A deep learning-based beam training and measurement pattern joint generation method

Abstract

This invention proposes a deep learning-based method for joint generation of beam training and measurement patterns, belonging to the field of communication technology. The method includes: acquiring a full codebook beam measurement quality dataset; performing self-supervised pre-training on the dataset using a high-ratio random masking strategy to obtain a general beam prediction model; inputting complete beam measurement data into the model and calculating importance scores through backpropagation; iteratively selecting the beam measurement points with the highest importance scores using a greedy selection strategy based on a given beam training overhead constraint; updating the importance scores of the remaining beams after each round of selection by penalizing them based on the spatial correlation between the selected beams and the remaining beams; stopping the iteration and outputting the final optimal beam measurement pattern when the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, guiding the base station to send the corresponding reference signal. It possesses strong generalization capabilities and can flexibly adapt to various measurement configurations and scenario changes.

Description

Technical Field

[0001] This invention belongs to the field of communication technology, specifically relating to a method for joint generation of beam training and measurement patterns based on deep learning. Background Technology

[0002] In high-frequency communication bands such as millimeter waves and terahertz, as well as in very large-scale antenna array systems, transmission based on high-gain directional beams is fundamental to overcoming high path loss and achieving high-speed communication. Beam training, the process of finding the optimal beam pair through measurement, is the primary and critical step in ensuring link performance.

[0003] As antenna apertures increase and communication scenarios shift from the far field to the near field, wavefronts exhibit spherical wave characteristics. This necessitates expanding beam training from traditional one-dimensional angle searches to a two-dimensional joint search of angle and distance, resulting in an exponential increase in the number of beams (codewords) that need to be measured. While exhaustive beam scanning and measurement based on a complete two-dimensional codebook can ensure theoretically optimal performance, the enormous time, signaling, and energy overhead it generates makes it unsuitable for meeting the low latency and high energy efficiency requirements of practical systems.

[0004] Existing deep learning-based beam prediction methods directly use fixed, uniform sampling patterns for partial measurements, without generating adaptive optimal observation patterns tailored to real-time channel characteristics and overhead constraints. Because the preset patterns are decoupled from specific scenarios, key feature information is lost under low-overhead conditions, and the model cannot flexibly adapt to dynamically changing measurement configurations. This rigid mechanism not only limits the theoretical upper limit of beam selection accuracy but also significantly increases system deployment costs due to the need to retrain the model for scene changes, severely restricting the practical effectiveness of intelligent beam management. Summary of the Invention

[0005] The present invention aims to at least partially solve one of the technical problems in the related art.

[0006] Therefore, the first objective of this invention is to propose a method for joint generation of beam training and measurement patterns based on deep learning.

[0007] The second objective of this invention is to propose a device for the joint generation of beam training and measurement patterns based on deep learning.

[0008] The third objective of this invention is to provide a computer device.

[0009] A fourth objective of this invention is to provide a non-transitory computer-readable storage medium.

[0010] To achieve the above objectives, a first aspect of the present invention proposes a method for joint generation of beam training and measurement patterns based on deep learning, comprising:

[0011] A full codebook beam measurement quality dataset is obtained, and a high-ratio random masking strategy is used to perform self-supervised pre-training on the full codebook beam measurement quality dataset to obtain a general beam prediction model that can reconstruct the complete beam spatial measurement quality based on sparse observations. The complete beam measurement data is input into the trained general beam prediction model, and the importance score characterizing the measurement value of each beam is calculated through backpropagation. Based on the given beam training overhead constraint, a greedy selection strategy is adopted to iteratively select the beam measurement points with the highest importance score and add them to the optimal pattern set. After each round of selection, the importance score of the remaining beams is updated by penalty based on the spatial correlation between the selected beams and the remaining beams. When the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, the iteration stops and the final optimal beam measurement pattern is output to guide the base station to send only the reference signal corresponding to the optimal beam measurement pattern.

[0012] In one embodiment of the present invention, the step of acquiring a full codebook beam measurement quality dataset and performing self-supervised pre-training on the full codebook beam measurement quality dataset using a high-ratio random masking strategy to obtain a general beam prediction model capable of reconstructing the complete beam spatial measurement quality based on sparse observations includes: The base station transmits a reference signal corresponding to the full codebook, and the user receives and measures all transmitted beams to construct a dataset containing beam identifiers and corresponding measurement quality. Apply a high proportion of random masking to the input data in the dataset, replace the masked locations with mask identifiers, and generate training samples containing only partial beam measurement information. The deep learning model is self-supervised and trained using the training samples, with the optimization objective of reconstructing the measurement quality of the complete beam space, to obtain the general beam prediction model.

[0013] In one embodiment of the present invention, the step of inputting complete beam measurement data into the trained general beam prediction model, calculating the influence of each beam measurement point on the prediction result through backpropagation, and obtaining an importance score characterizing the value of each beam measurement includes: The complete beam measurement data without masking is input into the general beam prediction model, and backpropagation is performed to obtain the gradient of each element of the input layer; The absolute value of the gradient is extracted as the importance score. The larger the absolute value of the gradient, the greater the influence of the corresponding beam measurement point on the optimal beam prediction and the higher its importance.

[0014] In one embodiment of the present invention, the step of iteratively selecting the beam measurement points with the highest importance scores and adding them to the optimal pattern set using a greedy selection strategy based on a given beam training overhead constraint, and updating the importance scores of the remaining beams with penalties based on the spatial correlation between the selected beams and the remaining beams after each round of selection, includes: Initialize an empty set of optimal patterns; In each iteration, the point with the highest importance score is selected from the remaining beams and added to the set, and the cosine similarity between the codeword corresponding to that point and all codewords in the full codebook is calculated; The importance score of the remaining beam is updated with a penalty using the cosine similarity; this process is repeated until the overhead constraint is met.

[0015] In one embodiment of the present invention, it further includes: The user reports the optimal beam measurement pattern to the base station; The base station transmits a corresponding number of reference signals according to the beam order in the pattern, wherein the number of transmitted reference signals is less than the total number of beams in the full codebook.

[0016] In one embodiment of the present invention, it further includes: The user measures the quality of the transmitted beam based on the optimal beam measurement pattern and constructs a sparse input vector with a length equal to the number of beams in the full codebook, where the beam positions in the pattern are filled with the measured values ​​and the remaining positions are filled with mask identifiers. The sparse input vector is input into a locally pre-trained general beam prediction model for inference, and the predicted quality value for all beams is output; the user selects at least one beam identifier with the best predicted quality and reports it to the base station.

[0017] In one embodiment of the present invention, the user's selection of at least one beam identifier with the best prediction quality and reporting it to the base station includes: Users use the reference signal received power or signal-to-interference-to-noise ratio as the quality evaluation criteria, sort all the beam prediction quality values ​​output by the model, select at least one beam identifier with the highest ranking as the best predicted beam, and report it to the base station through the uplink channel. The base station determines the transmission beam based on the reported information and instructs the user on that beam.

[0018] To achieve the above objectives, a second aspect of the present invention provides a deep learning-based beam training and measurement pattern joint generation apparatus, comprising: The dataset construction and pre-training module is used to obtain a full codebook beam measurement quality dataset and to perform self-supervised pre-training on the full codebook beam measurement quality dataset using a high-ratio random masking strategy, so as to obtain a general beam prediction model that can reconstruct the complete beam spatial measurement quality based on sparse observations. The importance score calculation module is used to input complete beam measurement data into the trained general beam prediction model and calculate the importance score characterizing the measurement value of each beam through backpropagation. The pattern generation module is used to iteratively select the beam measurement points with the highest importance scores and add them to the optimal pattern set based on a given beam training overhead constraint and a greedy selection strategy. After each round of selection, the importance scores of the remaining beams are updated with penalties based on the spatial correlation between the selected beams and the remaining beams. The pattern output module is used to stop iterating and output the final optimal beam measurement pattern when the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, so as to guide the base station to send only the reference signal corresponding to the optimal beam measurement pattern.

[0019] The method and apparatus for jointly generating beam training and measurement patterns based on deep learning according to an embodiment of the present invention have the following beneficial effects: Achieving higher beam prediction accuracy with low overhead: Under the same measurement overhead, this method can significantly improve the accuracy of beam selection by performing beam prediction based on the configuration of the optimal beam measurement pattern, and requires fewer measurements to achieve the same performance.

[0020] The trained model is adaptable to various measurement configurations and can be deployed flexibly: the beam prediction model trained based on random masks does not depend on any specific, fixed measurement mode, and can be directly used to generate different optimal patterns under different measurement overhead requirements without retraining. It can also adapt to different scene changes. This significantly reduces the cost and complexity of model training and maintenance, and improves the practicality and scalability of the system.

[0021] This invention provides a systematic method for joint optimization of beam training: it organically integrates two key aspects—deep learning model training and beam measurement resource allocation. A general model is obtained through random mask pre-training, and then this model is used to obtain the optimal measurement pattern, forming a complete joint design process.

[0022] To achieve the above objectives, a third aspect of this application provides a computer device comprising a processor and a memory; wherein the processor runs a program corresponding to the executable program code stored in the memory, for implementing a deep learning-based beam training and measurement pattern joint generation method as described in the first aspect embodiment.

[0023] To achieve the above objectives, a fourth aspect of this application provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements a deep learning-based method for joint generation of beam training and measurement patterns as described in the first aspect embodiment.

[0024] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0025] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for jointly generating beam training and measurement patterns based on deep learning according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the interaction between a base station and a user module according to an embodiment of the present invention; Figure 3 A schematic diagram of the overall process of a deep learning-based beam training and beam measurement pattern configuration joint design method provided in an embodiment of the present invention; Figure 4 This invention provides a schematic diagram of the near-field beam space codebook (angle-range domain) ID for a transmitted beam, as provided in an embodiment of the invention. Figure 5 This invention provides a pre-training process for a deep learning model based on a random mask, as exemplified by this invention. Figure 6 This is a schematic diagram of an optimal measurement beam pattern generation process provided by an embodiment of the present invention; Figure 7 A schematic diagram illustrating the inference process of a beam prediction model provided in an embodiment of the present invention; Figure 8 This is a structural diagram of a beam training and measurement pattern joint generation device based on deep learning according to an embodiment of the present invention; Figure 9 It is a computer device according to an embodiment of the present invention. Detailed Implementation

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions 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 embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0028] The following description, with reference to the accompanying drawings, describes an adaptive optimal beam measurement pattern generation method and apparatus based on random masking according to an embodiment of the present invention.

[0029] Example 1 Figure 1 This is a flowchart of a method for jointly generating beam training and measurement patterns based on deep learning according to an embodiment of the present invention, such as... Figure 1 As shown, it includes: S1. Obtain the full codebook beam measurement quality dataset and perform self-supervised pre-training on the full codebook beam measurement quality dataset using a high-ratio random masking strategy to obtain a general beam prediction model that can reconstruct the complete beam spatial measurement quality based on sparse observations. S2, Input the complete beam measurement data into the trained general beam prediction model, and calculate the importance score characterizing the measurement value of each beam through backpropagation; S3, based on the given beam training overhead constraint, adopts a greedy selection strategy to iteratively select the beam measurement points with the highest importance score and add them to the optimal pattern set, and after each round of selection, the importance score of the remaining beams is updated by penalty according to the spatial correlation between the selected beams and the remaining beams. S4. When the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, the iteration stops and the final optimal beam measurement pattern is output to guide the base station to send only the reference signal corresponding to the optimal beam measurement pattern.

[0030] An embodiment of the present invention provides a method for joint generation of beam training and measurement patterns based on deep learning, which can adaptively generate the optimal beam measurement pattern according to arbitrary overhead constraints, significantly reducing training latency and resource consumption while improving beam prediction accuracy; moreover, the pre-trained model does not depend on a fixed sampling mode, has strong generalization ability, and can flexibly adapt to various measurement configurations and scene changes.

[0031] Example 2 The following describes in detail, with reference to the accompanying drawings, an embodiment of the present invention, a method for jointly generating beam training and measurement patterns based on deep learning.

[0032] This invention proposes a systematic technical solution comprising three core stages: "offline model pre-training," "adaptive optimal beam measurement pattern generation," and "online low-overhead application." Figure 2 This is a schematic diagram illustrating the interaction between the base station and the user module.

[0033] The first stage, the offline model pre-training stage, aims to obtain a deep learning model with strong generalization ability that does not depend on any fixed measurement pattern. Its core is to adopt a self-supervised learning strategy with a high proportion of random masks, thereby training a backbone model that is universally robust to random incomplete observations.

[0034] The second stage, the adaptive optimal beam measurement pattern generation stage, aims to dynamically generate the optimal measurement pattern for any given measurement overhead constraint. Its core is an algorithm of "importance-based greedy selection + correlation penalty" to find the beam measurement pattern with the best performance.

[0035] The third stage involves online low-overhead applications, applying the results of the first two stages to actual beam scanning. The base station only needs to send a small amount of reference signals according to the generated optimal pattern. Users can then use a pre-trained general model to predict the globally optimal beam with high accuracy based on sparse measurements, thereby significantly reducing training overhead while ensuring performance.

[0036] To enable those skilled in the art to better understand the objectives, technical solutions, and advantages of this invention, the invention will be further described clearly and in detail below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are merely some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0037] refer to Figure 3 The following is a schematic diagram of the overall process of a deep learning-based beam training and beam measurement pattern configuration joint design method provided by an embodiment of the present invention. The steps are as follows: Step 1: The user measures all transmitted beams and builds a beam measurement quality dataset to train the model for beam prediction.

[0038] Step 101: The base station sends the reference signal corresponding to the full codebook to the user within a certain period of time, providing complete beam scanning information.

[0039] Step 102: The user determines the scanning and measurement sequence of the transmitted beam.

[0040] In one embodiment, the user receives and measures all transmitted beams, and sorts all transmitted beams in order to form a specific beam sequence table.

[0041] In one embodiment, reference Figure 4 This invention provides a sequence table of transmitted beams in a near-field codebook. The table is grouped according to range rings. Under each range ring, the transmitted beams in all angular directions are traversed to form a transmitted beam ID. Each ID corresponds to a transmitted beam with a specific angle and distance. At the same time, each beam ID corresponds to a beam measurement quality.

[0042] Step 103: The user measures the reference signals carried in all beams transmitted by the base station and obtains the measurement results for all beams.

[0043] In one embodiment, beam quality measurement is based on a CSI-RS reference signal. During the base station's transmit beam scanning process, the base station sends a CSI-RS reference signal to the user. The user measures the reference signal and obtains the beam quality in the transmit beam direction by measuring the signal quality of the reference signal.

[0044] In one embodiment, the user selects L1-RSRP or L1-SINR as the criterion for evaluating the quality of the reference signal.

[0045] Step 104: The user constructs a beam measurement quality dataset from all beam measurements.

[0046] Step 105: The full codebook measurement data collected in this stage is used for offline training of a general-purpose model. The purpose is to allow the model to learn the complete structure and inherent laws of the beam space. After actual system deployment, this full scan step is usually no longer required.

[0047] Step 2: The deep learning model is pre-trained on the input using a high proportion of random masks. The learning objective of the model is to reconstruct the beam quality of the complete beam space.

[0048] Furthermore, step 2 includes the following steps, see reference. Figure 5 The present invention provides a pre-training process for a deep learning model based on a random mask, comprising the following steps: Step 201: The user processes all transmitted beam measurement quality information collected over a period of time to form a model training dataset.

[0049] In one embodiment, the user processes the beam measurement quality information to form a dataset that can be used for model training and derivation. The dataset includes measurement timestamps, a table of transmitted beam IDs, and the measurement quality corresponding to the transmitted beam IDs.

[0050] In one embodiment, the measurement timestamp represents the time when the user performs beam measurement, the transmitted beam ID table represents the relative position of each beam pair among all the user's beam pairs, and the measurement quality corresponding to the transmitted beam ID represents the reference signal measured by the user on each beam pair.

[0051] Step 202: The input is masked with a configurable high-ratio random mask. Different mask ratios represent different degrees of information loss, and the corresponding training overhead is (1 - mask ratio) of the full codebook beam scan.

[0052] Applying a high proportion of random masking to the input data is equivalent to randomly masking most of the data during training, requiring the model to reconstruct all information using only a very small portion of the data. This training method forces the model to learn deep, non-local correlation features in the beam space, rather than simply memorizing them.

[0053] In one embodiment, it is assumed that the user's measurement beam pairs total N S samples are used, and the user randomly masks the dataset according to a configurable masking ratio. Assuming the masking ratio is ρ, and each position has an equal probability of being masked, then each sample contains N samples. S The training overhead for (1-ρ) beam measurement information is (1-ρ) times the full codebook beam scan.

[0054] In one embodiment, the numerical value at the masked location is replaced with a specific mask representation (such as 0 or a learnable vector) upon input.

[0055] In one embodiment, for each different mask ratio, the user trains a beam prediction neural network model; for example, the user can train beam prediction models with mask ratios of 80% and 95%. When the base station allows very low beam scanning overhead (e.g., 3%–10% of full codebook scanning), a model with a mask ratio of 95% can be used for prediction. When the base station allows high beam scanning overhead (e.g., 15%–30% of full codebook scanning), a model with a mask ratio of 80% can be used for prediction.

[0056] Step 203: The deep learning model determines the input length, model structure, and model parameters based on the input data. The learning objective of the model is to reconstruct complete beam spatial measurement quality information.

[0057] In one embodiment, the user recovers the measurement quality of all transmitted beams based on the measurement quality of a portion of the transmitted beams.

[0058] In one embodiment, the model structure of the training model can be determined as follows.

[0059] In one embodiment, the deep learning model can employ a Transformer or a Convolutional Neural Network (CNN). The specific hyperparameters of the model (such as the number of layers and nodes) can be adjusted according to the actual data scale and complexity. Its core design principle is to effectively process sequential or spatially structured data and complete the mapping task from parts to the whole. The design of the model's input structure must fully consider its ability to process sequential data and its computational efficiency.

[0060] For determining the inter-layer connection method, you can refer to the following: if the hidden layer and the input layer are fully connected, the activation function can be the ReLU function; if the hidden layers are fully connected, the activation function can be the ReLU function; if the hidden layer and the output layer are fully connected, the activation function can be the ReLU function.

[0061] The process of determining the loss function can employ methods such as mean squared error (MSE) loss function or mean absolute error (MAE) loss function.

[0062] For determining the hyperparameters of the network model, you can refer to setting the number of learning epochs to T. The setting of the number of learning epochs needs to consider the impact on model training speed, training cost, and model training accuracy; the learning rate is set to le-3; and random weight initialization is selected as the weight initialization method.

[0063] In one embodiment, the user calculates the training loss value based on the actual measurement data of the transmitted beam, for example, using the mean square error loss function:

[0064] Where represents the amount of data used for model training, y i This represents the actual measured value of data i. This represents the output result of the model after data i is processed.

[0065] After steps 1 and 2, corresponding to the first stage of offline training, the description of the deep learning-based beam training model and dataset generation in this invention is complete.

[0066] Step 3: After the beam prediction model is trained, freeze the model parameters, input the complete dataset into the model, and obtain the importance scores of the input points.

[0067] In one embodiment, the backpropagation gradient matrix of each element in the input layer can be used to model the importance score of the codeword. The absolute value of each element in the gradient matrix intuitively reflects the importance of that input location to the model's correct prediction. The larger the absolute value of the gradient at a beam measurement point, the greater the impact of the measurement at that point on the model's correct prediction of the optimal beam, and therefore its higher importance. Specifically, for any beam measurement point... Its importance score The calculation formula is as follows:

[0068] in, Indicates the beam measurement point Importance score Indicating in the evaluation dataset Mathematical expectation on This represents the reconstruction loss function (such as mean square error loss, MSE) of the general beam prediction model when predicting the complete beam space. Indicates the location of the beam measurement point in the input model. The corresponding beam measurement value (such as the amplitude of the reference signal). Represents the loss function For a specific input measurement value The partial derivatives (i.e., the gradients calculated by backpropagation).

[0069] In one embodiment, an unmasked training set in the current scene can be used as input to a trained model for backpropagation to obtain offline codeword importance scores.

[0070] In one embodiment, the base station can update the importance score online periodically or based on changes in channel statistical characteristics using a small number of full-scan samples. The user receives and measures all transmitted beams, inputs the measurements of all transmitted beams into a trained model for backpropagation, and obtains the online codeword importance score.

[0071] Step 4: Based on the importance scores obtained in Step 3, the adaptive pattern generation algorithm is executed. Given the training overhead, the optimal pattern set is selected through greedy selection and relevance penalty.

[0072] Furthermore, step 4 includes the following steps, see reference. Figure 6 This is a schematic diagram of an optimal measurement beam pattern generation process provided by an embodiment of the present invention, which includes the following steps: Step 401: Determine the beam training overhead and initialize the optimal pattern set.

[0073] In one embodiment, the beam training overhead can be delegated from the base station.

[0074] In one embodiment, the beam training overhead can be determined by the user based on a comprehensive consideration of the model performance under different mask probabilities in step 2.

[0075] Step 402: Select the most important point and add the corresponding transmit beam ID to the optimal pattern set, and calculate the similarity between the codeword corresponding to the transmit beam ID and all codewords in the full codebook.

[0076] In one embodiment, the importance score is measured by the absolute value of the gradient, and the point with the largest absolute value of the gradient is the most important point. Its corresponding beam ID is calculated from the shape input in step 2.

[0077] In one embodiment, the similarity between the codeword corresponding to the transmitted beam ID and all codewords in the full codebook can be calculated using cosine similarity. Specifically, let the codeword vector corresponding to the currently selected most important point be... Any remaining candidate codeword vector in the full codebook is The similarity between the two The calculation formula is: | , in, This represents the conjugate transpose operation of a matrix. This represents the modulus (absolute value) of a complex number. Since the codeword vectors (i.e., beam steering vectors) in the codebook are usually normalized, the modulus of this inner product is equivalent to the cosine similarity between two codewords in space. In the codebook, the cosine similarity of a codeword with itself is 1, and the cosine similarity with other codewords is between [0,1). Similarity measures the correlation between codewords; the higher the correlation, the less information it provides.

[0078] Step 403: Apply a similarity penalty to the importance score, update the importance matrix, and proceed to the next iteration.

[0079] If two beams are spatially very close, their correlation is high. Therefore, a significant portion of the information obtained from measuring one beam is redundant with that obtained from measuring the other. To obtain the maximum non-redundant information within limited overhead, after selecting an important location, the importance scores of other highly correlated locations need to be reduced, allowing for the exploration of more spatially informative locations in subsequent selections.

[0080] In one embodiment, the penalty update for importance scores employs an iterative decay mechanism based on spatial correlation. Specifically, during the iteration process, it is assumed that the beam measurement points newly selected into the optimal pattern set are... For any remaining candidate beam points that were not selected , let the feature correlation (i.e., spatial similarity) between the two be . If this remaining point has an importance score in the current iteration round of , then the updated importance score has an iterative update expression as follows: .

[0081] Among them, is a preset redundancy suppression penalty factor used to control the penalty strength for highly correlated beams. This complete iterative formula shows that the higher the correlation between the remaining beam and the selected beams, the greater the attenuation amplitude of its importance score; by adjusting the size of , the amount of information and spatial diversity of the beam points in the pattern can be flexibly balanced, so as to effectively explore points with more independent information in space in subsequent selections.

[0082] Step 404, repeat the above selection, calculation, and update steps until the optimal measurement beam pattern set is constructed after meeting the training overhead determined in step 302, and the loop ends. Based on this optimal measurement beam pattern, the model can make the most accurate global prediction.

[0083] After steps 3 and 4, corresponding to the optimal beam measurement pattern generation stage in the second phase, the determination, selection, and generation of the optimal measurement pattern for the important points in this invention are completed.

[0084] Step 5, the user reports the corresponding optimal beam measurement pattern to the base station.

[0085] In one embodiment, the optimal beam measurement pattern stores a series of codeword IDs corresponding to the transmit beams.

[0086] Step 6, the base station sends a reference signal according to the optimal measurement pattern during the beam scanning stage.

[0087] In one embodiment, the base station sends the CSI-RS reference signal in the corresponding time slot according to the codeword order in the pattern. Assuming that the traditional full codebook scan needs to send B = N×S different reference signals, in this invention, only L (L << B) reference signals need to be sent in this stage, significantly reducing the air interface resource consumption and scanning delay.

[0088] Step 7, the user makes an optimal beam prediction based on the sparse beam measurement values and reports the prediction result to the base station.

[0089] Further, step 7 includes the following steps. Refer to Figure 7 , which is a schematic flowchart of the inference process of a beam prediction model provided by an embodiment of this invention, including the following steps: Step 701: The user performs quality measurements on a portion of the transmitted beams based on the determined optimal beam measurement pattern to obtain the measurement quality of the portion of the transmitted beams, such as L1-RSRP.

[0090] Step 702: Beam prediction is performed on the user side using the trained model.

[0091] In one embodiment, the user constructs a sparse input vector (or matrix) with a length equal to the total number of codewords based on a determined optimal beam measurement pattern. Let the index corresponding to the full codebook beam set be... ( (Total number of beams), the set of beam indices corresponding to the optimal beam measurement pattern is: The sparse input constructed The Each element can be represented as:

[0092] in, Indicates the position of the beam The quality of the reference signal obtained from actual measurements (such as received amplitude or RSRP); This represents the mask identifier (which is usually set to a numerical value in practice). Using the above formula, the positions corresponding to the codewords in the pattern are filled with the actual measured values, while other unmeasured positions are masked. Subsequently, this sparse input... Input a locally pre-trained beam prediction model, which outputs prediction quality values ​​for all codewords based on the sparse observations.

[0093] Step 703: The user reports the IDs of one or more beams with the best predicted measurement quality to the base station.

[0094] In one embodiment, model inference is performed on the user side. The user then reports at least one transmit beam ID (reference signal) with the best predicted measurement quality to the base station. The base station records the predicted reference signal ID reported by the user.

[0095] Example 3 To achieve the above embodiments, such as Figure 8 As shown, this embodiment also provides a beam training and measurement pattern joint generation device 10 based on deep learning. The device 10 includes a dataset construction and pre-training module 100, an importance score calculation module 200, a pattern generation module 300, and a pattern output module 400.

[0096] The dataset construction and pre-training module 100 is used to obtain a full codebook beam measurement quality dataset and to perform self-supervised pre-training on the full codebook beam measurement quality dataset using a high-ratio random masking strategy to obtain a general beam prediction model that can reconstruct the complete beam spatial measurement quality based on sparse observations. Importance score calculation module 200 is used to input complete beam measurement data into the trained general beam prediction model and calculate the importance score characterizing the measurement value of each beam through backpropagation; The pattern generation module 300 is used to iteratively select the beam measurement points with the highest importance scores and add them to the optimal pattern set based on a given beam training overhead constraint and a greedy selection strategy. After each round of selection, the importance scores of the remaining beams are updated with penalties based on the spatial correlation between the selected beams and the remaining beams. The pattern output module 400 is used to stop the iteration and output the final optimal beam measurement pattern when the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, so as to guide the base station to send only the reference signal corresponding to the optimal beam measurement pattern.

[0097] Furthermore, the aforementioned dataset construction and pre-training module 100 is also used for: The base station transmits a reference signal corresponding to the full codebook, and the user receives and measures all transmitted beams to construct a dataset containing beam identifiers and corresponding measurement quality. Apply a high proportion of random masking to the input data in the dataset, replace the masked locations with mask identifiers, and generate training samples containing only partial beam measurement information. The deep learning model is self-supervised and trained using the training samples, with the optimization objective of reconstructing the measurement quality of the complete beam space, to obtain the general beam prediction model.

[0098] Furthermore, the aforementioned importance score calculation module 200 is also used for: The complete beam measurement data without masking is input into the general beam prediction model, and backpropagation is performed to obtain the gradient of each element of the input layer; The absolute value of the gradient is extracted as the importance score. The larger the absolute value of the gradient, the greater the influence of the corresponding beam measurement point on the optimal beam prediction and the higher its importance.

[0099] Furthermore, the above-mentioned drawing generation module 300 is also used for: Initialize an empty set of optimal patterns; In each iteration, the point with the highest importance score is selected from the remaining beams and added to the set, and the cosine similarity between the codeword corresponding to that point and all codewords in the full codebook is calculated; The importance score of the remaining beam is updated with a penalty using the cosine similarity; this process is repeated until the overhead constraint is met.

[0100] An embodiment of the present invention provides a beam training and measurement pattern joint generation device based on deep learning, which can adaptively generate the optimal beam measurement pattern according to arbitrary overhead constraints, significantly reducing training latency and resource consumption while improving beam prediction accuracy; moreover, the pre-trained model does not depend on a fixed sampling mode, has strong generalization ability, and can flexibly adapt to various measurement configurations and scene changes.

[0101] To implement the methods of the above embodiments, the present invention also provides a computer device, such as... Figure 9 As shown, the computer device 600 includes a memory 601 and a processor 602; wherein, the processor 602 reads executable program code stored in the memory 601 to run a program corresponding to the executable program code, so as to implement the various steps of the method described above.

[0102] To implement the above embodiments, this application also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method described in the foregoing embodiments.

[0103] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0104] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for joint generation of beam training and measurement patterns based on deep learning, characterized in that, include: A full codebook beam measurement quality dataset is obtained, and a high-ratio random masking strategy is used to perform self-supervised pre-training on the full codebook beam measurement quality dataset to obtain a general beam prediction model that can reconstruct the complete beam spatial measurement quality based on sparse observations. The complete beam measurement data is input into the trained general beam prediction model, and the importance score characterizing the measurement value of each beam is calculated through backpropagation. Based on the given beam training overhead constraint, a greedy selection strategy is adopted to iteratively select the beam measurement points with the highest importance score and add them to the optimal pattern set. After each round of selection, the importance score of the remaining beams is updated with a penalty based on the spatial correlation between the selected beams and the remaining beams. When the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, the iteration stops and the final optimal beam measurement pattern is output to guide the base station to send only the reference signal corresponding to the optimal beam measurement pattern.

2. The method as described in claim 1, characterized in that, The process of acquiring a full codebook beam measurement quality dataset and performing self-supervised pre-training on the dataset using a high-ratio random masking strategy to obtain a general beam prediction model capable of reconstructing complete beam spatial measurement quality from sparse observations includes: The base station transmits a reference signal corresponding to the full codebook, and the user receives and measures all transmitted beams to construct a dataset containing beam identifiers and corresponding measurement quality. Apply a high proportion of random masking to the input data in the dataset, replace the masked locations with mask identifiers, and generate training samples containing only partial beam measurement information. The deep learning model is self-supervised and trained using the training samples, with the optimization objective of reconstructing the measurement quality of the complete beam space, to obtain the general beam prediction model.

3. The method as described in claim 1, characterized in that, The process involves inputting complete beam measurement data into the trained general beam prediction model, calculating the influence of each beam measurement point on the prediction result through backpropagation, and obtaining an importance score characterizing the value of each beam measurement, including: The complete beam measurement data without masking is input into the general beam prediction model, and backpropagation is performed to obtain the gradient of each element of the input layer; The absolute value of the gradient is extracted as the importance score. The larger the absolute value of the gradient, the greater the influence of the corresponding beam measurement point on the optimal beam prediction and the higher its importance.

4. The method as described in claim 1, characterized in that, Based on a given beam training overhead constraint, a greedy selection strategy iteratively selects the beam measurement points with the highest importance scores to add to the optimal pattern set. After each round of selection, the importance scores of the remaining beams are updated with a penalty based on the spatial correlation between the selected beams and the remaining beams, including: Initialize an empty set of optimal patterns; In each iteration, the point with the highest importance score is selected from the remaining beams and added to the set, and the cosine similarity between the codeword corresponding to that point and all codewords in the full codebook is calculated; The importance score of the remaining beam is updated with a penalty using the cosine similarity; this process is repeated until the overhead constraint is met.

5. The method according to claim 1, characterized in that, Also includes: The user reports the optimal beam measurement pattern to the base station; The base station transmits a corresponding number of reference signals according to the beam order in the pattern, wherein the number of transmitted reference signals is less than the total number of beams in the full codebook.

6. The method according to claim 5, characterized in that, Also includes: The user measures the quality of the transmitted beam based on the optimal beam measurement pattern and constructs a sparse input vector with a length equal to the number of beams in the full codebook, where the beam positions in the pattern are filled with the measured values ​​and the remaining positions are filled with mask identifiers. The sparse input vector is input into a locally pre-trained general beam prediction model for inference, and the predicted quality values ​​for all beams are output. The user selects at least one beam identifier with the best prediction quality and reports it to the base station.

7. The method as described in claim 6, characterized in that, The user selects at least one beam identifier with the best prediction quality and reports it to the base station, including: Users use the reference signal received power or signal-to-interference-to-noise ratio as the quality evaluation criteria, sort all the beam prediction quality values ​​output by the model, select at least one beam identifier with the highest ranking as the best predicted beam, and report it to the base station through the uplink channel. The base station determines the transmission beam based on the reported information and instructs the user on the beam.

8. A device for jointly generating beam training and measurement patterns based on deep learning, characterized in that, include: The dataset construction and pre-training module is used to obtain a full codebook beam measurement quality dataset and to perform self-supervised pre-training on the full codebook beam measurement quality dataset using a high-ratio random masking strategy, so as to obtain a general beam prediction model that can reconstruct the complete beam spatial measurement quality based on sparse observations. The importance score calculation module is used to input complete beam measurement data into the trained general beam prediction model and calculate the importance score characterizing the measurement value of each beam through backpropagation. The pattern generation module is used to iteratively select the beam measurement points with the highest importance scores and add them to the optimal pattern set based on a given beam training overhead constraint and a greedy selection strategy. After each round of selection, the importance scores of the remaining beams are updated with penalties based on the spatial correlation between the selected beams and the remaining beams. The pattern output module is used to stop iterating and output the final optimal beam measurement pattern when the number of beam measurement points in the optimal pattern set reaches the beam training overhead constraint, so as to guide the base station to send only the reference signal corresponding to the optimal beam measurement pattern.

9. A computer device, characterized in that, Including processor and memory; The processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to implement the deep learning-based beam training and measurement pattern joint generation method as described in any one of claims 1-7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a method for joint generation of beam training and measurement patterns based on deep learning as described in any one of claims 1-7.

Images