO codebook search method based on deep learning. The codebook search problem is converted into a multi-classification problem by introducing a deep learning method. Two model structures are adopted to specifically convert the problem into a multi-label classification problem and a single-label multi-classification problem. In the training stage of the neural network, the collected channel information is used to maximize the spectrum efficiency. The codebook traversal method is used to obtain the best analog precoder and combiner index, and this index is the label of the corresponding channel information. The channel information is used as the input of the deep learning model, and the corresponding analog precoder and combiner index are used as the output of the model. The deep learning model is trained until the training is completed. The present invention does not rely on a specific codebook, achieves a spectrum efficiency close to that of the codebook traversal method, and can effectively reduce the computational complexity and improve the real-time performance of the system.

Description

Technical Field

[0001] The present invention relates to the technical field of millimeter wave MIMO codebook search methods, and in particular to a millimeter wave MIMO codebook search method based on deep learning. Background Art

[0002] In recent years, the millimeter wave communication field has attracted extensive research due to the rich spectrum resources in the millimeter wave band. Millimeter waves have the problem of large transmission loss. To provide sufficiently large coverage, beamforming gain is required. Fortunately, the smaller wavelength of millimeter waves reduces the space required for large antenna arrays. In addition, if a fully digital antenna array is used to provide beamforming gain, it will bring problems of high cost and high power consumption. The introduction of analog antenna arrays has solved the above problems well, so it has an important position in the field of millimeter wave communications. A large number of existing works in the field of millimeter wave communications are based on hybrid antenna architectures and analog antenna architectures.

[0003] In order to solve the problem of high computational complexity of exhaustive beam search schemes in millimeter wave communications, which leads to long time delays, the present invention provides a millimeter wave MIMO codebook search method based on deep learning. Unlike many hierarchical codebook search methods that need to rely on specific hierarchical codebooks, this method is suitable for most codebook sets. Summary of the invention

[0004] In view of this, the purpose of the present invention is to provide a millimeter wave MIMO codebook search method based on deep learning to solve the problem of high computational complexity in the traversal codebook search algorithm. The present invention achieves a spectral efficiency close to that of the codebook traversal method, and can effectively reduce the computational complexity and improve the real-time performance of the system.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] A millimeter wave MIMO codebook search method based on deep learning, the codebook search method comprising the following steps:

[0007] Step S1, constructing a training set for training a codebook search deep neural network model, the training set includes: channel information between a transmitter and a receiver collected at a transmitter, and a label corresponding to the channel information, wherein the label is the collected channel information, with the goal of maximizing spectrum efficiency, using a codebook traversal method to obtain the best index of the analog precoder and combiner; the trained codebook search deep neural network model includes: a single-label multi-classification model and a multi-label multi-classification model;

[0008] Step S2, performing preprocessing operations on the training set constructed in step S1;

[0009] Step S3, using the preprocessed training set to train the codebook search deep neural network model, wherein the preprocessed channel information is used as the model input, and the index form of the corresponding analog precoder and combiner is used as the model output; the single-label multi-classification model and the multi-label multi-classification model are trained respectively;

[0010] Step S4: Put the trained codebook search deep neural network model online for operation.

[0011] Further, in the step S1, for a millimeter wave MIMO communication system, a radio frequency circuit and Nt transmitting antennas are provided at its transmitting end, and correspondingly, a radio frequency circuit and Nr receiving antennas are provided at its receiving end;

[0012] In the wireless communication scenario of the millimeter wave frequency band, the narrowband cluster channel of the Saleh-Valenzuela model is used to simulate the transmission environment between the transmitter and the receiver. The channel matrix from the transmitter to the receiver is expressed as:

[0013]

[0014] In the formula, L represents the number of scattering paths, α l represents the complex gain of the lth path, the gain is an independent and uniformly distributed complex Gaussian random variable with zero mean and unit variance; and Represent the arrival angle and departure angle of the lth path respectively, and They represent the antenna array response vectors of the transmitting and receiving ends respectively.

[0015] Furthermore, assuming that both the transmitter and receiver are equipped with uniform linear arrays, the spacing between antennas is half the wavelength; then the array response vector is expressed as:

[0016]

[0017] Furthermore, in step S1, the collected channel information is used to maximize spectrum efficiency by using a codebook traversal method to obtain the best analog precoder and combiner index, which specifically includes:

[0018] Step S101, initialize the spectrum efficiency value to R = 0, the index of the best analog precoder and combiner is p1 = 0, p2 = 0; the analog precoding codeword set is There are M code words in total, and the index of the currently selected code word is m; judging m<M, if it is true, executing step S102, if it is false, executing step S104;

[0019] Step S102: The codeword set of the simulated combiner codebook is There are N code words in total, and the index of the currently selected code word is n; if n < N, execute step S103 if it is true, and execute step S101 if it is false;

[0020] Step S103: Calculate Among them, γ is the signal-to-noise ratio, H is the channel matrix from the transmitter to the receiver, and f n is the simulated precoding vector, w m is a simulated combiner vector; execute judgment Rn<R, if it is false, record the simulated precoding codeword index and the simulated combiner codebook index at this time, so that p1=m, p2=n, and R=Rn, and then execute step S102; if it is true, execute step S104;

[0021] Step S104: Obtain the best indexes p1 and p2 of the analog precoder and combiner.

[0022] Furthermore, in step S2, the following preprocessing is specifically performed on the channel information:

[0023] Step S201: Take out the real part and the imaginary part of the channel vector H between the transmitting end and the receiving end respectively, and then horizontally splice and merge them into a new vector, which is expressed as:

[0024] H n =[Re(H n ),Im(H n )]

[0025] In the formula, Re(·) and Im(·) represent the real and imaginary parts respectively, and H n is the new matrix after the first step of preprocessing.

[0026] Step S202: H n After the Flatten function, H n From multidimensional to one-dimensional h n .

[0027] Furthermore, in step S2, the following preprocessing is specifically performed on the tag:

[0028] Assume that the total number of codewords in the precoder codebook is T, and the total number of codewords in the combiner codebook is R.

[0029] In the single-label multi-classification model, it has two output channels, the number of output channels is T and R respectively;

[0030] The settings of the precoder and combiner of the single-label multi-classification model are the same. For the precoder, the specific settings are as follows:

[0031] Each codeword in the codebook corresponds to one label, that is, the total number of labels is the number T of codewords in the codebook at the transmitting end, and the number of output channels of the deep learning model is equal to the number of codewords in the codebook; the position of the simulated precoder index obtained in step S1 is set to 1, and the others are set to 0;

[0032] In the multi-label classification model, the number of output channels is T+R, and the specific settings are as follows:

[0033] The positions of the simulated precoder and combiner indexes obtained in step S1 are set to 1, and the other positions are set to 0, that is, only one position is 1 in the first T sequences and only one position is 1 in the last R sequences.

[0034] Furthermore, the single-label multi-classification model is a single-input multi-output model, the input of which is the channel state information after preprocessing, and there are two outputs, which are the settings of the precoder and the combiner of the single-label multi-classification model during the training phase;

[0035] The network consists of three layers in total. The activation function of the first two layers is Relu, and their outputs are output through two output layers, and their activation functions are both Softmax functions.

[0036] The Loss function of the model is categorical_crossentropy:

[0037]

[0038] In the formula, N represents the total number of samples, y n and are the true value and the predicted value respectively.

[0039] Furthermore, the multi-label multi-classification model is a single-input single-output model, the input is the channel state information after preprocessing, and during the training phase, the output is the settings of the multi-label classification model precoder and combiner;

[0040] The network consists of three layers. The activation function of the first two layers is relu, and the activation function of the last layer is sigmiod function.

[0041] The Loss function of the model is binary_crossentropy:

[0042]

[0043] In the formula, N represents the total number of samples, y n and are the true value and the predicted value respectively.

[0044] Furthermore, in step S3, the following method is used to train the single-label multi-classification model and the multi-label multi-classification model, including:

[0045] Step S301: Initialize the deep learning model parameters, and then train the collected training set, where the data of the training set is the channel state information after preprocessing and the index settings of the analog precoder and the combiner;

[0046] Step S302, setting the total number of training times K for the data of the entire training set and the number of times the current training is performed k, and initializing k=0;

[0047] Step S303: determine whether k is less than K. If it is true, execute step S304; if it is false, execute step S305;

[0048] Step S304, calculate the Loss function and update the parameters of the network;

[0049] Step S305: complete model training.

[0050] The beneficial effects of the present invention are:

[0051] The present invention designs a millimeter wave MIMO codebook search method based on deep learning. Compared with the traditional partial layered codebook search method, this method does not rely on a specific codebook.

[0052] A millimeter-wave MIMO codebook search method based on deep learning designed by the present invention achieves a spectral efficiency close to that of a codebook traversal method, and can effectively reduce computational complexity and improve system real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a schematic diagram of the structure of the single-label multi-classification model provided in Example 1;

[0054] Figure 2 This is a schematic diagram of the structure of the multi-label classification model provided in Example 1;

[0055] Figure 3 The output setting diagram of the single-label multi-classification model provided in Example 1;

[0056] Figure 4 The output setting diagram of the multi-label classification model provided in Example 1;

[0057] Figure 5 This is a spectrum efficiency diagram of the method in Example 1. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] Example 1

[0060] See also Figure 1-Figure 5 This embodiment provides a millimeter wave MIMO codebook search method based on deep learning, and the method specifically includes:

[0061] Step S1, constructing a training set for training a codebook search deep neural network model, the training set includes: channel information between a transmitter and a receiver collected at a transmitter, and a label corresponding to the channel information, wherein the label is the collected channel information, with the goal of maximizing spectrum efficiency, using a codebook traversal method to obtain the best index of the analog precoder and combiner; the trained codebook search deep neural network model includes: a single-label multi-classification model and a multi-label multi-classification model;

[0062] Specifically, in this embodiment, step S1 includes:

[0063] For a millimeter wave MIMO communication system, a radio frequency circuit and Nt transmitting antennas are provided at its transmitting end, and correspondingly, a radio frequency circuit and Nr receiving antennas are provided at its receiving end.

[0064] The transmitter collects channel information between the transmitter and the receiver for training the deep neural network.

[0065] In the wireless communication scenario of the millimeter wave frequency band, the narrowband cluster channel of the SV (Saleh-Valenzuela) model can be used to simulate the transmission environment between the transmitter and the receiver.

[0066] The collected channel information is used to maximize the spectrum efficiency by traversing the codebook to obtain the index of the best analog precoder and combiner. This index is the label of the corresponding channel information. The channel information and the corresponding index of the best analog precoder and combiner constitute the training set of the codebook search deep neural network.

[0067] More specifically, the above-mentioned collected channel information specifically includes:

[0068] The transmitter can continuously collect channel information between the transmitter and the receiver in real time, in which case information about the specific application scenario environment can be collected. In addition, in the wireless communication scenario of the millimeter wave frequency band, the narrowband cluster channel of the SV (Saleh-Valenzuela) model can be used to simulate the transmission environment between the transmitter and the receiver.

[0069] The channel matrix from the transmitter to the receiver can be expressed as:

[0070]

[0071] In formula (1), L represents the number of scattering paths, α l represents the complex gain of the lth path, the gain is an independent and uniformly distributed complex Gaussian random variable with zero mean and unit variance; and Represent the arrival angle and departure angle of the lth path respectively, and They represent the antenna array response vectors of the transmitting and receiving ends respectively.

[0072] In this embodiment, it is assumed that both the transmitting and receiving ends are equipped with a uniform linear array (ULA), and the spacing between antennas is half the wavelength. Therefore, the array response vector can be expressed as:

[0073] More specifically, the above-mentioned channel information is collected, with the goal of maximizing spectrum efficiency, and a codebook traversal method is used to obtain the best analog precoder and combiner index, which specifically includes:

[0074] Step S101, initialize the spectrum efficiency value to R = 0, the index of the best analog precoder and combiner is p1 = 0, p2 = 0; the analog precoding codeword set is There are M code words in total, and the index of the currently selected code word is m; judging m<M, if it is true, executing step S102, if it is false, executing step S104;

[0075] Step S102: The codeword set of the simulated combiner codebook is There are N code words in total, and the index of the currently selected code word is n; if n < N, execute step S103 if it is true, and execute step S101 if it is false;

[0076] Step S103: Calculate Among them, γ is the signal-to-noise ratio, H is the channel matrix from the transmitter to the receiver, and f n is the simulated precoding vector, w mis a simulated combiner vector; execute judgment Rn<R, if it is false, record the simulated precoding codeword index and the simulated combiner codebook index at this time, so that p1=m, p2=n, and R=Rn, and then execute step S102; if it is true, execute step S104;

[0077] Step S104: Obtain the best indexes p1 and p2 of the analog precoder and combiner.

[0078] Step S2, performing preprocessing operations on the training set constructed in step S1;

[0079] Specifically, in step S2, the following preprocessing is performed on the channel information:

[0080] Step S201: Take out the real part and the imaginary part of the channel vector H between the transmitting end and the receiving end respectively, and then horizontally splice and merge them into a new vector, which can be expressed as:

[0081] H n＝ [Re(H n ),Im(H n )] (2)

[0082] In formula (2), Re(·) and Im(·) represent the real part and imaginary part respectively, and H n is the new matrix after the first step of preprocessing.

[0083] Step S202: H n After the Flatten function, H n From multi-dimensional to one-dimensional h n .

[0084] Specifically, in step S2, the following preprocessing is performed on the label:

[0085] Assume that the total number of codewords in the precoder codebook is T, and the total number of codewords in the combiner codebook is R.

[0086] In the single-label multi-classification model, it has two output channels, and the number of output channels is T and R respectively.

[0087] The settings of the precoder and combiner of the single-label multi-classification model are the same. The precoder is used as an example to illustrate:

[0088] like Figure 3 As shown, each codeword in the codebook corresponds to one label, that is, the total number of labels is the number of codewords in the transmitter codebook T, and the number of output channels of the deep learning model is equal to the number of codewords in the codebook. The position of the simulated precoder index obtained in step S1 is set to 1, and the others are set to 0.

[0089] In the multi-label classification model, the number of output channels is T+R, such as Figure 4 As shown, specifically including:

[0090] The positions of the simulated precoder and combiner indexes obtained in step S1 are set to 1, and the other positions are set to 0, that is, only one position is 1 in the first T sequences and only one position is 1 in the last R sequences.

[0091] Step S3, using the preprocessed training set to train the codebook search deep neural network model, wherein the preprocessed channel information is used as the model input, and the index form of the corresponding analog precoder and combiner is used as the model output; the single-label multi-classification model and the multi-label multi-classification model are trained respectively;

[0092] Specifically, in this embodiment, the above two deep learning model structures are as follows:

[0093] The single-label multi-classification model structure is a single-input multi-output model, such as Figure 1 As shown, the input is the channel state information after preprocessing, and there are two outputs, which are the settings of the precoder and combiner of the single-label multi-classification model during the training phase;

[0094] During the prediction phase, the label with the highest output probability is selected.

[0095] The network consists of a total of three layers. The activation function of the first two layers is Relu, and their outputs are output through two output layers, and their activation functions are both Softmax functions.

[0096] The Loss function of the model is categorical_crossentropy:

[0097]

[0098] In formula (3), N represents the total number of samples, y n and are the true value and the predicted value respectively.

[0099] The structure of the multi-label multi-classification model is a single-input single-output model, such as Figure 2 As shown, the input is the channel state information after preprocessing, and during the training phase, the output is the settings of the precoder and combiner of the multi-label classification model;

[0100] In the prediction stage, the position with the highest probability in the first T sequences and the position with the highest probability in the last R sequences are selected. The network also consists of a three-layer structure, the activation function of the first two layers is relu, and the activation function of the last layer is sigmiod function.

[0101] The Loss function of the model is binary_crossentropy:

[0102]

[0103] In formula (4), N represents the total number of samples, y n and are the true value and the predicted value respectively.

[0104] Specifically, in this embodiment, the two deep learning models are trained offline by the following method:

[0105] Step S301: Initialize the deep learning model parameters, and then train the collected training set, where the data of the training set is the channel state information after preprocessing and the index settings of the analog precoder and the combiner;

[0106] Step S302, setting the total number of training times K for the data of the entire training set and the number of times the current training is performed k, and initializing k=0;

[0107] Step S303: determine whether k is less than K. If it is true, execute step S304; if it is false, execute step S305;

[0108] Step S304, calculate the Loss function and update the parameters of the network;

[0109] Step S305: complete model training.

[0110] Step S4: Put the trained codebook search deep neural network model online for operation.

[0111] In addition, it is found through experiments that the codebook traversal method is related to the number of codebooks. This embodiment intuitively observes the time complexity of the three methods through the time the program runs in Python, which is specifically shown in Table 1.

[0112] Table 1 Calculation time

[0113]

[0114] It can be seen that when the number of codebooks is large, the calculation time of the method of this embodiment is superior to that of the method of traversing the codebooks.

[0115] The matters not described in detail in the present invention are all known technologies to those skilled in the art.

[0116] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A millimeter wave MIMO codebook search method based on deep learning, characterized in that: The codebook search method comprises the following steps: Step S1, constructing a training set for training a codebook search deep neural network model, the training set includes: channel information between a transmitter and a receiver collected at a transmitter, and a label corresponding to the channel information, wherein the label is the collected channel information, with the goal of maximizing spectrum efficiency, using a codebook traversal method to obtain the best index of the analog precoder and combiner; the trained codebook search deep neural network model includes: a single-label multi-classification model and a multi-label multi-classification model; Step S2, performing preprocessing operations on the training set constructed in step S1; Step S3, using the preprocessed training set to train the codebook search deep neural network model, wherein the preprocessed channel information is used as the model input, and the index form of the corresponding analog precoder and combiner is used as the model output; the single-label multi-classification model and the multi-label multi-classification model are trained respectively; Step S4: Put the trained codebook search deep neural network model online for operation; In step S1, for a millimeter wave MIMO communication system, a radio frequency circuit and Nt transmitting antennas are provided at its transmitting end, and correspondingly, a radio frequency circuit and Nr receiving antennas are provided at its receiving end; In the wireless communication scenario of the millimeter wave frequency band, the narrowband cluster channel of the Saleh-Valenzuela model is used to simulate the transmission environment between the transmitter and the receiver. The channel matrix from the transmitter to the receiver is expressed as: In the formula, L represents the number of scattering paths, α l represents the complex gain of the lth path, the gain is an independent and uniformly distributed complex Gaussian random variable with zero mean and unit variance; and Represent the arrival angle and departure angle of the lth path respectively, and They represent the antenna array response vectors of the transmitting and receiving ends respectively; In step S1, the collected channel information is used to maximize spectrum efficiency by using a codebook traversal method to obtain the best analog precoder and combiner index, which specifically includes: Step S101, initialize the spectrum efficiency value to R = 0, the index of the best analog precoder and combiner is p1 = 0, p2 = 0; the analog precoding codeword set is W = {w0, ..., w m ,…,w M-1 }, there are M code words in total, and the currently selected code word index is m; judge m<M, if it is true, execute step S102, if it is false, execute step S104; Step S102: The codeword set of the simulated combiner codebook is F = {f0, ..., f n ,…,f N-1 }, there are N code words in total, and the currently selected code word index is n; determine n<N, if it is true, execute step S103, if it is false, execute step S101; Step S103: Calculate Among them, γ is the signal-to-noise ratio, H is the channel matrix from the transmitter to the receiver, and f n is the simulated precoding vector, w m is a simulated combiner vector; execute judgment Rn<R, if it is false, record the simulated precoding codeword index and the simulated combiner codebook index at this time, so that p1=m, p2=n, and R=Rn, and then execute step S102; if it is true, execute step S104; Step S104, obtaining the best indexes p1 and p2 of the analog precoder and combiner; In step S2, the following preprocessing is specifically performed on the channel information: Step S201: Take out the real part and the imaginary part of the channel vector H between the transmitting end and the receiving end respectively, and then horizontally splice and merge them into a new vector, which is expressed as: H n＝ [Re(H n ),Im(H n )] In the formula, Re(·) and Im(·) represent the real and imaginary parts respectively, and H n is the new matrix after the first step of preprocessing; Step S202: H n After the Flatten function, H n From multidimensional to one-dimensional h n ; In step S2, the following preprocessing is specifically performed on the tag: Assume that the total number of codewords in the precoder codebook is T, and the total number of codewords in the combiner codebook is R; In the single-label multi-classification model, it has two output channels, the number of output channels is T and R respectively; The settings of the precoder and combiner of the single-label multi-classification model are the same. For the precoder, the specific settings are as follows: Each codeword in the codebook corresponds to one label, that is, the total number of labels is the number T of codewords in the codebook at the transmitting end, and the number of output channels of the deep learning model is equal to the number of codewords in the codebook; the position of the simulated precoder index obtained in step S1 is set to 1, and the others are set to 0; In the multi-label classification model, the number of output channels is T+R, and the specific settings are as follows: The positions of the simulated precoder and combiner indexes obtained in step S1 are set to 1, and the other positions are set to 0, that is, only one position in the first T sequences is 1, and only one position in the last R sequences is 1; The single-label multi-classification model is a single-input multi-output model, the input is the channel state information after preprocessing, and there are two outputs, which are the settings of the precoder and combiner of the single-label multi-classification model during the training stage; The network consists of three layers in total. The activation function of the first two layers is Relu, and their outputs are output through two output layers, and their activation functions are both Softmax functions. The Loss function of the model is categorical_crossentropy: In the formula, N represents the total number of samples, y n and are the true value and the predicted value respectively; The multi-label multi-classification model is a single-input single-output model, the input is the channel state information after preprocessing, and during the training stage, the output is the settings of the multi-label classification model precoder and combiner; The network consists of a three-layer structure. The activation function of the first two layers is relu, and the activation function of the last layer is sigmiod function. The Loss function of the model is binary_crossentropy: In the formula, N represents the total number of samples, y n and are the true value and the predicted value respectively.

2. The millimeter wave MIMO codebook search method based on deep learning according to claim 1, characterized in that: Assuming that both the transmitter and receiver are equipped with uniform linear arrays and the spacing between antennas is half the wavelength, the array response vector is expressed as:

3. The millimeter wave MIMO codebook search method based on deep learning according to claim 1, characterized in that: In step S3, the following method is used to train the single-label multi-classification model and the multi-label multi-classification model, including: Step S301: Initialize the deep learning model parameters, and then train the collected training set, where the data of the training set is the channel state information after preprocessing and the index settings of the analog precoder and the combiner; Step S302, setting the total number of training times K for the data of the entire training set and the number of times the current training is performed k, and initializing k=0; Step S303: determine whether k is less than K. If it is true, execute step S304; if it is false, execute step S305; Step S304, calculate the Loss function and update the network parameters; Step S305: complete model training.

Images