CN115037344A - Personalized wave beam selection method based on fine adjustment - Google Patents
Personalized wave beam selection method based on fine adjustment Download PDFInfo
- Publication number
- CN115037344A CN115037344A CN202210663686.7A CN202210663686A CN115037344A CN 115037344 A CN115037344 A CN 115037344A CN 202210663686 A CN202210663686 A CN 202210663686A CN 115037344 A CN115037344 A CN 115037344A
- Authority
- CN
- China
- Prior art keywords
- model
- user
- global
- data
- local data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000010187 selection method Methods 0.000 title abstract description 10
- 238000000034 method Methods 0.000 claims abstract description 18
- 238000013136 deep learning model Methods 0.000 claims abstract description 14
- 238000012549 training Methods 0.000 claims description 39
- 230000006870 function Effects 0.000 claims description 27
- 238000013528 artificial neural network Methods 0.000 claims description 18
- 238000005457 optimization Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 6
- 238000001514 detection method Methods 0.000 claims description 6
- 238000010606 normalization Methods 0.000 claims description 5
- 230000003287 optical effect Effects 0.000 claims description 5
- 230000006399 behavior Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 239000000126 substance Substances 0.000 claims description 3
- 238000013480 data collection Methods 0.000 claims description 2
- 230000008014 freezing Effects 0.000 claims 1
- 238000007710 freezing Methods 0.000 claims 1
- 238000013135 deep learning Methods 0.000 abstract description 4
- 238000004891 communication Methods 0.000 description 3
- 230000007613 environmental effect Effects 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000009826 distribution Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
Images
Classifications
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station
- H04B7/0686—Hybrid systems, i.e. switching and simultaneous transmission
- H04B7/0695—Hybrid systems, i.e. switching and simultaneous transmission using beam selection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/0413—MIMO systems
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B7/00—Radio transmission systems, i.e. using radiation field
- H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
- H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
- H04B7/08—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the receiving station
- H04B7/0868—Hybrid systems, i.e. switching and combining
- H04B7/088—Hybrid systems, i.e. switching and combining using beam selection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
- G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses a personalized wave beam selection method based on fine tuning, which specifically comprises the following steps: users collect data and establish respective local data sets; the user uploads respective local data sets to the central server to construct a global data set; the central server uses the global data to train a global deep learning model in an off-line manner, and then sends the global deep learning model to the user; the user freezes the weight parameters of the global model and only allows updating the bias parameters. The user only uses the local data to continuously train the global model so as to fine tune the global model to be more suitable for the local data, and therefore an individualized beam selection model is established; and in the prediction stage, the user uses a personalized model to predict the optimal beam index. The invention solves the problem that the existing method is difficult to simultaneously show excellent performance on the user data with different statistical characteristics, and improves the prediction accuracy of the beam selection model based on deep learning on the user local data.
Description
Technical Field
The invention relates to the technical field of wireless communication, in particular to a beam selection method based on deep learning.
Background
Massive multiple input multiple output (Massive MIMO) and beamforming have become key technologies in current and future wireless communication systems. In recent years, a method based on deep learning is introduced into a beam management problem, and the method utilizes a large amount of data resources owned by a wireless communication system and an intelligent terminal, directly predicts an optimal beam by training a deep neural network, greatly saves channel estimation or beam search overhead, and is widely researched. However, data in the intelligent terminal often has statistical heterogeneity, and a single global model in the existing method is difficult to adapt to different user data distributions, and as the diversity of local data of different users is continuously increased, the problem is further worsened. In addition, due to the statistical deviation between the local data of the user and the global data, the accuracy of the beam selection model on the local data of the user is likely to be limited by taking the expected loss of the global data as a training target.
Disclosure of Invention
The invention aims to provide a personalized beam selection method based on fine adjustment. A personalized beam selection model is provided for each user in a targeted mode through a fine-tuning strategy, and the personalized model is required to obtain the optimal generalization performance on the local data (but not the global data) of the corresponding user, so that the prediction accuracy of the local data of the user is improved, and the user experience is maximized. The invention solves the problems that the prior method is difficult to simultaneously show excellent performance on user data with different statistical characteristics and sacrifice individual performance of a user in the process of pursuing global performance.
The technical scheme of the invention is as follows: a method for personalized beam selection based on fine tuning, comprising the steps of:
Further, the data collection step is independently performed by a plurality of users, and respective local data sets D are respectively established n . Due to differences of behavior habits, environmental geometries and the like of different users, statistical heterogeneity exists in local data of different users.
Further, the user integrates the position, the optical detection and ranging data and the base station position collected at the same time into the input of the neural network, and the optimal beam index at the time is used as the corresponding label. A pair of input and label constitutes one sample in the dataset.
And 3, performing off-line training by the central server by using the global data to establish a global deep learning model, wherein the training target is the expected minimum loss of the global data on the model, and performing iterative training until the model is converged.
Further, the beam selection problem is modeled as a received signal power maximization problem under the constraint of a fixed codebook, and specifically, the following is:
wherein the content of the first and second substances,
denotes an optimal beam index, M denotes the number of subcarriers, H denotes a channel, f p Representing base station side fixed codebooks
The p-th codeword in (1).
Further, the global deep learning model is a deep neural network with w as a parameter, and the optimization problem is solved by training the neural network to learn the mapping relation between the input information and the optimal beam index. The training objective is that the expected loss of global data on the model is minimal, as follows:
where the function F represents the optimization objective, w represents the parameters of the global model, and the function F represents the loss function.
Further, the neural network structure is specifically as follows:
the neural network is formed by connecting 6 convolutional layers and 3 fully-connected layers, the convolutional layers extract input implicit characteristics, and the fully-connected layers map the characteristics into predicted beam indexes;
the neural network introduces nonlinear components using activation functions, increasing the expressive power of the model.
Preferably, the activation function in step 3 is a ReLu function.
And 4, performing offline fine tuning training on the basis of the trained global deep learning model by using the local data only by the user, wherein the training target is that the expected loss of the local data on the model is minimum to establish a personalized beam selection model w n ,n∈{1,...,N}。
Further, the fine tuning training specifically includes:
when the user fine-tunes on the global deep learning model, the weight parameters of the global model are frozen (except for the batch normalization layer), and only the bias parameters are allowed to be updated. In particular, both the weights and bias parameters of the batch normalization layer allow for updating;
the user only uses the local data to perform fine tuning training, the training target is that the expected loss of the local data on the model is minimum, and for the user n, the following is specifically performed:
wherein the function F represents the optimization objective, w n Parameters representing a personalized model, function f n Representing the local loss function.
And 5, in a prediction stage, the user performs inference by using the personalized beam selection model to predict the optimal beam index.
The invention has the beneficial effects that:
the invention provides a personalized beam selection method based on fine adjustment, which focuses on the problems that the existing method is difficult to simultaneously show excellent performance on user data with different statistical characteristics, and the individual performance of a user is sacrificed in the process of pursuing global performance. Specifically, the invention provides personalized beam selection models for different users in a targeted manner through a fine-tuning strategy, and the prediction accuracy of the beam selection method based on deep learning on the local data of the users is improved.
Drawings
Fig. 1 is a flow chart of a method for personalized beam selection based on fine tuning according to the present invention.
FIG. 2 is a graph of the average accuracy of the model over the local data for each user.
FIG. 3 shows the accuracy of the model on the local data of each user.
Detailed Description
The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.
The present embodiment considers a beam selection problem in a downlink millimeter wave system. The system consists of 1 base station and 10 users, wherein the base station is provided with 64 antennas, a code book at the end of the base station comprises 64 code words, and a user side is a single antenna.
As shown in fig. 1, the detailed implementation steps of the personalized beam selection method based on fine tuning provided by the present invention include:
Specifically, step 1 comprises:
step 1.1, a plurality of users independently collect data and respectively establish respective local data sets D n . Due to differences of behavior habits, environmental geometries and the like of different users, statistical heterogeneity exists in local data of different users.
And step 1.2, integrating the position, the optical detection and ranging data and the base station position collected at the same time into the input of a neural network by a user, and taking the optimal beam index at the time as a label corresponding to the optimal beam index. A pair of input and label constitutes one sample in the dataset.
And 3, performing off-line training by using the global data by the central server to establish a global deep learning model, wherein the training target is that the expected loss of the global data on the model is minimum, and performing iterative training until the model is converged.
Specifically, step 3 includes:
step 3.1, modeling the beam selection problem as a received signal power maximization problem under the constraint of a fixed codebook, specifically as follows:
wherein the content of the first and second substances,
denotes an optimal beam index, M denotes the number of subcarriers, H denotes a channel, f p Representing base station side fixed codebooks
The p-th codeword in (1).
And 3.2, establishing and initializing a deep neural network with w as a parameter as a global deep learning model, and learning a mapping relation between input information and an optimal beam index by training the neural network so as to solve the optimization problem.
Specifically, the neural network structure is specifically as follows:
the neural network is formed by connecting 6 convolutional layers and 3 fully-connected layers, the convolutional layers extract input implicit characteristics, and the fully-connected layers map the characteristics into predicted beam indexes;
the neural network introduces nonlinear components by using an activation function to increase the expression force of the model;
the activation function adopts a ReLU function.
Step 3.3, the training target is that the expected loss of global data on the model is minimum, and the specific steps are as follows:
where the function F represents the optimization objective, w represents the parameters of the global model, and the function F represents the loss function.
And 3.4, performing iterative training until the global model is converged.
And 4, performing offline fine tuning training on the basis of the trained global deep learning model by using the local data only by the user, wherein the training target is that the expected loss of the local data on the model is minimum to establish a personalized beam selection model w n ,n∈{1,...,N}。
Specifically, step 4 includes:
and 4.1, the central server sends the global model trained to be converged to the user.
Step 4.2, the user freezes the weight parameters of the global model (except for the batch normalization layer), and only allows updating the bias parameters. In particular, both the weights and bias parameters of the batch normalization layer allow for updating.
Step 4.3, the user only uses the local data to perform fine tuning training, the training target is that the expected loss of the local data on the model is minimum, and for the user n, the following concrete steps are performed:
wherein the function F represents the optimization objective, w n Parameters representing a personalized model, function f n Representing a local loss function.
And 4.4, carrying out iterative training until the personalized model converges.
And 5, in a prediction stage, the user performs inference by using the personalized beam selection model to predict the optimal beam index.
In this embodiment, the global model is already converged during the 50 th training round, and then the global model after the 50 th training round is taken and sent to the user as the basis for the local fine tuning of the user. The user fine-tunes 50 training rounds on the basis of this global model, at which point the fine-tuning training has converged.
The accuracy curve of the whole training process is shown in fig. 2. Specifically, the accuracy rate refers to an average value of accuracy rates of the model on 10 user local data sets, and the comparison method is a traditional global model obtained by a central server through 100 rounds of global data training. The personalized wave beam selection method based on fine tuning obviously improves the accuracy of the traditional global model.
FIG. 3 illustrates the highest accuracy achieved by the model on each user's dataset, respectively. For each user, the personalized beam selection method based on fine tuning provided by the invention can provide higher accuracy for each user.
The above description is only an embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification or equivalent replacement of the technical solution of the present invention is covered in the scope of the claims of the present invention without departing from the spirit and scope of the technical solution of the present invention.
Claims (5)
1. A method for personalized beam selection based on fine tuning, comprising the steps of:
step 1, users independently collect user side data including user positionsPosition, optical detection and ranging data, base station position and optimal beam index, establishing respective local data sets D n N is equal to {1,. eta., N }, and N represents the number of users;
step 2, uploading the local data set D by the user n To the central server, the central server merges the data of the individual users to create a global data set D ═ D 1 ,...,D N }；
Step 3, the central server performs off-line training by using the global data to establish a global deep learning model w, the training target is the expected minimum loss of the global data on the model, and iterative training is performed until the model is converged;
and 4, performing offline fine tuning training on the basis of the trained global deep learning model by using the local data only by the user, wherein the training target is that the expected loss of the local data on the model is minimum to establish a personalized beam selection model w n ,n∈{1,...,N}；
And step 5, in a prediction stage, the user performs reasoning by using the personalized beam selection model to predict the optimal beam index.
2. The method of claim 1, wherein: in the step 1, data collection is independently performed by a plurality of users, and respective local data sets D are respectively established n (ii) a Due to the fact that behavior habits and wireless environments of different users are different, statistical heterogeneity exists in local data of different users;
integrating the position, the optical detection and ranging data and the base station position collected at the same moment into an input x of a model by a user, and taking the optimal beam index at the moment as a label y corresponding to the optimal beam index; a pair of input and label constitutes one sample (x, y) in the dataset.
3. The method of claim 1, wherein: in step 3, the beam selection problem is modeled as a received signal power maximization problem under the constraint of a fixed codebook, which is specifically as follows:
wherein the content of the first and second substances,
denotes an optimal beam index, M denotes the number of subcarriers, H denotes a channel, f p Representing base station side fixed codebooks
The p-th codeword in (a);
the global deep learning model is a deep neural network with w as a parameter, and the mapping relation between input information and the optimal beam index is learned by training the neural network so as to solve the optimization problem; the training objective is that the expected loss of global data on the model is minimal, as follows:
wherein, the function F represents an optimization target, w represents parameters of a global model, and the function F represents a loss function;
the neural network structure is specifically as follows:
the neural network is formed by connecting 6 convolutional layers and 3 fully-connected layers, the convolutional layers extract input implicit characteristics, and the fully-connected layers map the characteristics into predicted beam indexes;
the neural network introduces nonlinear components using activation functions, increasing the expressive power of the model.
4. The method of claim 3, wherein: the activation function adopts a ReLu function.
5. The method of claim 1, wherein: in step 4, the user continues offline training on the global deep learning model using the local data, and the method further includes:
when a user fine tunes on the global deep learning model, freezing the weight parameters of the global model, and only allowing to update the bias parameters; both the weight and bias parameters of the batch normalization layer are allowed to be updated;
the user only uses the local data to perform fine tuning training, the training target is that the expected loss of the local data on the model is minimum, and for the user n, the following is specifically performed:
wherein the function F represents the optimization objective, w n Parameters representing a personalized model, function f n Representing the local loss function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210663686.7A CN115037344A (en) | 2022-06-10 | 2022-06-10 | Personalized wave beam selection method based on fine adjustment |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210663686.7A CN115037344A (en) | 2022-06-10 | 2022-06-10 | Personalized wave beam selection method based on fine adjustment |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115037344A true CN115037344A (en) | 2022-09-09 |
Family
ID=83125154
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210663686.7A Pending CN115037344A (en) | 2022-06-10 | 2022-06-10 | Personalized wave beam selection method based on fine adjustment |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115037344A (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113783593A (en) * | 2021-07-30 | 2021-12-10 | 中国信息通信研究院 | Beam selection method and system based on deep reinforcement learning |
| US20220027764A1 (en) * | 2020-07-27 | 2022-01-27 | Thales Canada Inc. | Method of and system for online machine learning with dynamic model evaluation and selection |
| CN114580498A (en) * | 2022-01-26 | 2022-06-03 | 华东师范大学 | Federal learning method with high communication efficiency in wireless communication scene |
-
2022
- 2022-06-10 CN CN202210663686.7A patent/CN115037344A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20220027764A1 (en) * | 2020-07-27 | 2022-01-27 | Thales Canada Inc. | Method of and system for online machine learning with dynamic model evaluation and selection |
| CN113783593A (en) * | 2021-07-30 | 2021-12-10 | 中国信息通信研究院 | Beam selection method and system based on deep reinforcement learning |
| CN114580498A (en) * | 2022-01-26 | 2022-06-03 | 华东师范大学 | Federal learning method with high communication efficiency in wireless communication scene |
Non-Patent Citations (1)
| Title |
|---|
| 马文焱;戚晨皓;: "基于深度学习的上行传输过程毫米波通信波束选择方法", 合肥工业大学学报(自然科学版), no. 12, 28 December 2019 (2019-12-28) * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112383922B (en) | Deep reinforcement learning frequency spectrum sharing method based on prior experience replay | |
| Ahad et al. | Neural networks in wireless networks: Techniques, applications and guidelines | |
| CN105790813B (en) | Code book selection method based on deep learning under a kind of extensive MIMO | |
| CN109447275B (en) | Switching method based on machine learning in UDN | |
| CN112902969B (en) | Path planning method of unmanned aerial vehicle in data collection process | |
| CN112153616B (en) | Power control method in millimeter wave communication system based on deep learning | |
| CN112564752A (en) | Intelligent surface auxiliary communication method for optimizing activation and reconfiguration of sparse antenna | |
| CN112836822B (en) | Federal learning strategy optimization method and device based on width learning | |
| CN110167204A (en) | A kind of relay transmission policy selection and power distribution method based on MS-BAS algorithm | |
| Lan et al. | Progressive feature transmission for split classification at the wireless edge | |
| CN115065728B (en) | Multi-strategy reinforcement learning-based multi-target content storage method | |
| CN114997737A (en) | Unmanned aerial vehicle small base station cluster RAN slicing method based on layered federal learning | |
| Chen et al. | One for all: Traffic prediction at heterogeneous 5g edge with data-efficient transfer learning | |
| CN115169575A (en) | Personalized federal learning method, electronic device and computer readable storage medium | |
| CN115037344A (en) | Personalized wave beam selection method based on fine adjustment | |
| CN112738849B (en) | Load balancing regulation and control method applied to multi-hop environment backscatter wireless network | |
| CN117295090A (en) | Resource allocation method for Unmanned Aerial Vehicle (UAV) through-sense integrated system | |
| CN114980156B (en) | AP switch switching method of honeycomb millimeter wave-free large-scale MIMO system | |
| CN115987406A (en) | Underwater acoustic communication link self-adaption method based on deep reinforcement learning | |
| CN116384470A (en) | Convolutional neural network model compression method and device combining quantization and pruning | |
| CN113518457B (en) | Power distribution strategy based on one-dimensional deep convolutional neural network | |
| Qiu et al. | Learning personalized end-to-end task-oriented dialogue for fast and reliable adaptation | |
| CN116419257A (en) | Communication method and device | |
| CN113033653A (en) | Edge-cloud collaborative deep neural network model training method | |
| CN112966812A (en) | Automatic neural network structure searching method for communication signal modulation recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |