Qos guarantee low overhead coordinated multi-point beam tracking method based on alternating training and feedback

Abstract

The application provides a QoS guarantee low-overhead multi-point cooperative beam tracking method based on alternating training and feedback, solves the problem that the beam training of a non-cell MIMO system has high overhead, and the tracking performance and resource utilization rate are difficult to balance. The method comprises two modules of AP training sequence decision and alternating training beam tracking. Path loss information is first collected and input into the AP training sequence decision module, and the optimal AP training sequence is obtained through offline training. Then the sequence is input into the alternating training beam tracking module, and through the iteration process of newly added scanning beams, bottleneck user identification and beam allocation, the optimal beam scheme is output by on-demand training. The application adopts a dynamic alternating training strategy, and double time scales are used to realize adaptive adjustment of the training length and optimal decision of the AP training sequence, thereby reducing the training overhead, guaranteeing the system QoS, having robustness and scalability, and being suitable for user mobile beam tracking in a high-frequency communication scene.

Description

Technical Field

[0001] This invention belongs to the field of wireless communication technology, and particularly relates to a low-overhead multi-point cooperative beam tracking method with QoS guarantee based on alternating training and feedback. Background Technology

[0002] In high-frequency communication systems such as millimeter-wave (mmWave) and terahertz (THz), beam tracking is a core technology for ensuring high-speed and stable links. High-frequency signals suffer from high path loss and poor penetration, necessitating large-scale antenna arrays to achieve high-gain directional transmission, thereby compensating for attenuation and improving spectral efficiency. However, this highly directional transmission places extremely high demands on the accuracy of channel state information (CSI): even minute angular deviations or environmental disturbances can cause beam misalignment or even link interruption. Therefore, beam tracking must rely on real-time and reliable channel estimation to dynamically adjust the beam direction.

[0003] Traditional channel estimation often employs pilot-assisted mechanisms, acquiring the channel response by transmitting known pilot signals. Theoretically, reconstructing a sparse millimeter-wave channel composed of a few dominant paths requires a training time slot number at least equal to the number of transmit antennas to meet the basic requirements of compressed sensing or least-squares estimation. However, in practical large-scale MIMO systems, the number of transmit antennas can reach tens or even hundreds, significantly increasing training overhead and consuming substantial time-frequency resources, thus compressing the effective data transmission time. Especially in high-speed mobile scenarios, where channel coherence time is extremely short, an excessively long training process can render the acquired CSI invalid before data transmission, further degrading beam tracking performance. Simultaneously, frequent full-scale training increases power consumption and computational complexity, posing challenges to the energy efficiency design of terminal devices. Therefore, reducing beam training overhead while ensuring high accuracy and low latency in channel estimation has become a critical issue that urgently needs to be addressed in the field of beam tracking. Summary of the Invention

[0004] The main objective of this invention is to address the technical problems of excessively high beam training overhead and difficulty in balancing tracking performance and resource utilization under dynamic channel changes in existing millimeter-wave non-cellular MIMO systems. This invention provides a low-overhead, multi-point cooperative beam tracking method based on interleaved training and feedback to ensure QoS, effectively reducing training overhead and reliably guaranteeing system QoS, thereby improving the robustness and scalability of beam tracking in dynamic channel environments.

[0005] The technical solution adopted in this invention is: Step 1: Construct a millimeter-wave non-cellular MIMO downlink communication system and establish an optimization problem that minimizes the system beam training overhead; Step 2: Model the access point (AP) order decision problem over a long time scale based on path loss information, design the global state, action and reward function in sequence, use a policy gradient agent for offline training, and build an AP training order decision module. Step 3: Based on the AP sequence decision module obtained in Step 2, design the bottleneck user-aware alternating training beam tracking method and construct the alternating training beam tracking module.

[0006] Further, in step 1.1, consider a millimeter-wave non-cellular MIMO downlink communication system, where M APs jointly serve U single-antenna UEs. Each AP is equipped with... Root antenna, among which and These represent the antenna array dimensions in the horizontal and vertical directions, respectively. Indicates the horizontal direction. Indicates the vertical direction. Assume each AP's antenna array is equipped with... A radio frequency (RF) link. Since effective communication requires allocating an independent beam for each user, it is necessary to meet certain requirements. ,set up .

[0007] Without loss of generality, This represents the channel vector from the m-th base station to the u-th user, and this channel vector is modeled as... ,in It is the large-scale fading coefficient. This is the small-scale fading matrix. Small-scale fading. Using a geometric channel model, it is represented as in: This represents the total number of propagation paths from AP m to user u; This represents the distance from AP m to user u. The complex gain of each path follows an independent and identically distributed Gaussian distribution; This represents the steering vector at the transmitting end. and These are the first two lines from AP m to user u. The azimuth and elevation angles of the path from which the path departs.

[0008] For UPA, its expression is: in: ; ; Represents the Kronecker product; Step 1.2: For non-cellular downlink transmission, considering the total system power constraint, Let be the total system power, then the reception equation for the receiver of the u-th user can be written as: in, This represents the channel matrix received at each user. This indicates the conjugate transpose. Represents the field of complex numbers; Represents the transmission symbol vector for all users. ,in Indicates a complex Gaussian distribution; Let be the noise vector, where Noise power density; To simulate the precoding matrix, This is the baseband precoding matrix. for The i-th column; In downlink transmission, a hybrid beamforming (BF) architecture is employed, where the analog beamforming vectors are selected from the standard Discrete Fourier Transform (DFT) codebook. And satisfy , Let N represent the identity matrix; Let m be the analog beam of the m-th AP facing the k-th user, where Let the m-th AP be the codeword index of the k-th user. ,in This represents the analog precoding matrix on the m-th AP. Let the block diagonal matrix be represented; then the effective channel estimation matrix on the user side is: ,in This represents the calibrated channel estimation matrix at the u-th user. Indicates transpose; The baseband precoding uses MMSE precoding. The total power scaling factor is ; Then, the SINR of user u is ; Step 1.3: List the optimization problem, namely, in millimeter-wave non-cellular MIMO downlink communication systems, timely and accurate CSI is required to achieve efficient beam tracking in user mobility scenarios. This paper adopts a downlink pilot + uplink feedback CSI acquisition scheme, dividing the coherence interval into training and data transmission phases.

[0009] Traditional full-training schemes suffer from excessively long training times, leading to data transmission time compression. To address this, a dynamic interleaving training scheme is proposed: training is only conducted until the target SINR requirement is met, allowing the training length to be adaptively adjusted according to the instantaneous CSI, with the optimization objective being to minimize the training length.

[0010] ; in, This represents the minimum value. This means minimizing the training length while ensuring that the SINR of all current users is greater than the SINR threshold; where, , This represents the SINR threshold.

[0011] Furthermore, the AP training order decision module includes the following specific steps: Step 2.1: Design the state based on path loss information At the same time, the AP sequence is used as the action. Training length As a reward, a policy gradient agent is designed, with the optimization objective being to minimize the expected training length: Step 2.2: The agent's goal is to maximize long-term expected reward, that is, to learn the optimal policy. To minimize the average training length, the optimization problem can be formalized as the following objective function: ; Where D represents the contextual empirical distribution sampled from the actual network environment. Indicates the strategy parameters The affected agents are in the state Select action strategy, Indicates the state Select action The reward obtained; in order to optimize strategy parameters A policy gradient method with baseline is adopted, and the unbiased gradient estimator of this policy gradient method is: ; in Indicates the parameter gradient operator, That is, solving the objective function Regarding strategy parameters gradient, Indicates the state Baseline rewards below; Step 2.3: To promote effective exploration in the early stages of training and avoid the policy converging to a local optimum too early, we adopt a temperature-controlled soft policy sampling mechanism: in, Indicates the introduction of temperature parameters The probability distribution of the subsequent random policy Indicates proportional to, This represents an exponential function.

[0012] Furthermore, based on the AP order obtained by the AP training order decision module, a bottleneck-aware alternation mechanism is proposed, the core of which is "training on demand and completing as quickly as possible": only some APs are activated, and the process is gradually expanded until the system QoS target is met. The alternating training of the beam tracking module includes the following specific steps: Step 3.1, Start the dual-loop training framework: Training is executed by nested outer and inner loops: Outer loop: Activate APs sequentially according to the order determined by the long-term strategy; Inner loop: For each newly activated AP, add beams one by one from its codebook for incremental training.

[0013] Step 3.2, Bottleneck User Identification: The observed channel power between user u, AP m, and beam n is defined as: ; in, Represents the square of the modulus; This represents the set of available beams on the m-th AP. For each remaining user ,in For the first For the set of users to be assigned in the round, calculate their aggregated observation power: ; in It is the set of available beams for the m-th AP that has not yet been assigned in the i-th round; subsequently, the system identifies the user with the worst current aggregation channel conditions, i.e., the bottleneck user. , This represents the index corresponding to the minimum value; This user is considered a "weak link" in system performance, so the best beam resources are allocated to them first to ensure that the overall service quality is not dragged down by a single weak user.

[0014] Step 3.3, Per-AP Beam Allocation for Bottleneck Users: After identifying the bottleneck users in the current round... Then, the system assigns the optimal beam to the user on each activated AP. For the m-th AP, its beam selection metric is defined as: Where: molecule The value represents the observed channel power of the bottleneck user on the $n$-th beam of the m-th AP; the denominator is the total interference power of that beam to other unserved users plus the noise variance. This is used to reflect the interference cost of resource reuse; subsequently, the system selects the optimal beam for the m-th AP, with the index as: Step 3.4: After the resource pool update completes beam allocation, the system updates the remaining users and beam sets. ; in, This represents the set difference operation (i.e., deleting a specified element from the original set).

[0015] That is, remove the bottleneck users that have been served from the candidate pool and remove the beams they occupy from the available beams of the corresponding AP to ensure that resources are not reused.

[0016] Step 3.5, System Performance Calculation and Training Termination: The above iterative process ultimately generates a beam assignment result, based on which a simulated beamforming matrix can be constructed. And calculate the effective channel. Based on this, a digital precoder (such as MMSE) is used to solve the SINR for each user, and the system bottleneck SINR is defined as: This value serves as the termination criterion for the outer interleaved training loop. When it satisfies... When all users' SINR is not lower than the preset threshold. When the training process is interrupted, the entire training process is immediately terminated, thereby achieving the goal of efficient beam management: "training on demand and stopping as early as possible".

[0017] The beneficial effects of this invention are as follows: On a short timescale, the proposed alternating training strategy significantly shortens the training length compared to the full training scheme, while achieving higher effective throughput compared to fixed-length partial training methods; and the bottleneck user-aware beam allocation criterion exhibits the best effective rate performance among various baseline algorithms. On a long timescale, the designed reinforcement learning algorithm can stably achieve a 5% reduction in training overhead under different user movement speeds, while simultaneously improving effective throughput. In summary, this invention effectively balances beam tracking accuracy, training overhead, and system energy consumption, possesses excellent robustness and scalability in dynamic channel environments, requires no complex hardware modifications, adapts to the distributed cooperative characteristics of millimeter-wave non-cellular MIMO systems, and provides an efficient and feasible solution for user movement beam tracking in high-frequency communication scenarios. Attached Figure Description

[0018] Figure 1 This is a flowchart of an embodiment of the present invention; Figure 2 This is a timestamp image of the present invention; Figure 3 A comparison of the training lengths of the bottleneck user algorithm and the direct allocation algorithm under different SINR thresholds; Figure 4 A comparison of the effective transmission rates of the bottleneck user algorithm and the direct allocation algorithm under different SINR thresholds; Figure 5 Comparison of training lengths for different AP transmit powers at different SINR thresholds; Figure 6 A comparison of the effective transmission rates of different AP transmit powers under different SINR thresholds; Figure 7 A comparison of training lengths for alternating training and full / partial training at different SINR thresholds; Figure 8 A comparison of the effective transmission rates of alternating training and full / partial training under different SINR thresholds; Figure 9 For the convergence of the policy gradient agent; Figure 10 A comparison of the training lengths of policy gradient and fixed AP order and random AP order over time at different user movement speeds; Figure 11 A comparison of the effective transmission rates of policy gradient, fixed AP order, and random AP order over time at different user movement speeds; Detailed Implementation

[0019] To better understand the purpose, structure, and function of this invention, the following detailed description of the QoS-guaranteed low-overhead multi-point cooperative beam tracking scheme based on alternating training and feedback is provided in conjunction with the accompanying drawings.

[0020] The following is a QoS-guaranteed, low-overhead, multi-point cooperative beam tracking scheme based on alternating training and feedback proposed in this invention, such as... Figure 1 As shown, the specific steps for implementing the entire process include the following: First, a downlink channel model for a cellless MIMO system is constructed. Consider a millimeter-wave cellless MIMO downlink communication system, in which M APs jointly serve U single-antenna UEs, and a hybrid beamforming structure is adopted at the APs.

[0021] Then, in the AP training order decision module: the AP training order problem is modeled based on path loss information; the state is designed based on path loss information. At the same time, the AP sequence is used as the action. Training length As a reward, a policy gradient method with baseline is used, as follows: Finally, a beam tracking module is constructed based on the optimal AP training order obtained from the sequential decision module, and the timestamp of the dual-loop training framework is referenced. Figure 2 Training is performed by nested outer and inner loops: outer loop: activates APs sequentially according to the order determined by the long-term strategy; inner loop: for each newly activated AP, adds beams one by one from its codebook for incremental training.

[0022] After identifying bottleneck users based on the bottleneck user criteria, the system iteratively allocates beams to bottleneck users in each round. Then, the system updates the remaining users and beam sets, removes the bottleneck users that have been served from the candidate pool, and removes the beams they occupy from the available beam set of the corresponding AP to ensure that resources are not reused.

[0023] The above iterative process ultimately generates a beam assignment result, based on which a simulated beamforming matrix can be constructed, the effective channel can be calculated, and the system's bottleneck SINR can be defined as: This value serves as the termination criterion for the outer interleaved training loop. When it satisfies... When all users' SINR is not lower than the preset threshold. When the training process is interrupted, the entire training process is immediately terminated, thereby achieving the goal of efficient beam management: "training on demand and stopping as early as possible".

[0024] Experimental results Define effective transmission rate for: like Figure 3 and Figure 4 As shown, the bottleneck algorithm achieves a shorter training length across all SINR thresholds. and higher effective throughput . Follow It exhibits a single-peak trend of first rising and then falling: initially affected by Driven by this, the training length increases, but once it exceeds a certain threshold, excessively long training times become the dominant factor, leading to... decline.

[0025] like Figure 5 and Figure 6 This indicates that appropriately increasing the AP's transmit power can shorten the signal strength at all SINR thresholds. and improve .

[0026] like Figure 7 and Figure 8 As shown, alternating training achieves the highest effective transmission rate. The next best is fixed-length partial training, while full training is the least effective. At low SINR thresholds, alternating training requires only shorter intervals. The goal can be achieved, therefore Higher; and as the SINR threshold increases, alternating training adaptively increases the training length and maintains a higher SINR, thus sustaining a higher SINR over a longer training period. .

[0027] like Figure 9 As shown, the convergence curve of the policy gradient agent drops significantly at the beginning and basically converges after about 20 time slots.

[0028] like Figure 10 and Figure 11 As shown, this paper demonstrates the process of completing beam training at time 0 without prior beam information, and then verifying robustness through experiments over 13 time slots. The experimental results confirm that the proposed scheme maintains low training length and high effective transmission rate throughout the 13 time slots, while also preserving performance over the extended time dimension.

[0029] Specifically, in scenarios with low user movement speeds (2.0 m / s), policy gradient agents achieve higher performance compared to fixed and stochastic policies. Furthermore, the training length is shorter. It is worth noting that at high movement speeds (8.0 m / s), the performance of all schemes degrades due to increased channel fluctuations; however, the policy gradient method still maintains a relatively short training time. and higher This further demonstrates its robust and adaptive characteristics despite the performance degradation caused by increased speed.

[0030] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A QoS-guaranteed, low-overhead, multi-point cooperative beam tracking method based on alternating training and feedback, characterized in that: Includes the following steps: Step 1: Construct a millimeter-wave non-cellular MIMO downlink communication system and establish an optimization problem that minimizes the system beam training overhead; Step 2: Model the access point (AP) order decision problem over a long time scale based on path loss information, design the global state, action and reward function in sequence, use a policy gradient agent for offline training, and build an AP training order decision module. Step 3: Based on the AP sequence decision module obtained in Step 2, design the bottleneck user-aware alternating training beam tracking method and construct the alternating training beam tracking module.

2. The QoS-guaranteed low-overhead multi-point cooperative beam tracking method based on alternating training and feedback according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Consider a millimeter-wave cellular-free MIMO downlink communication system, where M APs jointly serve U single-antenna UEs; each AP is equipped with Root antenna, among which and These represent the antenna array dimensions in the horizontal and vertical directions, respectively. Indicates the horizontal direction. Indicates the vertical direction; the antenna array of each AP is equipped with One radio frequency (RF) link; configuration ; make This represents the channel vector from the m-th AP to the u-th user, and this channel vector is modeled as... ,in It is the large-scale fading coefficient. It is a small-scale fading matrix; small-scale fading A geometric channel model is adopted; Step 1.2: For non-cellular downlink transmission, considering the total system power constraint, Let be the total system power. Then, the reception equation for the receiver of the u-th user can be written as: ； in, This represents the channel matrix received at each user. This indicates the conjugate transpose. Represents the field of complex numbers; Represents the transmission symbol vector for all users. ,in Indicates a complex Gaussian distribution; Let be the noise vector, where Noise power density; To simulate the precoding matrix, This is the baseband precoding matrix. for The i-th column; In downlink transmission, a hybrid beamforming (BF) architecture is employed, where the analog beamforming vectors are selected from the standard Discrete Fourier Transform (DFT) codebook. And satisfy , Let N represent the identity matrix; Let m be the analog beam of the m-th AP facing the k-th user, where Let the m-th AP be the codeword index of the k-th user. ,in This represents the analog precoding matrix on the m-th AP. Let the block diagonal matrix be represented; then the effective channel estimation matrix on the user side is: ,in This represents the calibrated channel estimation matrix at the u-th user. Indicates transpose; The baseband precoding uses MMSE precoding. The total power scaling factor is ; Then, the SINR of user u is ； Step 1.3: In a millimeter-wave cellular-free MIMO downlink communication system, a CSI acquisition scheme of downlink pilot + uplink feedback is adopted, dividing the coherent interval into training and data transmission phases; a dynamic interleaving training scheme is proposed: training is only performed until the target SINR requirement is met, so that the training length is adaptively adjusted with the instantaneous CSI, and the optimization objective is to minimize the number of training beams or the training length. : ； in, This represents the minimum value. This means minimizing the training length while ensuring that the SINR of all current users is greater than the SINR threshold; where, , This represents the SINR threshold.

3. The QoS-guaranteed low-overhead multi-point cooperative beam tracking method based on alternating training and feedback according to claim 2, characterized in that, The AP training order decision module includes the following specific steps: Step 2.1: Design the state based on path loss information At the same time, the AP sequence is used as the action. Training length As a reward, a policy gradient agent is designed, with the optimization objective being to minimize the expected training length: ； in The strategy for the agent; The expectation operator is used to calculate the statistical average of the expression within the parentheses. Indicates the state Select action The obtained training length; Step 2.2: The agent's goal is to maximize the long-term expected reward, which can be formalized as the following objective function: ； Where D represents the contextual empirical distribution sampled from the actual network environment. Indicates the strategy parameters The affected agents are in the state Select action strategy, Indicates the state Select action The reward obtained; in order to optimize strategy parameters A policy gradient method with baseline is adopted, and the unbiased gradient estimator of this policy gradient method is: ； in Indicates the parameter gradient operator, That is, solving the objective function Regarding strategy parameters gradient, Indicates the state Baseline rewards below; Step 2.3: To promote effective exploration in the early stages of training and avoid the policy converging to a local optimum too early, a temperature-controlled soft policy sampling mechanism is adopted: ； in, Indicates the introduction of temperature parameters The probability distribution of the subsequent random policy Indicates proportional to, This represents an exponential function.

4. The QoS-guaranteed low-overhead multi-point cooperative beam tracking method based on alternating training and feedback according to claim 3, characterized in that, The alternating training beam tracking module includes the following specific steps: Step 3.1: Start the dual-loop training framework Training is performed using nested outer and inner loops: Outer loop: Activates APs sequentially according to the order determined by the long-term strategy; Inner loop: For each newly activated AP, add beams one by one from its codebook for incremental training; Step 3.2, Bottleneck User Identification The observation channel power between user u, AP m, and beam n is defined as: ； in, Represents the square of the modulus; This represents the set of available beams on the m-th AP. For each remaining user ,in For the first For the set of users to be assigned in the round, calculate their aggregated observation power: ； in It is the set of available beams for the m-th AP that has not yet been assigned in the i-th round; subsequently, the system identifies the user with the worst current aggregation channel conditions, i.e., the bottleneck user. , This represents the index corresponding to the minimum value; Step 3.3: Per-AP beam allocation for bottleneck users After identifying the bottleneck users in the current round Then, the system assigns the optimal beam to the user on each activated AP; for the m-th AP, its beam selection metric is defined as: ； Where: molecule The value represents the observed channel power of the bottleneck user on the $n$-th beam of the m-th AP; the denominator is the total interference power of that beam to other unserved users plus the noise variance. This is used to reflect the disruption cost of resource reuse; Subsequently, the system selects the optimal beam for the m-th AP, with the index being... ; Step 3.4, Resource Pool Update After beam assignment is completed, the system updates the remaining users and beam sets: ； in, This represents the set difference operation; Step 3.5: System performance calculation and training termination After beams are assigned to each user, a beam assignment result is generated, and a simulated beamforming matrix is ​​constructed based on this result. And calculate the effective channel. Based on this, a digital precoder is used to solve for the SINR of each user, and the system SINR is calculated: ; The termination criterion for the outer interleaved training loop is: when the condition is met... When all users' SINR is not lower than the preset threshold. If this happens, the entire training process will be terminated immediately.

Images