CN106851553A

CN106851553A - According to the soft self adaptation fingerprint positioning method for accepting and believing order

Info

Publication number: CN106851553A
Application number: CN201610875685.3A
Authority: CN
Inventors: 刘剑锋; 陈曦; 李泰聪; 杨铖; 林庆丰
Original assignee: Guangzhou Feng Shi Technology Co Ltd
Current assignee: Guangzhou Feng Shi Technology Co Ltd
Priority date: 2016-09-30
Filing date: 2016-09-30
Publication date: 2017-06-13
Anticipated expiration: 2036-09-30
Also published as: CN106851553B

Abstract

Present patent application is related to the wireless traffic in moving communicating field to support field, and in particular to a kind of according to the soft self adaptation fingerprint positioning method for accepting and believing order, including：The first step, offline drive test sampling and model training stage；Second step, tuning on-line stage, the 3rd step, location fingerprint data storehouse adaptive updates stage and SVR location model adaptive updates stages；By the self-renewing iterator mechanism in location fingerprint storehouse, will cause that the precision of whole localization method is improved constantly to a certain extent.

Description

According to the soft self adaptation fingerprint positioning method for accepting and believing order

Technical field

Field is supported the present invention relates to the wireless traffic in moving communicating field, and in particular to a kind of to accept and believe order according to soft Self adaptation fingerprint positioning method.

Background technology

In commercial application field, mobile communications network location technology is mainly used in location Based service (Location Based Service, LBS) aspect.Mobile positioning technique for the exploitation of its business application, at present only It is at the early-stage, except some specific tracking or monitoring purposes, waits further to open as civilian huge business potential Hair.According to the position of mobile subscriber, the respective services related to position are provided the user, such as user's positioning, service promotion will As the main trend in following mobile phone business.

Existing mobile terminal locating method mainly has following several：

(1) by satellite fix (with GPS as representative).

(2) by time of arrival (toa) (TOA), reaching time-difference, direction of arrival degree (AOA) or several indexs of the above The hybrid locating method for comprehensively using.

(3) cell ID+ Timing Advance (CellID+TA) localization method is used, i.e., using the service shared by mobile station The TA of location information of cell and Serving cell is positioned.

There is problems with existing mobile terminal location scheme：

(1) by satellite fix (with GPS as representative), (error exists can to obtain positioning precision higher in outdoor

10-50 meters).However, GPS location needs the mobile phone terminal to have GPS location function, typically only smart mobile phone is expired Foot, common mobile phone cannot meet；GPS cannot be positioned indoors, and GPS location is third-party soft based on smart mobile phone Part, network side is difficult to obtain the location information of user, even if obtaining the data of third party software, also relates to data deciphering etc. Step and be difficult to apply.

(2) by time of arrival (toa) (TOA), reaching time-difference, direction of arrival degree (AOA) or mixed positioning side Method, position error is typically all more than 200 meters, it is impossible to meet high-precision positioning requirements.

(3) cell ID+ Timing Advance (CellID+TA) localization method is used, i.e., using the service shared by mobile station The TA of location information of cell and Serving cell is positioned, and this mode burden increased to communication network is smaller, but fixed Position precision is relatively low.Also have in addition by measurement signal intensity, the technology that the method for distance is then calculated with propagation model to position, But due to the difference of wireless propagation environment, this method there is also the big problem of application condition.

In consideration of it, prior art proposes various fingerprint matchings based on SVR (support vector regression) for GSM network determining Position technology.Covered by the measurement report (Measurement Report, MR) and heterogeneous networks that are uploaded to base station to terminal device The sample fingerprint storehouse of cover area carries out correlation computations, obtains best match position.

With the evolution of mobile communications network, new LTE network starts popularization；However, after LTE network flattening, Measurement report (MR) data that originally can be directly gathered in 2,3G epoch, it is impossible to easily collect, this gives fingerprint positioning method Implementation in the lte networks causes obstruction；And the soft acquisition mode of signaling for passing through LTE, this can be got from main equipment A little MR data and Uu mouthfuls of related news；Therefore, realizing that fingerprint location technology becomes in the soft collection signalling analysis systems of LTE can Energy.

Now, be badly in need of be related to one kind, according to LTE it is soft accept and believe order carry out high accuracy fingerprint location method.

The content of the invention

For defect present in prior art, it is an object of the invention to provide a kind of according to the soft self adaptation for accepting and believing order Fingerprint positioning method, positioning precision can be effectively improved by the method.

The scheme that uses of the present invention for：Comprised the following steps according to the soft self adaptation fingerprint positioning method for accepting and believing order：First Step, offline drive test sampling and model training stage, specifically include：

(1) offline drive test acquisition phase：Drive test information is gathered by drive test terminal, measurement report is obtained；The measurement report Announcement includes the identification information and measurement report of involved cell in the position coordinates of drive test terminal, measurement report at it Field intensity data, delay data and direction angular data in each involved cell；

(2) the model training stage：Using each measurement report an as training sample, location fingerprint storehouse is built, and according to Location fingerprint storehouse generates tuning on-line model, including：

(2-1) the mobile network overlay area is divided into the grid that the length of side is L, and by grid numbering according to drive test end The position coordinates at end, training sample is assigned in corresponding grid；

(2-2) using grid numbering as be located at the grid training sample class indication, class indication and train sample Originally finger print data record is collectively constituted to be stored in location fingerprint storehouse；

(2-3) is processed the training sample in same grid using support vector regression algorithm, obtains each grid SVR location models, for calculate band position mobile terminal residing for longitude and latitude；

(2-4) is stored the SVR location models as tuning on-line model；

Second step, in the tuning on-line stage, specifically includes：

The measurement report of mobile terminal to be positioned is obtained by the soft collection of signaling, obtains involved small in measurement report Field intensity data of the identification information and measurement report in area in each cell involved by it, delay data and deflection number According to；

According to the identification information of cell involved in the measurement report, screening is related to identical small from location fingerprint storehouse The finger print data record in area；

Each finger print data that will be filtered out records the measurement of included training sample and online incoming mobile terminal Report calculates Euclidean distance；And therefrom screening obtains the minimum training sample of Euclidean distance, the grid according to corresponding to it is selected The grid is target grid；

The parameter that sample data under target grid and SVR location model off-line learnings are obtained, carries out convolutional calculation, obtains To final prediction longitude and latitude.

3rd step, location fingerprint data storehouse adaptive updates stage and SVR location model adaptive updates stages, specific bag Include：

The location fingerprint data storehouse adaptive updates stage：New measurement report is gathered according to step (1), and according to step (2-1) and (2-2) forms standby finger print data record；SVR location models according to residing grid quantify each finger print data The influence degree for location Calculation result is recorded, and marks each entry time of finger print data record；According to what is quantified Influence degree and entry time update location fingerprint storehouse, including：

According to the entry time that finger print data is recorded, the duration for screening storage exceedes the finger print data note of setting time section Record；

The influence degree quantified in the record that will be filtered out is rejected less than the record of setting value；

If the quantity of finger print data record supplements freshly harvested standby less than setting quantity in location fingerprint data storehouse Finger print data is recorded into fingerprint database；

The SVR location model adaptive updates stages, including：

If the finger print data record in a certain grid is removed and/or fills into, according to the finger after rejecting and/or filling into Line data record updates the SVR location models of the grid.

The beneficial effect of this programme is：This programme realizes being accurately positioned for mobile terminal by soft the adopting of signaling, is adapted to In being applied to current newest LTE communication network；

Grid is carried out by minimum Eustachian distance value and deletes choosing, implement simple and effective, high working efficiency；

Because SVR location algorithms output accuracy is influenceed larger, therefore is being reached a fixed number by the quality of sample data, quantity Need lasting to fill into meet quantitative needs before amount；Simultaneously need to (it be for positioning result according to the quality of sample data Quantization after influence degree) reject the less sample data of influence and update SVR location models, it is to avoid SVR location models have Excessive invalid computation and increase effectively calculating；Therefore by the self-renewing iterator mechanism in location fingerprint storehouse, will cause whole The precision of localization method is improved constantly to a certain extent.

Electromagnetic environment in other grid as in grid the change of the barrier such as building, trees and constantly occur Change, causes the sample data loss of accuracy of original collection, this method, the step of take continuous collecting new samples data, leads to Cross and periodically reject old sample data and insert freshly harvested sample data, it is ensured that the effect of location model is not because of grid There is excessive change and decline in lattice environmental factor.

By constantly updating fingerprint base so that fingerprint base sample reaches relatively preferred quantity, while by updating iteration The quality of sample data, accuracy is kept to maintain higher level, and then with the SVR location models produced by renewal every time more Accurately, so as to ensure that the positioning precision of whole method.

In sum, the self-renewing iterator mechanism that this programme passes through location fingerprint storehouse, will cause whole localization method Precision is improved constantly to a certain extent.

Optionally, in step (2), before building location fingerprint storehouse, the measurement report to collecting is screened, and will be met It is required that measurement report data be used for build location fingerprint storehouse；The standard is：

1. cell number involved in measurement report data is more than or equal to 4；

2. delay data numerical value is less than or equal to the maximum for using measurement report data standard to allow；

3. the cell involved by measurement report data is less than or equal to 3000 meters with the distance of the data sampling point.

So, it is ensured that each measurement report is reliable and carries enough information, it is ensured that exist in fingerprint base The literary data record of system quality.

Optionally, in step (2), before building location fingerprint storehouse, each measured value in measurement report is normalized Pretreatment, using the measurement report after treatment as training sample, builds location fingerprint storehouse.

So each dimension is balanced for the influence degree of prediction, will not be larger because of the absolute value of certain numerical value And more influences are produced on location Calculation；Each parameter value after balance causes that positioning result is more accurate.

Optionally, the rejecting step of fingerprint data record specifically also includes in the 3rd step：

Delete/retain for each finger print data record addition and mark；

And, recording mark of the influence degree less than setting value quantified in record will be filtered out according to entry time to delete Remove；

When the specified time for updating location fingerprint storehouse is reached, all marks are rejected from the location fingerprint storehouse is Finger print data record.

Optionally, the location fingerprint data storehouse adaptive updates stage in the third step, if in location fingerprint data storehouse The quantity of finger print data record is less than setting quantity, then when the specified time in arrival renewal location fingerprint storehouse, the new collection of supplement Standby finger print data record into fingerprint database.

Optionally, also including for each finger print data record addition version flag the step of, and in the third step The SVR location model adaptive updates stages, by being marked by version flag for the record of new addition；

When the specified time for updating location fingerprint storehouse is reached, by version flag and deletion/reservation grid of marker for judgment one Whether the finger print data record in lattice is removed and/or fills into, if it is, according to the finger print data after rejecting and/or filling into Record updates the SVR location models of the grid.

Such way, is convenient to be managed each fingerprint recording by database, and intelligence degree is high.

Optionally, the setting time section is 24 hours.

Such renewal frequency is enough to ensure that positioning precision, while also will not be that the system for performing this method brings excessive Live load.

Optionally, the sample sets quantity as 2000/grid.

Sample number is then positioned greatly more accurately, but also to pay the increased cost of location model average workout times, the setting Quantity causes that the maximum average workout times and average positioning precision of model reach a balance, and the application for being suitable for this method is real Apply.

Brief description of the drawings

Fig. 1 carries out the flow chart of the learning training of model to use SMO heuritic approaches in the embodiment of the present invention；

Fig. 2 is the flow chart in tuning on-line stage in the embodiment of the present invention；

Fig. 3 is the flow chart in SVR location model adaptive updates stages in the embodiment of the present invention；

Fig. 4 is prediction mean error and average workout times when being positioned using the embodiment of the present invention relative to every The curve map of the sample size set point change of individual grid.

Specific embodiment

The present embodiment, it is for the LTE network of China Mobile and in the way of computer program that plan implementation is soft in signaling Adopt in analysis system；

The scheme for using for：The first step, MR data samplings and model training stage, specifically include：

(1) MR data acquisition phases：Using ATU (the Auxiliary Test for meeting group of China Mobile standard Unit), that is, automatic road measuring instrument is tested, the collection of MR data is carried out, the Tool integration has GPS module, therefore, the MR data Except the field intensity data of each cell including cell identity information and involved by it, delay data and direction angular data Outward, the longitude and latitude of each sampled point is also included.

(2) the model training stage：By importing equipment, to the storage device of system host in the data that will be gathered in ATU It is interior；Qualified MR data are screened, the present embodiment employs following standard：

1st, involved cell numbers are more than or equal to 4；

2nd, TA numerical value is less than or equal to 37；

3rd, involved by data cell and the distance of data sampling need to be less than or equal to 3000 meters；

Using each by the qualified MR data of screening an as training sample, structure location fingerprint storehouse；And according to Location fingerprint storehouse generates tuning on-line model, including：

The mobile network overlay area is divided into the length of side and is the grid of L by (2-1), and grid is numbered, according to every The sampling point position coordinate of MR data, training sample is distributed in all corresponding grid；

(2-2) using grid numbering as be located at the grid training sample class indication, class indication and train sample Originally finger print data record is collectively constituted to be stored in location fingerprint storehouse.

The present embodiment uses PostgreSQL Database fingerprint bases, total primary word of every finger print data record sheet Section, be respectively：

areaid：Grid is numbered, Long types；

sampleid：Record number under same grid, Int types；

longitude：The longitude of this record sampling, Double types；

latitude：The latitude of this record sampling, Double types；

servingCellid：The identification information of Serving cell, Int types；

cellid0-7：7 identification informations of non-service cell (adjacent cell), can be sky, Int types；

servingCellRsrp:The identification information of Serving cell, corresponding field intensity data (level), Double types；

rsrp0-7：The corresponding field intensity data (level) in 7 non-service cell (adjacent cell) non-service cells, non-serving It is then sky, Double types that cell identification information is empty；

ta：With the time delay of Serving cell, Int types；

aoa：With the deflection of Serving cell, Int types.

(2-3) is processed the training sample in same grid using support vector regression algorithm (SVR), obtains each The SVR location models of grid, for calculating with the longitude and latitude residing for positioning mobile terminal；

Carrying out before processing, the way being more highly preferred to be by 10 finger print datas (servingCellRsrp, rsrp0-7, 10 vector dimensions needed for ta, aoa, i.e. algorithm) the standard formulated according to China Mobile of value be normalized, have The way of body is：By the value of servingCellRsrp, rsrp0-7 divided by 140, ta value divided by 37, aoa value divided by 360, Last algorithm input data is obtained, so each dimension is balanced for the influence degree of prediction.

(2-4) is stored the SVR location models as tuning on-line model；

The mapping that input data passes through core, realizes that SVR is returned in the space of latitude high；In further test assessment, It was found that Radial basis kernel function (RBF, also known as gaussian kernel function) shows preferable effect, therefore the positioning of middle use is calculated in this implementation SVR training is mapped using Radial basis kernel function in method.

For example, it is assumed that following sample：

4 samples, then 4 dimensions of each sample, the core convolution for working as the sample of j=1, i.e., first is output as：

Wherein (x_i-x₁)²It is inner product of vectors, σ values are 2 herein.

During architecture model, convolution is done to each training sample, ultimately forms kernel value matrixes,.

In the present embodiment, the insensitive regression estimates of ε are constructed under the latitude space high as follows：

Wherein ε is fault tolerance, and ξ represents distance of the actual sample beyond ε, and C is to discipline parameter as a warning, if with subscript * Represent that the point is located at tropic either above or below.

The present embodiment carries out the learning training of model using SMO heuritic approaches, matches the parameter of each sample, specifically For：

Above kernel value matrixes are represented with K, according to SVR models, in iterative process after i-th sample input, Error E between its prediction output f (recurrence result of calculation) and prediction output f and actual value y is calculated with as follows：

E_i=f_i-y_i (30)

Therefore, it is that each training sample matches optimal lagrange multipliers that the purpose of learning training is thenAnd b, to minimize E；E is specifically, the longitude that is gone out according to training sample data prediction of model in the present embodiment With the sampling longitude and latitude of the training sample recorded in latitude and finger print data record.

Because in heuristic SMO learning trainings, we only optimize two lagrange multipliers every time, have selected

Outer loop：

1. first in λ_iThe λ for violating KKT is found in ∈ (- C, 0) ∪ (0, C)_i, violate KKT conditions and be represented by：

|y_i-f_i|≠ε (31)

Wherein ε is traditionally arranged to be 10^-3。

2. λ is worked as_iAll λ in ∈ (- C, 0) ∪ (0, C)_iKKT conditions are not all violated, then travels through whole sample

The λ of middle violation KKT_i, violate KKT and be represented by：

So KKT conditions are followed until all samples.

Interior loop：

1. same in λ_iCorrespondence λ is found in ∈ (- C, 0) ∪ (0, C)_iUpdate the maximum λ of step-length_j, wherein updating

Step-length is to the maximum finds maximum Δ E：

2. target λ is obtained_i,λ_j, next do further iteration and update：

WhereinFor the λ that outer circulation and interior circulation are obtained_i,λ_j。

η=K_vv+K_uu-2K_uv (35)

Further update for the first time

λ_i=s^*-λ_j (38)

ForSituation further adjusted：

IfAnd | λ_i|≥Δ∧|λ_j| >=Δ, then：

IfAnd | λ_i|≥Δ∧|λ_j| >=Δ the condition is unsatisfactory for, then：

Wherein step is to take 1 when 0.

After above-mentioned (39) (40) differentiate adjustment, then do further editing, it is ensured that λ_jFall within feasible zone, its In：

Can then be obtained after editing：

Final updating threshold value b：

Work as λ_uAnd λ_vAll in the range of [- C, C], then：

B=b^u=b^v (43)

Work as λ_uAnd λ_vIn there is any one not fall when in the range of [- C, C], then：

Wherein：

In the present embodiment, the specific implementation of the algorithm is as shown in figure 1, herein, with alpha represent above it is every The vector that a corresponding to individual sample is constituted (its dimension is the number of training sample)；Initialization data first, according to all of Sample Establishing kernelvalue matrixes, then set alpha and b initial value be 0, and specifically search for including：

1non-bounded heuristic searches

I. traversal alpha first meets condition field (- C, 0) ∪ (0, C), and those points are called non-bounded, because those Point plays leading role in final prediction output.

Ii. secondly traversal violates the alpha of KKT conditions, on KKT conditions, formula (31) is specifically shown in, if being unsatisfactory for this in addition Judge, then jump out condition, return to alpha, b.

Iii. then find and update second maximum alpha of step-length, traveled through in non-bounded set, while depositing Current the first, the second alpha is store, to do further calculating below；

Iv. further more fresh target alpha, overall process correspondence formula (34)~(42) above, are SMO cores；

V. the alpha for obtaining above is done and is differentiated, judge whether it there are enough renewal step-lengths for the first time, second judgement Whether the alpha set of bounded is empty, be it is empty need not then continue renewal and go down, third time judges, same as above, right (39) in formula (34)~(42) are answered, (40), the 4th judgement still judges whether alpha there are enough step-lengths, so far, The heuristic renewal alpha of non-bounded complete an iteration；

Vi. final updating b values return to alpha and b simultaneously.

In this way, the complete non-bounded set of iteration, have updated the alpha for violating KKT, it is encapsulated in In nonBoundedLoop functions.

2.Bounded heuristic searches

Bounded heuristic searches are similar to non-bounded above, and difference is that the former is 0, in ± C The value of alpha each dimension is updated, result is played an important role due to during is supporting vector (i.e. above each dimension Alpha of the value within (- C, 0) ∪ (0, C) gathers), therefore each dimension value of alpha is 0, the category of ± C is secondary, but it updates Can still influence supporting vector that respective change occurs, therefore it still needs to do further traversal iteration, process is as follows：

I. traversal alpha first meets condition field 0, ± C, and those points be called bounded, and those points are exported in final prediction In play secondary role；

Ii. secondly traversal violates the alpha of KKT conditions, if being unsatisfactory for the judgement in addition, jumps out condition, returns Alpha, b；

Iii. then find and update second maximum alpha of step-length, traveled through in full sample set, while storage is current The first, the second alpha, to do further calculating below；

Iv. alpha is further updated, because the search to internal layer alpha is in bulk sample sheet, therefore without being judged again Whether search result set is empty, simultaneously because the alpha on bounded collection plays a secondary role in prediction, therefore not further Judge its whether have enough update (also some reason is among these, and the point search in bulk sample sheet, though renewal step-length not It is enough big, also cannot again change other points), corresponding formula (34)~(42) above of overall process；

V. final updating b values return to alpha and b simultaneously；

In this way, the complete bounded set of iteration, have updated the alpha for violating KKT, BoundedLoop is encapsulated in In function.

3. total traversal and judgement

Total traversal and Rule of judgment be：Highest step number and this knots modification of bulk sample, both are to be divided into again with relation, the latter： Whether knots modification number of times travels through bulk sample sheet, and both are or relation.

Because alpha is initialized as 0 vector in algorithm, after being traveled through by 1 time, judge whether to have traveled through bulk sample sheet, if It is no, then judges that its alpha changes whether number of times is 0, not for 0 traversal bounded collection until alpha knots modifications are 0, then turns To non-bounded collection traversal, finally when reaching highest step number, or whole samples have been traveled through and on non-bounded Knots modification is 0, jumps out circulation, output result.

In result output, because each grid needs each self-training once, so each grid is to that should have respective ginseng Number output, below for the citing of certain grid, the parameter list output of the grid is as shown in table 1, and parameter list includes 6 fields, its Middle Area_ID (grid numbering) is unlisted, and id is the numbering of each sample under the grid, alpha_lng, b_lng correspondence longitude The learning outcome of model, the learning outcome of alpha_lat, b_lat corresponding latitude model.

id	alpha_lng	b_lng	alpha_lat	b_lat
					1	-7.07E-07	113.1146114	-1.15E-06	23.11104412
2	-7.95E-07	113.1146114	1.72E-07	23.11104412
					3	-7.42E-07	113.1146114	-1.77E-06	23.11104412
4	-7.10E-07	113.1146114	-1.98E-07	23.11104412
					5	1.03E-07	113.1146114	-1.24E-06	23.11104412
6	8.93E-07	113.1146114	2.24E-07	23.11104412
					7	2.40E-07	113.1146114	5.48E-08	23.11104412
8	-1.19E-06	113.1146114	-1.74E-07	23.11104412
					9	-6.08E-07	113.1146114	2.93E-07	23.11104412
10	-5.92E-07	113.1146114	3.79E-07	23.11104412
					11	-5.95E-07	113.1146114	-1.12E-06	23.11104412
12	-7.36E-07	113.1146114	-1.20E-06	23.11104412
					13	2.45E-06	113.1146114	-4.03E-07	23.11104412
14	-1.13E-06	113.1146114	-1.53E-07	23.11104412
					15	3.91E-07	113.1146114	-1.07E-06	23.11104412
16	1.17E-06	113.1146114	-2.33E-07	23.11104412
					17	2.08E-07	113.1146114	-1.03E-05	23.11104412
18	3.32E-07	113.1146114	-1.17E-06	23.11104412
					19	9.36E-07	113.1146114	-3.25E-07	23.11104412
20	-1.62E-07	113.1146114	4.72E-07	23.11104412
					21	-4.82E-08	113.1146114	2.60E-07	23.11104412
22	4.07E-07	113.1146114	-1.04E-07	23.11104412
					23	-9.68E-07	113.1146114	1.49E-07	23.11104412
24	-6.53E-06	113.1146114	-1.20E-05	23.11104412
					25	2.14E-07	113.1146114	-1.17E-06	23.11104412

The parameter list example of the output of table 1

Second step, in the tuning on-line stage, specifically includes：

The real-time MR data of mobile terminal to be positioned are taken by the soft harvest of signaling, it is involved in the real-time MR data of acquisition Field intensity data in each cell involved by it of the identification information of cell and real-time MR data, delay data and direction Angular data；

According to the identification information of cell involved in the real-time MR data, screening is related to identical from location fingerprint storehouse The finger print data record of cell；

According to the field intensity data in each cell being related to, delay data and direction angular data, real-time MR data are calculated The Euler's distance recorded with each finger print data filtered out in previous step；So as to by the mobile terminal Primary Location to be positioned In grid in previous step where the minimum job data record of Euler's distance value；

Then, all finger print datas record and the parameter list under the grid are transferred, following operation is done：

(1) when all cell in training sample and examining report incoming in real time are identical

Table 2

As shown in table 2, it is assumed that packet incoming in real time is containing covering under 4 cell, and this 4 cell and the 7th samples Cell it is consistent, then the core convolution is directly handled as follows by rsrp, TA, AOA data：

Wherein x_{In real time}It is real-time incoming data, x₇It is the 7th sample data in fingerprint base, (- 0.06,0.5, -0.13, 0.5) it is x_{In real time}With x₇Corresponding rsrp data are subtracted each other and are obtained under 4 cells, and 0.2 is x_{In real time}With x₇TA data subtract each other and obtain, 0.3x_{In real time}With x₇AOA data subtract each other and obtain, inner product of vectors is done afterwards, obtain 0.6505, dx substitutions formula below is obtained：

(2) when sample is identical with data division cell incoming in real time, then take same section and subtract each other, different piece mends 0 pair Should, then subtract each other, then inner product is sought, core convolution is exemplified below：

Table 3

By table 3 it can be seen that data incoming in real time have three cells identical for the 1st with sample data, respectively： 6723891st, 6723893,6700472, for this 3, only need to directly subtract each other, and for the cell in real time data 6723895, without occurring in sample 1, then when subtracting each other, the correspondence position of sample 1 is mended 0, equally, cell in sample 1 6719203, do not have in real-time incoming data, then when subtracting each other, real-time incoming data correspondence position mend 0, while TA with AOA calculates as above part, is equally calculated as follows:

Wherein x_{In real time}It is real-time incoming data, x₁It is the 1st sample data in fingerprint base, (- 0.93, -0.05, -0.12, - 0.056,0.8887) it is x_{In real time}With x₁Corresponding rsrp data are subtracted each other and are obtained under 4 cells, and 0.4 is x_{In real time}With x₁Middle TA data are subtracted each other Obtain, 0.3 is x_{In real time}With x₁Middle AOA data are subtracted each other and are obtained, and inner product of vectors is done afterwards, obtain 1.924724, and dx is substituted into following public affairs Formula is obtained：

This is the convolutional calculation when sample is identical with data division cell incoming in real time.

(3) when sample is entirely different with incoming data cell in real time, then all take benefit 0 to subtract each other, calculating process with it is upper State identical, TA and AOA calculate as above part.

After all samples under core convolution traversal target grid, just form the kernel_value of target sample to Amount；Taken further according to formula (30) in the item and parameter list in the corresponding kernel_value of each sample under alpha_lng row Item do after multiplication read group total again plus b, you can obtain the predicted value output of longitude, the prediction output of dimension is also same road Reason, difference is that its parameter for using is classified as alpha_lat.

The location fingerprint data storehouse adaptive updates stage：Method of the timing in the first step gathers new MR data, and Form standby finger print data record；SVR location models according to residing grid quantify each finger print data record for positioning The influence degree of result of calculation, and for/reservation mark, version flag and the fingerprint are deleted in each finger print data record addition The entry time of data record, and update location fingerprint data storehouse according to deleting/retaining mark and entry time.

In the present embodiment, every finger print data record also includes three tag fields：

Version:Version number, integer types；

Delete:Delete/retain mark, Boolean types；

Updatetime：Entry time, timestamp without time zone types.

The renewal process of fingerprint data record is as follows in the stage：

According to the entry time that finger print data is recorded, the duration for screening storage exceedes the finger print data note of setting time section Record, setting time is 24 hours in the present embodiment；

Because the size of Alpha values represents influence degree size of the sample to prediction of result in SVR location models, On the basis of previous step, any one in Alpha values belonging to mark is less than 10^-7Sample to delete, i.e. Delete fields Value be ture；

Fingerprint data record updates once daily in the present embodiment, when reach specify daily renewal time point when, from fixed The value of all Delete fields is rejected in the fingerprint database of position for the finger print data of TURE is recorded；If now location fingerprint data When the quantity that the finger print data record of a certain grid is belonged in storehouse is less than 2000, then it is fixed to be added to standby finger print data record In the fingerprint database of position.

In the SVR location model adaptive updates stages, the stage is as shown in Figure 3 in the present embodiment：

Whether the Rule of judgment for updating SVR location models is：

Whether the data that Version fields are NULL there are in specific bit fingerprint base under the grid；

Whether the data that Delete field marks are TRUE there are in specific bit fingerprint base under the grid；

If the finger print data record under a certain grid meets the finger print data first represented in the grid of above-mentioned condition Record is updated, and then the SVR location models of the grid are updated according to the finger print data record after renewal；

The renewal of SVR location models, the same flow using shown in Fig. 1 finds optimized parameter, and output parameter again Table.

After renewal, the version flag of each finger print data record in the grid is set as new compared with current version time one-level Version, i.e. the value of Version fields is NUll, will be reset to 1, and the non-NULL of value of other Version fields is then Directly Jia 1, so, the finger print data note of multiple versions (value of Version fields is 1,2,3 ...) is there has been after repeatedly updating Record.

The locating effect of the present embodiment is illustrated below by certainty of measurement：Specifically, the length of side L for using grid is 100 meters, The sample number setting value of each grid is 2000, and as a result of principle is updated, the sample number of each grid is stepped up, finally It is maintained within 1500 to 2000.Parameter setting is as follows：In longitude SVR models：

Discipline parameter C=100 as a warning,

Maximum searches step number maxStep=10000,

σ values guassian_delta=0.055；

Latitude SVR model kinds are only not to be all using another σ guassian_delta=0.08；

4 1000 forecast sample position error statistics of table

As can be seen from Table 4, in 1000 forecast samples, the ratio in 10 error of meter is about 84%, 20 error of meter Interior ratio is 91%.

Table 5

In addition, defined herein prediction mean error is：

Longitude mean error is all longitude predicted values with the difference sum of actual longitude divided by forecast sample quantity；

Latitude mean error is all latitude predicted values with the difference sum of actual longitude divided by forecast sample quantity；

Prediction mean error=(latitude mean error+longitude mean error)/2.

The curve declined from left to right in Fig. 4 represents prediction mean error, and the curve for rising from left to right represents average instruction Practice the time, as shown in table 5 and Fig. 4, by changing the sample size setting value of each grid, make for the calculating time and The contrast of positioning precision is learnt, although sample size increase can make knot predict that mean error is smaller, but averagely to train Time is cost；And work as sample size control (1500,2000] in the range of when, average workout times are only 895 seconds, prediction Mean error has reached the level of 5.46663E-05；Sample size is promoted to (2000,3000] it is interior when, average workout times Surge to 2557, increased close to twice, prediction mean error only drops to 4.12434E-05 levels, and the decline for bringing is very It is limited；Therefore the selection in the present embodiment by sample size control (1500,2000] in the range of.

Above-described is only embodiments of the invention, and the general knowledge such as known concrete structure and characteristic is not made herein in scheme Excessive description, technical field that the present invention belongs to is all of before one skilled in the art know the applying date or priority date Ordinary technical knowledge, can know all of prior art in the field, and with normal experiment hand before the application date The ability of section, under the enlightenment that one skilled in the art can be given in the application, improves and implements with reference to self-ability This programme, some typical known features or known method should not implement the application as one skilled in the art Obstacle.It should be pointed out that for a person skilled in the art, on the premise of structure of the present invention is not departed from, can also make Go out some deformations and improvement, these should also be considered as protection scope of the present invention, these effects implemented all without the influence present invention Fruit and practical applicability.This application claims protection domain should be defined by the content of its claim, the tool in specification Body implementation method etc. records the content that can be used for explaining claim.

Claims

1. according to the soft self adaptation fingerprint positioning method for accepting and believing order, it is characterised in that comprise the following steps：The first step, offline road Sampling and model training stage are surveyed, is specifically included：

(1) offline drive test acquisition phase：Drive test information is gathered by drive test terminal, measurement report is obtained；In the measurement report

The identification information and measurement report of involved cell are at it in position coordinates, measurement report including drive test terminal Field intensity data, delay data and direction angular data in each involved cell；

(2) the model training stage：Using each measurement report an as training sample, location fingerprint storehouse is built, and according to positioning Fingerprint base generates tuning on-line model, including：

The mobile network overlay area is divided into the length of side and is the grid of L by (2-1), and grid is numbered, according to drive test end The position coordinates at end, training sample is distributed in all corresponding grid；

(2-2), using the class indication numbered as the training sample positioned at the grid of grid, class indication and training sample are common It is stored in location fingerprint storehouse with composition finger print data record；

(2-3) uses support vector regression algorithm（SVR）Training sample in same grid is processed, each grid is obtained SVR location models, for calculate band position mobile terminal residing for longitude and latitude；

(2-4) is stored the SVR location models as tuning on-line model；

Second step, in the tuning on-line stage, specifically includes：

The measurement report of mobile terminal to be positioned is obtained by the soft collection of signaling, involved cell in acquisition measurement report Identification information and, field intensity data of the measurement report in each cell involved by it, delay data and direction angular data；

According to the identification information of cell involved in the measurement report, screening is related to same cells from location fingerprint storehouse Finger print data is recorded；

Each finger print data that will be filtered out records the measurement report of included training sample and online incoming mobile terminal Calculate Euclidean distance；And therefrom screening obtains the minimum training sample of Euclidean distance, the grid according to corresponding to it selectes the grid Lattice are target grid；

The parameter that sample data under target grid and SVR location model off-line learnings are obtained, carries out convolutional calculation, obtains most Whole prediction longitude and latitude；

3rd step, location fingerprint data storehouse adaptive updates stage and SVR location model adaptive updates stages, specifically include：

The location fingerprint data storehouse adaptive updates stage：According to step（1）The new measurement report of collection, and according to step（2-1） With（2-2）Form standby finger print data record；SVR location models according to residing grid quantify each finger print data and record right In the influence degree of location Calculation result, and mark each entry time of finger print data record；According to the influence journey for quantifying Degree and entry time update location fingerprint storehouse, including：

According to the entry time that finger print data is recorded, the duration for screening storage exceedes the finger print data record of setting time section；

If the quantity of finger print data record supplements freshly harvested standby fingerprint less than setting quantity in location fingerprint data storehouse Data record enters fingerprint database；

The SVR location model adaptive updates stages, including：

If the finger print data record in a certain grid is removed and/or fills into, according to the fingerprint number after rejecting and/or filling into The SVR location models of the grid are updated according to record.

2. according to claim 1 according to the soft self adaptation fingerprint positioning method for accepting and believing order, it is characterised in that：Step（2） In, before building location fingerprint storehouse, the measurement report to collecting is screened, and satisfactory measurement report data is used for Build location fingerprint storehouse；The standard is：

3. according to claim 1 according to the soft self adaptation fingerprint positioning method for accepting and believing order, it is characterised in that：It is fixed building Before the fingerprint base of position, each measured value in measurement report is normalized pretreatment, using the measurement report after treatment as instruction Practice sample, build location fingerprint storehouse.

4. according to claim 1 according to the soft self adaptation fingerprint positioning method for accepting and believing order, it is characterised in that：In the 3rd step The rejecting step of middle fingerprint data record specifically also includes：

Delete/retain for each finger print data record addition and mark；

And, recording mark of the influence degree less than setting value quantified in record will be filtered out according to entry time to delete；

When the specified time for updating location fingerprint storehouse is reached, the finger that all marks are is rejected from the location fingerprint storehouse Line data record.

5. according to claim 1 according to the soft self adaptation fingerprint positioning method for accepting and believing order, it is characterised in that：In the 3rd step In the location fingerprint data storehouse adaptive updates stage, if the quantity of finger print data record is less than setting in location fingerprint data storehouse Fixed number amount, then when the specified time for updating location fingerprint storehouse is reached, supplement freshly harvested standby finger print data and record into fingerprint Database.

6. according to claim 4 according to the soft self adaptation fingerprint positioning method for accepting and believing order, it is characterised in that：Also including being The step of each finger print data record addition version flag；

And the SVR location model adaptive updates stages in the third step, by being entered by version flag for the record of new addition Line flag；

When the specified time for updating location fingerprint storehouse is reached, by version flag and deletion/reservation grid of marker for judgment one Finger print data record whether be removed and/or fill into, if it is, according to after rejecting and/or filling into finger print data record Update the SVR location models of the grid.

7. according to any one of claim 1-6 according to the soft self adaptation fingerprint positioning method for accepting and believing order, its feature exists In：The setting time section is 24 hours.

8. according to any one of claim 1-6 according to the soft self adaptation fingerprint positioning method for accepting and believing order, its feature exists In：The sample sets quantity as 2000/grid.