High-speed rail mobile communication 6G full-decoupling network downlink transmission method and system based on geographic position

Abstract

The invention discloses a high-speed rail mobile communication 6G full-decoupling network downlink transmission method and system based on a geographic position, and the method comprises the steps: achieving the precoding design and MAC layer resource distribution based on a train position through designing a two-stage machine learning algorithm; in the first stage, a deep neural network is designed and fused, the geographic position of a train is used as input, and multi-user joint precoding meeting power constraint is output; in the second stage, a deep reinforcement learning network is designed, train positions and data rate requirements are input into the network, and a resource allocation strategy meeting total data rate-fairness balance is output; performing offline training by using the historical channel data of each position; in the actual deployment stage, the actual position of a train is input, and the network outputs precoding parameters and resource allocation strategies of corresponding positions. According to the invention, the problems of untimely channel feedback and large resource overhead caused by high-speed movement in the high-speed rail scene are effectively solved, the frequency efficiency of wireless communication in the high-speed rail scene is improved, and the balance between the total data rate and the fairness is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of wireless mobile communications, and relates to multi-user multiple-input multiple-output transmission and MAC layer resource allocation in machine learning and high-speed rail scenarios. Specifically, it relates to a downlink transmission method and system for a 6G fully decoupled network for high-speed rail mobile communications based on geographical location. Background Art

[0002] With the booming development of the railway system, high-speed train communication has received extensive attention. In the fifth-generation mobile network (5G), a significant performance improvement has been achieved through joint precoding design and massive multiple-input multiple-output (mMIMO) of multi-connection technology, making full use of diversity and multiplexing gains. However, high speeds will cause problems such as channel aging, shorter coherence time, and Doppler frequency offset (DFO), resulting in a serious decline in the performance of the high-speed rail communication system. In addition, since 5G uses higher-frequency spectrums, handovers between base stations will be more frequent in high-speed mobile scenarios.

[0003] The solutions adopted by existing 5G railway technology (5G-R) use more frequent pilot transmissions and CSI feedback to cope with the reduced channel coherence time, and utilize coordinated multi-point (CoMP) technology to improve the spectral efficiency (SE) in the handover area. However, the increasing pilot transmissions and feedback will generate a large amount of time-frequency overhead, and CoMP highly depends on real-time and accurate global CSI, which is difficult to obtain in high-mobility scenarios. In addition, feedback latency further exacerbates the difficulty of obtaining accurate CSI, resulting in an error between the actual channel and the obtained CSI. Moreover, the total data rate and fairness are conflicting goals and cannot be maximized simultaneously. Existing systems require feedback to obtain the actual data rate of each terminal and allocate spectrum resources to meet the data rate requirements of the terminals. However, as mentioned above, accurate and timely feedback is still limited by high mobility and latency. The above challenges will ultimately reduce the overall performance and reliability of the high-speed rail communication system.

[0004] The sixth-generation (6G) mobile communication network is expected to provide more flexible resource allocation and enhance the user experience. Academician Yu Quan et al. proposed a 6G fully-decoupled radio access network (FD-RAN) architecture in "A Fully-Decoupled RAN Architecture for 6G Inspired by Neurotransmission", which decouples the traditional base station into an uplink base station (UBSs), a downlink base station (DBSs), and a control base station. FD-RAN has significant advantages in high-speed railway communication systems. On the one hand, all control-plane signals are transmitted by the control base station. Since the amount of control information data is small, the control base station can use a lower frequency band to cover a larger area, and the handover overhead can be greatly reduced. On the other hand, FD-RAN adopts a location-based non-feedback transmission scheme. Through this scheme, not only can the negative impacts of channel variations and resource overhead be mitigated, but also feedback delay can be avoided through location prediction. Since the train trajectory is fixed, its location can be easily predicted in advance. However, it is still challenging to determine appropriate precoding and resource allocation strategies solely from location information.

[0005] In summary, the problems existing in the prior art are as follows: (1) Due to feedback delay, it is difficult for the base station transmitter in the high-speed railway scenario to obtain real-time channel estimation and other feedback information, resulting in a decline in the performance of feedback-based precoding design and resource allocation strategies. (2) In the 6G fully-decoupled network, the high feedback latency caused by the hardware decoupling and isolation architecture results in a significant performance loss for transmission methods that rely on the existing feedback mechanism. Summary of the Invention

[0006] Object of the Invention: The object of the present invention is to provide a downlink transmission method and system for a 6G fully-decoupled network for high-speed railway mobile communication based on geographical location, which can directly select precoding and transmission parameters based on the geographical location of the train, achieve a balance between the total data rate and fairness, and replace the traditional channel estimation and channel feedback processes.

[0007] Technical Solution: To achieve the above object of the invention, the technical solution adopted by the present invention is: A downlink transmission method for a 6G fully-decoupled network for high-speed railway mobile communication based on geographical location, which uses staged machine learning to fit the mapping relationship between the geographical location of the train and the precoding and resource allocation strategies, mainly including the following steps:

[0008] Construct a deep neural network with the train geographical location information as the input and the multi-user joint precoding information that satisfies the power constraint as the output, design the loss function of the deep neural network with the system frequency efficiency as the optimization target, and use the precoding output by the deep neural network for data transmission;

[0009] Construct a deep reinforcement learning network with the train's geographical location information and data rate requirements as input states and the MAC layer resource allocation strategy as output actions. Design a deep reinforcement learning algorithm with the weighted sum of the total data rate and the demand satisfaction rate as the reward, and use the output policy for MAC layer resource allocation;

[0010] After multiple deep neural networks (DNNs) are offline trained using the historical channel data of each location sample to update the parameters until convergence, model fusion is performed based on the performance of multiple networks on the evaluation set to obtain the fused DNN (mDNN) in the first stage; mDNN selects the DNN with the minimum loss on the evaluation set for each input location;

[0011] The deep reinforcement learning network uses the historical channel data of each location sample and combines the trained mDNN for offline training to update the parameters until convergence to obtain the deep reinforcement learning network in the second stage;

[0012] During the actual deployment phase, the train location information is input, the mDNN outputs the joint precoding at the corresponding location, and the D3QN outputs the resource allocation strategy to achieve feedback-free transmission.

[0013] Furthermore, in multi-user joint precoding, the precoding vector of the l-th downlink base station for the k-th terminal on the train roof is Satisfying the normalization condition The allocated power is p kl , satisfying the power constraint where N represents the number of antennas of the downlink base station, and p max is the maximum power limit. The channel Η kl between the k-th train terminal and the l-th downlink base station has a line-of-sight path and satisfies the Rice channel distribution; the achievable signal-to-interference-plus-noise ratio of the k-th terminal is where respectively represent the joint channel and joint precoding of the k-th terminal, represents the total power allocated to the k-th terminal, L represents the number of base stations, K represents the number of terminals, represents the noise power. The spectral efficiency of the k-th terminal can be expressed as R k = log2(1 + SINR k ).

[0014] Furthermore, in MAC layer resource allocation, the available bandwidth is divided into A resource blocks (RBs) with a bandwidth of F Hz, and use to indicate whether the a-th resource block serves the terminal k, then the data rate of the k-th terminal can be expressed as Use to represent the satisfaction rate of k terminals, where represents the data rate requirement.

[0015] The optimization problem of the system is modeled as

[0016]

[0017] where μ1 and μ2 represent the weighted coefficients of the data rate and the satisfaction rate respectively, and W, p, and i represent the joint precoding, power, and resource block allocation of all terminals; the optimization objective is to maximize the weighted sum of the total system data rate and the overall terminal satisfaction rate, achieving a balance between maximum performance and fairness.

[0018] Furthermore, for the precoding design and power allocation in the physical layer of the first stage, considering that the number of terminals served within a resource block a can be freely selected between 1 and K, the optimization problem can be modeled as

[0019]

[0020] where K a ′ represents the number of terminals using resource block a, and R k,a represents the spectral efficiency of the k-th terminal using resource block a. The loss function of the deep neural network is defined as the negative of the sum of the system spectral efficiencies, T represents the number of samples, and i k,a,t , R k,a,t represent the service index and spectral efficiency of the a-th resource block for the k-th terminal in the t-th sample respectively; the output of the deep neural network is the precoding result including power allocation According to it is transformed into a precoding vector that satisfies the constant modulus constraint.

[0021] Furthermore, the deep neural network adopts a fully connected deep learning neural network or a residual network; during the training process, first, the entire dataset is divided into a training set and an evaluation set. Then, multiple DNN sets with the same loss function are trained to convergence using different learning rates. Finally, according to the Loss value on the evaluation set, the DNN with the smallest loss at each input position is selected and fused to form mDNN.

[0022] Furthermore, for the resource allocation problem in the MAC layer of the second stage, after the spectral efficiency has been determined by precoding and power, the optimization problem can be modeled as

[0023]

[0024] Furthermore, for each step j in the deep reinforcement learning training episode, the state of the deep reinforcement learning network It consists of three parts, which respectively represent the positions of all terminals at the j-th step, the data rate still required for each terminal to meet the data rate requirement, and the remaining number of resource blocks; the action is defined as a resource block allocation strategy, and a resource block can be allocated to 1-K terminals; the reward is defined as the weighted sum of two components: 1) the total increased data rate 2) the increased satisfaction rate of the requirements of all terminals, which can be expressed as

[0025]

[0026] Among them, represents the increased data rate of the k-th terminal after the j-th step, satisfying

[0027] Furthermore, after the first-stage repeated deep learning algorithm training, K' = [1,..., K] mDNN models for different numbers of terminals are fused, and then, the mDNN models under different K' are fixed for the next stage. During the training process of the second-stage deep reinforcement learning network, an iterative loop containing all position sampling points is defined, and the training of each position corresponds to one episode. The position input within one episode remains fixed (i.e., all steps j within one episode are used to train one position). The parameters of the deep reinforcement learning algorithm are updated after each iterative loop.

[0028] The present invention also provides a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. The computer program is loaded into the processor to implement the steps of the method for downlink transmission of a 6G fully decoupled network for high-speed rail mobile communication based on geographical location.

[0029] The present invention also provides a computer program product, including a computer program, and the computer program is loaded into the processor to implement the steps of the method for downlink transmission of a 6G fully decoupled network for high-speed rail mobile communication based on geographical location.

[0030] Beneficial effects: The present invention designs a two-stage machine learning algorithm to achieve precoding design based on train position and MAC layer resource allocation. In the first stage, a fusion deep neural network is designed, which takes the geographical location of the train as input and outputs multi-user joint precoding that satisfies the power constraint. In the second stage, a deep reinforcement learning network is designed, which inputs the train position and data rate requirements into the network and outputs a resource allocation strategy that satisfies the balance of total data rate - fairness. Offline training is carried out using historical channel data at each location. In the actual deployment stage, the actual position of the train is input, and the network outputs the precoding parameters and resource allocation strategy at the corresponding position, realizing feedback-free transmission. Compared with the prior art, the present invention has the following advantages: 1. The present invention provides a location-based high-speed rail downlink communication system under an FD-RAN architecture, avoiding the problem of degraded transmission performance caused by channel aging and feedback delay in the high-speed rail scenario. 2. The present invention realizes multi-user multiple-input multiple-output precoding selection and resource allocation based on the train's geographical location, saving pilot estimation and feedback overhead, and proposes a fusion DNN network to further improve the system frequency efficiency. 3. The present invention uses staged deep learning to fit the mapping relationship between train geographical location information and precoding and resource allocation strategies, achieving a balance between the improvement of the system's total data rate and on-demand service fairness. Description of the Drawings

[0031] Figure 1 is a schematic diagram of the scenario applicable to the embodiment of the present invention.

[0032] Figure 2 is a schematic diagram of the staged machine learning training process of the embodiment of the present invention.

[0033] Figure 3 is a curve graph of the comparison of frequency efficiency under different speeds and delays in the embodiment of the present invention.

[0034] Figure 4 is a curve graph of the comparison of frequency efficiency under different Rice factors R in the embodiment of the present invention.

[0035] Figure 5 is a curve graph of the influence of input position error on frequency efficiency in the embodiment of the present invention.

[0036] Figure 6 is a curve graph of the comparison and change of frequency efficiency under different architectures in the embodiment of the present invention.

[0037] Figure 7 is a curve graph of the average reward under different throughput requirements in the embodiment of the present invention.

[0038] Figure 8 is a curve graph of the data rate under different throughput requirements in the embodiment of the present invention.

[0039] Figure 9It is the overall QoS satisfaction rate curve diagram of the embodiments of the present invention under different throughput requirements. Specific embodiments

[0040] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the following detailed description of the embodiments of the present invention is given with reference to the accompanying drawings: These embodiments are implemented on the premise of the technical solutions of the present invention, and detailed implementation manners and specific operation processes are given. It should be understood that the specific examples described herein are only used to explain the present invention, but the protection scope of the present invention is not limited to the following embodiments.

[0041] A downlink transmission method for a 6G fully decoupled high-speed rail mobile communication network based on geographical location disclosed in an embodiment of the present invention mainly includes the following steps:

[0042] Step 1: Construct a deep neural network with train geographical location information as input and multi-user joint precoding information that satisfies power constraints as output, and design a loss function of the deep neural network with system spectral efficiency as the optimization objective, and use the precoding output by the deep neural network for data transmission;

[0043] Step 2: Construct a deep reinforcement learning network with the state of train geographical location information and data rate requirements as input and the MAC layer resource allocation strategy as output action, and design a deep reinforcement learning algorithm with the weighted sum of the total data rate and demand satisfaction rate as the reward, and use the output strategy for MAC layer resource allocation;

[0044] Step 3: After multiple deep neural networks (DNNs) are offline trained and updated with the historical channel data of each location sample until convergence, model fusion is performed according to the performance of multiple networks on the evaluation set to obtain the fused DNN (mDNN) in the first stage;

[0045] Step 4: The deep reinforcement learning network uses the historical channel data of each location sample, combines the trained mDNN for offline training and updates the parameters until convergence to obtain the deep reinforcement learning network in the second stage;

[0046] Step 5: In the actual deployment stage, the train location information is input, the mDNN outputs the joint precoding at the corresponding location, and the deep reinforcement learning network outputs the resource allocation strategy to achieve feedback-free transmission.

[0047] The following combines Figure 1 The high-speed rail communication scenario shown is used to illustrate the specific steps of the embodiments of the present invention. It is assumed that the downlink base station and the mobile relay are in the same plane, there are L base stations serving K train mobile terminals, and the precoding vector of the l-th downlink base station for the k-th terminal at the top of the train is Satisfy the normalization condition The allocated power is pkl , satisfying the power constraint where N represents the number of antennas of the downlink base station, and p max is the maximum power limit.

[0048] Assume that the channel Η between the mobile relay at the top of the k-th train and the l-th downlink base station kl has a line-of-sight path and satisfies the Rice channel distribution; the received signal at the k-th train terminal is where s k represents the signal sent to the k-th terminal and satisfies represents the noise at, and represents the noise power. The achievable signal-to-interference-plus-noise ratio at the k-th terminal is where represent the combined channel and combined precoding of the k-th terminal respectively, represents the total power allocated to the k-th terminal; the spectral efficiency of the k-th terminal can be expressed as R k = log2(1 + SINR k ).

[0049] The following gives the specific process of modeling the system total data rate and fairness trade-off problem.

[0050] According to the 3GPP protocol, the available bandwidth is divided into A resource blocks (RBs) with a bandwidth of F Hz, and use to represent whether the a-th resource block serves terminal k; then the data rate of k terminals can be expressed as

[0051] Use to represent the satisfaction rate of k terminals, where represents the data rate requirement; the optimization problem of the system is modeled as

[0052]

[0053] where μ1 and μ2 represent the weighting coefficients of the data rate and satisfaction rate respectively, and W, p, and i represent the combined precoding, power, and resource block allocation of all terminals; the optimization goal is to maximize the weighted sum of the system total data rate and the overall terminal satisfaction rate, and achieve the balance of maximum performance and fairness.

[0054] The following introduces the specific process of problem decomposition and the phased machine learning algorithm for the targeted optimization problem. For the above optimization problem, it is decomposed into two sub-problems according to the decision level: 1) precoding design and power allocation at the physical layer; 2) resource block allocation at the MAC layer; and a two-stage machine learning algorithm is designed to solve the above sub-problems.

[0055] For the precoding design and power allocation in the first - stage physical layer, considering that the number of terminals served within a resource block a can be freely selected between 1 and K, the optimization problem can be modeled as

[0056]

[0057] where, represents K a ′ represents the number of terminals using resource block a; the loss function of the deep neural network is defined as the negative of the sum of the system spectral efficiencies, T represents the number of samples, i k,a,t ,R k,a,t respectively represent the service index and spectral efficiency of the ath resource block for the kth terminal in the tth sample; the output of the deep neural network is the precoding result including power allocation According to it is transformed into a precoding vector that satisfies the constant - modulus constraint.

[0058] In this embodiment, the deep - learning neural network uses a fully - connected deep - learning neural network or a residual network; during the training process, first, the entire dataset is divided into a training set and an evaluation set; then, multiple DNN sets with the same loss function are trained to convergence using different learning rates; finally, according to the Loss value on the evaluation set, the DNN with the smallest loss at each input position is selected and fused to form mDNN.

[0059] For the resource allocation problem in the second - stage MAC layer, after the spectral efficiency has been determined by precoding and power, the optimization problem can be modeled as

[0060]

[0061] For the resource allocation in the MAC layer, in this embodiment, it is proposed to use the location information of the terminals and directly map to the optimal RB allocation strategy using deep reinforcement learning and the D3QN network; for each step j in the deep - reinforcement - learning training episode, the state of the D3QN consists of three parts, which respectively represent the positions x of all terminals at the jth step j , the data rate that each terminal still needs to meet the data - rate requirement and the remaining number of resource blocks A j ; the action is defined as an allocation strategy for a resource block, and a resource block can be allocated to 1 to K terminals; the reward is defined as the weighted sum of two components: 1) the total increased data rate 2) the increased satisfaction rate of all terminals' requirements, which can be expressed as

[0062]

[0063] where, represents the increased data rate of the kth terminal after the jth step, satisfying

[0064] The specific training process of the two-stage machine learning algorithm is as follows Figure 2 As shown, after repeating the above deep learning algorithm training multiple times in the first stage, K' = [1,...,K] mDNN models for different numbers of terminals are fused. Then, the mDNN models under different K' are fixed for the next stage. During the training process of D3QN in the second stage, an iterative loop (iteration) is defined to include all position sampling points. The training of each position corresponds to an episode, and the input position within an episode is kept fixed (i.e., all steps j within an episode are used to train one position). Parameters such as the exploration rate λ of the proposed deep reinforcement learning algorithm are updated after each iteration, rather than after each episode as in traditional DRL algorithms.

[0065] To make this embodiment more intuitive and compare the advantages and disadvantages of the no-feedback scheme and the feedback-based scheme, Figure 3 A comparison schematic diagram showing the simulation results of frequency efficiency comparison at different speeds and latencies is presented. The results show that the average frequency efficiency of the no-feedback, location-based DNN, and mDNN remains stable at different speeds, while the performance of the MMSE scheme that relies on accurate real-time CSI feedback degrades as the speed and latency increase. The proposed mDNN network has a higher performance than the traditional DNN, and its performance reaches 95% of the theoretical optimal value of the feedback-based scheme at a speed of 0. Although the feedback-based scheme shows better performance at lower speeds, once the speed exceeds 125 km / h, the proposed mDNN shows the best performance, indicating the advantage of the no-feedback method in high-speed mobile scenarios.

[0066] Figure 4 A comparison schematic diagram showing the simulation results of frequency efficiency comparison at different Rice factors R when the speed is 350 km / h is presented. The results show that the performance of each scheme improves as the R coefficient increases. Among them, the SE of the location-based mDNN increases the most, indicating its stronger ability to utilize the line-of-sight component of the channel. Although the performance of mDNN is lower when the line-of-sight path is not the main component of the channel, when this coefficient exceeds 10 dB, the performance of mDNN reaches the best among all schemes. Considering that the environment with a strong line-of-sight component is the main scenario of high-speed rail communication, these simulation results further verify the applicability of the proposed mDNN to the high-speed rail environment.

[0067] Considering the possible error between the input position and the actual position of the train, Figure 5It shows the impact of the input position error on the frequency efficiency degradation. Since the train trajectory is fixed, the input position error only affects a single direction. In contrast, the CSI used in the MMSE-FPA scheme is based on the channel at the corresponding incorrect position. As Figure 5 shown, the proposed feedback-free mDNN scheme is more robust to such inaccuracies. When the input position error reaches 5 m, the SE of the feedback-free scheme drops by about 15%, while the performance of the feedback-based MMSE-FPA scheme drops by about 50%. This is because the accuracy of position-based precoding is lower than that of CSI-based precoding, which reduces its sensitivity to position errors and emphasizes the robustness of the mDNN scheme in high-speed railway communication applications.

[0068] To make this embodiment more intuitive and compare the advantages and disadvantages of the FD-RAN architecture with other architectures, Figure 6 it shows a comparison diagram of the high-speed railway relay frequency efficiency simulation results under different architectures. Figure 6 (a) shows the frequency efficiency at different positions when v = 216 km / h. At this speed, MMSE-FPA exhibits an average frequency efficiency similar to that of the proposed method. Research shows that the performance of the single-connection SC-MMSE scheme is the worst because it cannot benefit from the diversity gain of multi-base station joint precoding. In addition, compared with the proposed mDNN, the feedback-based scheme shows significant performance fluctuations near the base station coverage edge and poor performance in the handover area. We can also observe that in the area where the DNN performance degrades, the proposed mDNN shows improved frequency efficiency, indicating that mDNN can compensate for the performance at positions where traditional DNNs cannot be effectively trained. Figure 6 (b) shows the cumulative distribution function (CDF) of the frequency efficiency under different transmission modes, further indicating that the mDNN proposed in the present invention obtains a more stable SE than MMSE-FPA and a higher average SE than DNN and SC-MMSE-FPA. The frequency efficiency and its CDF at different positions when v = 500 km / h are as Figure 6 (c) and (d) shown, where the proposed mDNN shows more stable performance and better average performance among all the schemes. This proves the superiority of the FD-RAN scheme in high-speed railway communication scenarios.

[0069] To make this embodiment more intuitive and compare the advantages and disadvantages of the feedback-free scheme and the feedback-based one, Figure 7 it shows a comparison diagram of the total data rate and the average terminal satisfaction rate simulation results under different data rate requirements and delays. As Figure 7As shown, the feedback-based Swap Matching (SM) method achieves near-theoretical optimal rewards under low data rate requirements, but its performance significantly degrades when the data rate requirement exceeds 50 Mbps. The performance of the proposed feedback-free method is close to the theoretical best performance of the feedback-based method (ES), and it outperforms the best feedback-based scheme at a latency of 1 ms. Considering that Figure 7 the reward in Figure 8 is defined as the weighted sum of the data rate and the data rate satisfaction rate, Figure 9 and

[0070] This embodiment of the present invention also discloses a computer system, including a memory, a processor, and a computer program stored on the memory and executable on the processor. The computer program is loaded onto the processor to implement the steps of the downlink transmission method of the 6G fully decoupled high-speed rail mobile communication network based on geographical location.

[0071] This embodiment of the present invention also discloses a computer program product, including a computer program, which is loaded onto the processor to implement the steps of the downlink transmission method of the 6G fully decoupled high-speed rail mobile communication network based on geographical location.

[0072] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location, characterized in that, It includes the following steps: Construct a deep neural network with the train's geographical location information as the input and the multi-user joint precoding information that satisfies the power constraint as the output, design the loss function of the deep neural network with the system frequency efficiency as the optimization goal, and use the precoding output by the deep neural network for data transmission; Construct a deep reinforcement learning network with the state of the train's geographical location information and data rate requirements as the input and the MAC layer resource allocation strategy as the output action, design the deep reinforcement learning algorithm with the weighted sum of the total data rate and the demand satisfaction rate as the reward, and use the output policy for MAC layer resource allocation; After multiple deep neural networks (DNNs) use the historical channel data of each location sample for offline training to update the parameters until convergence, model fusion is performed according to the performance of multiple networks on the evaluation set to obtain the fused DNN (mDNN) in the first stage; The mDNN selects the DNN with the smallest loss on the evaluation set for each input location; The deep reinforcement learning network uses the historical channel data of each location sample and combines the trained mDNN for offline training to update the parameters until convergence to obtain the deep reinforcement learning network in the second stage; In the actual deployment stage, the train location information is input, the mDNN outputs the joint precoding at the corresponding location, and the deep reinforcement learning network outputs the resource allocation strategy to achieve feedback-free transmission.

2. The downlink transmission method of a 6G fully decoupled network for high-speed rail mobile communication based on geographical location according to claim 1, wherein: In multi - user joint precoding, the precoding vector of the \(l\) - th downlink base station for the \(k\) - th terminal on the top of the train is satisfying the normalization condition The allocated power is \(p\) kl , satisfying the power constraint where \(N\) represents the number of antennas of the downlink base station, \(\text{tr}(\cdot)\) represents the trace of the matrix, and \(p\) max is the maximum power limit; the channel \(\mathbf{H}\) between the \(k\) - th train terminal and the \(l\) - th downlink base station kl has a line - of - sight path and satisfies the Rician channel distribution; the achievable signal - to - interference - plus - noise ratio (SINR) of the \(k\) - th terminal is where represent the joint channel and joint precoding of the \(k\) - th terminal respectively, represents the total power allocated to the \(k\) - th terminal, \(L\) represents the number of base stations, \(K\) represents the number of terminals, represents the noise power; the spectral efficiency of the \(k\) - th terminal is denoted as \(R\) k \(=\log_2(1 + \text{SINR}\) k ).

3. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location according to claim 1, characterized in that: In the MAC layer resource allocation, the available bandwidth is divided into A resource blocks (RBs) with a bandwidth of F Hz, and is used to indicate whether the a-th resource block serves the terminal k; then the data rate of the k-th terminal is expressed as Using to represent the satisfaction rate of k terminals, where represents the data rate requirement; the optimization problem of the system is modeled as Among them, μ1 and μ2 respectively represent the weighted coefficients of the total data rate and the satisfaction rate; W, p, and i respectively represent the joint precoding, power, and resource block allocation of all terminals, L represents the number of base stations, K represents the number of terminals, and W kl and p kl respectively represent the precoding vector and the allocated power of the l-th downlink base station to the k-th terminal on the top of the train, and p max is the maximum power limit.

4. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location according to claim 1, characterized in that: For the precoding design and power allocation in the physical layer of the first stage, considering that the number of terminals served within a resource block a can be freely selected between 1 and K, the optimization problem is modeled as Among them, W and p respectively represent the joint precoding and power allocation of all terminals, and K a ′ represents the number of terminals using resource block a, and R k,a represents the spectral efficiency of the k-th terminal using resource block a, W kl and p kl respectively represent the precoding vector and the allocated power of the l-th downlink base station to the k-th terminal on the top of the train. p max is the maximum power limit, L represents the number of base stations, and K represents the number of terminals; the loss function of the deep neural network is defined as the negative of the sum of the system spectral efficiencies, T represents the number of samples, and i k,a,t , R k,a,t respectively represent the service index and spectral efficiency of the a-th resource block to the terminal k in the t-th sample; the output of the deep neural network is the precoding result including power allocation According to it is transformed into a precoding vector that satisfies the constant modulus constraint.

5. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location according to claim 1, characterized in that: The deep neural network adopts a fully connected deep learning neural network or a residual network; during the training process, first divide the entire dataset into a training set and an evaluation set; then, train multiple DNN sets with the same loss function using different learning rates until convergence; finally, according to the Loss value on the evaluation set, select the DNN with the smallest loss for each input location and fuse them to form the mDNN.

6. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location according to claim 1, characterized in that: For the resource allocation problem in the MAC layer of the second stage, after the frequency efficiency has been determined by precoding and power, the optimization problem is modeled as where, i k,a indicates whether the a-th resource block serves terminal k, i represents the resource block allocation for all terminals, K represents the number of terminals, and A represents the number of resource blocks. and respectively represent the data rate and satisfaction rate of the k-th terminal, and μ1 and μ2 are the corresponding weighting coefficients.

7. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location according to claim 6, characterized in that: For each step j in a deep reinforcement learning training episode, the state of the deep reinforcement learning network consists of three parts, representing the positions of all terminals at step j, the data rate still required for each terminal to meet the data rate requirement, and the remaining number of resource blocks; the action a j is defined as an allocation policy for a resource block, and a resource block can be allocated to 1 - K terminals; the reward is defined as the weighted sum of two components: 1) the total increased data rate 2) the increased satisfaction rate of the demands of all terminals, expressed as Among them, represents the data rate improved by the k-th terminal after the j-th step, satisfying 8. A downlink transmission method for a 6G fully decoupled network of high-speed rail mobile communication based on geographical location according to claim 1, characterized in that: In the first stage, for different numbers of terminals, after repeating the training of the deep learning algorithm and fusing, K′ = [1,..., K] mDNN models are obtained, and then the mDNN models under different K′ are fixed for the next stage; during the training process of the deep reinforcement learning network in the second stage, define an iterative loop that includes all location sampling points, the training of each location corresponds to one episode, and the input location remains fixed within one episode; the parameters of the deep reinforcement learning algorithm are updated after each iterative loop.

9. A computer system, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that: The computer program is loaded into the processor to implement the steps of the geographical location-based 6G fully decoupled network downlink transmission method for high-speed rail mobile communication according to any one of claims 1-8.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the geographical location-based 6G fully decoupled network downlink transmission method for high-speed rail mobile communication according to any one of claims 1-8.

Images