A downlink adaptive method and system for 6G fully decoupled network

Abstract

The present invention discloses a downlink link adaptation method and system for a 6G fully decoupled network, which can realize effective link adaptation under long-delay feedback. The present invention first constructs a low-dimensional channel feature vector based on the signal-to-interference-noise ratio to represent the state of the current frequency selective fading channel; then predicts the channel feature vector at the future transmission moment based on the historical data of the low-dimensional channel feature vector; then constructs a frame error rate FER estimation model based on a neural network, which takes the channel feature vector and the modulation coding scheme MCS as input and the corresponding FER as output; finally, according to the FER estimated by the model, the MCS scheme with the maximum expected throughput is selected, and reported to the downlink base station through a long feedback link to realize downlink link adaptive transmission. The present invention can alleviate the degradation of link adaptation performance caused by long feedback delay in 6G fully decoupled networks, and provide a feasible solution for adaptive transmission of fully decoupled networks.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless mobile communications, and in particular relates to a downlink adaptation method and system for a 6G fully decoupled network. Background Art

[0002] To meet the growing traffic demand of consumers and industrial users, the fifth generation cellular system (5G) is currently being deployed worldwide. However, 5G still faces the problems of spectrum scarcity and high network operation costs. In this context, Academician Yu Quan et al. proposed a 6G fully decoupled radio access network architecture in "A Fully-Decoupled RAN Architecture for 6G Inspired by Neurotransmission" to address these challenges. In a fully decoupled network, the base station is physically decoupled into a control base station, an uplink base station, and a downlink base station. The fully physically decoupled architecture can achieve flexible spectrum utilization and network collaboration, and improve frequency efficiency and energy efficiency. However, the decoupling of the control base station and the downlink base station in the fully decoupled network makes it difficult for the downlink base station to obtain real-time user feedback information, which affects the application of some advanced communication technologies such as link adaptation in the fully decoupled network. Link adaptation is a key technology to improve link throughput. It dynamically adjusts transmission parameters, such as modulation and coding scheme (MCS), according to the time-varying channel between the base station and the user. In a fully decoupled network, due to the full physical decoupling, the MCS is reported to the downlink base station through a long feedback link. The delay caused by the long feedback link causes a mismatch between the MCS received at the downlink BS and the channel state, which reduces the link adaptation performance. We can alleviate the performance degradation caused by the delay through channel prediction. The traditional linear predictor is only effective when the feedback delay is small, while the prediction method based on neural networks such as LSTM has high computational complexity and is not suitable for link adaptation mechanisms with high real-time requirements. In addition, link adaptation requires dynamic selection of MCS. The traditional link adaptation method based on the lookup table uses a static mapping between the one-dimensional channel quality and the frame error rate (FER) of each MCS to select the best MCS. However, in a fully decoupled network, the one-dimensional channel quality cannot effectively express the state of the multi-carrier MIMO channel, and the static mapping method cannot cope with complex and dynamic channels and multi-antenna scenarios.

[0003] In summary, the problems faced by link adaptation in 6G fully decoupled networks include: (1) The long feedback delay of full decoupling reduces the performance of link adaptation. (2) The traditional link adaptation method based on lookup tables cannot cope with the complex channel conditions of fully decoupled networks. Therefore, it is necessary to study efficient link adaptation methods suitable for fully decoupled networks. Summary of the invention

[0004] Purpose of the invention: In view of the shortcomings of the prior art, the purpose of the present invention is to provide an efficient downlink adaptation method and system suitable for 6G fully decoupled networks, which can achieve significant throughput performance improvement under the long feedback delay of the fully decoupled network and has a good guiding significance for the practical application of the fully decoupled network.

[0005] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution: A downlink adaptation method for a 6G fully decoupled network, comprising the following steps:

[0006] Step 1: Based on the SINR set on all subcarriers and spatial streams, a channel feature vector is constructed to represent the state of the frequency selective fading channel; wherein the dimension of the channel feature vector is lower than the number of elements in the SINR set;

[0007] Step 2: Based on the historical data sequence of the channel feature vector, predict the channel characteristics at the future transmission time;

[0008] Step 3: Construct a neural network-based frame error rate (FER) estimation model, which takes the channel feature vector and modulation and coding scheme (MCS) as input and FER as output; select the MCS scheme with the maximum expected throughput as the downlink adaptive transmission scheme based on the FER estimated by the model.

[0009] Furthermore, in step 1, the SINRs on all subcarriers and spatial streams of the current transmission frame are calculated to form an SINR set, and sampling is performed to construct a channel feature vector. The element distribution characteristics in the channel feature vector are similar to the SINR set distribution characteristics, which can better characterize the channel state while reducing the feature dimension.

[0010] Preferably, by sorting all SINRs and sampling d values ​​of the equally spaced elements of the sorted SINR set, a low-dimensional channel feature vector based on SINR is constructed. To characterize the current channel status.

[0011] Furthermore, in step 2, considering the long feedback delay caused by the physical decoupling of the control plane and the data plane in the fully decoupled network, a KRLS-based method is used to predict the channel characteristics at future moments based on the historical moment data sequence of each dimension of the channel feature vector. KRLS is a kernel adaptive filtering algorithm that is applied to nonlinear and non-stationary time series prediction. KRLS is also an online algorithm that can iteratively update model parameters with relatively low computational complexity, and is suitable for real-time link adaptation mechanisms.

[0012] Furthermore, in step 2, the KRLS prediction of a single dimension channel feature is expressed as:

[0013]

[0014] in, is the weight vector The ith element of x(t) represents the input vector consisting of the channel feature information of the past p moments. is the independent vector stored in the kernel dictionary, m is the number of independent vectors stored in the kernel dictionary, is the kernel function. The KRLS prediction algorithm iteratively updates the weight vector based on historical training data to improve prediction accuracy and adapt to channel changes. The prediction of each dimension of the channel feature is calculated separately, and then combined into a predicted channel feature vector It will be used to select the best MCS to compensate for the channel changes caused by the feedback delay.

[0015] Furthermore, the neural network in step 3 adopts a multi-layer perceptron network (MLP), the activation functions in the hidden layer and the output layer of the network are ReLU and Sigmoid respectively, the input of the network is the channel feature vector z and a specific MCSm, and the output is the probability of transmission failure using MCSm under the current channel characteristics, that is, the estimated FER; the optimal network parameters are obtained by training with training data collected in real time.

[0016] Furthermore, the optimal MCS selection in step 3 is implemented by the following process: the trained neural network inputs the predicted channel feature vector and different MCS values, and then outputs the corresponding FER. The MCS with the maximum expected throughput is selected and fed back. Therefore, the optimal MCS selection based on the neural network can be written as:

[0017]

[0018] in, is the FER estimated by the neural network, θ is the network parameter, T m represents the number of bits transmitted using MCSm, is a given value, and M represents the candidate MCS set. In this way, we can obtain an MCS with the maximum expected throughput.

[0019] Furthermore, the user terminal selects the best MCS according to the estimated channel and feeds back the best MCS to the downlink base station. The downlink base station adjusts the transmission parameters according to the fed-back MCS.

[0020] Based on the same inventive concept, the present invention provides a downlink adaptive system for a 6G fully decoupled network, including: a channel feature extraction module, used to construct a channel feature vector based on the SINR set on all subcarriers and spatial streams to represent the state of a frequency selective fading channel; the dimension of the channel feature vector is lower than the number of elements in the SINR set; a channel feature prediction module, used to predict the channel feature vector at a future transmission moment according to a historical moment data sequence of the channel feature vector; and an optimal MCS selection module, used to construct a neural network-based FER estimation model, which takes the channel feature vector and the modulation coding scheme MCS as input and takes FER as output; the MCS scheme with the maximum expected throughput is selected as the link adaptive transmission scheme according to the FER estimated by the model.

[0021] Based on the same inventive concept, the present invention provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the link adaptation method for a 6G fully decoupled network are implemented.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention can effectively realize effective downlink adaptation in the 6G fully decoupled network, alleviating the performance degradation caused by long feedback delay; 2. The present invention constructs a low-dimensional feature vector based on the signal-to-interference-noise ratio for the multi-antenna channel scenario of the fully decoupled network, which can effectively characterize the channel state and balance the complexity and accuracy of the calculation; 3. The present invention expresses the channel prediction problem as a time series prediction problem, uses the historical data of the channel characteristics to predict the channel characteristics at the future transmission time, and through the predicted channel characteristics, realizes the optimal MCS selection based on the neural network, which can significantly improve the performance of link adaptation compared with the traditional table lookup method. 4. In the face of the long feedback delay problem, the present invention further adopts a low-dimensional channel feature prediction scheme based on KRLS, which can effectively predict the rapidly changing channel; 5. The neural network designed by the present invention takes the predicted channel characteristics and MCS as input and FER as output, and adopts a lightweight network model, such as a simple multi-layer perceptron network, which can be quickly retrained to adapt to changes in the communication environment, and has the advantages of low complexity and high flexibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of the 6G fully decoupled network downlink adaptation method according to an embodiment of the present invention.

[0024] Figure 2 It is a schematic diagram of an applicable scenario of an embodiment of the present invention.

[0025] Figure 3Schematic diagram of an MLP network for frame error rate estimation in an embodiment of the present invention.

[0026] Figure 4 It is a throughput performance comparison diagram of the embodiment of the present invention and the traditional table lookup link adaptation method.

[0027] Figure 5 It is a throughput performance comparison chart of the embodiment of the present invention and the traditional prediction method.

[0028] Figure 6 It is a comparison chart of the frame error rate performance of the embodiment of the present invention and the traditional prediction method. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the embodiments of the present invention are described in detail below in conjunction with the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given. It should be understood that the specific examples described here are only used to explain the present invention, but the protection scope of the present invention is not limited to the following embodiments.

[0030] like Figure 1 As shown, an embodiment of the present invention discloses a downlink adaptation method for a 6G fully decoupled network, which mainly includes: first, based on the signal-to-interference-noise ratio (SINR) set on all subcarriers and spatial streams, a low-dimensional channel feature vector is constructed to represent the state of the frequency selective fading channel; then, according to the historical moment data sequence of the channel feature vector, the channel feature vector at the future transmission moment is predicted; then, a FER estimation model based on a neural network is constructed, with the channel feature vector and the modulation coding scheme MCS as input and the FER as output; finally, the MCS scheme with the maximum expected throughput is selected as the downlink adaptive transmission scheme according to the FER estimated by the model.

[0031] Specifically, the embodiment of the present invention considers a single-base-station single-user downlink adaptation scenario in a fully decoupled network. The basic principle of link adaptation is to dynamically adjust the transmission parameters according to the channel state. The specific process is as follows: Figure 2 As shown in Figure 1, the user selects the best MCS based on the estimated channel and provides feedback by controlling the long link between the base station, edge cloud and downlink base station. The downlink base station adjusts the transmission parameters based on the feedback MCS. Assume that the base station is configured with N T Root transmitting antennas, users are equipped with N R Root receiving antenna. Given subcarrier k∈1,…K, the channel matrix Precoding matrix Then the symbol received on the kth subcarrier is Expressed as

[0032] y k =Hk W k x k +n k

[0033] in is the transmission symbol, N L is the number of spatial streams, The variance is σ 2 At the receiver, the received signal is passed through the linear equalization matrix G k Therefore, the post-processing SINR on the k-th subcarrier and the l-th spatial stream is expressed as

[0034]

[0035] in,

[0036] Based on this system model, the link adaptation method proposed in the embodiment of the present invention includes the following steps:

[0037] S1: Construct a low-dimensional channel feature vector based on signal-to-interference-noise ratio to represent the state of the current frequency selective fading channel;

[0038] S2: predict the channel characteristics at future transmission times based on historical data of low-dimensional channel feature vectors;

[0039] S3: Construct a neural network-based FER estimation model that takes the predicted low-dimensional channel feature vector and MCS as input and the corresponding FER as output. Select the MCS scheme with the maximum expected throughput based on the FER estimated by the model and report it to the downlink base station through a long feedback link to achieve downlink link adaptive transmission.

[0040] In step S1: the sorted post-processing SINR set on all subcarriers and spatial streams fully characterizes the current channel state. The channel feature dimension based on all post-processing SINRs is too high, which is not conducive to subsequent channel prediction and MCS selection. In order to balance complexity and accuracy, we hope to extract a low-dimensional channel feature vector. The distribution characteristics of the elements in the extracted low-dimensional channel feature vector are similar to the distribution characteristics of the SINR set. In order to reduce the distribution feature loss of the SINR set, an embodiment of the present invention first sorts all post-processing SINRs in descending order (or ascending order), and then selects d values ​​from these sorted SINR values ​​by equally spaced sampling. In this way, we obtain a d-dimensional channel feature z, which is an approximation of the sorted SINR set.

[0041] In step S2: One method to solve the MCS feedback delay problem is to compensate for channel changes through channel prediction. In the present invention, we predict the channel characteristics at future moments based on historical values. Then, the reported MCS is selected based on the predicted channel characteristics. The channel prediction method can use traditional linear prediction such as Kalman filtering or neural networks. In order to better apply to long feedback delays and reduce computational complexity, an embodiment of the present invention uses a nonlinear kernel adaptive filter KRLS. Specifically, we formulate the channel feature z prediction problem as a time series prediction problem and use the KRLS method with a Gaussian kernel function for prediction. This method is an online algorithm that iteratively updates the relevant model parameters based on the input training data to adapt to the changing channel environment. Specifically, for the current moment, firstly, the training data and input data are constructed based on the channel information, then the model parameters are updated using the training data, and finally the predicted channel value is obtained using the updated model and input data. It should be noted that since the channel feature z is multidimensional, each dimension needs to be predicted.

[0042] The specific steps for dimension prediction are given below.

[0043] Step 1: Construct the KRLS model. The relevant parameters include the kernel dictionary D(t), the kernel matrix Intermediate matrices A(t) and P(t), weight vector Linear correlation threshold v.

[0044] Step 2: Construct input and training data: Assume that τ is the feedback delay, z is a dimension of the channel feature, and the history length used for prediction is p, then the input data at time t is The ideal output is y(t) = z(t+τ), and the training data {x * (t),y * (t)}={x(t-τ),y(t-τ)}.

[0045] Step 3: Calculate the correlation of training data: Among them, a(t) is the optimal coefficient vector, is a kernel dictionary containing m vectors, is a high-dimensional mapping function, and v is a given threshold. If it is satisfied, go to step 4, if not, go to step 5.

[0046] Step 4: Model parameter update:

[0047] The kernel dictionary and kernel matrix are not updated:

[0048] Update the intermediate matrix: P(t)=P(t-1)-q(t)a(t) T P(t-1)

[0049] Calculate the prediction error:

[0050] Update the weight vector:

[0051] Step 5: Model parameter update:

[0052] Update kernel dictionary: m = m + 1, D(t) = D(t-1) ∪ x * (t), that is, x * (t) Add to the kernel dictionary to update the kernel matrix:

[0053] Update the intermediate matrix:

[0054] Calculate the prediction error:

[0055] Update the weight vector:

[0056] Step 6: For the input channel data x(t) at the current moment, use the updated model to calculate its predicted output This method predicts each dimension of the channel feature, and then combines it to obtain the predicted channel feature

[0057] In step S3: After obtaining the predicted channel characteristics, we need to select the best MCS. Let m∈{0,1,…,M} represent the index of MCS. The MCS selection problem can be expressed as:

[0058] m * = argmax m∈M {T m *(1-FER(m,z))}

[0059] Among them, T m is the number of bits when using MCSm for transmission, and FER(m,z) represents the frame error rate when using MCSm for transmission when the channel state is z. It is not difficult to see that the selection of MCS requires accurate FER estimation, and the exact FER function is not available in actual communication. Since neural networks can approximate highly nonlinear functions, we use a simple multilayer perceptron network to fit the FER function, such as Figure 3 shown.

[0060] Specifically, the MLP network has 2 hidden layers, and the activation functions in the hidden layer and the output layer are ReLU and Sigmoid, respectively. This lightweight model can be quickly retrained to adapt to changes in the communication environment. The FER function estimated by the MLP network can be expressed as ρ(z,m;θ), where MCSm and channel features z are the inputs of the network, and θ represents the trainable network weights and biases. We collect training data for the network in real time and construct a training set by observing N transmission frames. Among them (z n ,m n ) is the low-dimensional channel feature and the MCS used for the nth transmission frame, e n ∈{0,1} indicates that the nth frame was successfully transmitted / unsuccessfully. Then, we update the network parameters by gradient descent and minimize the cross entropy loss function of the MLP network on the training set. The cross entropy loss function is expressed as follows:

[0061]

[0062] To select the best MCS based on the MLP network, we first extract the predicted channel features Then, by traversing all candidate MCS values ​​and inputting them into the trained MLP network, the MLP network outputs the FER of different MCSs. Finally, the MCS that provides the maximum expected throughput is selected and fed back, which can be expressed as:

[0063]

[0064] in, is the FER estimated by the MLP network, and θ is the trained network parameter.

[0065] In order to make this embodiment more intuitive and reflect the performance advantages of the method, we evaluate the method through link-level simulation based on MATLAB. The basic simulation parameters are shown in Table 1. We consider 15 MCSs defined by the LTE physical layer specification. Figure 4 It can be seen that the proposed MCS selection method of the MLP network can achieve higher throughput than the traditional table lookup method, especially in the medium SNR region. Figure 5Represents the relationship between the link throughput performance of different prediction methods and the length of the feedback delay. It can be seen from the figure that if channel prediction is not performed, the system performance will rapidly decrease with the increase of the feedback delay. When the feedback delay is less than 10ms, all prediction schemes can provide prediction gain. However, when it is greater than 10ms, the correlation between the input and output of the predictor is insufficient. The traditional linear prediction method cannot provide prediction gain, and its performance is even worse than when there is no prediction. The proposed KRLS-based method can always provide the highest performance gain. The average throughput can be increased by 5%-25%, and the greater the feedback delay, the greater the performance gain. Figure 6 Similar results can be observed in the FER performance comparison in . In summary, this method can select a more accurate MCS under long feedback delay in a fully decoupled network, thereby achieving significant performance gain.

[0066] Based on the same inventive concept, an embodiment of the present invention discloses a downlink adaptive system for a 6G fully decoupled network, including: a channel feature extraction module, used to construct a channel feature vector based on a set of signal-to-interference-and-noise ratios (SINRs) on all subcarriers and spatial streams to represent the state of a frequency selective fading channel; a channel feature prediction module, used to predict the channel feature vector at a future transmission moment according to a historical moment data sequence of the channel feature vector; and an optimal MCS selection module, used to construct a neural network-based frame error rate (FER) estimation model, which takes the channel feature vector and the modulation coding scheme (MCS) as input and takes the FER as output; and selects the MCS scheme with the maximum expected throughput as the link adaptive transmission scheme according to the FER estimated by the model.

[0067] Based on the same inventive concept, an embodiment of the present invention discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the steps of the link adaptation method for a 6G fully decoupled network are implemented.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A downlink adaptation method for a 6G fully decoupled network, characterized in that: The following steps are involved: Step 1: Based on the SINR set on all subcarriers and spatial streams, construct a channel feature vector to represent the state of the frequency selective fading channel; the dimension of the channel feature vector is lower than the number of elements in the SINR set; Step 2: Based on the historical data sequence of the channel feature vector, predict the channel feature vector at the future transmission time; Step 3: Construct a frame error rate (FER) estimation model based on a multi-layer perceptron network (MLP), which takes the channel feature vector and the modulation coding scheme (MCS) as input and the FER as output; select the MCS scheme with the maximum expected throughput as the downlink adaptive transmission scheme based on the FER estimated by the model; construct the MLP network training set by observing N transmission frames (z n ,m n ) is the low-dimensional channel feature and the MCS used for the nth transmission frame, e n ∈{0,1} indicates whether the nth frame is transmitted successfully; the network parameters are updated by gradient descent, and the MLP network cross entropy loss function on the training set is minimized θ is the network parameter, ρ(z n ,m n ; θ) is the FER of the nth frame estimated by the MLP network; the optimal MCS selection is expressed as is the predicted channel feature vector, is the FER of the predicted channel estimated by the MLP network, T m represents the number of bits transmitted using MCSm, and M represents the candidate MCS set.

2. A downlink adaptation method for a 6G fully decoupled network according to claim 1, characterized in that: In step 1, the SINRs on all subcarriers and spatial streams of the current transmission frame are calculated to form an SINR set, and sampling is performed to construct a channel feature vector, in which the distribution characteristics of the elements in the channel feature vector are similar to the distribution characteristics of the SINR set.

3. A downlink adaptation method for a 6G fully decoupled network according to claim 2, characterized in that: The channel feature vector is constructed by sorting all SINRs and sampling the sorted SINR set at equal intervals.

4. A downlink adaptation method for a 6G fully decoupled network according to claim 1, characterized in that: In step 2, different dimensions of the channel feature vector are predicted independently, and the channel characteristics at future moments are predicted based on the historical moment data sequence of each dimension of the channel feature vector using the kernel recursive least squares (KRLS) method.

5. A downlink adaptation method for a 6G fully decoupled network according to claim 4, characterized in that: The KRLS prediction of a single dimension channel feature is expressed as: in, is the weight vector The ith element of x(t) represents the input vector consisting of the channel feature information of the past p moments. is the predicted channel characteristic information, is the independent vector stored in the kernel dictionary, m is the number of independent vectors stored in the kernel dictionary, is the kernel function; the KRLS prediction algorithm is iteratively updated based on historical training data To improve prediction accuracy and adapt to channel changes.

6. A downlink adaptation method for a 6G fully decoupled network according to claim 1, characterized in that: In the multi-layer perceptron network MLP in step 3, the activation functions in the hidden layer and the output layer of the network are ReLU and Sigmoid respectively, the input of the network is the channel feature vector and a specific MCS, and the output is the estimated FER; the optimal network parameters are obtained by training with training data collected in real time.

7. A downlink adaptation method for a 6G fully decoupled network according to claim 1, characterized in that: The user terminal selects the best MCS according to the estimated channel and feeds back to the downlink base station. The downlink base station adjusts the transmission parameters according to the fed-back MCS.

8. A downlink adaptive system for a 6G fully decoupled network, characterized in that: include: A channel feature extraction module, configured to construct a channel feature vector based on a set of signal-to-interference-plus-noise ratios (SINRs) on all subcarriers and spatial streams to represent the state of a frequency selective fading channel; the dimension of the channel feature vector is lower than the number of elements in the SINR set; A channel feature prediction module is used to predict the channel feature vector at a future transmission time based on the historical time data sequence of the channel feature vector; And, the best MCS selection module is used to construct a frame error rate FER estimation model based on a multi-layer perceptron network MLP, which takes the channel feature vector and the modulation coding scheme MCS as input and the FER as output; the MCS scheme with the maximum expected throughput is selected as the link adaptive transmission scheme according to the FER estimated by the model; the MLP network training set is constructed by observing N transmission frames (z n ,m n ) is the low-dimensional channel feature and the MCS used for the nth transmission frame, e n ∈{0,1} indicates whether the nth frame is transmitted successfully; the network parameters are updated by gradient descent, and the cross entropy loss function of the MLP network on the training set is minimized θ is the network parameter, ρ(z n ,m n ; θ) is the FER of the nth frame estimated by the MLP network; the optimal MCS selection is expressed as is the predicted channel feature vector, is the FER of the predicted channel estimated by the MLP network, T m represents the number of bits transmitted using MCSm, and M represents the candidate MCS set.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the computer program is loaded into the processor, the steps of the link adaptation method for a 6G fully decoupled network are implemented according to any one of claims 1 to 7.

Images