CN106599908A - Field intensity prediction method based on modularized neural network - Google Patents
Field intensity prediction method based on modularized neural network Download PDFInfo
- Publication number
- CN106599908A CN106599908A CN201611078350.5A CN201611078350A CN106599908A CN 106599908 A CN106599908 A CN 106599908A CN 201611078350 A CN201611078350 A CN 201611078350A CN 106599908 A CN106599908 A CN 106599908A
- Authority
- CN
- China
- Prior art keywords
- sample
- neural network
- point
- field intensity
- training
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
- G06F18/23213—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions with fixed number of clusters, e.g. K-means clustering
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04B—TRANSMISSION
- H04B17/00—Monitoring; Testing
- H04B17/30—Monitoring; Testing of propagation channels
- H04B17/309—Measuring or estimating channel quality parameters
- H04B17/318—Received signal strength
Abstract
The invention relates to the technology of electromagnetic field intensity prediction, so as to enable a prediction model to be more suitable for a complex electromagnetic propagation environment and solve a problem that a conventional scheme is poor in prediction precision and low in convergence speed. The technical scheme employed in the invention is that a field intensity prediction method based on a modularized neural network comprises the following steps: 1, building a radio propagation scene, selecting a number of receiving sample points from the scene, and obtaining the intensity values of the points through measurement or simulation; 2, carrying out the clustering of all sample points through a K-mean clustering method according to the distribution characteristics of field intensity data of received signals, so as to achieve the decomposition of an input sample space, and build corresponding sub-neural-network modules; 3, carrying out the training of the sub-neural-network modules of the modularized neural network through the above sample points; 4, carrying out the prediction through the trained modularized neural network. The method is mainly used in an electromagnetic field prediction occasion.
Description
Technical field
The present invention relates to a kind of radio field intensity Predicting Technique, specifically, is related to based on the field intensity of modular neural network
Forecasting Methodology.
Background technology
The prediction of wireless electromagnetic wave propagation characteristic for cordless communication network planning, design and optimize most important.Pass
The electric wave propagation prediction of system mainly includes empirical model and deterministic models.The precision of prediction of empirical model is not high；And
Deterministic models need accurate scene environment information, and will be through big computing；And it is based on the reception signal field of neutral net
Strong forecast model need not provide accurate environmental information, just can obtain the field intensity prediction value of enough accuracy, and for not
There is good generalization ability with communication environments.In existing neutral net field-strength prediction model, generally use single
Neutral net.Due to the complexity of electromagnetic propagation environment, predicting signal field intensity problem is received generally also more complicated.Single
BP neural network has that precision of prediction is poor, convergence rate when challenge is solved.For this purpose, the present invention proposes base
In the field-strength prediction model of modular neural network.
The content of the invention
To overcome the deficiencies in the prior art, it is contemplated that forecast model more adapts to the electromagnetic wave propagation ring of complexity
Border, solves the problems, such as that existing scheme precision of prediction difference convergence rate is slow.The technical solution used in the present invention is, based on modularity god
The field intensity prediction method of Jing networks, step is as follows
Step 1, sets up radio propagation scene, and a number of reception sample point is chosen from scene, by measurement
Or emulation obtains the field intensity value of the point；
Step 2, it is poly- to whole sample points using K mean cluster method according to the characteristic distributions for receiving signal strength data
Class, to realize the decomposition to input sample space, and sets up corresponding sub-neural network module；
Step 3, is trained using sample above point to the sub-network module of modular neural network；
Step 4, is predicted using the modular neural network for training.
Training set sample space is classified specifically using K mean cluster algorithm, not only consider the coordinate of sampled point, also
To consider the field intensity of sampled point simultaneously, if whole training set sample points are divided into into Ganlei, the close point minute in position in same class,
Classification results equally imply field intensity situation of change；Wherein, width refer to the maximum of all sample point distance centers of each apoplexy due to endogenous wind away from
From, that is to say, that all sample points of apoplexy due to endogenous wind are both less than equal to width with the distance of central point, it is determined that the center of a sample of each class
With width, the parameters of Task-decomposing device have been determined that.
Step 3 specifically, needs first to distribute sub- training set to each subnet before training subnet, and sub- training set includes basis
Sample point and extension sample point, whole sample points of a certain class are the basis of corresponding sub- training set in previous step cluster result
Sample point, in order that network also has preferable performance in the edge in each subtask space, needs to make each training
Have certain overlap between collection, that is, select suitable overlap coefficient, to determine the extension width of subset, overlap coefficient be one very
Little positive number；Other apoplexy due to endogenous wind also serve as the sample point of this class corresponding subset with this class centre distance less than the point of extension width,
Sample point is namely extended, is distributed after corresponding sub- training set to each subnet, be respectively trained each sub-neural network, trained
Method is identical with single neural network training method, using Levenberg-Marquardt training algorithms.
Step 4 specifically, for each sample in forecast set, when being predicted using the modular neural network after training
Firstly the need of Task-decomposing is carried out, that is, the distance of sample coordinate and each class center is calculated, find out closest class, correspondence is subordinate to
Category degree highest subnet, exports decision strategy, i.e., using the output of degree of membership highest subnet as whole net using selection type
The output of network, therefore, only the output of its degree of membership highest subnet need to be calculated, and the output of whole subnet need not be calculated, network
Output result be field intensity prediction value at the coordinate position.
Of the invention the characteristics of and beneficial effect are：
The invention provides the field intensity prediction method based on modular neural network, compared with the conventional method, using this
Bright provided method predictablity rate has a certain degree of lifting.
Description of the drawings：
Fig. 1 is used modular neural network structure chart.
Fig. 2 is test zone schematic diagram.
Fig. 3 is sample space division result.
Specific embodiment
To overcome the deficiencies in the prior art, it is contemplated that not significantly improving on the basis of computation complexity, nerve is lifted
The precision of prediction of network field-strength prediction model.The technical solution used in the present invention is as follows：
Step 1, sets up radio propagation scene, and a number of reception sample point is chosen from scene, by measurement
Or emulation obtains the field intensity value of the point；
Step 2, it is poly- to whole sample points using K mean cluster method according to the characteristic distributions for receiving signal strength data
Class, to realize the decomposition to input sample space, and sets up corresponding sub-neural network module；
Step 3, is trained using sample above point to the sub-network module of modular neural network；
Step 4, is predicted using the modular neural network for training, and is compared with the prior art.
The present invention sets up the model of field intensity prediction using modular neural network structure as shown in Figure 1.Model is built
It is vertical to include：Radio propagation scene is set up, a number of reception sample point is therefrom chosen, is somebody's turn to do by measurement or emulation
The field intensity value of point；According to the characteristic distributions for receiving signal strength data, whole sample points are clustered using K mean cluster method,
To realize the decomposition to input sample space, and set up corresponding sub-neural network module；Using sample above point to modularity
The sub-network module of neutral net is trained；It is predicted using the modular neural network for training.
1) scene is set up.Select certain dormitory area as test zone, set up modular neural network field-strength prediction model.God
The |input paramete of Jing networks is the horizontal, vertical coordinate figure of receiving point, and output parameter is corresponding reception signal strength value.Fig. 2 is
The plane structure chart of test zone.Wireless signal transmitter is located at the position of Tx in figure, and dash area is the sample chosen herein
Point pickup area, on the road between two row's dormitory buildings shown in solid line closed figures.Using ray trace computed in software
Go out sample point field intensity theoretical value.In order to verify the accuracy of proposed model, split sample set using random algorithm, from sample set
In randomly selected 50 points as test sample collection, for testing the effectiveness of method provided by the present invention, remaining sample
Put as training set, for training modular neural network involved in the present invention.
2) input sample space is decomposed.Task-decomposing device is set up according to training set.Here K mean cluster algorithm pair is adopted
Training set sample space is classified, and determines the cluster centre and width of each class.The characteristics of being distributed according to wireless signal field, no
Belt edge signal field intensity change with propagation regions is violent, therefore, cluster will also according to the coordinate for not only considering sampled point
The field intensity of sampled point is considered simultaneously.Fig. 3 represents the result i.e. division of sample space of cluster.Whole training set sample points are divided into
7 classes, 7 black represents the center of a sample of each class in figure.It can be seen that the close point in position has been in together substantially
One apoplexy due to endogenous wind, and it can be seen from field intensity data cases, classification results equally imply field intensity situation of change.Width refers to each class
In all sample point distance centers ultimate range, that is to say, that the distance of all sample points of apoplexy due to endogenous wind and central point is both less than etc.
In width.Center of a sample and the width of each class are determined, the parameters of Task-decomposing device have been determined that.
3) subnet is trained.Need first to distribute sub- training set to each subnet before training subnet, sub- training set includes basis
Sample point and extension sample point.Whole sample points of a certain class are the basis of corresponding sub- training set in previous step cluster result
Sample point.In order that network also has preferable performance in the edge in each subtask space, need to make each training
There is certain overlap between collection, that is, suitable overlap coefficient is selected, to determine the extension width of subset.The value model of overlap coefficient
Enclose for (0,1), recommend value be 0.5, this is because, when overlap coefficient value is less will affect network training effect, take
The training speed of network will be affected when being worth larger.Other apoplexy due to endogenous wind also serve as this with this class centre distance less than the point of extension width
The sample point of class corresponding subset, that is, extend sample point.Distribute after corresponding sub- training set to each subnet, be respectively trained
Each sub-neural network.Training method is identical with single neural network training method, is trained using Levenberg-Marquardt
Algorithm (LM algorithms).Compare with gradient descent method with the gradient descent method with momentum, LM algorithm the convergence speed is fast, and is difficult to be absorbed in
Local best points.
4) field intensity is predicted.For each sample in forecast set, when being predicted using the modular neural network after training
Firstly the need of Task-decomposing is carried out, that is, the distance of sample coordinate and each class center is calculated, find out closest class, correspondence is subordinate to
Category degree highest subnet.The present invention using selection type export decision strategy, i.e., using degree of membership highest subnet output as
The output of whole network, therefore, only the output of its degree of membership highest subnet need to be calculated, and the defeated of whole subnet need not be calculated
Go out.The output result of network is the field intensity prediction value at the coordinate position.
Claims (4)
1. a kind of field intensity prediction method based on modular neural network, is characterized in that, step is as follows：
Step 1, sets up radio propagation scene, and a number of reception sample point is chosen from scene, by measuring or imitating
Really obtain the field intensity value of the point；
Whole sample points, according to the characteristic distributions for receiving signal strength data, are clustered by step 2 using K mean cluster method, with
The decomposition to input sample space is realized, and sets up corresponding sub-neural network module；
Step 3, is trained using sample above point to the sub-network module of modular neural network；
Step 4, is predicted using the modular neural network for training.
2. the field intensity prediction method based on modular neural network as claimed in claim 1, is characterized in that, step 2 is equal using K
Value clustering algorithm is classified specifically to training set sample space, not only considers the coordinate of sampled point, and sampling is also considered simultaneously
The field intensity of point, if whole training set sample points are divided into into Ganlei, the close point minute in position is in same class, and classification results are equally hidden
Situation of change containing field intensity；Wherein, width refers to the ultimate range of all sample point distance centers of each apoplexy due to endogenous wind, that is to say, that class
In the distance of all sample points and central point be both less than equal to width, it is determined that the center of a sample of each class and width, determine that
The parameters of Task-decomposing device.
3. the field intensity prediction method based on modular neural network as claimed in claim 1, is characterized in that, step 3 specifically,
Need first to distribute sub- training set to each subnet before training subnet, sub- training set includes basic sample point and extension sample point,
Whole sample points of a certain class are the basic sample point of corresponding sub- training set in previous step cluster result, in order that network
Also there is preferable performance in the edge in each subtask space, need to make have certain overlap between each training subset,
Suitable overlap coefficient is selected, to determine the extension width of subset, overlap coefficient is the positive number of a very little；Other apoplexy due to endogenous wind with
This class centre distance also serves as the sample point of this class corresponding subset less than the point of extension width, that is, extends sample point, gives
Each subnet distributes after corresponding sub- training set, is respectively trained each sub-neural network, and training method is with single neutral net
Training method is identical, using Levenberg-Marquardt training algorithms.
4. the field intensity prediction method based on modular neural network as claimed in claim 1, is characterized in that, step 4 specifically,
For each sample in forecast set, firstly the need of the task of carrying out point when being predicted using the modular neural network after training
Solution, that is, calculate the distance of sample coordinate and each class center, finds out closest class, and correspondence degree of membership highest subnet is adopted
Export decision strategy with selection type, i.e., using degree of membership highest subnet output as whole network output, therefore, only need
The output of its degree of membership highest subnet is calculated, and the output of whole subnet need not be calculated, the output result of network is the seat
Field intensity prediction value at cursor position.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611078350.5A CN106599908A (en) | 2016-11-29 | 2016-11-29 | Field intensity prediction method based on modularized neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201611078350.5A CN106599908A (en) | 2016-11-29 | 2016-11-29 | Field intensity prediction method based on modularized neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN106599908A true CN106599908A (en) | 2017-04-26 |
Family
ID=58595812
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201611078350.5A Pending CN106599908A (en) | 2016-11-29 | 2016-11-29 | Field intensity prediction method based on modularized neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN106599908A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109217955A (en) * | 2018-07-13 | 2019-01-15 | 北京交通大学 | Wireless environment electromagnetic parameter approximating method based on machine learning |
| CN111147163A (en) * | 2019-12-17 | 2020-05-12 | 南京航空航天大学 | Wireless communication link loss prediction method based on DNN neural network |
| CN116226696A (en) * | 2023-03-12 | 2023-06-06 | 南通大学 | Electric field strength prediction method, electric field strength prediction device, electronic equipment and storage medium |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102662040A (en) * | 2012-04-20 | 2012-09-12 | 辽宁工程技术大学 | Ammonian online soft measuring method for dynamic modularized nerve network |
| CN105930916A (en) * | 2016-04-07 | 2016-09-07 | 大连理工大学 | Parallel modular neural network-based byproduct gas real-time prediction method |
-
2016
- 2016-11-29 CN CN201611078350.5A patent/CN106599908A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102662040A (en) * | 2012-04-20 | 2012-09-12 | 辽宁工程技术大学 | Ammonian online soft measuring method for dynamic modularized nerve network |
| CN105930916A (en) * | 2016-04-07 | 2016-09-07 | 大连理工大学 | Parallel modular neural network-based byproduct gas real-time prediction method |
Non-Patent Citations (3)
| Title |
|---|
| 刘红欣: "基于神经网络的TD-SCDMA基站电磁场强强度预测", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
| 李界家等: "基于模块化神经网络的铁水硅含量预测系统", 《沈阳建筑大学学报(自然科学版)》 * |
| 马千里等: "基于模糊边界模块化神经网络的混沌时间序列预测", 《物理学报》 * |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109217955A (en) * | 2018-07-13 | 2019-01-15 | 北京交通大学 | Wireless environment electromagnetic parameter approximating method based on machine learning |
| CN109217955B (en) * | 2018-07-13 | 2020-09-15 | 北京交通大学 | Wireless environment electromagnetic parameter fitting method based on machine learning |
| CN111147163A (en) * | 2019-12-17 | 2020-05-12 | 南京航空航天大学 | Wireless communication link loss prediction method based on DNN neural network |
| CN116226696A (en) * | 2023-03-12 | 2023-06-06 | 南通大学 | Electric field strength prediction method, electric field strength prediction device, electronic equipment and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Zhang et al. | Air-to-air path loss prediction based on machine learning methods in urban environments | |
| Wen et al. | Path loss prediction based on machine learning methods for aircraft cabin environments | |
| CN106452629B (en) | A kind of radio channel multi-path cluster-dividing method based on core power density | |
| CN111147163A (en) | Wireless communication link loss prediction method based on DNN neural network | |
| CN105610528B (en) | A kind of sub-clustering and tracking for time varying channel multipath component | |
| CN107370521A (en) | A kind of cognitive radio multi-user Cooperation frequency spectrum sensing method | |
| CN106599908A (en) | Field intensity prediction method based on modularized neural network | |
| CN112512069A (en) | Network intelligent optimization method and device based on channel beam pattern | |
| CN108040368A (en) | A kind of unmanned plane distribution method of time frequency resources declined based on block coordinate | |
| CN112182961A (en) | Large-scale fading modeling prediction method for wireless network channel of converter station | |
| He et al. | Random forests based path loss prediction in mobile communication systems | |
| CN103384174A (en) | Method based on cooperation of multiple users and multiple antennas for optimizing spectrum sensing detection probability | |
| Li et al. | The way to apply machine learning to IoT driven wireless network from channel perspective | |
| CN106713190A (en) | MIMO (Multiple Input Multiple Output) transmitting antenna number blind estimation algorithm based on random matrix theory and feature threshold estimation | |
| CN104734793B (en) | The energy detection method of wireless cooperation frequency spectrum perception based on p powers | |
| CN113242066B (en) | Multi-cell large-scale MIMO communication intelligent power distribution method | |
| CN113037668B (en) | Millimeter wave point-to-point communication channel equalization method | |
| CN104079359A (en) | Cooperative spectrum sensing threshold optimization method in cognitive wireless network | |
| CN103036626B (en) | Wireless communication method based on cognitive radio cooperation users and threshold testing combined selection | |
| Li et al. | On the modeling of near-field scattering of vehicles in vehicle-to-x wireless channels based on scattering centers | |
| Chenguang et al. | Analysis of the channel capacity with shadowing fading in VANET | |
| Liu et al. | Multi-criteria coverage map construction based on adaptive triangulation-induced interpolation for cellular networks | |
| CN102651071B (en) | Support vector machine-based cabin interior path loss prediction method | |
| CN106788818A (en) | Based on the CRSN frequency spectrum sensing methods that cognitive function and sensor node are separate | |
| CN113453148A (en) | Indoor position fingerprint positioning method combining deep learning and weighted K-neighbor algorithm |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170426 |
|
| WD01 | Invention patent application deemed withdrawn after publication |