An ai-based cross-dataset cross-scenario robustness beamforming design method

Abstract

The application discloses an AI-based cross-dataset cross-scene robustness beamforming design method, channel data is collected in different scenes, a beamforming network is established, channel vectors in the channel vectors of different scenes are selected respectively to be mixed, the mixed channel data vectors are divided into support sources and query sources, channel vectors in the query sources and the support sources are extracted each time to form a task, and multiple tasks are repeated to form a training batch; the channel vectors in the support sources are input into a pre-deeply copied network to be updated to obtain a transition model, data in the query sources are input into the transition model, a loss is obtained, and the beamforming network is back propagated; after multiple times of training, a model with the highest final spectral efficiency is taken as a production model to obtain a beamforming vector. Compared with a traditional model, the method provided by the application not only has robustness in a single scene, but also is flexible and adaptive to different physical scenes, training is convenient, and the method has cross-scene robustness.

Description

Technical Field

[0001] This invention relates to the fields of communication and machine learning technology, and in particular to an AI-based robust beamforming design method across datasets and scenarios. Background Technology

[0002] In recent years, beamforming technology for massive MIMO antenna arrays has been widely applied and is widely recognized as an important component of 5G and the upcoming 6G. Therefore, various techniques have been used to improve its performance, such as deep learning, advanced optimization algorithms, and so on. While the performance of deep learning has been quite attractive in many previous research scenarios, its performance often degrades rapidly when the environment or dataset changes. Therefore, designing efficient beamforming networks with strong robustness remains an open problem in intelligent wireless communication. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by providing an AI-based robust beamforming design method that spans multiple datasets and scenarios.

[0004] The technical solution adopted in this invention is as follows: A robust beamforming design method based on AI across datasets and scenarios, the method comprising the following steps:

[0005] S1. Channel data is collected in different scenarios. M scenario channel vectors are collected for each scenario, where each scenario channel vector includes N... r One user channel vector;

[0006] S2. Establish a beamforming network: This network consists of an image input layer, a flattening layer, a feature extraction layer, and a custom function layer. The input is the scene channel vector and the signal-to-noise ratio, and the output is the same number of beamforming vectors as the user.

[0007] S3. Constructing Beamforming Network Training Data: Select channel vectors from channel vectors in different scenarios and mix them. Divide the mixed channel data vector into support source and query source. Each time, extract channel vectors from the query source and support source to form a task, repeating N times. b A training batch consists of N tasks, with each task containing N tasks. s Channel vectors of the support sources and N q The channel vector of the query source;

[0008] S4. Training the beamforming network: Initialize the network parameters, perform a deep copy of the beamforming network parameters, input the channel vectors from the support sources into the deep copy network for forward and backward propagation, update the transition model, then input the channel vectors from the query sources into the transition model, obtain the loss, and perform backward propagation on the beamforming network.

[0009] S5. Train the beamforming network step times, and finally test the obtained step+1 models on the test set of the hybrid channel. Take the model with the highest spectral efficiency as the production model, and save the corresponding training hybrid channel data as a new training data source.

[0010] S6. Input the user's channel vector and signal-to-noise ratio into the production model, and output the beamforming vector.

[0011] Furthermore, the number of different scenarios is two. When the number of different scenarios is greater than two, for each additional scenario, the data of the newly added scenario and the original corresponding training mixed channel data are used as the data of the two scenarios. During training, the beamforming network parameters inherit the parameters when no scenario is added, until the channel data of the scenario to be deployed are used.

[0012] Furthermore, in the beamforming network, the loss function is calculated from the output beamforming vector and the signal-to-noise ratio.

[0013] Furthermore, the loss function is specifically as follows:

[0014]

[0015] Where R is the spectral efficiency obtained from the i-th query source sample and the k-th user. k,i (v k Specifically:

[0016]

[0017] Where, σ 2 This refers to the variance of the circular complex Gaussian noise, where the channel vector for the k-th user is h. k The beamforming vector is v k Specifically:

[0018]

[0019] Where P represents the constraint power magnitude; the vector output of the last random deactivation layer is represented as...

[0020] Furthermore, in S3, the percentages of data entries for different scenarios are q1 and q2, with q1 set to 0 and q2 set to 1 during the initial training.

[0021] Furthermore, the support source and the query source are obtained by dividing the mixed channel data vector in half.

[0022] Furthermore, the number of transition models obtained in S4 is the same as the batch size; the number of losses obtained by inputting the channel vector in the query source into the transition model is also the same as the batch size; and the loss for backpropagation of the beamforming network is the average of the losses of all transition models.

[0023] Furthermore, in S5, the training steps of the beamforming network are specifically defined as follows: each training step increases q1 by step_size and decreases q2 by step_size; where step*step_size = 1.

[0024] Furthermore, when the S5 model reaches the stage of deployment, it needs online fine-tuning. Small samples that can quickly adapt are collected in the deployment scenarios and input into the model to adjust the parameters. adapt After the second update, the final model is obtained, where N adapt This is a hyperparameter.

[0025] On the other hand, the present invention also provides an AI-based cross-dataset and cross-scene robust beamforming system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it is used to implement an AI-based cross-dataset and cross-scene robust beamforming design method.

[0026] Beneficial Effects: This invention provides an AI-based robust beamforming design method across datasets and scenarios. The trained AI model exhibits robustness across datasets and scenarios. Even when the dataset or environment changes, this invention can still achieve optimized spectral efficiency and is easy to train and extend. The special training method proposed in this invention makes the beamforming network highly sensitive to input distribution, allowing for fine-tuning with fewer samples and fewer updates to adapt to different scenarios. Adding support for a new scenario only requires sampling new channel data for that scenario, recalculating, and retraining, enabling widespread deployment and application in reality. Attached Figure Description

[0027] Figure 1 It is a flowchart of the overall beamforming design;

[0028] Figure 2 It is a diagram of the designed neural network structure;

[0029] Figure 3 This is a flowchart of a neural network training method;

[0030] Figure 4 This is the simulation verification and comparison results when q1=0;

[0031] Figure 5 This is the simulation verification and comparison results when q1=1;

[0032] Figure 6 This is the simulation verification and comparison results when q1 = 0.5. Detailed Implementation

[0033] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0034] This invention provides an AI-based robust beamforming design method across datasets and scenarios. The core of this invention lies in the design of the beamforming network, a special training method, and the mixing of training data from multiple scenarios. The beamforming network directly outputs beamforming vectors and employs a self-supervised learning mechanism, eliminating the need to pre-calculate the optimal beamforming vector. The proposed special training method makes the beamforming network highly sensitive to input distribution, allowing for fine-tuning with fewer samples and fewer updates to adapt to different scenarios. The mixing of training data from multiple scenarios enables the model to perceive the channel characteristics of various application scenarios, and selecting the optimal data ratio helps the model perform well across multiple scenarios. The AI-based robust beamforming design method provided by this invention produces an AI model with robustness across datasets and scenarios. When the dataset or environment changes, the system still achieves optimized spectral efficiency and is easy to train and expand. Adding support for a new scenario only requires sampling new channel data in that scenario, recalculating, and training, enabling widespread deployment in reality. The overall beamforming flowchart is shown below. Figure 1 As shown, the specific training and deployment methods are as follows:

[0035] Step 1, first test N in both scenarios. r Channel data is collected from each user, and the channel vector for each scenario includes N. r M user channel vectors are collected for each scenario, and M channel vectors are collected in other scenarios where this technology needs to be deployed.

[0036] Step 2, design the beamforming network, the specific architecture of which is as follows: Figure 2As shown. The input includes the scene channel vector and the signal-to-noise ratio (SNR), and the output is the beamforming vectors of the same number of users. The loss function is then calculated from the beamforming vectors and the SNR. In addition to the traditional structure, the neural network substructure should also include a custom output layer. The traditional structure, besides the input layer, should contain multiple identical sub-modules connected in series to extract features. The input layer mainly separates and concatenates the real and imaginary parts of the received channel vector. For example, if the dimension before splitting is (batchsize, n1, n2), a complex vector, then the dimension after splitting is (batchsize, 2, n1, n2), a real vector. There is also a flattening layer, resulting in a flattened dimension of (batchsize, 2*n1*n2), where n1 is the total number of users and n2 is the number of antennas. The identical sub-modules are sequentially composed of a fully connected layer, a linear rectifier unit layer, a batch normalization layer, and a random deactivation layer connected in series. The custom output layer uses a direct output beamforming vector method, primarily to constrain the beamforming vector power. Given N... r For a given number of users, the vector output by the last random deactivation layer can be represented as: The final output beamforming vector can be represented as The purpose of the custom function layer is to output v from u. P represents the constrained power magnitude, and the beamforming vector for the k-th user is...

[0037]

[0038] Assume the channel vector of the k-th user is h k Beamforming is v k Assume that the spectral efficiency obtained from the i-th query source sample and the k-th user is R. k,i (v k ), can be represented as

[0039]

[0040] Where σ 2 If we refer to the variance of the circular complex Gaussian noise, then the output loss function can be expressed as:

[0041]

[0042] Where N q This refers to the number of query source channel data samples input into the neural network at one time. This loss function takes the negative of the average spectral efficiency. It does not require pre-calculated beamforming vectors to compare and calculate the loss function, saving computational resources and time, and allowing the neural network to learn how to output the optimal beamforming vector on its own.

[0043] The designed neural network training model is robust within a single scene. Then, through training with channel data sampled from multiple scenes in a certain proportion and special training methods, cross-scene robustness and rapid adaptability are achieved.

[0044] Step 3, Initial training of the beamforming network: as follows Figure 3 As shown, the first training iteration utilizes the channel vectors and signal-to-noise ratio information from the first two scenarios in step 1, along with the neural network designed in step 2. This training does not require labels, falling under the category of self-supervised learning. The loss function is simply the negative of the objective function, and it automatically outputs optimized beamforming vectors.

[0045] Before training, the training data needs to be structured appropriately: First, select a certain proportion of channel vectors from each of the two scenarios, resulting in a total of M channel vectors, with the percentage of data points from each scenario being q1 and q2. In the initial training, set q1 to 0 and q2 to 1. Then, divide the mixed channel data vectors in half again, with one half called the support source and the other half called the query source. Assume the batch size is N. b Then, N will be randomly selected from the support source. s N channel vectors, and N randomly selected from the query source. q Channel vectors, repeated N times b Each task contains N times. s +N q Channel vectors, N b These tasks together form a batch during training. There are a total of K such batches. In each batch, the channel vector of the i-th task belonging to the support source is denoted as SS(i), and the channel vector of the i-th task belonging to the query source is denoted as QS(i).

[0046] Next, using the data structure defined above and the network architecture from step two, we will proceed as follows: Figure 2 The training process shown here, specifically this step is as follows:

[0047] (3.1) First, randomly initialize the parameters of the neural network model designed in step two, and then perform N... b N deep copies correspond to N... b This is a task. The neural network copied here is just an intermediate step, designed to calculate intermediate gradients, and will be destroyed after a batch of training is completed.

[0048] (3.2) Input the SS channel vector corresponding to each task into the deep copy of the neural network, and perform N... u The forward and backward propagation is performed twice to update the parameters of this deep copy model. Then, the QS channel vector of each task is input into the corresponding updated copy of the neural network, and the i-th task obtains the corresponding loss Loss(i).

[0049] (3.3) Take the average of each loss obtained in (3.2) and use it as the true loss to update the parameters of the original neural network model. Then destroy the copied N... b Each neural network model stores the parameters of the original neural network model before copying, which are then used as the initialization parameters for the next batch of models.

[0050] (3.4) Each batch is trained according to the process from (3.1) to (3.3) until K batches are trained, which completes one epoch. After multiple epochs, training ends and the model is saved.

[0051] Step 4: Train the model using the same method as in Step 3, increasing the step size for each training iteration (q1) and decreasing it for each iteration (q2); where step * step size = 1. Finally, test these + 1 models on a mixed-channel test set. Select the model with the highest spectral efficiency as the production model, and save the corresponding mixed-channel training data as a new training data source. The mixed-channel test set can be a proportional mix of small samples from the two scenarios in Step 1, ensuring balanced model performance across both scenarios. This unique training method makes the neural network highly sensitive to input distribution, giving it strong adaptability and small-sample learning capabilities, facilitating fine-tuning and adaptation for online deployment.

[0052] If step_size is 0.1, then step is 10. q1 will be trained according to the following ratio [0, 0.1, 0.2, 0.3, ..., 0.9, 1], where the ratio 0 for q1 is already trained in S4, and the subsequent steps are done additionally.

[0053] Step 5: When there are more than two scenarios N to be deployed, for each additional scenario, the channel data of the new scenario and the original training mixed channel data are used as the data for the two scenarios. The training is performed according to Step 3 and Step 4. The model parameters inherit the parameters when no scenarios are added, and the channel data of the scenarios to be deployed are used.

[0054] The proportion of each dataset is not determined all at once, but rather gradually. When there are more than two datasets, it gradually expands from mixing channel data from the two datasets to mixing multiple datasets, avoiding the huge search space and training cost of determining the optimal proportion all at once.

[0055] Step 6, the model from Step 5 reaches the stage where it needs to be deployed online, requiring online fine-tuning. In the deployment scenario, data including M... o A small sample of channel vectors is input into the model, and the parameters are processed by N. adapt This update yields the final model. N adapt This is a hyperparameter, and it is generally selected to be 5 to 10 times.

[0056] The user's channel vector and signal-to-noise ratio are input into the production model, and the output is the beamforming vector.

[0057] In this embodiment, spectral efficiency tests were conducted under different signal-to-noise ratios. Two datasets were tested: DeepMIMO and WAIR-D. Two scenarios, O1_60 and scenario 2, were selected from these datasets. DeepMIMO_n represents the nth batch of 1000 users between rows 1000 and 1300 in the O1_60 scenario of DeepMIMO. WAIR-D_n represents the nth case in scenario 2 of the WAIR-D dataset. The simulation is set to N. r =1, M=1000, σ 2 =1, the number of antennas is set to 64, the communication bandwidth is set to 50MHz, and the signal-to-noise ratio is between -20dB and 20dB, then Figure 4 , Figure 5 and Figure 6 Simulation results for q1=0, q1=1, and q1=0.5 are presented respectively.

[0058] Figure 4 and Figure 5 This demonstrates the robustness of the proposed model within the dataset; its test performance is similar to its training set performance on the same type of test set. Figure 6 After selecting an appropriate mixing ratio, the model's robustness across datasets was demonstrated, with its performance across datasets approaching that of the training set. Here, the horizontal axis represents the signal-to-noise ratio (SNR) of the input channel vector, and the vertical axis represents the average spectral efficiency of the test samples; as the SNR increases, the spectral efficiency increases across all test samples. Figure 4 and Figure 5 For performance evaluation of models trained on a single dataset, the test performance and spectral efficiency are similar on the same type of test set as the training set, but the spectral efficiency is lower on test sets with different types of channel data. Figure 6The performance evaluation of the model trained by mixing the two datasets in equal proportions shows that the spectral efficiency on different types of datasets is close to that of the spectral efficiency of the training set test, demonstrating the robustness of the model across datasets.

[0059] Corresponding to the aforementioned embodiment of an AI-based cross-dataset and cross-scene robust beamforming design method, this invention also provides an embodiment of an AI-based cross-dataset and cross-scene robust beamforming system.

[0060] The present invention provides an AI-based cross-dataset and cross-scenario robust beamforming system, which includes one or more processors for implementing the graph neural network-based channel decoding method in the above embodiments.

[0061] An embodiment of the AI-based cross-dataset and cross-scene robust beamforming system of the present invention can be applied to any device with data processing capabilities, such as a computer or other equipment or apparatus.

[0062] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0063] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the channel decoding method based on graph neural networks described in the above embodiments.

[0064] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., mounted on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0065] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. An AI-based, cross-dataset, cross-scenario robust beamforming design method, characterized in that, The method includes the following steps: S1. Channel data is collected in different scenarios. M scenario channel vectors are collected for each scenario, where each scenario channel vector includes N... r One user channel vector; S2. Establish a beamforming network: This network consists of an image input layer, a flattening layer, a feature extraction layer, and a custom function layer. The input is the scene channel vector and the signal-to-noise ratio, and the output is the same number of beamforming vectors as the user. S3. Constructing Beamforming Network Training Data: Select channel vectors from channel vectors in different scenarios and mix them. Divide the mixed channel data vector into support source and query source. Each time, extract channel vectors from the query source and support source to form a task, repeating N times. b A training batch consists of N tasks, with each task containing N tasks. s Channel vectors of the support sources and N q The channel vector of the query source; S4. Training the beamforming network: Initialize the network parameters, perform a deep copy of the beamforming network parameters, input the channel vectors from the support sources into the deep copy network for forward and backward propagation, update the transition model, then input the channel vectors from the query sources into the transition model, obtain the loss, and perform backward propagation on the beamforming network. S5. Train the beamforming network step times, and finally test the obtained step+1 models on the test set of the hybrid channel. Take the model with the highest spectral efficiency as the production model, and save the corresponding training hybrid channel data as a new training data source. S6. Input the user's channel vector and signal-to-noise ratio into the production model, and output the beamforming vector.

2. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 1, characterized in that, The number of different scenarios is two. When the number of different scenarios is greater than two, for each additional scenario, the data of the newly added scenario and the original training mixed channel data are used as the data of the two scenarios. During training, the beamforming network parameters inherit the parameters when no scenarios are added, until the channel data of the scenarios that need to be deployed are used.

3. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 1, characterized in that, In the beamforming network, the loss function is calculated from the output beamforming vector and the signal-to-noise ratio.

4. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 3, characterized in that, The loss function is specifically as follows: Where R is the spectral efficiency obtained from the i-th query source sample and the k-th user. k,i (v k Specifically: Where, σ 2 This refers to the variance of the circular complex Gaussian noise, where the channel vector for the k-th user is h. k The beamforming vector is v k Specifically: Where P represents the constraint power magnitude; the vector output of the last random deactivation layer is represented as...

5. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 2, characterized in that, In S3, the percentages of data entries for different scenarios are q1 and q2. During the initial training, q1 is set to 0 and q2 is set to 1.

6. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 1, characterized in that, The support source and query source are obtained by dividing the mixed channel data vector in half.

7. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 1, characterized in that, The number of transition models obtained in S4 is the same as the batch size; the number of losses obtained by inputting the channel vector in the query source into the transition model is also the same as the batch size; and the loss of backpropagation of the beamforming network is the average of the losses of all transition models.

8. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 5, characterized in that, In step S5, the training steps of the beamforming network are specifically defined as follows: each training step increases q1 by step_size and decreases q2 by step_size; where step*step_size = 1.

9. The AI-based cross-dataset, cross-scene robust beamforming design method according to claim 1, characterized in that, When the S5 model reaches the deployment stage, it needs online fine-tuning. Small samples that can quickly adapt are collected for the deployment scenario, and these samples are input into the model to adjust the parameters. adapt After the second update, the final model is obtained, where N adapt This is a hyperparameter.

10. An AI-based cross-dataset, cross-scene robust beamforming system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the AI-based cross-dataset, cross-scene robust beamforming design method as described in any one of claims 1-9.

Images