CN104540221B - WLAN indoor orientation methods based on semi-supervised SDE algorithms - Google Patents
WLAN indoor orientation methods based on semi-supervised SDE algorithms Download PDFInfo
- Publication number
- CN104540221B CN104540221B CN201510020485.5A CN201510020485A CN104540221B CN 104540221 B CN104540221 B CN 104540221B CN 201510020485 A CN201510020485 A CN 201510020485A CN 104540221 B CN104540221 B CN 104540221B
- Authority
- CN
- China
- Prior art keywords
- formula
- regions
- sub
- point
- 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.)
- Active
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
-
- H04W4/04—
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04W—WIRELESS COMMUNICATION NETWORKS
- H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
- H04W4/02—Services making use of location information
- H04W4/023—Services making use of location information using mutual or relative location information between multiple location based services [LBS] targets or of distance thresholds
Abstract
Based on the WLAN indoor orientation methods of semi-supervised SDE algorithms, it is related to indoor positioning field.The present invention is to solve the problem of that the high mobile terminal location real-time of tuning on-line complexity present in existing WiFi indoor orientation methods is poor.The present invention is by introducing semi-supervised SDE dimension-reduction algorithms, by using the Unlabeled data for being easy to acquisition, the low dimensional manifold of the high dimensional data of characterization location information is found out, the calculation amount of position fixing process is effectively reduced while the positioning accuracy for ensureing WLAN indoor locating systems.Reduce the workload acquired with reference to point data simultaneously, the real-time update for database provides easy-to-use approach.The tuning on-line complexity of the present invention is low, and mobile terminal location is real-time.The present invention is suitable for WLAN indoor positionings.
Description
Technical field
The present invention relates to indoor positioning fields, and in particular to a kind of location fingerprint indoor orientation method.
Background technology
It is widely available with wireless network, mobile communication and general fit calculation, location based service (LBS,
Location-based Services) also increasingly get more and more people's extensive concerning, wherein how to determine that the position of user is real
The key of existing LBS.It is well known that global positioning satellite (GPS, Global Positioning System) system passes through reception
Device measures the reaching time-difference estimated location from 5~24 satellite-signals, can provide the location estimation of degree of precision.But
It is that the city intensive with high building can not realize positioning to GPS due to the non line of sight of satellite-signal indoors.With IEEE
The it is proposed of 802.11 standards, WLAN (WLAN, Wireless Local Area Networks) have been widely distributed in school
Garden, office block and family.And because having, deployment is convenient, at low cost, is not required to based on the indoor locating system of received signal strength
Add positioning measurement specialized hardware the features such as and it is in widespread attention.
Under WLAN environment, by the wireless network card and corresponding software of mobile terminal measure from wireless access point (AP,
Access Point) received signal strength (RSS, Received Signal Strength) value, acquisition be equipped with corresponding positions
The information of pass, and then mobile subscriber present position is predicted by matching algorithm.Wherein based on the location algorithm of location fingerprint because
For positioning accuracy height, existing utility can be made full use of, upgrade and safeguards the advantages that small to customer impact and is used widely.Position
Put fingerprinting localization algorithm and be divided into two steps of off-line measurement stage and tuning on-line stage, off-line phase mainly establish position with
Correspondence between received signal strength passes through witness mark in area to be targeted by certain rule setting reference point
The different AP signal strength values that place receives establish corresponding location fingerprint database Radio Map.The tuning on-line stage leads to
The RSS values that test point receives are crossed, using corresponding matching algorithm, mainly including nearest neighbor method, k-nearest neighbor, probabilistic method and god
Through network technique.Wherein k-nearest neighbor (KNN, K Nearest Neighbors) all has one on algorithm complexity and positioning accuracy
Determine advantage, be widely used in tuning on-line matching, find in location fingerprint database with its immediate position, as final
Location estimation result.The Radio Map that off-line phase is established include a large amount of data information, and with localization region expands,
The increase of reference point, leading to Radio Map information content, exponentially situation increases.
WLAN indoor locating systems based on location fingerprint acquire data information as much as possible by off-line phase, can
To effectively improve the positioning accuracy of system.And a large amount of data information of tuning on-line phase process, increase the number of position fixing process
According to operand, its processing capacity is limited for mobile terminal, causes location algorithm operation difficult.While certain characteristic informations are not
Effective location information cannot be only provided or even can also influence the accuracy of positioning result.
When AP numbers increase, the dimension information of expression AP numbers has reformed into high dimensional data in Radio Map, can pass through dimension
Number about subtracts the burden to mitigate processing high dimensional data.High dimensional data may include many features, these features are all same in description
A things, these features are closely coupled to a certain extent.Such as when taking pictures simultaneously to same object from multiple angles, obtain
Information of the data arrived just containing overlapping.If nonoverlapping expression that some that can obtain these data simplify, it will greatly
It improves the efficiency of data processing operation and improves accuracy to a certain extent in ground.The purpose of dimension-reduction algorithm is also exactly to be to improve height
The treatment effeciency of dimension data.
There is the dimension-reduction algorithm much based on different purposes at present, including linearity and non-linearity dimension-reduction algorithm.Wherein PCA and
LDA is typical linear dimension-reduction algorithm.This kind of algorithm have good dimensionality reduction for the high dimensional data with linear structure as a result,
But the high dimensional data of nonlinear organization is not suitable for it.Nonlinear reductive dimension algorithm is then based on manifold learning arithmetic.SDE algorithms are bases
LDE algorithms propose in manifold learning arithmetic, it is a kind of manifold learning arithmetic of typical feature based extraction.
Invention content
The present invention is to solve the height of tuning on-line complexity present in existing WiFi indoor orientation methods, mobile terminal
The problem of real-time is poor is positioned, so as to provide a kind of WLAN indoor orientation methods based on semi-supervised SDE algorithms.
Based on the WLAN indoor orientation methods of semi-supervised SDE algorithms, it is realized by following steps：
Step 1: arrange m access point AP (AP for indoor environmentj, 1≤j≤m), it is ensured that appoint in the indoor environment
The signal that meaning is a little sent out by two or more wireless access point APs covers；M is positive integer；
Step 2: being uniformly arranged reference point in environment indoors, choose a reference point and establish rectangular coordinate system for origin,
Coordinate position of each reference point in the rectangular coordinate system is obtained, and is acquired simultaneously using signal receiver in each reference point
Received signal strength RSS value of the record from each AP k times, and carry out data processing；K is positive integer；
It is the ginseng per sub-regions Step 3: indoor positioning environment is divided into Q sub-regions according to K mean cluster algorithm
Examination point marks respective classification information；
Step 4: the unmarked RSS data that acquisition is random, the feature vector of each sub-regions obtained with step 3 carries out
Compare, that is, ask for the distance with the feature vector of all subregion, by random data category division with its feature vector distance most
In near subregion；
Step 5: to using SDE algorithms in K sub-regions per sub-regions, eigentransformation matrix is obtained；
The input parameter of SDE algorithms, i.e., the value of intrinsic dimension, by existing intrinsic dimension algorithm for estimating to Radio
Map partition datas are estimated, provide the intrinsic dimension estimated value of each area data, determine the eigentransformation square in each region
Battle array, and generate the location fingerprint database Radio Map after dimensionality reduction*；
Step 6: the feature vector for each sub-regions that the signal strength RSS values that tested point obtains are obtained with step 3
Be compared, that is, ask for the distance of the feature vector of test point and the feature vector of all subregion, by test point be located in and its
In the closest subregion of feature vector；
Step 7: in the subregion positioned, using the eigentransformation matrix that step 5 obtains to the RSS values of tested point
Dimensionality reduction is carried out, obtains the RSS of low-dimensional*, with fingerprint database Radio Map*It is matched, is positioned and calculated using k nearest neighbor location fingerprint
Method is accurately positioned test point.
Being acquired in each reference point using signal receiver described in step 2 simultaneously records the reception from each AP
Signal strength RSS values k times, and carry out data processing the specific steps are：
Step 2 one obtains each reference point in one k × m rank matrix, and the i-th row jth row of matrix represent ith acquisition
In the RSS values from j-th of AP that receive；K, m, i, j are positive integer；
Element all in k × m rank matrix column vectors that each reference point obtains is added to obtain one by step 2 two
Value, then this value divided by k, then each reference point is obtained for the vector of a 1 × m, for each reference point, the vector
The referred to as feature vector of the reference point, vector in j-th of element as the reference point j-th of feature；An if reference
The RSS values of certain AP can't detect on point, then is assigned a value of the minimum signal value -100dBm that can be received under the environment, therefore
And ranging from -100dBm≤v≤0dBm of the received signal strength RSS values v of arbitrary reference point；This group vector is used to implement step
Three cluster subregion.
The specific steps for being divided into Q sub-regions to indoor positioning environment according to K mean cluster algorithm described in step 3
For：
The feature vector for all reference points that step 3 one, input step two or two measure and subregion number Q；
Step 3 two, the RSS for obtaining choosing in data K reference point from step 2 two at random, i.e.,：The spy of each reference point
Levy cluster centre of the vector value as K sub-regions；
Step 3 three, the Euclidean distance for calculating each reference point and K cluster centre feature vector, by each reference point point
With the subregion for giving its Euclidean distance minimum；
Step 3 four averages the RSS values of reference point each in every sub-regions, obtains new cluster centre；
Step 3 five repeats step 3 three and step 3 four until the center of every sub-regions no longer changes；
Step 3 six obtains K sub-regions and the corresponding cluster centre vector of all subregion, i.e.,：The vector of one 1 × m,
The vector is referred to as the feature vector of this sub-regions, j-th of feature of j-th of element representation this sub-regions of the vector,
It is the RSS mean values from j-th of AP that this sub-regions obtains.
The detailed process of the category division of the random RSS data of step 4 is：
Step 4 one, input random acquisition Unlabeled data, the Unlabeled data only have signal strength values without
Location information；
Unlabeled data and the feature vector of each sub-regions that step 3 obtains are compared by step 4 two, that is, are asked
The distance with the feature vector of all subregion is taken, by random data distribution in the subregion closest with its feature vector,
As the classification belonging to it.
Dimensionality reduction is carried out using SDE algorithms to every sub-regions in K sub-regions described in step 5, determines each region
Eigentransformation matrix, and the specific method for generating new location fingerprint database is：
Step 5 one, construction adjacent map：
Directionless figure G and G' is constructed according to the class label information of high dimensional data point and its neighbor relationships；Wherein neighbor relationships
It is the criterion provided using KNN algorithms, that is, K point for selecting data point nearest works as x as its neighbour, G expressionsiWith xjCategory
Remember information yi=yjWhen and xi、xjK nearest neighbor relationship each other；G', which shows, works as xiWith xjClass label information yi≠yjWhen and xi、xjK each other
Neighbor relationships；
Step 5 two calculates weight matrix：
According to the adjacent map that step 5 one constructs, the calculating of weight matrix is carried out using class Gaussian function, expression formula is
(1) shown in：
Wherein：wijRepresent Neighbor Points xiWith xjBetween weights, | | xi-xj||2For Neighbor Points xiWith xjBetween norm away from
From using matrix-style calculating norm distance, t is weights normalized parameter；
Step 5 three determines object function and its solution：
According to the target of SDE algorithms：Divergence in class is minimized while maximizing class scatter；Divergence is similar using expression
And the norm distance at non-like numbers strong point represents；
Its corresponding optimization object function is obtained by the target of SDE algorithms, as shown in formula (2), eigentransformation matrix P is treats
The optimal solution asked：
According to the optimization object function that formula (2) provides：
The calculating formula of known matrix normThe calculating formula of the mark of the formula and matrix | | A | |2=tr (AAT)
Unanimously, therefore formula (2) is expressed as the mark of matrix：
Formula (3) is further simplified as：
It is real number by the calculating scalar nature and weights element of trace of a matrix, formula (4) is reduced to：
J (V)=2tr { PT[X(D′-W′)XT]P} (5)
Similarly, formula (2) is simplified to：
In formula (6), X is input high dimensional data, and W, W' are respectively G weight matrixs corresponding with G'；D and D' is diagonal matrix,
Its diagonal element is acquired by formula (7)：
To formula (6) using lagrange's method of multipliers, obtain shown in formula (8)：
X(D′-W′)XTP=λ X (D-W) XTP (8)
Generalized eigenvalue decomposition is carried out to formula (8), obtains eigenvalue λ=[λ of its Eigenvalues Decomposition1,λ2,…,λn]TAnd
Feature vector p=[p1,p2,…,pn]T；
Step 5 four, the estimation of intrinsic dimension：
The characteristic value and its feature vector being obtained according to step 5 three estimate intrinsic dimension according to formula (9)：
Wherein：η*It is the threshold value that projector space retains information, usual value is more than 80%, that is, d maximum feature before choosing
The ratio between the sum of value and All Eigenvalues summation meet the preferable low-dimensional insertion to primary data information (pdi) not less than 80%；
Step 5 five calculates embedded result：
Intrinsic dimension estimation threshold value is set according to step 5 four, the corresponding feature vector of d characteristic value of selection, which is formed, to be become
Change matrix P=[p1,p2,…,pd], calculating input high dimensional data point x by formula (10)iData Z after dimensionality reductioniFor：
Zi=PTxi (10)
The received signals fingerprint data of low-dimensional and eigentransformation matrix are obtained by SDE algorithms, are denoted as Radio Map respectively*And P.
Described in step 7 to every sub-regions in K sub-regions, be utilized respectively the low-dimensional Radio that step 5 acquires
Map*And eigentransformation matrix, use k neighbor positions fingerprinting localization algorithm to the specific method that test point is positioned for：
Step 7 two, the low-dimensional feature vector of test pointWith region low-dimensional Radio Map*In
I-th of reference pointThe distance between acquired by formula (11)：
Step 7 three chooses the k reference points closest with test point feature vector from small to large from result, by public affairs
Formula (12) calculates the location estimation coordinate of test point
Complete the positioning to test point.
The present invention, by using the Unlabeled data for being easy to acquisition, finds out characterization by introducing semi-supervised SDE dimension-reduction algorithms
The low dimensional manifold of the high dimensional data of location information efficiently reduces while the positioning accuracy for ensureing WLAN indoor locating systems
The calculation amount of position fixing process.Reduce the workload acquired with reference to point data simultaneously, the real-time update for database provides
Easy-to-use approach.The tuning on-line complexity of the present invention is low, and mobile terminal location is real-time.
Description of the drawings
Fig. 1 is the indoor scene schematic diagram described in the specific embodiment three of the present invention.
Specific embodiment
The positioning of the WLAN indoor orientation methods of semi-supervised SDE algorithms described in specific embodiment one, present embodiment
Process is：
Step 1: arrange m AP (AP for indoor environmentj, 1≤j≤m), it is ensured that any point is by two in the environment
The signal covering that a or more than two AP are sent out；
Step 2: being uniformly arranged reference point in environment indoors, choose a reference point and establish rectangular coordinate system for origin,
Coordinate position of each reference point in the rectangular coordinate system is obtained, and is acquired simultaneously using signal receiver in each reference point
It records received signal strength RSS values k times from each AP and carries out corresponding data processing；
It is the ginseng per sub-regions Step 3: indoor positioning environment is divided into Q sub-regions according to K mean cluster algorithm
Examination point marks respective classification information.The received signal strength RSS values of each reference point have similar in every sub-regions
Feature, i.e., the feature vector of each reference point are similar；
Step 4: the unmarked RSS data of acquisition at random is (with reference point difference lies in only signal strength values without position
Confidence ceases), the feature vector of each sub-regions obtained with step 3 is compared, that is, asks for the feature vector with all subregion
Distance, by random data category division in the subregion closest with its feature vector；
Step 5: to SDE algorithms are used in K sub-regions per sub-regions.Input parameter as SDE algorithms：It is intrinsic
The value of dimension (Intrinsic Dimensionality) divides Radio Map by existing intrinsic dimension algorithm for estimating
Area's data are estimated, provide the intrinsic dimension estimated value of each area data.Determine the eigentransformation matrix in each region, and
Generate location fingerprint database (the Radio Map after dimensionality reduction*)；
Step 6: the feature vector for each sub-regions that the signal strength RSS values that tested point obtains are obtained with step 3
Be compared, that is, ask for the distance of the feature vector of test point and the feature vector of all subregion, by test point be located in and its
In the closest subregion of feature vector；
Step 7: in the subregion positioned, using the eigentransformation matrix that step 5 obtains to the RSS values of tested point
Dimensionality reduction is carried out, obtains the RSS of low-dimensional*, with fingerprint database Radio Map*It is matched, is positioned and calculated using k nearest neighbor location fingerprint
Method is accurately positioned test point.
Specific embodiment two, present embodiment are the WLAN to the semi-supervised SDE algorithms described in specific embodiment one
The further explanation of indoor orientation method, being connect in each reference point using signal described in step 2 in specific embodiment one
Receipts machine acquires and records received signal strength RSS values k times from each AP and carry out the specific step of corresponding data processing
Suddenly it is：
Step 2 one obtains each reference point in one k × m rank matrix, and the i-th row jth row of matrix represent ith acquisition
In the RSS values from j-th of AP that receive；
Element all in k × m rank matrix column vectors that each reference point obtains is added to obtain one by step 2 two
Value, then this value divided by k, reference point each in this way is obtained for the vector of a 1 × m, for each reference point, this to
Amount be known as the reference point feature vector, vector in j-th of element (i.e. from APjThe signal strength RSS mean values of acquisition) it can be with
As j-th of feature of the reference point.Sometimes the RSS values of certain AP can't detect in a reference point, then is assigned
It is worth the minimum signal value -100dBm for that can be received under the environment, so the received signal strength RSS values v of arbitrary reference point
Ranging from -100dBm≤v≤0dBm.The cluster subregion that this group vector will be used to implement step 3.
Present embodiment provides fingerprint database sample for follow-up specific embodiment.
Specific embodiment three, present embodiment are the WLAN to the semi-supervised SDE algorithms described in specific embodiment one
The further explanation of indoor orientation method, in specific embodiment one described in step 3 according to K mean cluster algorithm to interior
Localizing environment be divided into Q sub-regions the specific steps are：
The feature vector for all reference points that step 3 one, input step two or two measure and subregion number Q；
Step 3 two, the RSS (spies of i.e. each reference point for obtaining choosing in data K reference point from step 2 two at random
Sign vector) cluster centre of the value as K sub-regions；
Step 3 three, the Euclidean distance for calculating each reference point and K cluster centre feature vector, by each reference point point
With the subregion for giving its Euclidean distance minimum；
Step 3 four averages the RSS values of reference point each in every sub-regions, obtains new cluster centre；
Step 3 five repeats step 3 three and step 3 four until the center of every sub-regions no longer changes；
Step 3 six, obtain K sub-regions and all subregion corresponding cluster centre vector (vector of i.e. one 1 × m,
The vector is referred to as the feature vector of this sub-regions, j-th of feature of j-th of element representation this sub-regions of the vector,
It is the RSS mean values from j-th of AP that this sub-regions obtains).
Present embodiment can guarantee carries out effective subregion to localizing environment, receives the reference point in every sub-regions
The signal strength RSS values from each AP, i.e., the feature vector similarity degree of a reference point is more than from two different sub-districts
The feature vector similarity of the reference point in domain, this also lays the foundation for the random RSS data category division in step 4.
It is tested in indoor scene shown in Fig. 1, possesses 19 laboratories, 1 meeting room and 1 table tennis room,Represent elevator, the material of wall is brick, aluminium alloy window and metallic door, and wireless access point AP is Linksys WAP54G-
CN, and with AP1, AP2 ..., AP27 indicate 1 to No. 27 AP, each AP is fixed on the position away from ground 2m height.Signal receiver
1.2m from the ground, the position that arrow mark is placed for 1 to No. 27 AP in figure, selects corridor as experimental place, i.e. net in figure
Trellis region, between neighboring reference point between be divided into 1m, totally 247 reference points.
It is connected and networked using the wireless network card of Intel PRO/Wireless 3945ABG network connection,
NetStumbler softwares on association's V450 notebooks are installed, acquire the signal strength RSS values from 27 access point AP；From
In the line stage, in four different orientation of all reference points, with 2/second sample frequencys, continuous sampling records 100 of AP
The relevant information of RSS values and AP.The physical coordinates of all reference points and RSS values are stored as what position fixing process was called
Data establish Radio Map.The RSS data of 580 points in random acquisition localization region, direction is random and with 2/second frequencies, often
A point sampling 10 seconds, taking mean value, these data tracer signal intensity values are without having as the Unlabeled data in SDE algorithms
Body position information.With Radio Map together as the input data of SDE algorithms.
Specific embodiment four, present embodiment are the WLAN to the semi-supervised SDE algorithms described in specific embodiment one
The further explanation of indoor orientation method, the specific mistake of the category division of the random RSS data of step 4 in specific embodiment one
Cheng Wei：
Step 4 one, the Unlabeled data for inputting random acquisition (only signal strength values are without location information)；
Unlabeled data and the feature vector of each sub-regions that step 3 obtains are compared by step 4 two, that is, are asked
The distance with the feature vector of all subregion is taken, by random data distribution in the subregion closest with its feature vector,
As the classification belonging to it.
Present embodiment can divide the classification belonging to the Unlabeled data of random acquisition, be the SDE algorithm structures in step 5
It builds total data adjacent map matrix and classification information is provided.
Specific embodiment five, present embodiment are the WLAN to the semi-supervised SDE algorithms described in specific embodiment one
The further explanation of indoor orientation method, in specific embodiment one described in step 5 to each sub-district in K sub-regions
Domain carries out dimensionality reduction using SDE algorithms, determines the eigentransformation matrix in each region, and generates new location fingerprint database and carry out
It illustrates：
SDE algorithms are based on a kind of maximized manifold learning arithmetic of divergence in class scatter and class.To SDE algorithms into
Input data is done before row theory analysis as described below：Input the marked data of higher-dimensionAnd Unlabeled dataThe two class data main distinctions are：The former corresponds to the data acquired at reference point, contains location information (xix,yiy)
And its class label yi∈{1,2,…,c}.High dimensional data is divided into c submanifold by wherein c expressions, the high dimension that will be inputted
According to being divided into c classes.Unlabeled data also obtains corresponding category information after step 4.
Step 5 one, construction adjacent map
Directionless figure G and G' is constructed according to the class label information of high dimensional data point and its neighbor relationships.Wherein neighbor relationships
It is the criterion provided using KNN algorithms, that is, K point for selecting data point nearest works as x as its neighbour, G expressionsiWith xjCategory
Remember information yi=yjWhen and xi、xjK nearest neighbor relationship each other；G', which shows, works as xiWith xjClass label information yi≠yjWhen and xi、xjK each other
Neighbor relationships.
Step 5 two calculates weight matrix
According to the adjacent map that step 5 one constructs, the calculating of weight matrix is carried out using class Gaussian function.Its expression formula is
(1) shown in.W in formula (1)ijRepresent Neighbor Points xiWith xjBetween weights, | | xi-xj||2For Neighbor Points xiWith xjBetween model
Number distance calculates norm distance using matrix-style, and t is weights normalized parameter.
Step 5 three determines object function and its solution
According to the target of SDE algorithms：Divergence in class is minimized while maximizing class scatter.Divergence is similar using expression
And the norm distance at non-like numbers strong point represents.Its corresponding optimization object function can be obtained by the target of SDE algorithms, such as formula
(2) shown in, eigentransformation matrix P is optimal solution to be asked.
In formula：Maximize is represented：It maximizes, subject to are represented：It obeys；
Analysis below is made according to the optimization object function that formula (2) provides：
The calculating formula of known matrix normThe calculating formula of the mark of the formula and matrix | | A | |2=tr (AAT)
Unanimously, therefore formula (2) can be expressed as the mark of matrix：
Formula (3) is further simplified as：
It is real number by the calculating scalar nature and weights element of trace of a matrix, formula (4) can be reduced to：
J (V)=2tr { PT[X(D′-W′)XT]P} (5)
Similarly, formula (2) is simplified to：
In formula (6), X is input high dimensional data, and W, W' are respectively G weight matrixs corresponding with G'.D and D' is diagonal matrix,
Its diagonal element can be acquired by formula (7)：
To formula (6) using lagrange's method of multipliers, it can be deduced that shown in formula (8)：
X(D′-W′)XTP=λ X (D-W) XTP (8)
Generalized eigenvalue decomposition is carried out to formula (8), obtains eigenvalue λ=[λ of its Eigenvalues Decomposition1,λ2,…,λn]TAnd
Feature vector p=[p1,p2,…,pn]T。
Step 5 four, the estimation of intrinsic dimension
Intrinsic dimension is the characteristic value retained during low-dimensional insertion and its number of corresponding feature vector.Feature vector corresponds to
Characteristic value is bigger, and the corresponding between class distance of the direction is bigger, also means that retained feature more has judgement index.Intrinsic dimension
Number d are an important parameters of algorithm, whether value estimation is accurate, determine low-dimensional data information contained amount number, so as to
Influence positioning accuracy.The characteristic value and its feature vector being obtained according to step 5 three estimate intrinsic dimension according to formula (9)：
Wherein η*It is the threshold value that projector space retains information, usual value is more than 80%, that is, d maximum eigenvalue before choosing
The sum of be not less than 80% with the ratio between All Eigenvalues summation, you can meet the preferable low-dimensional insertion to primary data information (pdi).
Step 5 five calculates embedded result
Intrinsic dimension estimation threshold value is set according to step 4, the corresponding feature vector of d characteristic value of selection forms transformation
Matrix P=[p1,p2,…,pd], calculating input high dimensional data point x by formula (10)iData Z after dimensionality reductioniFor：
Zi=PTxi (10)
The theory deduction of SDE algorithms is provided by formula (2)~(8).The received signals fingerprint number of low-dimensional can be obtained by SDE algorithms
According to and eigentransformation matrix, be denoted as Radio Map respectively*And P.
Specific embodiment six, present embodiment are the WLAN to the semi-supervised SDE algorithms described in specific embodiment one
The further explanation of indoor orientation method, in specific embodiment one described in step 7 to each sub-district in K sub-regions
Domain is utilized respectively the low-dimensional Radio Map that step 5 acquires*And eigentransformation matrix, using k neighbor positions fingerprinting localization algorithms
Positioning is carried out to test point to be specifically described：
Test point is located in subregion by step 7 one, step 6, and the RSS signals that test point receives are believed in real time for higher-dimension
Number, it is expressed as Rtest=[r1,r2,…,rn].Formula is utilized with the eigentransformation matrix P (having been obtained in step 5 five) in the region
(10) it is multiplied, calculates the signal value after dimensionality reduction
Step 7 two, the low-dimensional feature vector of test pointWith region low-dimensional Radio Map*(step
Obtained in rapid 5 5) in i-th of reference pointThe distance between can be acquired by formula (11)：
Step 7 three chooses the k reference points closest with test point feature vector from small to large from result, by public affairs
Formula (12) calculates the location estimation coordinate of test point
Complete the positioning to test point.
Claims (5)
1. based on the WLAN indoor orientation methods of semi-supervised SDE algorithms, realized by following steps：
Step 1: arrange m access point AP for indoor environment, it is ensured that any point is by two or two in the indoor environment
The signal covering that above wireless access point AP is sent out；M is positive integer；
Step 2: being uniformly arranged reference point in environment indoors, choose a reference point and establish rectangular coordinate system for origin, obtain
Coordinate position of each reference point in the rectangular coordinate system, and acquire and record using signal receiver in each reference point
Received signal strength RSS values from each AP k times, and carry out data processing；K is positive integer；
It is the reference point per sub-regions Step 3: indoor positioning environment is divided into Q sub-regions according to K mean cluster algorithm
Mark respective classification information；
Step 4: the unmarked RSS data that acquisition is random, the feature vector of each sub-regions obtained with step 3 is compared
Compared with the distance with the feature vector of all subregion being asked for, by random data category division closest with its feature vector
Subregion in；
Step 5: to using SDE algorithms in K sub-regions per sub-regions, eigentransformation matrix is obtained；
The input parameter of SDE algorithms, i.e., the value of intrinsic dimension, by existing intrinsic dimension algorithm for estimating to RadioMap points
Area's data are estimated, provide the intrinsic dimension estimated value of each area data, determine the eigentransformation matrix in each region, and
Generate the location fingerprint database Radio Map after dimensionality reduction*；
Step 6: the feature vector for each sub-regions that the signal strength RSS values that tested point obtains are obtained with step 3 carries out
Compare, that is, ask for the distance of the feature vector of test point and the feature vector of all subregion, test point is located in and its feature
In the nearest subregion of vector distance；
Step 7: in the subregion positioned, the RSS values of tested point are carried out using the eigentransformation matrix that step 5 obtains
Dimensionality reduction obtains the RSS of low-dimensional*, with fingerprint database Radio Map*It is matched, using k nearest neighbor location fingerprint location algorithm pair
Test point is accurately positioned；It is characterized in that, being acquired simultaneously using signal receiver in each reference point described in step 2
Record received signal strength RSS values k times from each AP, and carry out data processing the specific steps are：
Step 2 one obtains each reference point in one k × m rank matrix, and the i-th row jth row of matrix represent to connect in ith acquisition
The RSS values from j-th of AP received；K, m, i, j are positive integer；
Element all in k × m rank matrix column vectors that each reference point obtains is added to obtain a value by step 2 two, then
This value divided by k, then each reference point is obtained for the vector of a 1 × m, and for each reference point, which is known as
The feature vector of the reference point, vector in j-th of element as the reference point j-th of feature；If in a reference point
The RSS values of certain AP can't detect, then is assigned a value of the minimum signal value -100dBm that can be received under the environment, so appoint
Ranging from -100dBm≤v≤0dBm of the received signal strength RSS values v for reference point of anticipating；This group vector is used to implement step 3
Cluster subregion.
2. the WLAN indoor orientation methods of semi-supervised SDE algorithms according to claim 1, it is characterised in that described in step 3
Indoor positioning environment is divided into according to K mean cluster algorithm Q sub-regions the specific steps are：
The feature vector for all reference points that step 3 one, input step two or two measure and subregion number Q；
Step 3 two, the RSS for choosing K reference point in the data that step 2 two obtains at random, i.e.,：The feature of each reference point to
Cluster centre of the magnitude as K sub-regions；
Step 3 three, the Euclidean distance for calculating each reference point and K cluster centre feature vector, each reference point is distributed to
With the subregion of its Euclidean distance minimum；
Step 3 four averages the RSS values of reference point each in every sub-regions, obtains new cluster centre；
Step 3 five repeats step 3 three and step 3 four until the center of every sub-regions no longer changes；
Step 3 six obtains K sub-regions and the corresponding cluster centre vector of all subregion, i.e.,：The vector of one 1 × m, claiming should
Vector is the feature vector of this sub-regions, j-th of feature of j-th of element representation this sub-regions of the vector and this
The RSS mean values from j-th of AP that sub-regions obtain.
3. the WLAN indoor orientation methods of semi-supervised SDE algorithms according to claim 1, it is characterised in that step 4 is random
The detailed process of the category division of RSS data is：
Step 4 one, the Unlabeled data for inputting random acquisition, the Unlabeled data only have signal strength values without position
Information；
Unlabeled data and the feature vector of each sub-regions that step 3 obtains are compared by step 4 two, that is, ask for
The distance of the feature vector of all subregion, by random data distribution in the subregion closest with its feature vector, as
Classification belonging to it.
4. the WLAN indoor orientation methods of semi-supervised SDE algorithms according to claim 1, it is characterised in that described in step 5
Dimensionality reduction is carried out using SDE algorithms to every sub-regions in K sub-regions, determine the eigentransformation matrix in each region, and
The specific method for generating new location fingerprint database is：
Step 5 one, construction adjacent map：
Directionless figure G and G' is constructed according to the class label information of high dimensional data point and its neighbor relationships；Wherein neighbor relationships are to adopt
The criterion provided with KNN algorithms, that is, K point for selecting data point nearest work as x as its neighbour, G expressionsiWith xjClass label letter
Cease yi=yjWhen and xi、xjK nearest neighbor relationship each other；G', which shows, works as xiWith xjClass label information yi≠yjWhen and xi、xjK nearest neighbor each other
Relationship；
Step 5 two calculates weight matrix：
According to the adjacent map that step 5 one constructs, the calculating of weight matrix is carried out using class Gaussian function, expression formula is (1) institute
Show：
Wherein：wijRepresent Neighbor Points xiWith xjBetween weights, | | xi-xj||2For Neighbor Points xiWith xjBetween norm distance, adopt
Norm distance is calculated with matrix-style, t is weights normalized parameter；
Step 5 three determines object function and its solution：
According to the target of SDE algorithms：Divergence in class is minimized while maximizing class scatter；Divergence is similar and non-using expression
The norm distance at like numbers strong point represents；
Its corresponding optimization object function is obtained by the target of SDE algorithms, as shown in formula (2), eigentransformation matrix P waits to ask
Optimal solution：
According to the optimization object function that formula (2) provides：
The calculating formula of known matrix normThe calculating formula of the mark of the formula and matrix | | A | |2=tr (AAT) unanimously,
Therefore formula (2) is expressed as the mark of matrix：
Formula (3) is further simplified as：
It is real number by the calculating scalar nature and weights element of trace of a matrix, formula (4) is reduced to：
J (V)=2tr { PT[X(D′-W′)XT]P} (5)
Similarly, formula (2) is simplified to：
In formula (6), X is input high dimensional data, and W, W' are respectively G weight matrixs corresponding with G'；D and D' is diagonal matrix, right
Angle element is acquired by formula (7)：
To formula (6) using lagrange's method of multipliers, obtain shown in formula (8)：
X(D′-W′)XTP=λ X (D-W) XTP (8)
Generalized eigenvalue decomposition is carried out to formula (8), obtains eigenvalue λ=[λ of its Eigenvalues Decomposition1,λ2,...,λn]TAnd feature
Vectorial p=[p1,p2,...,pn]T；
Step 5 four, the estimation of intrinsic dimension：
The characteristic value and its feature vector being obtained according to step 5 three estimate intrinsic dimension according to formula (9)：
Wherein：η*It is the threshold value that projector space retains information, usual value is more than 80%, that is, the sum of d maximum eigenvalue before choosing
With the ratio between All Eigenvalues summation not less than 80%, that is, meet the preferable low-dimensional insertion to primary data information (pdi)；
Step 5 five calculates embedded result：
Intrinsic dimension estimation threshold value is set according to step 5 four, the corresponding feature vector of d characteristic value of selection forms transformation square
Battle array P=[p1,p2,...,pd], then calculate input high dimensional data point x by formula (10)iData Z after dimensionality reductioniFor：
Zi=PTxi (10)
The received signals fingerprint data of low-dimensional and eigentransformation matrix are obtained by SDE algorithms, are denoted as Radio Map respectively*And P.
5. the WLAN indoor orientation methods of semi-supervised SDE algorithms according to claim 4, it is characterised in that described in step 7
To every sub-regions in K sub-regions, be utilized respectively the low-dimensional Radio Map that step 5 acquires*And eigentransformation matrix,
Use k neighbor positions fingerprinting localization algorithm to the specific method that test point is positioned for：
Test point is located in subregion by step 7 one, step 6, and the RSS signals that test point receives are higher-dimension live signal,
It is expressed as Rtest=[r1,r2,...rn]；It is multiplied with the eigentransformation matrix P in the region using formula (10), after calculating dimensionality reduction
Signal value
Step 7 two, the low-dimensional feature vector of test pointWith region low-dimensional Radio Map*In i-th ginseng
Examination pointThe distance between acquired by formula (11)：
Step 7 three chooses the k reference points closest with test point feature vector from small to large from result, by formula
(12) the location estimation coordinate of test point is calculated
Complete the positioning to test point.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510020485.5A CN104540221B (en) | 2015-01-15 | 2015-01-15 | WLAN indoor orientation methods based on semi-supervised SDE algorithms |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510020485.5A CN104540221B (en) | 2015-01-15 | 2015-01-15 | WLAN indoor orientation methods based on semi-supervised SDE algorithms |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104540221A CN104540221A (en) | 2015-04-22 |
| CN104540221B true CN104540221B (en) | 2018-06-22 |
Family
ID=52855644
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510020485.5A Active CN104540221B (en) | 2015-01-15 | 2015-01-15 | WLAN indoor orientation methods based on semi-supervised SDE algorithms |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104540221B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105160690B (en) * | 2015-08-18 | 2018-05-25 | 武汉大学 | A kind of reference point recognition methods applied to the positioning of video projection target |
| CN106604217B (en) * | 2015-10-20 | 2019-11-05 | 北斗导航位置服务（北京）有限公司 | A kind of indoor orientation method based on error semi-supervised learning |
| CN106851550B (en) | 2015-12-04 | 2020-02-14 | 华为技术有限公司 | Method for positioning terminal and baseband unit |
| CN107423762A (en) * | 2017-07-26 | 2017-12-01 | 江南大学 | Semi-supervised fingerprinting localization algorithm based on manifold regularization |
| WO2019033282A1 (en) * | 2017-08-15 | 2019-02-21 | 深圳市大疆创新科技有限公司 | Positioning system and building method therefor |
| WO2019090527A1 (en) * | 2017-11-08 | 2019-05-16 | 上海诺基亚贝尔股份有限公司 | Indoor positioning method and apparatus using reconfigurable antenna |
| CN108307306A (en) * | 2017-12-26 | 2018-07-20 | 武汉创驰蓝天信息科技有限公司 | Region division and localization method based on space partition zone and system |
| CN108627798B (en) * | 2018-04-04 | 2022-03-11 | 北京工业大学 | WLAN indoor positioning algorithm based on linear discriminant analysis and gradient lifting tree |
| CN111328012B (en) * | 2018-12-14 | 2021-08-27 | 中国移动通信集团广东有限公司 | Stereo layered positioning method and device based on building WiFi |
| CN110972094B (en) * | 2019-12-06 | 2020-12-18 | 中国科学院测量与地球物理研究所 | L0 norm positioning method under indoor mixed sparse LOS/NLOS scene |
| CN111537950B (en) * | 2020-04-14 | 2023-04-21 | 哈尔滨工业大学 | Satellite position prediction tracking method based on position fingerprint and two-step polynomial fitting |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103476118A (en) * | 2013-09-29 | 2013-12-25 | 哈尔滨工业大学 | WLAN indoor location fingerprint positioning method used for real-time monitoring |
| CN103648106A (en) * | 2013-12-31 | 2014-03-19 | 哈尔滨工业大学 | WiFi indoor positioning method of semi-supervised manifold learning based on category matching |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP3794413B2 (en) * | 2003-02-28 | 2006-07-05 | セイコーエプソン株式会社 | Positioning system and positioning terminal |
-
2015
- 2015-01-15 CN CN201510020485.5A patent/CN104540221B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103476118A (en) * | 2013-09-29 | 2013-12-25 | 哈尔滨工业大学 | WLAN indoor location fingerprint positioning method used for real-time monitoring |
| CN103648106A (en) * | 2013-12-31 | 2014-03-19 | 哈尔滨工业大学 | WiFi indoor positioning method of semi-supervised manifold learning based on category matching |
Non-Patent Citations (1)
| Title |
|---|
| 《基于半监督流形学习的WLAN室内定位算法》;夏颖等;《系统工程与电子技术》;CNKI;20140731;第36卷(第7期);第1422-1426页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104540221A (en) | 2015-04-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN104540221B (en) | WLAN indoor orientation methods based on semi-supervised SDE algorithms | |
| CN106793086B (en) | Indoor positioning method | |
| CN106646338B (en) | A kind of quickly accurate indoor orientation method | |
| Niu et al. | WicLoc: An indoor localization system based on WiFi fingerprints and crowdsourcing | |
| He et al. | Tilejunction: Mitigating signal noise for fingerprint-based indoor localization | |
| CN105813194B (en) | Indoor orientation method based on fingerprint database secondary correction | |
| Dong et al. | A calibration-free localization solution for handling signal strength variance | |
| Roos et al. | A probabilistic approach to WLAN user location estimation | |
| CN103648106B (en) | WiFi indoor positioning method of semi-supervised manifold learning based on category matching | |
| CN109951807A (en) | Fusion RSS and CSI indoor orientation method based on WiFi signal | |
| Zou et al. | A mutual information based online access point selection strategy for WiFi indoor localization | |
| CN104581945B (en) | The WLAN indoor orientation methods of semi-supervised APC clustering algorithms based on distance restraint | |
| CN110557716A (en) | Indoor positioning method based on lognormal model | |
| CN109327797A (en) | Mobile robot indoor locating system based on WiFi network signal | |
| CN103596267A (en) | Fingerprint map matching method based on Euclidean distances | |
| Majeed et al. | Indoor localization and radio map estimation using unsupervised manifold alignment with geometry perturbation | |
| CN109041206A (en) | A kind of indoor positioning floor method of discrimination based on improvement fuzzy kernel clustering | |
| Li et al. | Location estimation in large indoor multi-floor buildings using hybrid networks | |
| Ali et al. | Harvesting indoor positioning accuracy by exploring multiple features from received signal strength vector | |
| CN108260076A (en) | Method, platform and the system of unmanned plane running orbit monitoring | |
| Bi et al. | A method of radio map construction based on crowdsourcing and interpolation for wi-fi positioning system | |
| Wei et al. | Efficient Wi-Fi fingerprint crowdsourcing for indoor localization | |
| Basiouny et al. | Enhancing Wi-Fi fingerprinting for indoor positioning system using single multiplicative neuron and PCA algorithm | |
| Zhang et al. | Dual-band wi-fi based indoor localization via stacked denosing autoencoder | |
| Wei et al. | Handling device heterogeneity in Wi-Fi based indoor positioning systems |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |