Multi-beam radio frequency oam positioning identification method and system based on deep learning

Abstract

The application discloses a kind of based on deep learning multi-beam radio frequency OAM positioning identification method and system, the method and system use optical orbital angular momentum as the basis of pre-training, using pre-training weight information trains radio frequency orbital angular momentum image, and realizes the positioning and identification function of radio frequency orbital angular momentum.The trained deep learning model of the application can automatically identify the real radio frequency orbital angular momentum image received in real time, without manual discrimination, thereby simplifying the operation process of technical personnel.The application reduces the training time, realizes the automation of identification, improves the spectrum utilization, and helps the demultiplexing system of radio frequency OAM system.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and communication, and in particular to a multi-beam radio frequency OAM positioning and identification method and system based on deep learning. Background Technology

[0002] Due to the orthogonality between different OAM modes and the theoretical infinity of available OAM modes, introducing OAM and multiplexing OAM beams for transmission will greatly improve the capacity and bandwidth utilization of optical communication systems, and significantly increase the transmission capacity of communication systems. Therefore, OAM communication has received widespread attention from researchers in recent years.

[0003] However, during the identification process, technicians need to manually determine the position and identify the mode number of the received radio frequency orbital angular momentum image, which wastes a lot of identification time and reduces the efficiency of identification. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-beam radio frequency OAM positioning and identification method and system based on deep learning, which can realize functions such as automatic positioning of orbital angular momentum position and identification of orbital angular momentum pattern, thereby reducing the time of manual identification and improving the efficiency of identification.

[0005] The technical solution to achieve the purpose of this invention is as follows: Firstly, this invention provides a multi-beam radio frequency OAM positioning and recognition method based on deep learning. It establishes a model based on a convolutional neural network, using optical orbital angular momentum images as a pre-training dataset, and trains the model on the radio frequency orbital angular momentum images using the pre-trained weight information. The method includes the following steps: collecting training images; labeling the images with mode numbers and constructing a dataset; building a network model and training it to obtain the corresponding model file; and using the trained deep learning model to automatically identify the received images in real time, obtaining the orbital angular momentum mode number and position information.

[0006] Furthermore, the deep learning model is pre-trained using an optical orbital angular momentum dataset, and the obtained pre-trained weight information is used to train the prepared dataset.

[0007] Furthermore, a radio frequency orbital angular momentum image dataset generated by simulated UCA was used and preprocessed.

[0008] Furthermore, the collected images are preprocessed by randomly adding Gaussian noise with a mean of 0 and a variance of 0-0.2; at the same time, the labeled images are divided into datasets according to the proportions.

[0009] Furthermore, during the training process, a preset judgment process is used to determine whether the recognition accuracy of the deep learning model has reached a preset threshold. If the threshold is not reached, iterative calculations are performed based on the adjusted weight parameters, and training continues until the recognition accuracy reaches the preset threshold. The weight parameter information of the deep learning model is then saved.

[0010] Furthermore, the trained network is used to predict real-world measured images in real time to evaluate model performance. The steps of acquiring images and making predictions include: mounting the transmitting antenna and horn in a dark room and aligning them horizontally; acquiring several images at different transmission frequencies, different receiving distances, or different deflection angles; inputting the acquired images into the trained convolutional neural network for prediction; and testing the recognition accuracy of images acquired under different conditions.

[0011] Secondly, the present invention provides a multi-beam radio frequency OAM positioning and identification system based on deep learning, comprising:

[0012] The data import module is used to import and read pre-collected and labeled datasets, normalize the data, and extract dataset feature information.

[0013] The parameter setting module is used to set the parameters of the convolutional layers, pooling layers, fully connected layers, batch normalization, loss function, and activation function used during training.

[0014] The training module trains the pre-trained network on the optical orbital angular momentum dataset onto the collected radio frequency orbital angular momentum dataset to obtain network model parameters, including forward propagation calculation of training samples, network error loss value, and recognition accuracy.

[0015] The prediction module saves the weights obtained from training, inputs the newly acquired radio frequency orbital angular momentum image into the model, sorts the data within the threshold in descending order of feature significance, and views the prediction results.

[0016] The dataset acquisition module is divided into two parts. The first part is to acquire simulated image data, create labels in a specific format for the image data, and combine them into a dataset. The second part is to acquire actual antenna images after training to test the recognition performance of the deep learning model.

[0017] The dataset training module takes the dataset collected in the first part of the dataset acquisition module and inputs it into the training module for training, generating the corresponding model file.

[0018] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in the first aspect.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The present invention realizes the functions of automatic identification, detection and positioning. The trained model can automatically extract features, locate and identify orbital angular momentum patterns, realize the automation and informatization of deep learning, and improve the recognition efficiency; (2) The features extracted by the deep convolutional neural network model can avoid the influence of noise interference and can extract complex features; (3) The effective combination of deep learning model and classifier can quickly and effectively perform positioning and recognition, making the recognition accuracy more accurate. Attached Figure Description

[0020] Figure 1 This is a structural diagram of the deep learning platform according to an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating the training and decision-making process according to an embodiment of the present invention.

[0022] Figure 3 This is a flowchart of the deep learning recognition process according to an embodiment of the present invention.

[0023] Figure 4 This is a flowchart illustrating the image acquisition and prediction process according to an embodiment of the present invention. Detailed Implementation

[0024] This invention provides a deep learning-based multi-beam radio frequency (RF) OAM positioning and recognition method. It establishes a model based on a convolutional neural network, using optical orbital angular momentum images as a pre-training dataset, and trains the model on the RF orbital angular momentum images using the pre-trained weight information. The method includes the following steps:

[0025] A graph of the radio frequency orbital angular momentum transmitted by a simulated antenna;

[0026] Gaussian noise was randomly added to the acquired images to increase the robustness of the model;

[0027] Create a tag in a specific format from the image;

[0028] The prepared dataset is divided proportionally into training, testing, and validation sets, and then input into a deep convolutional neural network for training.

[0029] After reading the dataset, convolution, pooling, and fully connected layer calculations are performed to obtain the feature vector.

[0030] Once the accuracy of the deep learning model reaches a set threshold, training is stopped, and the weights associated with the deep learning model and the feature information of the dataset are saved for real-time prediction of the received radio frequency orbital angular momentum image.

[0031] This method first uses an optical orbital angular momentum dataset to pre-train a deep learning model, and then uses the obtained pre-trained weight information to train the prepared dataset.

[0032] The radio frequency orbital angular momentum image dataset generated by simulated UCA was used and preprocessed.

[0033] To increase the robustness of the deep learning model, the collected images are preprocessed by randomly adding Gaussian noise with a mean of 0 and a variance of 0-0.2 to the images; at the same time, the labeled images are divided into datasets according to the proportion.

[0034] During training, a pre-defined judgment process is used to determine whether the recognition accuracy of the deep learning model has reached a preset threshold. If the threshold is not reached, iterative calculations are performed based on the adjusted weight parameters, and training continues until the recognition accuracy reaches the preset threshold. The weight parameter information of the deep learning model is then saved.

[0035] The trained network can be used to predict real-world measured images in real time to evaluate model performance.

[0036] The steps for acquiring and predicting images include: mounting the transmitting antenna and horn in a dark room and aligning them horizontally; acquiring several images at different transmission frequencies, receiving distances, or deflection angles; inputting the acquired images into a trained convolutional neural network for prediction; and verifying that the model's performance meets the standards by testing the recognition accuracy of the images acquired under different conditions. After this verification, we can apply this deep learning model to the field of communication to improve communication efficiency.

[0037] Based on the same inventive concept, the present invention also provides a multi-beam radio frequency OAM positioning and identification system based on deep learning, comprising:

[0038] The data import module is used to import and read pre-collected and labeled datasets, normalize the data, and extract dataset feature information.

[0039] The parameter setting module is used to set the parameters of the convolutional layers, pooling layers, fully connected layers, batch normalization, loss function, and activation function used during training.

[0040] The training module will train the pre-trained network on the optical orbital angular momentum dataset on the radio frequency orbital angular momentum dataset we have collected to obtain the network model parameters, including the forward propagation calculation of the training samples, the network error loss value, and the recognition accuracy.

[0041] The prediction module saves the weights obtained from training, inputs the newly acquired radio frequency orbital angular momentum image into the model, sorts the data within the threshold in descending order of feature significance, and views the prediction results.

[0042] The dataset acquisition module is divided into two parts. The first part is to acquire simulated image data, create labels in a specific format for the image data, and combine them into a dataset. The second part is to acquire some actual antenna images after training to test the recognition performance of the deep learning model.

[0043] The dataset training module takes the dataset collected in the first part of the dataset acquisition module and inputs it into the training module for training, generating the corresponding model file.

[0044] Furthermore, in the training module, the backpropagation of the convolutional neural network continuously updates and adjusts the weights and biases by minimizing the residuals, and determines whether the recognition accuracy can reach a preset threshold. If it does not reach the threshold, it iterative calculation is performed based on the adjusted weight parameters until the recognition accuracy reaches the preset threshold.

[0045] The implementation methods of each module in the above recognition system are the same as those of the aforementioned recognition method.

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions and embodiments of this invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0047] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0048] Example

[0049] like Figure 1 The diagram shown is a structural diagram of the deep learning platform provided in an embodiment of the present invention, which includes a data import module S100, a parameter setting module S200, a training module S300, a prediction module S400, a dataset acquisition module S500, and a dataset training module S600.

[0050] The data import module S100 is used to read the input dataset and obtain the dataset feature information, which includes the data type and expected result attributes.

[0051] The dataset is pre-collected and labeled, containing images and label information. The dataset images use three-channel RGB images, and the dataset label format is converted to a specific format according to the requirements of the deep learning algorithm framework for easy reading.

[0052] The parameter setting module S200 is an image processing method and system, involving the design and construction of convolutional neural networks. A specific embodiment of the present invention includes the following steps: establishing a deep convolutional neural network using convolutional layers, pooling layers, fully connected layers, batch normalization processing, loss functions, and activation function structures.

[0053] Convolutional neural networks extract raw features from the input dataset, and after processing these features through a combination of convolution, pooling, and fully connected layers, they transform the input raw image into a feature vector output.

[0054] Specifically, this module uses 49 convolutional layers, employing different convolutional kernels to convolve the input image, generating multiple feature maps. Each feature map consists of arranged rectangular neurons, with neurons within the same feature map sharing weights, i.e., the convolutional kernel. The convolutional kernel is initialized as a random fractional matrix, and during network training, it learns appropriate weights.

[0055] In addition, this module includes a pooling layer for downsampling the input image to reduce its size, feature dimension, and the number of data and parameters, thereby improving the model's fault tolerance. In this embodiment, an average pooling layer is used to obtain the pooled value by calculating the average value of the image region.

[0056] Furthermore, this module includes a fully connected layer for classifying the processed image. The feature maps generated by the preceding layers are compressed into a single vector, which is then input into the fully connected layer for classification using the softmax function. The softmax function maps each value in the vector to the range (0,1), aiming to combine all local features into global features to calculate the score for each class, thus outputting a multi-class classification result.

[0057] To calculate the loss, this module uses a combination of classification loss and localization loss. The classification loss is calculated using multi-class cross-entropy loss, while the localization loss uses Smoooh L1 Loss. Furthermore, commonly used activation functions in the network include ReLU, sigmoid, and tanh. Given that ReLU performs better in deep convolutional neural networks, this embodiment chooses ReLU as the network's activation function.

[0058] The training module S300 uses the convolutional neural network set according to S200 to train the dataset. During the training process, the loss function curve and intermediate prediction results can be viewed.

[0059] See Figure 2This is a training and determination process in this embodiment. Based on the prediction result, it is determined whether the recognition accuracy of the deep learning model has reached a preset threshold. If it has not reached the preset threshold, iterative calculation is performed based on the adjusted weight parameters, and training continues until the recognition accuracy reaches the preset threshold. The weight parameter information of the deep learning model is then saved.

[0060] The prediction module S400 uses the deep learning model set in S200 to extract features from the radio frequency orbital angular momentum image collected and input to the terminal computing platform.

[0061] After feature extraction is completed, the prediction module S400 uses the weight parameters provided by S300 to further process the features, thereby obtaining the category of the orbital angular momentum mode and the corresponding position information.

[0062] Finally, the prediction module S400 returns the processed prediction results to the user interface so that users can view and analyze them.

[0063] The dataset acquisition module S500 includes dataset creation and acquisition of predicted images;

[0064] For details on the creation of the dataset, please refer to steps S501 to S503 in the subsequent applications of this patent. These steps describe in detail the specific content of the dataset creation.

[0065] For the acquisition of predicted images, please refer to... Figure 4 It includes the following steps:

[0066] Step S506: Install the transmitting antenna and speaker in the dark room and align them horizontally;

[0067] Step S507: Under different transmission conditions, such as different transmission frequencies, different receiving distances, and different deflection angles, acquire several images respectively;

[0068] In step S508, the acquired image is input into the trained convolutional neural network for prediction. By testing the recognition accuracy of images acquired under different conditions, and verifying that the performance of the deep learning model under different conditions meets the standards, we can apply it to the field of communication to improve communication efficiency and quality.

[0069] The dataset training module S600 uses a deep convolutional neural network to train the dataset;

[0070] The data acquisition module S500 collects the prepared dataset and inputs it into the deep learning model set by the parameter setting module S200 for training, converting the input raw image into a feature vector output.

[0071] Therefore, the present invention provides an image processing method and system that achieves feature extraction and classification of images through a combination of convolutional layers, pooling layers and fully connected layers, while improving the performance and accuracy of the model by using appropriate activation functions and loss functions.

[0072] By training the deep learning model, newly acquired orbital angular momentum patterns can be located and identified in real time. This method and system can be widely applied in fields such as computer vision, pattern recognition, and image classification, and provide a reliable method for demultiplexing orbital angular momentum patterns.

[0073] Figure 3 A flowchart of deep learning recognition provided for embodiments of the present invention.

[0074] Step S501: Collect the dataset and perform preprocessing;

[0075] To enhance the robustness of the deep learning model, the radio frequency orbital angular momentum image generated by simulating UCA was preprocessed by randomly adding Gaussian noise with a mean of 0 and a variance of 0-0.2 to the image.

[0076] Step S502: Add labels to the collected dataset;

[0077] Each processed image is labeled with its corresponding location and pattern number category. The label format is .xml, which is created according to the requirements of the deep learning platform. Together with the images, a dataset with orbital angular momentum location and pattern number is created.

[0078] Step S503: Split the dataset;

[0079] The dataset needs to be divided into training, testing, and validation sets. In this embodiment, the dataset is divided according to a pre-defined ratio, which is 8:1:1. The prepared dataset includes image files and .xml files, which are then input into the data import module S100 for reading.

[0080] Step S504: Train the read dataset;

[0081] In this embodiment, the optical orbital angular momentum dataset is used first to pre-train the deep convolutional neural network set by the parameter setting module S200 before training, and the obtained weight parameters are saved.

[0082] Using the weight parameters, the read radio frequency orbital angular momentum dataset is input into the deep convolutional neural network for training, which can improve recognition accuracy while reducing training time.

[0083] During training, you can view the loss function curve and prediction results. As the weights are continuously updated, the error of the convolutional neural network decreases and eventually stabilizes.

[0084] Once the recognition accuracy reaches the preset threshold, training stops and the weight parameters are saved.

[0085] Step S505: The trained network makes real-time predictions on the detected images.

[0086] The detected radio frequency orbital angular momentum image is input into a deep learning model for feature extraction to obtain a feature vector.

[0087] The deep learning model weights trained in step S504 are used to predict the feature vector, predicting the number of orbital angular momentum modes and the corresponding position information, and returning the results to the user interface.

[0088] Once the test results meet the standards, the deep learning model can be applied to the field of communication to determine the number of orbital angular momentum modes and locate positions in real time, thereby automating and informatizing deep learning and improving communication efficiency.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-beam radio frequency OAM positioning and identification system based on deep learning, characterized in that, include: The data import module is used to import and read pre-collected and labeled datasets, normalize the data, and extract dataset feature information. The parameter setting module is used to set the parameters of the convolutional layers, pooling layers, fully connected layers, batch normalization, loss function, and activation function used during training. The training module uses a pre-trained network based on the optical orbital angular momentum dataset to train on the acquired radio frequency orbital angular momentum dataset, obtaining network model parameters, including forward propagation calculation of training samples, network error loss value, and recognition accuracy. The prediction module saves the weights obtained from training, inputs the newly acquired radio frequency orbital angular momentum image into the model, sorts the data within the threshold in descending order of feature significance, and views the prediction results. The dataset acquisition module is divided into two parts. The first part is to acquire simulated image data, create labels in a specific format for the image data, and combine them into a dataset. The second part is to acquire actual antenna images after training to test the recognition performance of the deep learning model. The dataset training module takes the dataset collected in the first part of the dataset acquisition module and inputs it into the training module for training, generating the corresponding model file.

2. The deep learning-based multi-beam radio frequency OAM positioning and identification system according to claim 1, characterized in that, A dataset of radio frequency orbital angular momentum images generated by simulated UCA was used and preprocessed by randomly adding Gaussian noise with a mean of 0 and a variance of 0-0.2 to the images. At the same time, the labeled images were divided into datasets according to the proportions.

3. The deep learning-based multi-beam radio frequency OAM positioning and identification system according to claim 1, characterized in that, During training, a preset judgment process is used to determine whether the recognition accuracy of the deep learning model has reached a preset threshold. If the threshold is not reached, iterative calculations are performed based on the adjusted weight parameters, and training continues until the recognition accuracy reaches the preset threshold. The weight parameter information of the deep learning model is then saved. The trained network is used to predict real-world measured images in real time to evaluate model performance; the steps of acquiring images and making predictions include: mounting the transmitting antenna and horn in a dark room and aligning them horizontally; Several images were acquired under different transmission frequencies, different receiving distances, or different deflection angles. The acquired images were then input into a trained convolutional neural network for prediction. The recognition accuracy of the images acquired under different conditions was tested.

4. A deep learning-based multi-beam radio frequency OAM positioning and identification method applied to the system of claim 1, characterized in that, A convolutional neural network is built using optical orbital angular momentum images as a pre-training dataset. The weight information obtained from the pre-training is used to train a model based on radio frequency orbital angular momentum images. The process includes the following steps: collecting training images; labeling the images with mode numbers and constructing a dataset; building a network model and training it to obtain the corresponding model file; and using the trained deep learning model to automatically identify the received images in real time to obtain the orbital angular momentum mode number and position information.

5. The deep learning based multi-beam radio frequency OAM positioning and identification method according to claim 4, characterized in that, The deep learning model is pre-trained using an optical orbital angular momentum dataset, and the resulting pre-trained weight information is used to train the model.

6. The deep learning-based multi-beam radio frequency OAM positioning and identification method according to claim 5, characterized in that, The radio frequency orbital angular momentum image dataset generated by simulated UCA was used and preprocessed.

7. The method according to claim 6, characterized in that, The collected images were preprocessed by randomly adding Gaussian noise with a mean of 0 and a variance of 0-0.

2. At the same time, the labeled images were divided into datasets according to the proportions.

8. The deep learning-based multi-beam radio frequency OAM positioning and identification method according to claim 4, characterized in that, During training, a preset judgment process is used to determine whether the recognition accuracy of the deep learning model has reached a preset threshold. If the threshold is not reached, iterative calculations are performed based on the adjusted weight parameters, and training continues until the recognition accuracy reaches the preset threshold. The weight parameter information of the deep learning model is then saved.

9. The deep learning-based multi-beam radio frequency OAM positioning and identification method according to claim 8, characterized in that, The trained network is used to predict real-world measured images in real time to evaluate model performance. The steps of acquiring images and making predictions include: mounting the transmitting antenna and horn in a dark room and aligning them horizontally; acquiring several images at different transmission frequencies, different receiving distances, or different deflection angles; inputting the acquired images into the trained convolutional neural network for prediction; and testing the recognition accuracy of images acquired under different conditions.

10. An electronic device 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 program, it implements the steps of the method as described in any one of claims 4-9.

Images