CN106793070A - A kind of WiFi localization methods and server based on reinforcement deep neural network - Google Patents
A kind of WiFi localization methods and server based on reinforcement deep neural network Download PDFInfo
- Publication number
- CN106793070A CN106793070A CN201611072479.5A CN201611072479A CN106793070A CN 106793070 A CN106793070 A CN 106793070A CN 201611072479 A CN201611072479 A CN 201611072479A CN 106793070 A CN106793070 A CN 106793070A
- Authority
- CN
- China
- Prior art keywords
- data
- training
- deep neural
- neural network
- wap
- 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
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W64/00—Locating users or terminals or network equipment for network management purposes, e.g. mobility management
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W84/00—Network topologies
- H04W84/02—Hierarchically pre-organised networks, e.g. paging networks, cellular networks, WLAN [Wireless Local Area Network] or WLL [Wireless Local Loop]
- H04W84/10—Small scale networks; Flat hierarchical networks
- H04W84/12—WLAN [Wireless Local Area Networks]
Landscapes
- Engineering & Computer Science (AREA)
- Computer Networks & Wireless Communication (AREA)
- Signal Processing (AREA)
- Mobile Radio Communication Systems (AREA)
Abstract
The invention discloses a kind of based on the WiFi localization methods for strengthening deep neural network, methods described includes step:Each WAP receives the network data of the signal that client to be detected sends in S100, acquisition detection zone;The network data is multidimensional data;S200, the input data layer by the location model after network data input training, enter the Internet of location model after the 1*1 convolutional layers of location model are calculated;S300, the Internet based on the location model after training calculate the network data of client position to be detected, and the position of client to be detected is determined by the output result that output layer is exported.Location model in the present invention is trained by a large amount of training sample data using the deep neural network after training to deep neural network, lifts Position location accuracy and precision.
Description
Technical field
Determine the present invention relates to wireless local area network technology field, more particularly to a kind of WiFi based on reinforcement deep neural network
Position method and server.
Background technology
Location technology worldwide mainly has GPS location, Wi-Fi positioning, bluetooth positioning etc., GPS location at present
Outdoor is mainly used in, Wi-Fi, bluetooth positioning can be not only used for interior, it can also be used to outdoor.Because Wi-Fi positions relative maturity,
Below particular content of the invention is introduced with Wi-Fi location technologies as background.With the popularization of wireless router, current big portion
Point public domain all has been carried out more than ten or even tens WiFi signals coverings, and these routers are propagated to surrounding
While WiFi signal, the information such as its physical address and signal intensity are also ceaselessly sent, as long as in its signal cover,
Even if not knowing the password of Wi-Fi, these information can be similarly obtained.
General WiFi indoor positioning technologies are mostly the WLANs (WLAN) based on IEEE802.11b/g agreements
Signal intensity location technology.Location technology general principle based on signal intensity is to calculate letter according to the intensity of the signal for receiving
The distance between number receiver and signal source, are largely divided into two classes:Triangle intensity algorithm and location fingerprint recognizer.Its
Intermediate cam shape intensity arithmetic accuracy is low, it is difficult to meet indoor positioning requirement;And there is receiving device in general fingerprint recognizer
It is different and cause to receive the defect that signal has error.
The content of the invention
In order to solve the above technical problems, the present invention provide it is a kind of based on strengthen deep neural network WiFi localization methods and
Server, 1*1 convolutional calculations are carried out by the network data of the client position to be detected to gathering, and realize the net of multidimensional
Network data are preferably merged, and improve the precision that the WiFi based on deep neural network is positioned.
The technical scheme that the present invention is provided is as follows:
The invention discloses a kind of based on the WiFi localization methods for strengthening deep neural network, methods described includes step:
Each WAP receives the network data of the signal that client to be detected sends in S100, acquisition detection zone;It is described
Network data is multidimensional data;S200, the input data layer by the location model after network data input training, by fixed
The 1*1 convolutional layers of bit model enter the Internet of location model after calculating;S300, the Internet based on the location model after training
The network data of client position to be detected is calculated, and visitor to be detected is determined by the output result that output layer is exported
The position at family end.
It is further preferred that the network data receives client to be detected in detection zone including each WAP
The input power of the signal strength data, channel number and correspondence WAP of the signal sent in domain.
It is further preferred that also including step before the step S100:S000, training in advance deep neural network, will
Deep neural network after training is used as the location model.
It is further preferred that the step S000 further includes step:S001, pre-set training location tags;
S002, each WAP is gathered successively receive training terminal in each training location tags institute in detection zone
The network data of the signal for sending;The network data includes that each WAP receives the training location tags in detection
The signal strength data of the signal that the training terminal on the correspondence position of region is sent out, channel number and correspondence WAP it is defeated
Enter power;S003, respectively by each it is described training location tags and its corresponding signal strength data, channel number and correspondence
The input power of WAP generates training dataset as one group of training sample data, and sends into deep neural network;
S004, the input data of deep neural network layer is defined as Three-channel data layer, the node of the Three-channel data layer with it is each
Individual WAP is corresponding, according to the node mode corresponding with WAP of Three-channel data layer respectively by every group of training
Each signal strength data combination channel number corresponding with WAP and corresponding WAP in sample data
Input power is input into three passages of the Three-channel data layer corresponding node;The training sample data are by 1*1 convolutional layers
After calculating into the deep neural network Internet, finally via the deep neural network output layer output with it is described
The corresponding training result of training location tags;S005, the corresponding training position of training result that will be exported successively
Label is compared, and deep neural network is trained according to comparative result, using the deep neural network after training as institute
State location model.
It is further preferred that also including step between the step S003 and the step S004:S035, respectively to each
The input work of the signal strength data for training the corresponding network data of location tags, channel number and correspondence WAP
Rate is normalized;Also include step between the step S100 and step S200:S150, the institute by client to be detected
The input power for stating the signal strength data in network data, channel number and correspondence WAP is normalized.
The invention also discloses a kind of based on the WiFi location-servers for strengthening deep neural network, including:Data acquisition
Module, the network data of the signal that client to be detected sends is received for obtaining each WAP in detection zone;
The network data is multidimensional data;The network data receives client to be detected in detection including each WAP
The input power of the signal strength data, channel number and correspondence WAP of the signal sent in region;Locating module, uses
In the multi-channel data layer of the network data input location model that will be collected, by the 1*1 convolutional layer meters of location model
Into the Internet of location model after calculation, the Internet based on location model calculates the network data, and by output layer
Output result determines the position of client to be detected.
It is further preferred that also including:Training module, for training in advance deep neural network, by the depth after training
Neutral net is used as the location model.
It is further preferred that the training module is further included:Label presets submodule, for pre-setting for instructing
Experienced training location tags;Training dataset generates submodule, and training is received eventually for gathering each WAP successively
The network data of the signal sent in detection zone in each training location tags is held, respectively by each training
Location tags and its corresponding network data generate training dataset, and send into depth god as one group of training sample data
Through in network;The network data includes that each WAP receives the training location tags in detection zone correspondence position
On the signal sent out of training terminal signal strength data, channel number and correspondence WAP input power;Input
Data Layer defines submodule, for the input data layer of deep neural network to be defined as into Three-channel data layer, the triple channel
The node of data Layer is corresponding with each WAP;Convolutional layer defines submodule, for the input in deep neural network
1*1 convolutional layers are defined between data Layer and Internet;Training prediction submodule, for node and nothing according to Three-channel data layer
The corresponding mode of line access point is respectively by each the signal intensity number corresponding with WAP in every group of training sample data
Be input into the Three-channel data layer corresponding node according to the input power for combining channel number and correspondence WAP three are led to
Road, the Internet by entering the deep neural network after the calculating of 1*1 convolutional layers, finally via the deep neural network
Output layer is exported and the training corresponding training result of location tags, successively by institute that the training result for exporting is corresponding
State training location tags to be compared, deep neural network is trained according to comparative result, by the depth nerve after training
Network is used as the location model.
It is further preferred that also including:Data processing module, for corresponding to training location tags each described respectively
The input power of the signal strength data, channel number and correspondence WAP of network data is normalized, and
For the defeated of the signal strength data in the network data by client to be detected, channel number and correspondence WAP
Enter power to be normalized.
Compared with prior art, what the present invention was provided is a kind of based on the WiFi localization methods and clothes of strengthening deep neural network
Business device, by collecting the network data of client position to be measured, and the client position to be detected to gathering net
Network data carry out 1*1 convolutional calculations, the location model that input is trained, you can determine client position to be measured.The present invention
Deep neural network is improved, increases 1*1 convolutional layers, the three kind data higher to the degree of correlation in the network data of multidimensional are rolled up
Product, can be such that data preferably merge, and improve the precision of WiFi positioning.Meanwhile, by using containing a large amount of training sample data
Training data set pair deep neural network training, using the deep neural network after improvement as location model, not only lifted
The lifting of positioning precision, while the accuracy of positioning result can be lifted in the case where locating speed is not influenceed, successfully will be fixed
Position problem is dissolved into the background of big data, and the property of real-time location-server is effectively improved using the advantage of big data
Energy.
Brief description of the drawings
Below by clearly understandable mode, preferred embodiment is described with reference to the drawings, the present invention is given furtherly
It is bright.
Fig. 1 is a kind of key step schematic diagram based on the WiFi localization methods for strengthening deep neural network of the present invention;
Fig. 2 is a kind of training deep neural network based on the WiFi localization methods for strengthening deep neural network of the present invention
Step schematic diagram;
Fig. 3 is a kind of main composition schematic diagram based on the WiFi location-servers for strengthening deep neural network of the present invention;
Fig. 4 is that the present invention is a kind of is fully composed schematic diagram based on the WiFi location-servers for strengthening deep neural network.
Reference:
100th, data acquisition module, 200, locating module, 300, training module, 311, the default submodule of label, 312, instruction
Practice data set generation submodule, 313, input data layer define submodule, 314, training prediction submodule, 315, convolutional layer definition
Submodule, 400, data processing module.
Specific embodiment
In order to illustrate more clearly about the embodiment of the present invention or technical scheme of the prior art, control is illustrated below
Specific embodiment of the invention.It should be evident that drawings in the following description are only some embodiments of the present invention, for
For those of ordinary skill in the art, on the premise of not paying creative work, other can also be obtained according to these accompanying drawings
Accompanying drawing, and obtain other implementation methods.
To make simplified form, part related to the present invention is only schematically show in each figure, they are not represented
Its as product practical structures.In addition, so that simplified form is readily appreciated, there is identical structure or function in some figures
Part, only symbolically depicts one of those, or has only marked one of those.Herein, " one " is not only represented
" only this ", it is also possible to represent the situation of " more than one ".
Fig. 1 is a kind of key step schematic diagram based on the WiFi localization methods for strengthening deep neural network of the present invention, such as
Shown in Fig. 1, a kind of based on the WiFi localization methods for strengthening deep neural network, methods described includes step:S100, acquisition detection
Each WAP receives the network data of the signal that client to be detected sends in region;The network data is multidimensional
Data;S200, the input data layer by the location model after network data input training, by 1*1 volumes of location model
Lamination enters the Internet of location model after calculating;S300, the Internet based on the location model after training calculate visitor to be detected
The network data of family end position, and the position of client to be detected is determined by the output result that output layer is exported.
Specifically, above-mentioned client to be detected (hereinafter referred to as STA) is with smart mobile phone, notebook computer or personal flat board
The intelligent terminals such as computer are carrier.
Network data described in the present embodiment be multidimensional data, wherein specifically include each WAP receive it is to be checked
Survey the input work of the signal strength data, channel number and correspondence WAP of the signal that client sends in detection zone
Rate.Because channel number and Frequency point positive correlation, and frequency range is related to space attenuation, and channel number is bigger, decays in spatial
It is bigger, while the collection of the input power of AP can also influence progress, because the AP power outputs design objective of different model can be with
Different, even if identical AP is also entirely possible different in the output-index of 2.4G frequency ranges and 5G frequency ranges, therefore the present invention is not
RSSI is gathered on cochannel different frequency range, the input power and signal strength data channel number, AP in frequency range are gathered in the lump
Input deep neural network is trained, and greatly improves the accuracy rate of positioning.
Specifically, each WAP receives the letter that client to be detected is sent on position in detection zone
Number signal strength data obtain in the following manner:STA sends detection frame in real time within a detection region, and WAP is received
The signal intensity of the detection frame is obtained afterwards, and each WAP reports signal intensity to home server or Cloud Server,
The RSSI field intensity message that server is reported according to each WAP generates signal strength data.For example, signal strength data
Form be<RSSI1, RSSI2, RSSI3, RSSI4, RSSI5>, the RSSI of the STA that wherein RSSI1 is received for AP1, RSSI2 are
The RSSI of the STA that AP2 is received, by that analogy.By taking real data as an example, the network data of present invention collection client to be detected
Form is as follows,<(- 30,6,20), (- 12,11,18), (- 14,1,20), (- 67,36,18), (- 54,149,23)>, wherein having
Body implication is as follows:
RSSI1=-30dBm, channel 6 exports 20dBm
RSSI2=-12dBm, channel 11 exports 18dBm
RSSI3=-14dBm, channel 1 exports 20dBm
RSSI4=-67dBm, channel 36 exports 18dBm
RSSI5=-54dBm, channel 149 exports 23dBm
Specifically, because above-mentioned network data is actually three-dimensional data, by the Three-channel data of deep neural network
Although layer input can solve the precision of input data, the data of triple channel are the data of entirely different dimension, are not melted.
Meanwhile, network data includes the input power of signal strength data, channel number and correspondence WAP, three in the present embodiment
Person has very big correlation, therefore in order that multidimensional network data just produce preferably fusion, the present invention in input data layer
A 3 passage convolutional layers of 1x1 convolution kernels are inserted directly into input data layer and Internet.
Legacy network:
Three-channel data layer ---->Internet ---->Output layer
Inventive network:
Three-channel data layer ---->3 passage convolutional layers of 1x1 convolution kernels ---->Internet ---->Output layer
The Three-channel data of original input data layer is:X, y, z
3 passage convolutional layer output datas of 1x1 convolution kernels are:
X1=w11*x+w12*y+w13*z
X2=w21*x+w22*y+w23*z
X3=w31*x+w32*y+w33*z
Data so on 3 passage convolutional layer each passage of 1x1 convolution kernels are original Three-channel data layer triple channels
The linear combination of upper data, is more beneficial for the training of deep neural network below.
Network data by collecting client position to be measured of the invention, and the client to be detected place to gathering
The network data of position carries out 1*1 convolutional calculations, the location model that input is trained, you can determine that client institute to be measured is in place
Put.The present invention is improved deep neural network, increases 1*1 convolutional layers, three kinds higher to the degree of correlation in the network data of multidimensional
Data carry out convolution, and data can be made preferably to merge, and improve the precision of WiFi positioning.
Preferably, step is also included before the step S100:S000, training in advance deep neural network, after training
Deep neural network as the location model.
Specifically, the output result difference obtained according to positioning in the present invention, can include two kinds of specific implementations, side
Formula one is used as location model and exports belonging to client position to be detected certain and pre-sets by deep neural network
The probable value of classification, mode two is to be used as location model by deep neural network directly to export client position to be detected
Preset position coordinates.The present invention is not construed as limiting to specific training method.
The method that the present invention is trained using the global parameter for having supervision:The physical location of the known network data for receiving,
Cause that the output of the Internet of deep neural network is identical with real result by constantly adjustment network parameter.
The step of Fig. 2 is a kind of one embodiment of WiFi localization methods based on reinforcement deep neural network of the present invention is shown
It is intended to.Preferably, as shown in Fig. 2 the step S000 further includes step:S001, pre-set training location tags;
S002, each WAP is gathered successively receive training terminal in each training location tags institute in detection zone
The network data of the signal for sending;The network data includes that each WAP receives the training location tags in detection
The signal strength data of the signal that the training terminal on the correspondence position of region is sent out, channel number and correspondence WAP it is defeated
Enter power;S003, respectively by each it is described training location tags and its corresponding signal strength data, channel number and correspondence
The input power of WAP generates training dataset as one group of training sample data, and sends into deep neural network;
S004, the input data of deep neural network layer is defined as Three-channel data layer, the node of the Three-channel data layer with it is each
Individual WAP is corresponding, according to the node mode corresponding with WAP of Three-channel data layer respectively by every group of training
Each signal strength data combination channel number corresponding with WAP and corresponding WAP in sample data
Input power is input into three passages of the Three-channel data layer corresponding node;The training sample data are by 1*1 convolutional layers
After calculating into the deep neural network Internet, finally via the deep neural network output layer output with it is described
The corresponding training result of training location tags;S005, the corresponding training position of training result that will be exported successively
Label is compared, and deep neural network is trained according to comparative result, using the deep neural network after training as institute
State location model.
Specifically, because the present embodiment in network data be multidimensional data, and specially signal data intensity, channel number with
And correspondence WAP input power, therefore definition as location model deep neural network input data layer be also
Three-channel data layer.
Exported belonging to client position to be detected as location model by deep neural network in mode one below
As a example by the probable value of certain classification for pre-setting, the specific process for introducing the present invention to deep neural network training.
1st, in training, each WAP is gathered first and receives training terminal in default training location tags in inspection
Survey the network data on correspondence position in region.It is self-defining that training location tags are preset in the present embodiment, can specifically be passed through
Mesh generation detection zone, detection zone is divided into the grid classification of predetermined number, and the classification distribution of each grid is corresponding
Default training location tags, also can set up plane right-angle coordinate by detection zone, set correspondence in a coordinate system respectively
Position coordinates be default training location tags.Illustrated in mode one in the present embodiment, this is explained with real data below
The network data of each the default training location tags correspondence position gathered in invention.
Assuming that a default network data for training location tags to be collected is as follows wherein:
<(- 30,6,20), (- 12,11,18), (- 14,1,20), (- 67,36,18), (- 54,149,23), 34>
Represent:
RSSI1=-30dBm, channel 6 exports 20dBm
RSSI2=-12dBm, channel 11 exports 18dBm
RSSI3=-14dBm, channel 1 exports 20dBm
RSSI4=-67dBm, channel 36 exports 18dBm
RSSI5=-54dBm, channel 149 exports 23dBm
Label=34, represents that this default training location tags is designated 34, represents the net that 34 are designated in monitored area
Lattice position.
2nd, the input data layer of training sample data training deep neural network is passed sequentially through.
The original network data of collection is three-dimensional data, therefore, deep neural network of the definition as location model
Input data layer is also Three-channel data layer, and the node of Three-channel data layer is corresponding with each WAP.According to threeway
The node mode corresponding with WAP of track data layer is respectively by each in every group of training sample data and wireless access
The input power of the corresponding signal strength data combination channel number of point and correspondence WAP is input into the Three-channel data
Three passages of layer corresponding node.
Table one
Passage 1 | Passage 2 | Passage 3 |
-30 | 6 | 20 |
-12 | 11 | 18 |
-14 | 1 | 20 |
-67 | 36 | 18 |
-54 | 149 | 23 |
As shown in Table 1, passage 1 represents the signal strength data corresponding with each WAP of collection in table one, leads to
Road 2 represents corresponding channel number, and passage 3 represents the incoming frequency of each WAP of correspondence.Successively by each default training
Three passages of the Three-channel data layer of the corresponding training sample data feeding deep neural network of location tags.
3rd, 1*1 convolutional calculations
Also increase a triple channel convolutional layer for 1*1 convolution kernels between Three-channel data layer and Internet in the present invention,
From Three-channel data layer input each signal strength data combination channel number corresponding with WAP and it is corresponding wirelessly
The input power of access point carries out 1*1 convolutional calculations, and Internet is input into after calculating;
4th, deep neural network is trained
Finally by output layer output training result and the error for training location tags, finally by percentage regulation nerve net
Parameter in network causes that the Loss i.e. error of whole network is minimum.
It should be noted that do not indicate design parameter in entire depth neutral net, because these parameters and specific
The number of space and AP is relevant, not in the range of this patent.
Preferably, step is also included between the step S003 and the step S004:S035, respectively to instruction each described
The input power for practicing the signal strength data, channel number and correspondence WAP of the corresponding network data of location tags is carried out
Normalized;Also include step between the step S100 and step S200:S150, the network by client to be detected
The input power of signal strength data, channel number and correspondence WAP in data is normalized.
Specifically, in order that the network data of the multidimensional of collection can be efficiently entering deep neural network treatment, it is necessary to every
Individual initial data is all normalized.Following normalization formula manipulation is used in the present embodiment:
X '=(X-u)/σ, u are average, and σ is standard deviation
Normalization is only normalized in respective passage in the present embodiment.
Fig. 3 is a kind of main composition schematic diagram based on the WiFi location-servers for strengthening deep neural network of the present invention.
As shown in figure 3, it is a kind of based on the WiFi location-servers for strengthening deep neural network, including:Data acquisition module 100, is used for
Each WAP receives the network data of the signal that client to be detected sends in acquisition detection zone;The network number
According to being multidimensional data;The network data receives client to be detected and is sent in detection zone including each WAP
The signal strength data of signal, channel number and correspondence WAP input power;Locating module 200, for that will adopt
The multi-channel data layer of the network data input location model for collecting, calculates laggard by the 1*1 convolutional layers of location model
Enter the Internet of location model, the Internet based on location model calculates the network data, and the output knot for passing through output layer
Fruit determines the position of client to be detected.
Specifically, above-mentioned client to be detected (hereinafter referred to as STA) is with smart mobile phone, notebook computer or personal flat board
The intelligent terminals such as computer are carrier.
Network data described in the present embodiment be multidimensional data, wherein specifically include each WAP receive it is to be checked
Survey the input work of the signal strength data, channel number and correspondence WAP of the signal that client sends in detection zone
Rate.Because channel number and Frequency point positive correlation, and frequency range is related to space attenuation, and channel number is bigger, decays in spatial
It is bigger, while the collection of the input power of AP can also influence progress, because the AP power outputs design objective of different model can be with
Different, even if identical AP is also entirely possible different in the output-index of 2.4G frequency ranges and 5G frequency ranges, therefore the present invention is not
RSSI is gathered on cochannel different frequency range, the input power and signal strength data channel number, AP in frequency range are gathered in the lump
Input deep neural network is trained, and greatly improves the accuracy rate of positioning.
Specifically, each WAP receives the signal of the signal that client to be detected is sent in detection zone
Intensity data is obtained in the following manner:STA sends detection frame in real time within a detection region, and WAP obtains institute after receiving
The signal intensity of detection frame is stated, each WAP reports signal intensity to home server or Cloud Server, server root
The RSSI field intensity message reported according to each WAP generates signal strength data.For example, the form of signal strength data is<
RSSI1, RSSI2, RSSI3, RSSI4, RSSI5>, the RSSI of the STA that wherein RSSI1 is received for AP1, RSSI2 receives for AP2
The RSSI of STA, by that analogy.By taking real data as an example, the form of the network data of present invention collection client to be detected is as follows,
<(- 30,6,20), (- 12,11,18), (- 14,1,20), (- 67,36,18), (- 54,149,23)>, wherein concrete meaning is such as
Under:
RSSI1=-30dBm, channel 6 exports 20dBm
RSSI2=-12dBm, channel 11 exports 18dBm
RSSI3=-14dBm, channel 1 exports 20dBm
RSSI4=-67dBm, channel 36 exports 18dBm
RSSI5=-54dBm, channel 149 exports 23dBm
Specifically, because above-mentioned network data is actually three-dimensional data, by the Three-channel data of deep neural network
Although layer input can solve the precision of input data, the data of triple channel are the data of entirely different dimension, are not melted.
Meanwhile, network data includes the input power of signal strength data, channel number and correspondence WAP, three in the present embodiment
Person has very big correlation, therefore in order that multidimensional network data just produce preferably fusion, the present invention in input data layer
A 3 passage convolutional layers of 1x1 convolution kernels are inserted directly into input data layer and Internet.
Legacy network:
Three-channel data layer ---->Internet ---->Output layer
Inventive network:
Three-channel data layer ---->3 passage convolutional layers of 1x1 convolution kernels ---->Internet ---->Output layer
The Three-channel data of original input data layer is:X, y, z
3 passage convolutional layer output datas of 1x1 convolution kernels are:
X1=w11*x+w12*y+w13*z
X2=w21*x+w22*y+w23*z
X3=w31*x+w32*y+w33*z
Data so on 3 passage convolutional layer each passage of 1x1 convolution kernels are original Three-channel data layer triple channels
The linear combination of upper data, is more beneficial for the training of deep neural network below.
Network data by collecting client position to be measured of the invention, and the client to be detected place to gathering
The network data of position carries out 1*1 convolutional calculations, the location model that input is trained, you can determine that client institute to be measured is in place
Put.The present invention is improved deep neural network, increases 1*1 convolutional layers, three kinds higher to the degree of correlation in the network data of multidimensional
Data carry out convolution, and data can be made preferably to merge, and improve the precision of WiFi positioning.
Fig. 4 is that the present invention is a kind of is fully composed schematic diagram based on the WiFi location-servers for strengthening deep neural network.
As shown in Figure 4, it is preferred that also include:Training module 300, for training in advance deep neural network, by the depth god after training
Through network as the location model.
Specifically, the output result difference obtained according to positioning in the present invention, can include two kinds of specific implementations, side
Formula one is used as location model and exports belonging to client position to be detected certain and pre-sets by deep neural network
The probable value of classification, mode two is to be used as location model by deep neural network directly to export client position to be detected
Preset position coordinates.The present invention is not construed as limiting to specific training method.
The method that the present invention is trained using the global parameter for having supervision:Known signal corresponding with each WAP is strong
The physical location of degrees of data belongs to certain grid, by constantly adjustment network parameter cause deep neural network Internet it is defeated
Go out identical with real result.
Preferably, the training module is further included:Label presets submodule 311, for pre-setting for training
Training location tags;Training dataset generates submodule 312, and training is received eventually for gathering each WAP successively
The network data of the signal sent in detection zone in each training location tags is held, respectively by each training
Location tags and its corresponding network data generate training dataset, and send into depth god as one group of training sample data
Through in network;The network data includes that each WAP receives the training location tags in detection zone correspondence position
On the signal sent out of training terminal signal strength data, channel number and correspondence WAP input power;Input
Data Layer defines submodule 313, for the input data layer of deep neural network to be defined as into Three-channel data layer, the threeway
The node of track data layer is corresponding with each WAP;Convolutional layer defines submodule 315, in deep neural network
1*1 convolutional layers are defined between input data layer and Internet;Training prediction submodule 314, for according to Three-channel data layer
Node mode corresponding with WAP is respectively by each letter corresponding with WAP in every group of training sample data
The input power of number intensity data combination channel number and correspondence WAP is input into the Three-channel data floor corresponding node
Three passages, by 1*1 convolutional layers calculating after enter the deep neural network Internet, finally via the depth god
Output layer output and the training corresponding training result of location tags through network, successively by the training result for exporting and its
The corresponding training location tags are compared, and deep neural network are trained according to comparative result, after training
Deep neural network is used as the location model.
It should be noted that referring to the present invention for based on network data for the training process of above-mentioned training module 300
WiFi localization methods in train deep neural network explanation, no longer repeat herein.Letter in book server between each module
The contents such as breath interaction, implementation procedure and above method embodiment are based on same design, and particular content can be found in the inventive method reality
The narration in example is applied, here is omitted.
Preferably, also include:Data processing module 400, for respectively to training the corresponding net of location tags each described
The input power of the signal strength data, channel number and correspondence WAP of network data is normalized, Yi Jiyong
The input of signal strength data, channel number and correspondence WAP in the network data by client to be detected
Power is normalized.
Specifically, in order that the network data of the multidimensional of collection can be efficiently entering deep neural network treatment, it is necessary to every
Individual initial data is all normalized.Following normalization formula manipulation is used in the present embodiment:
X '=(X-u)/σ, u are average, and σ is standard deviation
Normalization is only normalized in respective passage in the present embodiment.
The contents such as information exchange, implementation procedure in book server between each module and above method embodiment are based on same
Design, particular content can be found in the narration in the inventive method embodiment, and here is omitted.
It should be noted that above-described embodiment can independent assortment as needed.The above is only of the invention preferred
Implementation method, it is noted that for those skilled in the art, is not departing from the premise of the principle of the invention
Under, some improvements and modifications can also be made, these improvements and modifications also should be regarded as protection scope of the present invention.
Claims (9)
1. it is a kind of based on the WiFi localization methods for strengthening deep neural network, it is characterised in that methods described includes step:
Each WAP receives the network data of the signal that client to be detected sends in S100, acquisition detection zone;
The network data is multidimensional data;
S200, the input data layer by the location model after network data input training, by the 1*1 convolution of location model
Layer enters the Internet of location model after calculating;
S300, the Internet based on the location model after training calculate the network data, and the output exported by output layer
Result determines the position of client to be detected.
2. it is as claimed in claim 1 based on the WiFi localization methods for strengthening deep neural network, it is characterised in that the network
Data include treating that each WAP receives the signal intensity number of the signal that client to be detected sends in detection zone
According to, channel number and the input power of correspondence WAP.
3. it is as claimed in claim 2 based on the WiFi localization methods for strengthening deep neural network, it is characterised in that the step
Also include step before S100:S000, training in advance deep neural network, using the deep neural network after training as described fixed
Bit model.
4. it is as claimed in claim 3 based on the WiFi localization methods for strengthening deep neural network, it is characterised in that the step
S000 further includes step:
S001, pre-set training location tags;
S002, gather successively each WAP receive training terminal each it is described training location tags in detection zone
The network data of interior sent signal;The network data includes that each WAP receives the training location tags and exists
The signal strength data of the signal that the training terminal on detection zone correspondence position is sent out, channel number and correspondence WAP
Input power;
S003, respectively by each it is described training location tags and its corresponding signal strength data, channel number and correspondence nothing
The input power of line access point generates training dataset as one group of training sample data, and sends into deep neural network;
S004, the input data of deep neural network layer is defined as Three-channel data layer, the node of the Three-channel data layer
It is corresponding with each WAP, according to the node mode corresponding with WAP of Three-channel data layer respectively by every group
Each signal strength data combination channel number corresponding with WAP and corresponding wireless access in training sample data
The input power of point is input into three passages of the Three-channel data layer corresponding node;The training sample data are by 1*1 volumes
Lamination calculate after enter the deep neural network Internet, finally via the deep neural network output layer output with
The corresponding training result of the training location tags;
S005, the corresponding training location tags of the training result of output are compared successively, according to comparative result
Deep neural network is trained, using the deep neural network after training as the location model.
5. as claimed in claim 4 based on the WiFi localization methods for strengthening deep neural network, it is characterised in that:
Also include step between the step S003 and the step S004:
S035, respectively to training the signal strength data of the corresponding network data of location tags, channel number and right each described
The input power of WAP is answered to be normalized;
Also include step between the step S100 and step S200:
S150, the signal strength data by the network data of client to be detected, channel number and correspondence wireless access
The input power of point is normalized.
6. a kind of based on the WiFi location-servers for strengthening deep neural network, it is characterised in that including:
Data acquisition module, for obtaining detection zone in each WAP receive the signal that client to be detected sends
Network data;The network data is multidimensional data;The network data receives to be detected including each WAP
The input work of the signal strength data, channel number and correspondence WAP of the signal that client sends in detection zone
Rate;
Locating module, the multi-channel data layer of the network data input location model for that will collect, by positioning mould
The 1*1 convolutional layers of type enter the Internet of location model after calculating, the Internet based on location model calculates the network data,
And the position of client to be detected is determined by the output result of output layer.
7. it is as claimed in claim 6 based on the WiFi location-servers for strengthening deep neural network, it is characterised in that also to wrap
Include:
Training module, for training in advance deep neural network, using the deep neural network after training as the location model.
8. it is as claimed in claim 7 based on the WiFi location-servers for strengthening deep neural network, it is characterised in that the instruction
Practice module to further include:
Label presets submodule, for pre-setting the training location tags for training;
Training dataset generates submodule, and training terminal is received in each instruction for gathering each WAP successively
Practice the network data of signal that location tags are sent in detection zone, respectively by each described training location tags and its
Corresponding network data generates training dataset as one group of training sample data, and sends into deep neural network;The net
Network data include that each WAP receives training terminal institute of the training location tags on detection zone correspondence position
The input power of the signal strength data, channel number and correspondence WAP of the signal of hair;
Input data layer defines submodule, for the input data layer of deep neural network to be defined as into Three-channel data layer, institute
The node for stating Three-channel data layer is corresponding with each WAP;
Convolutional layer defines submodule, for defining 1*1 convolutional layers between the input data of deep neural network layer and Internet;
Training prediction submodule, for the node mode corresponding with WAP according to Three-channel data layer respectively by every group
Each signal strength data combination channel number corresponding with WAP and corresponding wireless access in training sample data
The input power of point is input into three passages of the Three-channel data layer corresponding node, by entering institute after the calculating of 1*1 convolutional layers
The Internet of deep neural network is stated, finally the output layer output via the deep neural network and the training location tags
, be compared for the corresponding training location tags of the training result of output successively by corresponding training result, according to
Comparative result is trained to deep neural network, using the deep neural network after training as the location model.
9. it is as claimed in claim 8 based on the WiFi location-servers for strengthening deep neural network, it is characterised in that also to wrap
Include:
Data processing module, for respectively to trained each described the corresponding network data of location tags signal strength data,
The input power of channel number and correspondence WAP is normalized, and for by described in client to be detected
The input power of signal strength data, channel number and correspondence WAP in network data is normalized.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611072479.5A CN106793070A (en) | 2016-11-28 | 2016-11-28 | A kind of WiFi localization methods and server based on reinforcement deep neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611072479.5A CN106793070A (en) | 2016-11-28 | 2016-11-28 | A kind of WiFi localization methods and server based on reinforcement deep neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106793070A true CN106793070A (en) | 2017-05-31 |
Family
ID=58905523
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611072479.5A Pending CN106793070A (en) | 2016-11-28 | 2016-11-28 | A kind of WiFi localization methods and server based on reinforcement deep neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106793070A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107529222A (en) * | 2017-09-15 | 2017-12-29 | 上海交通大学 | A kind of WiFi indoor locating systems based on deep learning |
WO2019036860A1 (en) * | 2017-08-21 | 2019-02-28 | Beijing Didi Infinity Technology And Development Co., Ltd. | Positioning a terminal device based on deep learning |
CN111044974A (en) * | 2020-03-18 | 2020-04-21 | 腾讯科技(深圳)有限公司 | Indoor positioning method and device based on WiFi signal and storage medium |
CN111654910A (en) * | 2020-04-15 | 2020-09-11 | 深圳新贝奥科技有限公司 | Indoor positioning calculation method based on neural network |
CN113015093A (en) * | 2021-01-29 | 2021-06-22 | 辽宁大学 | Indoor wireless positioning method based on three-dimensional depth residual error neural network |
WO2021129634A1 (en) * | 2019-12-24 | 2021-07-01 | 北京嘀嘀无限科技发展有限公司 | Network positioning method and system |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1920593A (en) * | 2005-08-25 | 2007-02-28 | 广州天润信息科技有限公司 | Position fingerprint identification location method |
CN103945533A (en) * | 2014-05-15 | 2014-07-23 | 济南嘉科电子技术有限公司 | Big data based wireless real-time position positioning method |
CN104039008A (en) * | 2014-05-16 | 2014-09-10 | 南京大学 | Hybrid positioning method |
CN105101408A (en) * | 2015-07-23 | 2015-11-25 | 常熟理工学院 | Indoor positioning method based on distributed AP selection strategy |
-
2016
- 2016-11-28 CN CN201611072479.5A patent/CN106793070A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1920593A (en) * | 2005-08-25 | 2007-02-28 | 广州天润信息科技有限公司 | Position fingerprint identification location method |
CN103945533A (en) * | 2014-05-15 | 2014-07-23 | 济南嘉科电子技术有限公司 | Big data based wireless real-time position positioning method |
CN104039008A (en) * | 2014-05-16 | 2014-09-10 | 南京大学 | Hybrid positioning method |
CN105101408A (en) * | 2015-07-23 | 2015-11-25 | 常熟理工学院 | Indoor positioning method based on distributed AP selection strategy |
Cited By (10)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019036860A1 (en) * | 2017-08-21 | 2019-02-28 | Beijing Didi Infinity Technology And Development Co., Ltd. | Positioning a terminal device based on deep learning |
CN110892760A (en) * | 2017-08-21 | 2020-03-17 | 北京嘀嘀无限科技发展有限公司 | Positioning terminal equipment based on deep learning |
TWI695641B (en) * | 2017-08-21 | 2020-06-01 | 大陸商北京嘀嘀無限科技發展有限公司 | Positioning a terminal device based on deep learning |
CN107529222A (en) * | 2017-09-15 | 2017-12-29 | 上海交通大学 | A kind of WiFi indoor locating systems based on deep learning |
WO2021129634A1 (en) * | 2019-12-24 | 2021-07-01 | 北京嘀嘀无限科技发展有限公司 | Network positioning method and system |
CN111044974A (en) * | 2020-03-18 | 2020-04-21 | 腾讯科技(深圳)有限公司 | Indoor positioning method and device based on WiFi signal and storage medium |
CN111044974B (en) * | 2020-03-18 | 2020-07-07 | 腾讯科技(深圳)有限公司 | Indoor positioning method and device based on WiFi signal and storage medium |
CN111654910A (en) * | 2020-04-15 | 2020-09-11 | 深圳新贝奥科技有限公司 | Indoor positioning calculation method based on neural network |
CN113015093A (en) * | 2021-01-29 | 2021-06-22 | 辽宁大学 | Indoor wireless positioning method based on three-dimensional depth residual error neural network |
CN113015093B (en) * | 2021-01-29 | 2024-04-05 | 辽宁大学 | Indoor wireless positioning method based on three-dimensional depth residual error neural network |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106793070A (en) | A kind of WiFi localization methods and server based on reinforcement deep neural network | |
CN106792507A (en) | A kind of WiFi localization methods and server based on network data | |
CN105653502B (en) | A kind of communication base station electromagnetic radiation correlation analysis based on genetic algorithm | |
CN106507476A (en) | A kind of WiFi localization methods and server and location model construction method | |
CN106535326A (en) | WiFi locating method based on depth neural network and server | |
CN106792769A (en) | A kind of WiFi localization methods and server and location model method for building up | |
CN110418354A (en) | It is a kind of that propagation model wireless network planning method is exempted from based on machine learning | |
CN106535134A (en) | Multi-room locating method based on wifi and server | |
CN106792553A (en) | A kind of many floor location methods and server based on wifi | |
CN106792506A (en) | A kind of WiFi localization methods and server | |
CN106793067A (en) | A kind of many floor indoor orientation methods and server based on joint network | |
CN106550331A (en) | A kind of indoor orientation method and equipment | |
CN106982436A (en) | The detection method of adjustment and device of a kind of indoor and outdoor network coverage | |
CN105610528B (en) | A kind of sub-clustering and tracking for time varying channel multipath component | |
CN106412973A (en) | Network coverage quality detection method and device | |
CN106160819B (en) | A kind of signaling method and equipment | |
Wang et al. | System capacity analysis and antenna placement optimization for downlink transmission in distributed antenna systems | |
CN106503846A (en) | Route calculation algorithm patrolled and examined by a kind of power equipment | |
CN104765016B (en) | Radio frequency identification and location method based on intelligent control over power | |
CN107968987A (en) | RSSI weighted mass center localization methods based on definite integral combining environmental parameter | |
CN105188034A (en) | Collaborative positioning method in wireless sensor network | |
Nyayu et al. | Analysis on swarm robot coordination using fuzzy logic | |
CN106488559A (en) | A kind of outdoor positioning method based on visibility and server | |
CN106610486A (en) | Node positioning method and device | |
CN105897321A (en) | SWIPT signal receiving method based on time slot assignment, and receiver |
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: 20170531 |
|
WD01 | Invention patent application deemed withdrawn after publication |