A beam alignment method and device based on quantized perception training
By optimizing the probe beam codebook and service beam prediction network under the hardware constraints of low-resolution phase shifters through a quantized sensing training method, the problems of decreased beam alignment accuracy and training instability are solved, and efficient beam alignment and low-complexity deployment are achieved.
Patent Information
- Application Number
- CN202611326629.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-30
- Publication Date
- 2026-11-10
AI Technical Summary
Under the hardware constraints of analog beamforming using low-resolution phase shifters in base stations, existing technologies struggle to perform end-to-end joint optimization of a small number of probe beam codebooks and service beam prediction networks, leading to decreased beam alignment accuracy and instability in the training process.
A quantization-based sensing training method is adopted. By introducing discrete phase constraints from a low-resolution phase shifter and combining stopping gradient processing and progressive annealing strategies, a probe beam that satisfies constant modulus and unity power constraints is generated. The propagation characteristics of specific deployment scenarios are then used for training to generate a probe beam codebook adapted to actual hardware.
It reduces the number of beam scans and reference signal overhead, improves beam alignment accuracy, reduces system deployment complexity, and adapts to different user locations and link gain conditions.
Smart Images

Figure CN123074547A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of wireless communication technology, and in particular to a beam alignment method and apparatus based on quantized sensing training. Background Technology
[0002] High-frequency communication technologies such as millimeter waves and terahertz waves offer high bandwidth and data rates, serving as crucial supporting technologies for 5G and future 6G mobile communication enhancements. Due to high-frequency path loss and weak diffraction capabilities, base stations typically require large-scale antenna arrays, utilizing highly directional narrow beams to achieve array gain. During initial access, link recovery, or user migration, the base station needs to determine and configure suitable serving beams from a candidate serving beam codebook to achieve beam alignment. This necessitates beam scanning and acquiring the received power of candidate beams. Traditional exhaustive beam scanning requires scanning all candidate serving beams one by one; the scanning overhead increases with the number of candidate serving beams, resulting in high initial access latency and high reference signal overhead, making it difficult to meet the requirements of low-latency, high-frequency communication systems.
[0003] To reduce beam scanning overhead, existing technologies have proposed a layered beam scanning method. This method typically scans a wide beam with a large coverage area first, then gradually narrows the scanning range to finally determine a narrow beam. Compared to exhaustive beam scanning, layered beam scanning can reduce scanning overhead, but it still has the following drawbacks: First, the structure of the layered codebook is usually pre-designed manually, making it difficult to adapt to the distribution of buildings, streets, obstructions, and reflection paths in specific deployment scenarios; second, the multi-level scanning process suffers from error propagation problems, where errors in upper-level beam identification directly affect the scanning results of subsequent levels; third, multi-level beam interaction introduces a non-negligible training latency, limiting its application in fast beam alignment scenarios.
[0004] Another type of existing technology utilizes auxiliary information for beam alignment, such as inferring the serving beam corresponding to the target user equipment based on sub-6GHz channel characteristics, terminal location information, visual images, radar perception information, or multimodal fusion information. This type of method can utilize additional environmental information to improve the accuracy of serving beam selection and beam alignment success rate in some scenarios, but it usually relies on additional sensors, additional positioning functions, cross-modal synchronization, or additional signaling interaction, increasing system deployment costs and engineering complexity.
[0005] A similar approach is the learning-based probe codebook design method tailored to specific deployment scenarios. This type of method uses training data to learn a small number of probe beams and infers the final serving beam index based on the received power of the reference signal obtained from these probe beams, thus completing beam alignment with fewer probes than the number of candidate serving beams. However, existing learning-based probe codebook methods typically assume that the analog phase shifter has continuous phase or sufficiently high phase resolution. In reality, high-frequency communication analog beamforming front-ends often use 2-bit, 3-bit, or other low-resolution phase shifters to reduce cost, power consumption, and hardware complexity. Due to the difference in hardware phase resolution, the probe beams learned under the continuous phase assumption cannot be directly deployed into low-resolution phase shifter hardware.
[0006] If the continuous phase detector codebook is directly quantized into low-resolution phase after training, it can easily cause distortion of the detector beam pattern, resulting in a shift in the distribution of detector power measurement features. This leads to inconsistencies between the input distribution received by the serving beam prediction network during deployment and training, resulting in a decrease in beam alignment accuracy. If the phase parameters to be learned are forcibly restricted to values from the discrete phase set at the beginning of training, and iterative optimization is directly performed within this discrete phase parameter space, the training process is prone to getting stuck in local optima or experiencing convergence instability due to the high dimensionality of the discrete phase parameter space, the non-convexity of the objective function, and the non-differentiability of the discrete quantization mapping. Conventional rounding quantization operations are also difficult to directly incorporate into the gradient backpropagation process.
[0007] Therefore, the core technical problem to be solved in this application is: under the hardware constraint of analog beamforming using low-resolution phase shifters in base stations, how to perform end-to-end joint optimization of a small number of probe beam codebooks and service beam prediction networks so that the obtained probe beams simultaneously satisfy constant mode constraints, unity power constraints and discrete phase constraints, and can utilize the propagation characteristics of specific deployment scenarios to complete beam alignment of target user equipment with fewer probes than the number of candidate service beams. Summary of the Invention
[0008] This invention addresses the problem of limited phase resolution of phase shifters in simulated beamforming hardware devices in practical millimeter-wave systems by proposing a beam alignment method and apparatus based on quantized sensing training, which can directly meet the constraints of low-resolution phase shifters.
[0009] To achieve the above objectives, the present invention provides the following technical solution:
[0010] The beam alignment method based on quantized perception training proposed in this invention includes a training preparation stage, an offline training stage, and an online deployment stage.
[0011] The training preparation phase is used to configure system and beam codebook parameters, construct a trainable simulated detector layer and a service beam prediction network, and construct training samples. The trainable simulated detector layer introduces discrete phase constraints from a low-resolution phase shifter, performs periodization and discrete quantization on the potential phase, and generates a detector beam that satisfies constant modulus and unity power constraints through a complex exponential method. The training samples are constructed by determining the optimal service beam label corresponding to each training sample based on the training channel matrix and the measurement results of the candidate service beams.
[0012] The offline training phase is used to perform offline quantization sensing training. The input probe beam power vector is used to jointly train the probe beam codebook and the service beam prediction network. The final probe beam codebook matrix and the trained beam prediction network parameters are saved, so that the trained probe beam can be directly adapted to the actual hardware and used for online beam alignment. The probe beam power vector is calculated based on the training channel matrix and the probe beam codebook matrix.
[0013] The online deployment phase obtains power measurement results through the final quantization of the probe beam codebook and uses the service beam prediction network to determine the service beam of the target user.
[0014] Furthermore, the system parameters include a candidate service beam codebook, the number of candidate service beams, the number of base station antenna elements, the number of detector beams, and the number of phase shifter quantization bits. The candidate service beam codebook is preset by the communication system. The number of candidate service beams is determined by the number of candidate service beam codewords contained in the candidate service beam codebook. The candidate service beam codebook contains a certain number of candidate service beam codewords. The number of base station antenna elements is determined by the base station antenna array hardware configuration. The number of detector beams is preset by the base station or network side according to the beam training overhead requirements. The number of phase shifter quantization bits is determined by the hardware specifications of the low-resolution phase shifter in the base station's simulated beamforming front end. The system calculates the phase quantization step size and constructs a discrete phase set based on the number of phase shifter quantization bits.
[0015] Furthermore, the trainable simulated detection layer includes a latent phase parameter module, a phase constraint processing module, a probe beam generation module, and a probe power calculation module. The latent phase parameter module is used to store trainable latent phase parameters; the phase constraint processing module is used to sequentially perform phase periodization, phase quantization, progressive annealing, and stopping gradient processing on the latent phase parameters; the probe beam generation module is used to generate complex probe beams that satisfy constant modulus constraints and unity power constraints based on the processed phase parameters; and the probe power calculation module is used to calculate the probe power corresponding to each probe beam based on the training channel matrix and the complex probe beams.
[0016] Furthermore, the progressive annealing and stopping gradient processing method is as follows: the probe beam gradually transitions from a continuous phase state to a fully quantized state, and the gradient is transferred to the potential phase parameters during backpropagation.
[0017] Furthermore, the detection beam power is calculated and processed as follows: the detection beam power is averaged using multiple resources and then log-centered to obtain the input data for the service beam prediction network.
[0018] Furthermore, the joint training method of the serving beam prediction network is as follows: the serving beam prediction result is output based on the preprocessed probe beam power, and the probe layer and prediction network parameters are jointly updated based on the prediction result and the optimal serving beam label.
[0019] Furthermore, the final quantization of the probe beam codebook is generated as follows: after training, the system performs final quantization on the potential phase and converts the obtained probe beam codebook into a low-resolution phase shifter control word.
[0020] Secondly, the present invention provides a beam alignment device based on quantization-sensing training, comprising the following modules for implementing the method described in any of the above-mentioned embodiments:
[0021] A trainable simulated detector layer includes a latent phase parameter module, a phase constraint processing module, a detector beam generation module, and a detector power calculation module. The latent phase parameter module stores trainable latent phase parameters. The phase constraint processing module sequentially performs phase periodization, phase quantization, progressive annealing, and stopping gradient processing on the latent phase parameters. The detector beam generation module generates complex detector beams that satisfy constant modulus constraints and unity power constraints based on the processed phase parameters. The detector power calculation module calculates the detector power corresponding to each detector beam based on the training channel matrix and the complex detector beams.
[0022] The storage module is used to store the final detection beam codebook matrix, the low-resolution phase shifter control word corresponding to the final detection beam codebook matrix, and the trained beam prediction network parameters. The final detection beam codebook matrix is generated based on the final quantized phase matrix. The final quantized phase matrix is obtained by uniformly quantizing the final periodicized phase matrix element by element according to the phase quantization step size. The final periodicized phase matrix is obtained by periodically processing the trained potential phase matrix element by element.
[0023] The prediction module is used to load the final detection beam codebook matrix, the low-resolution phase shifter control word corresponding to the final detection beam codebook matrix, and the trained beam prediction network parameters, and to calculate the target user candidate service beam prediction probability vector.
[0024] Thirdly, the present invention provides an electronic device, including a processor and a memory, wherein the memory stores program instructions, characterized in that the program instructions, when executed by the processor, implement the method described in any of the above-mentioned embodiments.
[0025] Fourthly, the present invention provides a computer-readable storage medium storing program instructions that, when executed by a processor, implement the method described in any of the preceding claims.
[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0027] 1. This invention provides a beam alignment method based on quantization-sensing training. During the probe beam training process, discrete phase constraints from a low-resolution phase shifter are directly introduced, enabling the trained probe beam to adapt to actual hardware. Compared to methods that quantize the continuous phase probe codebook after training, this invention reduces beam pattern distortion and performance degradation caused by quantization and supports direct deployment of the probe beam codebook.
[0028] 2. This invention combines stopping gradient processing and progressive annealing strategies to incorporate the non-differentiable phase quantization process into end-to-end training, and gradually transitions the probe beam from a continuous phase state to a fully quantized state. Compared with methods that use completely discrete phase from the start of training, this invention can alleviate problems such as gradient non-propagation, local optima, and unstable training convergence.
[0029] 3. This invention generates the probe beam using a complex exponential method and fixes the amplitude of each antenna element, ensuring that the probe beam naturally satisfies constant mode and unity power constraints. Therefore, this invention reduces additional beam projection, power normalization, and hardware calibration operations, which helps to reduce the difference between the model output and the actual simulated beamforming hardware.
[0030] 4. This invention jointly learns a probe beam codebook and a service beam prediction network based on channel data from the target deployment scenario, enabling the probe beam to utilize propagation paths, occlusion relationships, and reflection characteristics within the specific scenario. Compared to manually designed fixed codebooks and hierarchical codebooks, this invention improves the probe beam's adaptability to scenario characteristics.
[0031] 5. This invention only needs to send a pre-set number of probe beams to determine the serving beam based on the probe beam power fed back by the target user, without needing to traverse all candidate serving beams online. Therefore, this invention can reduce the number of beam scans, reference signal overhead, and beam alignment delay.
[0032] 6. This invention uses the probe beam power as the main input to the serving beam prediction network online, without relying on additional sensors such as cameras, radar, and positioning devices. Simultaneously, logarithmic centralization reduces the impact of overall path loss and signal amplitude variations. Therefore, this invention can reduce system deployment complexity and improve adaptability under different user locations and link gain conditions.
[0033] 7. This invention generates corresponding discrete phase levels and probe beam codebooks by setting different quantization bit numbers for phase shifters, and converts the training results into phase shifter control words for storage and loading. Therefore, this invention can adapt to low-resolution phase shifters with different phase resolutions and reduce the computational and configuration overhead during online deployment. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0035] Figure 1 A flowchart of a beam alignment method based on quantization-sensing training provided in an embodiment of the present invention. Detailed Implementation
[0036] To better understand this technical solution, the technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described examples are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art based on this application are within the scope of protection of the present invention.
[0037] The core technical problem to be solved in this application is: under the hardware constraint of analog beamforming using low-resolution phase shifters in base stations, how to perform end-to-end joint optimization of a small number of probe beam codebooks and service beam prediction networks so that the obtained probe beams simultaneously satisfy constant mode constraints, unity power constraints, and discrete phase constraints, and can utilize the propagation characteristics of specific deployment scenarios to complete beam alignment of target user equipment with fewer probes than the number of candidate service beams.
[0038] To address the aforementioned core technical issues, this application further aims to resolve the following problems: First, how to embed the discrete phase quantization process into the training process so that the probe beam generated during the training phase can reflect the actual phase constraints of the low-resolution phase shifter; second, how to use gradient approximation methods to solve the problem of gradient propagation difficulties caused by the non-differentiability of discrete quantization operations; third, how to avoid being directly constrained by discrete phase at the beginning of training, thereby alleviating the problems of local optima and convergence instability in the discrete phase parameter space; and fourth, how to narrow the beam alignment performance gap between the continuous phase ideal model, the model quantized after training, and the actual low-resolution hardware-deployable model.
[0039] To address the limited phase resolution of phase shifters in simulated beamforming hardware in practical millimeter-wave systems, this invention designs a probe beam codebook learning method and an online beam alignment process that can directly satisfy the constraints of low-resolution phase shifters. Furthermore, to solve the problems of beam pattern distortion caused by quantization of continuous-phase probe codebooks after training, non-differentiability of discrete-phase quantization operations, and optimization difficulties and convergence instability caused by directly applying discrete-phase constraints for optimization in the initial training stage, this invention proposes a quantization-sensing beam alignment method for low-resolution phase shifters in millimeter-wave communication systems.
[0040] like Figure 1 As shown, the technical solution of the present invention includes a training preparation stage, an offline training stage, and an online deployment stage. The detailed steps are described below.
[0041] 1. Training Preparation Phase
[0042] Step 1: Configure system and codebook parameters
[0043] The base station pre-configures a candidate serving beam codebook, denoted as: Candidate service beam codebook Pre-set by the communication system. The system acquires the number of candidate serving beams, denoted as: Number of candidate service beams From candidate service beam codebook The number of candidate service beam codewords included is determined, and the number of candidate service beams is... satisfy Candidate service beam codebook This includes the number of candidate service beams. The first candidate service beam codeword. The system will... The candidate service beam codewords are denoted as follows: ,in, Candidate service beam codeword It is a complex vector. Its length is equal to the number of base station antenna elements, denoted as . Candidate service beam codebook Discrete Fourier transform narrow beam codebooks, standard protocol codebooks, or other beam codebooks pre-configured by the communication system can be used.
[0044] The system obtains the number of base station antenna elements. This value is determined by the base station antenna array hardware configuration and typically ranges from 8 to 256. The system acquires the number of detection beams, denoted as: Number of detection beams The beam training overhead is preset by the base station or network side according to the requirements, and the value usually ranges from 2 to 32, and the number of detection beams... satisfy The system obtains the number of quantization bits of the phase shifter, denoted as: Phase shifter quantization bits The value is determined by the hardware specifications of the low-resolution phase shifter in the base station's analog beamforming front-end, typically ranging from 1 to 6. The system uses the number of quantization bits of the phase shifter as the basis for its value. The phase quantization step size is calculated and denoted as: Phase quantization step size Calculate using the following formula:
[0045]
[0046] The system quantizes the number of bits based on the phase shifter. Construct the discrete phase set as follows: Discrete phase set Determine using the following formula:
[0047]
[0048] Step 2: Construct the detection layer and prediction network
[0049] The system constructs a trainable simulated detection layer and a service beam prediction network. The trainable simulated detection layer includes a latent phase parameter module, a phase constraint processing module, a probe beam generation module, and a probe power calculation module. The latent phase parameter module stores trainable latent phase parameters; the phase constraint processing module sequentially performs phase periodization, phase quantization, progressive annealing, and stopping gradient processing on the latent phase parameters; the probe beam generation module generates complex probe beams that satisfy constant mode constraints and unity power constraints based on the processed phase parameters; and the probe power calculation module calculates the probe power corresponding to each probe beam based on the training channel matrix and the complex probe beams.
[0050] During the offline training phase, the trainable simulated probe layer takes the training channel matrix and the annealing factor corresponding to the current training round as inputs and outputs the probe beam power vector; the phase processing, probe beam generation and probe power calculation processes are described in detail in subsequent steps 4 and 5, respectively.
[0051] The system sets the potential phase matrix, denoted as: Potential phase matrix The dimension is The system will use the potential phase matrix The Middle Line number The elements of a column are denoted as: ,in, , The system has a latent phase matrix. Initialization is performed, and the initialization method is preset by the training device, including random initialization or loading of pre-trained phase parameters of the continuous phase detection model.
[0052] The system sets the total number of training rounds, denoted as: Total number of training rounds Pre-set by the training equipment, typically ranging from 10 to 200. The system sets the number of annealing cycles, denoted as: Number of annealing rounds The values are preset by the training equipment and typically range from 1 to the total number of training rounds. And satisfy The system sets the training batch size, denoted as: Training batch size Pre-set by the training equipment, typically ranging from 16 to 1024. The system sets a numerical stability constant, denoted as: Numerical stability constant Pre-set by the training equipment, typically with a value range of [value range missing]. to The system sets the initial learning rate, denoted as: Set the full-quantization fine-tuning learning rate, denoted as: Initial learning rate The typical value range is: to Fully quantized fine-tuning of the learning rate The typical value range is: to The system configures an optimizer, which is pre-configured by the training device, including the Adam optimizer, the AdamW optimizer, or the stochastic gradient descent optimizer.
[0053] Step 3: Construct training samples
[0054] Before offline training begins, the system collects a training sample set in the target's real-world communication scenario, denoted as: The system obtains the number of training samples, denoted as: Number of training samples The value is determined by the scale of on-site data collection, and typically ranges from [value range missing]. to Training sample set include Training samples.
[0055] For the The training user device, the first Each training user equipment transmits an uplink channel sounding reference signal on a preset uplink channel sounding resource; the base station receives the uplink channel sounding reference signal through an antenna array and performs channel estimation based on the known reference signal sequence to obtain the first... The training channel matrix of the training samples is denoted as: ,in, .
[0056] The number of channel sampling resources acquired by the system is denoted as: Number of channel sampling resources The value is determined by the amount of frequency domain resources, time domain resources, or measurement resources configured during training, and typically ranges from [value range missing]. to . No. Training channel matrix of training samples The dimension is Training the channel matrix The Middle Line number The element of the column represents the first element. The base station antenna element in the first Complex channel estimates on a preset channel sampling resource, where... , After the base station completes the uplink and downlink channel reciprocity calibration, it will... Training channel matrix of training samples This serves as channel characterization data for calculating the power of the probe beam.
[0057] Base stations sequentially use candidate serving beam codebooks All of them The candidate service beam codeword, directed to the first The training user equipment transmits the service beam reference signal; the... Each training user device measures and reports with the above. The candidate service beam codewords correspond to The system calculates the measurement values of each candidate service beam according to the candidate service beam codeword in the candidate service beam codebook. The order of arrangement in the above The candidate service beam measurements are arranged to obtain the nth... The candidate service beam measurement vectors of the training samples are denoted as: . No. Candidate service beam measurement vectors of training samples The Middle Each element is denoted as: ,in, .
[0058] The system measures the candidate service beam vector. Determine the first The optimal serving beam label for each training sample is denoted as: ,have
[0059]
[0060] During training, the input probe beam power vector of the serving beam prediction network is based on the training channel matrix. The codebook matrix of the detection beam is calculated. The system will then... Training channel matrix of training samples With optimal service beam tag Corresponding storage, forming the first The training samples. Each training sample is represented as:
[0061]
[0062] 2. Offline training phase
[0063] Step 4: Offline Quantization Perception Training
[0064] The system records the current training round as: ,in, In the current training round In the middle, the system has a latent phase matrix After periodization, a periodic phase matrix is obtained, denoted as: The system will periodize the phase matrix. The Middle Line number The elements of a column are denoted as: Periodic phase Calculate using the following formula:
[0065]
[0066] The system is based on the phase quantization step size For the periodized phase matrix Uniform quantization is performed to obtain the quantization phase matrix, denoted as: The system will quantize the phase matrix. The Middle Line number The elements of a column are denoted as: Quantization phase Calculate using the following formula:
[0067]
[0068] in, This indicates rounding operations. The system calculates the rounding value based on the current training round. and number of annealing rounds Calculate the annealing factor, denoted as: Annealing factor Calculate using the following formula:
[0069]
[0070] The system is based on the periodic phase matrix Quantization phase matrix and annealing factor Calculate the mixed phase matrix, denoted as: The system will mix the phase matrix. The Middle Line number The elements of a column are denoted as: Mixed phase Calculate using the following formula:
[0071]
[0072] The system sets the stopping gradient operator, denoted as: The system is based on the hybrid phase matrix. Construct the output phase matrix, denoted as: The system will output a phase matrix. The Middle Line number The elements of a column are denoted as: Output phase Calculate using the following formula:
[0073]
[0074] The system based on the output phase matrix Generate the probe beam codebook matrix, denoted as: Detection beam codebook matrix Including the number of detection beams The system will detect the beam vector. The probe beam vector is denoted as: ,in, Detection beam codebook matrix Calculate using the following formula:
[0075]
[0076] Among them, the imaginary unit For the complex number computation constants used by the system, This represents an element-wise complex exponential mapping.
[0077] Step 5: Calculate and process the detection power characteristics
[0078] The system starts from the training sample set Select the training batch size The training sample. For the th training sample in the training batch The system, based on the training sample, ... Training channel matrix for each training sample and detection beam codebook matrix , generate the first The detection beam power vector of each training sample is denoted as: . No. The detection beam power vector of each training sample The Middle The power value of each detection beam is denoted as: The system will... Training channel matrix for each training sample The Middle The training channel sampling vectors are denoted as: The detection beam codebook matrix The Middle The probe beam vector is denoted as: . No. Each detection beam power value Calculate using the following formula:
[0079]
[0080] The system for the first The detection beam power vector of each training sample Perform log-centering to obtain the th Each training sample is input as a probe beam power vector, denoted as: . No. Input the probe beam power vector from each training sample The Middle The power values of the input probe beams are denoted as: It is calculated using the following formula:
[0081]
[0082] Step 6: Input the detected features into the prediction model
[0083] The system will Input the probe beam power vector from each training sample Input the beam prediction network to obtain the first... The logical score vectors of the candidate service beams for each training sample are denoted as: It is represented as:
[0084]
[0085] No. Each training sample candidate service beam logic score vector The Middle The logical score of each candidate service beam is denoted as: The system calculates the first... The probability vector of candidate service beam prediction for each training sample is denoted as: . No. A candidate service beam prediction probability vector for each training sample The Middle The prediction probability values for each candidate service beam are denoted as: It is calculated using the following formula:
[0086]
[0087] Step 7: Calculate the loss and update the model
[0088] The system is based on the first in the training batch Optimal service beam label for each training sample and the A candidate service beam prediction probability vector for each training sample Calculate the training loss function, denoted as: Training loss function Calculate using the following formula:
[0089]
[0090] In the current training round At that time, the system uses an initial learning rate. ; When the current training round At that time, the system uses full quantization to fine-tune the learning rate. The system uses the training loss function. Calculate the gradient and update the latent phase matrix. Beam prediction network parameters In the current training round At that time, the system will enable the current training round. Increase by 1, and repeat steps 4 through 7; in the current training round Training ends when the time is right. After training, the system obtains the latent phase matrix, denoted as: And obtain the trained beam prediction network parameters, denoted as: .
[0091] Step 8: Save the probe codebook and prediction network
[0092] The system performs training on the latent phase matrix. After performing element-wise periodization, the final periodic phase matrix is obtained, denoted as: Final periodized phase matrix The Middle Line number The elements of a column are denoted as: The calculation formula is as follows:
[0093]
[0094] in, Represents the latent phase matrix after training. The Middle Line number The elements of the column.
[0095] The system is based on the phase quantization step size For the final periodized phase matrix Element-wise uniform quantization is performed to obtain the final quantized phase matrix, denoted as: Final quantization phase matrix The Middle Line number The elements of a column are denoted as: The calculation formula is as follows:
[0096]
[0097] in, For step 1, the number of quantization bits based on the phase shifter A defined phase quantization step size, This indicates rounding operations.
[0098] The system is based on the final quantized phase matrix The final detection beam codebook matrix is generated, denoted as: Final detection beam codebook matrix include The final detection beam vector is calculated using the following formula:
[0099]
[0100] The base station will eventually detect the beam codebook matrix. Converted to low-resolution phase shifter control words. Final detection beam codebook matrix. The corresponding low-resolution phase shifter control word is used for transmitting the detection reference signal during the online deployment phase.
[0101] The system saves the final detection beam codebook matrix. Final detection beam codebook matrix The corresponding low-resolution phase shifter control word, and the trained beam prediction network parameters. The data is stored in the base station's local storage, the network-side training server, the model management server, or other storage media that can be read by the base station.
[0102] 3. Online Deployment Phase
[0103] Step 9: Load the probe codebook and prediction network
[0104] Before online deployment, the base station loads the final detection beam codebook matrix from the storage medium. Final detection beam codebook matrix The corresponding low-resolution phase shifter control word, and the trained beam prediction network parameters. The base station configures the low-resolution phase shifter control word to the analog beamforming front end and applies the trained beam prediction network parameters. Load it into the beam prediction network.
[0105] Step 10: Perform online beam detection
[0106] During the online beam alignment phase, the base station sequentially uses the final probe beam codebook matrix. In Each final detection beam vector sends a detection reference signal to the target user equipment. The target user equipment measures and reports accordingly. The target user's probe beam power value. The base stations are arranged according to the final probe beam vector transmission order. The target user's detected beam power value is used to obtain the target user's detected beam power vector, denoted as: , of which The target user detection beam power value is denoted as: .
[0107] Step 11: Process online features and make predictions
[0108] Base station detects beam power vector of target user Perform the same log-centering process as in step 5 to obtain the target user input probe beam power vector, denoted as: The base station will input the target user's probe beam power vector. The input is loaded with the parameters of the trained beam prediction network. The beam prediction network is used, and the target user candidate service beam prediction probability vector is calculated according to step 6, denoted as: , of which The probability values for predicting the candidate service beam for each target user are denoted as: .
[0109] Step 12: Select serving beam and transmit
[0110] The base station predicts the probability vector based on the candidate serving beam for the target user. Select the target user prediction service beam index, denoted as: Target User Prediction Service Beam Index Predicting probability vectors for candidate service beams for target users The index corresponding to the largest element in the middle, that is:
[0111]
[0112] Base station predicts service beam index based on target user From the candidate service beam codebook Select candidate service beam codeword and the candidate service beam codeword Data transmission service beam for target users.
[0113] The above description is merely a detailed explanation of preferred embodiments and principles of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention, based on the ideas provided by the present invention, should be considered within the scope of protection of the present invention.
Claims
1. A beam alignment method based on quantization-sensing training, characterized in that: It includes the training preparation phase, the offline training phase, and the online deployment phase; The training preparation phase is used to configure system and beam codebook parameters, construct a trainable simulated detector layer and a service beam prediction network, and construct training samples. The trainable analog detector layer introduces discrete phase constraints from a low-resolution phase shifter, periodizes and discretely quantizes the potential phase, and generates a detector beam that satisfies constant modulus and unity power constraints through a complex exponential method. The training samples are constructed by determining the optimal service beam label corresponding to each training sample based on the training channel matrix and the measurement results of the candidate service beams. The offline training phase is used to perform offline quantization sensing training, input the probe beam power vector to jointly train the probe beam codebook and the service beam prediction network, and save the final probe beam codebook matrix and the trained beam prediction network parameters, so that the trained probe beam can be directly adapted to the actual hardware and used for online beam alignment. The probe beam power vector is calculated based on the training channel matrix and the probe beam codebook matrix; The online deployment phase obtains power measurement results through the final quantization of the probe beam codebook and uses the service beam prediction network to determine the service beam of the target user.
2. The beam alignment method based on quantization-sensing training according to claim 1, characterized in that, System parameters include candidate service beam codebook, number of candidate service beams, number of base station antenna elements, number of detector beams, and number of phase shifter quantization bits; The candidate service beam codebook is pre-set by the communication system; The number of candidate service beams is determined by the number of candidate service beam codewords contained in the candidate service beam codebook; the candidate service beam codebook contains the number of candidate service beam codewords; the number of base station antenna elements is determined by the base station antenna array hardware configuration; the number of detection beams is preset by the base station or network side according to the beam training overhead requirements. The number of quantization bits for the phase shifter is determined by the hardware specifications of the low-resolution phase shifter in the base station's analog beamforming front end. The system calculates the phase quantization step size and constructs a discrete phase set based on the number of quantization bits for the phase shifter.
3. The beam alignment method based on quantization-sensing training according to claim 1, characterized in that, The trainable simulated detection layer includes a latent phase parameter module, a phase constraint processing module, a detection beam generation module, and a detection power calculation module. The latent phase parameter module is used to store trainable latent phase parameters. The phase constraint processing module is used to sequentially perform phase periodization, phase quantization, progressive annealing, and stopping gradient processing on the latent phase parameters. The detection beam generation module is used to generate a complex detection beam that satisfies constant mode constraints and unity power constraints based on the processed phase parameters. The detection power calculation module is used to calculate the detection power corresponding to each detection beam based on the training channel matrix and the complex detection beams.
4. The beam alignment method based on quantization-sensing training according to claim 3, characterized in that, The progressive annealing and stopping gradient processing method is to gradually transition the probe beam from a continuous phase state to a fully quantized state, and to transfer the gradient to the potential phase parameters during backpropagation.
5. The beam alignment method based on quantization-sensing training according to claim 3, characterized in that, The detection beam power calculation and processing method is as follows: the detection beam power is averaged by multiple resources and log-centered to obtain the input data of the service beam prediction network.
6. The beam alignment method based on quantization-sensing training according to claim 1, characterized in that, The joint training method for the service beam prediction network is as follows: output the service beam prediction result based on the preprocessed probe beam power, and jointly update the probe layer and prediction network parameters based on the prediction result and the optimal service beam label.
7. The beam alignment method based on quantization-sensing training according to claim 1, characterized in that, The final quantization of the probe beam codebook is generated as follows: after training, the system performs final quantization on the potential phase and converts the obtained probe beam codebook into a low-resolution phase shifter control word.
8. A beam alignment device based on quantization-sensing training, characterized in that, The following modules are included for implementing the method according to any one of claims 1-7: The trainable simulated detector layer includes a latent phase parameter module, a phase constraint processing module, a detector beam generation module, and a detector power calculation module. The latent phase parameter module is used to store trainable latent phase parameters. The phase constraint processing module is used to sequentially perform phase periodization, phase quantization, progressive annealing, and stopping gradient processing on the latent phase parameters. The detector beam generation module is used to generate a complex detector beam that satisfies constant mode constraints and unity power constraints based on the processed phase parameters. The detection power calculation module is used to calculate the detection power corresponding to each detection beam based on the training channel matrix and the complex detection beams. The storage module is used to store the final detection beam codebook matrix, the low-resolution phase shifter control word corresponding to the final detection beam codebook matrix, and the trained beam prediction network parameters. The final detection beam codebook matrix is generated based on the final quantized phase matrix. The final quantized phase matrix is obtained by uniformly quantizing the final periodicized phase matrix element by element according to the phase quantization step size. The final periodicized phase matrix is obtained by periodically processing the trained latent phase matrix element by element. The prediction module is used to load the final detection beam codebook matrix, the low-resolution phase shifter control word corresponding to the final detection beam codebook matrix, and the trained beam prediction network parameters, and to calculate the target user candidate service beam prediction probability vector.
9. An electronic device, characterized in that, The method includes a processor and a memory, wherein the memory stores program instructions, and the program instructions, when executed by the processor, implement the method according to any one of claims 1-7.
10. A computer-readable storage medium storing program instructions, characterized in that, When the program instructions are executed by the processor, they implement the method described in any one of claims 1-7.
