CN109581280A - The adaptive tuning on-line method, system and device of terminal - Google Patents
The adaptive tuning on-line method, system and device of terminal Download PDFInfo
- Publication number
- CN109581280A CN109581280A CN201811161901.3A CN201811161901A CN109581280A CN 109581280 A CN109581280 A CN 109581280A CN 201811161901 A CN201811161901 A CN 201811161901A CN 109581280 A CN109581280 A CN 109581280A
- Authority
- CN
- China
- Prior art keywords
- model
- matrix
- data
- terminal
- initial
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 55
- 230000003044 adaptive effect Effects 0.000 title claims abstract description 21
- 238000006467 substitution reaction Methods 0.000 claims abstract description 7
- 239000011159 matrix material Substances 0.000 claims description 78
- 238000012549 training Methods 0.000 claims description 36
- 230000008569 process Effects 0.000 claims description 11
- 238000012545 processing Methods 0.000 claims description 3
- -1 hidden node quantity Substances 0.000 claims description 2
- 238000004422 calculation algorithm Methods 0.000 abstract description 24
- 230000008901 benefit Effects 0.000 abstract description 5
- 238000012360 testing method Methods 0.000 description 11
- 230000004807 localization Effects 0.000 description 8
- 238000013528 artificial neural network Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 6
- 230000006870 function Effects 0.000 description 6
- 238000013507 mapping Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 5
- 238000013480 data collection Methods 0.000 description 4
- 238000009826 distribution Methods 0.000 description 4
- 238000002474 experimental method Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 241000208340 Araliaceae Species 0.000 description 2
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 2
- 235000003140 Panax quinquefolius Nutrition 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000008859 change Effects 0.000 description 2
- 238000004891 communication Methods 0.000 description 2
- 230000007423 decrease Effects 0.000 description 2
- 235000008434 ginseng Nutrition 0.000 description 2
- 230000006872 improvement Effects 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 230000006378 damage Effects 0.000 description 1
- 230000007812 deficiency Effects 0.000 description 1
- 238000009795 derivation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 235000013399 edible fruits Nutrition 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 230000035800 maturation Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000001568 sexual effect Effects 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
- 230000003442 weekly effect Effects 0.000 description 1
- 238000005303 weighing Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S5/00—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations
- G01S5/02—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using radio waves
- G01S5/0252—Radio frequency fingerprinting
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S5/00—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations
- G01S5/02—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using radio waves
- G01S5/0278—Position-fixing by co-ordinating two or more direction or position line determinations; Position-fixing by co-ordinating two or more distance determinations using radio waves involving statistical or probabilistic considerations
Abstract
The present invention provides a kind of adaptive tuning on-line method, system and device of terminal, this method comprises: the data set according to a certain amount of calibration fingerprint, establishes initial model by trained mode；The initial model is applied in different positioning terminals；Positioning terminal in use, acquires the signal data without calibration；The initial model in positioning terminal is updated by the signal data of the no calibration, model after being updated, the master mould before model substitution updates after the update, and then realizes that the continuous iteration of terminal updates.Technical solution of the present invention inherits the advantages of ELM algorithm, it is updated simultaneously using the iteration that the signal data of no calibration carries out model, to solve the imeliness problem and data problem of calibrating of model, and with the continuous renewal of model, the otherness feature of device end hardware will be protruded constantly, to which model will increasingly adapt to the terminal, positioning accuracy will be also continuously improved.
Description
Technical field
The present invention relates to the indoor positioning fields of terminal more particularly to one kind can use calibrated data set and not
The data set of calibration carries out adaptive location method, system and the device of accurate indoor positioning.
Background technique
In recent years, Internet of Things (Internet of Things, IoT) development is more and swifter and more violent, and location technology is as Internet of Things
One of key technology of net receives the extensive concern of researcher.GPS positioning system, the maturations such as Beidou satellite alignment system
Positioning system have been achieved with that coverage area is big, the higher outdoor positioning of precision.However since shielding of building and interior are multiple
Miscellaneous environment, GPS signal can not carry out effective indoor positioning.Thus, many indoor positioning technologies continue to bring out, such as based on low
The indoor positioning technologies of the short range, wireless signals such as power consumption bluetooth, the indoor positioning skill combined based on motion sensor or both
Art, wherein location fingerprint algorithm increases that hardware, positioning is at low cost, orientation range is wide without additional, without knowing the definite position of AP
It sets and positioning can be realized with transmission power.Currently, the research based on location fingerprint algorithm has become indoor positioning area research
Mainstream.
However existing fingerprinting localization algorithm still remains three critical issues: data scaling difficulty problem, model timeliness are asked
Topic and equipment variability issues.
(1) data scaling difficulty problem
Data scaling in location fingerprint localization method refers to: in off-line training step, needing to acquire multiple reference points
Training fingerprint, every group of fingerprint require the location information of addition current reference point.More accurately locating effect in order to obtain, usually
It needs under the premise of rationally distributed, minimizes the distribution spacing of reference point.The density of area to be targeted internal reference examination point setting
The duration and frequency of size and the acquisition of each reference point locations training sample data, determine the workload of data scaling.Reduce
The distribution spacing of reference point, the more crypto set that reference point can be made to be distributed are conducive to improve spatial signal properties to physical bit
The accuracy for setting mapping guarantees the accuracy of positioning.In large-scale indoor positioning region, the quantity of reference point will very
More, each reference point also needs to acquire multiple samples to reduce error；In order to obtain the position coordinates of reference point, it is also necessary to establish
The map in indoor positioning region marks the location information of all reference points on unified map.This off-line training process will
Huge workload is brought, with the increase of localization region, this workload also increases index.Meanwhile for having built up
Good location fingerprint database, complicated and changeable due to indoor environment, signal base station is likely to occur quantity increase and decrease, change in location etc.
Situation causes the mapping relations of signal characteristic and physical location in original database to make destruction.Therefore in order to guarantee location fingerprint number
According to the validity in library, the work of data scaling also needs periodically to carry out, this undoubtedly considerably increases the work of fingerprint positioning method again
It measures.
(2) model imeliness problem
Location fingerprint localization method based on RSS makes according to the RSS feature of localization region and the mapping relations of physical location
Obtaining this method can be by receiving RSS signal characteristic come estimated position information.The premise of this method is the mapping of training stage
Relationship and the mapping relations of positioning stage are consistent.However, indoor radio signal communication environments are complicated and changeable, same physical bit
Received RSS signal is set with high dynamic and randomness, this characteristic of signal makes above-mentioned mapping relations be difficult to stablize
Unanimously, position estimation is caused error occur.
(3) equipment variability issues
The off-line training step of fingerprint positioning method in position, it would be desirable to go acquisition position to refer to by certain terminal device
Then line constructs location fingerprint database, later in the tuning on-line stage, go acquisition positioning to refer to using terminal device to be positioned
Fingerprint compares in line, with database, to complete to position.In the process, the key for influencing positioning accuracy is offline
Whether training data and on-line prediction data meet same model distribution.However training equipment and positioning device overwhelming majority situation
Be it is different, positioning user it is different, positioning terminal equipment is also many kinds of.In this background, due to different terminals
The radio hardware scheme of equipment is difficult to unification, and different radio chips is different the definition of RSSI, so even not
Signal acquisition is carried out in the same place of synchronization with terminal, RSSI value may also be different.
Summary of the invention
Aiming at the problems existing in the prior art, the present invention propose a kind of adaptive tuning on-line method of terminal, system and
Device, to trained location model utilizes what is acquired in real time to be updated without nominal data in terminal, to solve above-mentioned ask
Topic.
The present invention proposes a kind of on-line study model that terminal is adaptive on the basis of existing fingerprinting localization algorithm, compares
In the prior art, technical solution of the present invention carries out the online of model without nominal data using real-time reception in increment method
It updates, greatly reduces staking-out work amount, simultaneously so that model increasingly adapts to terminal in continuous renewal, effectively avoids
Equipment variability issues, so that positioning accuracy be continuously improved.
In order to realize the incremental learning based on no nominal data, the present invention proposes improvement for equipment variability issues
Semi-supervised Incremental Learning Algorithm (Semi-supervised Online Sequential ELM, SOS-ELM).
Specifically, the present invention provides the following technical scheme that
On the one hand, the present invention provides a kind of adaptive tuning on-line methods of terminal, which comprises
Step 1, according to it is a certain amount of calibration fingerprint data set, initial model is established by trained mode；
The initial model is applied in different positioning terminals by step 2；
Step 3, the positioning terminal in use, acquire the signal data without calibration；
The initial model in step 4, the positioning terminal is updated by the signal data of the no calibration, according to
According to newly-increased sample data, model after being updated, the master mould before model substitution updates after the update；
Step 5 repeats step 3, step 4, so that model increasingly adapts to the positioning terminal after the update.
Preferably, the initial model that the step 1 is tracked are as follows:
β=K-1HTJT
In formula,
H indicates hidden layer output matrix, and β is expressed as output weight matrix, and T indicates output classification matrix；J=diag (1,
1 ..., 0,0) it is interim intermediate parameters matrix that, 1 number, which is flag data number l, K,.
Preferably, the step 1 further include: by the data set of a certain amount of calibration fingerprint, obtain the first of model
Beginning parameter, the initial parameter include β0、K0。
Preferably, by the data set of a certain amount of calibration fingerprint, output weight matrix β trained for the first time is obtained0
And K0Are as follows:
Preferably, the initial parameter of model is obtained, comprising:
Step 101 provides input weight a by random fashioniWith biasing biAssignment, i=1,2 ..., L；
Step 102 calculates initial matrix J0Laplacian MatrixAnd hidden layer output matrix H0；
Step 103, calculating matrix K0With initial output weight matrix β0。
Preferably, in the step 101, the input weight and biasing are hidden for the forward direction list with L hidden node
Layer neural network, output are as follows:
In formula, G (ai,bi, x) be i-th of hidden node output, β=[βi1,βi2,L,βim]TI-th of expression connection hidden
The weight of node layer and output node；RnIndicate the real number set of n dimension；aiIt is the input weight that random fashion provides, biThen for
The biasing that machine mode provides；L is the quantity of hidden node.
Preferably, the step 4 further comprises:
Step 401 is based on nominal data number l and non-nominal data number u in newly-increased sample, the parameter of more new model；
The parameter includes hidden layer output matrix, hidden node quantity, matrix J, wherein J=diag (1,1 ..., 0,0), 1 number
For flag data number l；
Step 402, the calculated result based on the step 401 updates output weight matrix β.
Preferably, it in the step 402, updates output weight matrix and carries out in the following manner:
In formula, βk+1Indicate that kth+1 time output weight matrix, λ are the constraint percentage contribution for reflecting Laplce's manifold
Weight coefficient.
On the other hand, the present invention also provides a kind of adaptive tuning on-line system of terminal, the system be can be set
In positioning terminal, the system comprises:
Signal acquisition module acquires the signal data without calibration in the positioning terminal use process；
Model memory module, at the beginning of the system use, storage is trained by the data set of a certain amount of calibration fingerprint
The initial model established of mode, and, the initial model is updated by the signal data of the no calibration, according to new
Increasing sample data, model after being updated, the master mould before model substitution updates after the update realize constantly changing to model
In generation, updates.
Preferably, the initial model are as follows:
β=K-1HTJT
In formula,
H indicates hidden layer output matrix, and β is expressed as output weight matrix, and T indicates output classification matrix；J=diag (1,
1 ..., 0,0) it is interim intermediate parameters matrix that, 1 number, which is flag data number l, K,.
Preferably, the initial model obtains the initial ginseng of model by the data set of a certain amount of calibration fingerprint
Number, the initial parameter includes β0、K0。
Preferably, the initial parameter battle array β0And K0It obtains in the following manner:
In formula, λ is the weight coefficient for reflecting the constraint percentage contribution of Laplce's manifold.
Preferably, acquisition output weight matrix β trained for the first time0And K0It is accomplished by the following way:
Input weight a is provided by random fashioniWith biasing biAssignment, i=1,2 ..., L；
Calculate initial matrix J0Laplacian MatrixAnd hidden layer output matrix H0；
Calculating matrix K0With initial output weight matrix β0。
Preferably, the input weight aiWith biasing biFor the forward direction neural networks with single hidden layer with L hidden node,
It is exported are as follows:
In formula, G (ai,bi, x) be i-th of hidden node output, β=[βi1,βi2,L,βim]TI-th of expression connection hidden
The weight of node layer and output node.
Preferably, the model memory module is in the following manner updated model:
Nominal data number l and non-nominal data number u in newly-increased sample is counted, then calculating matrix Jk+1,And
Hk+1；Then the output weight matrix β of update is calculatedk+1。
Preferably, the βk+1It is iterated in the following manner:
In formula, βk+1Indicate that kth+1 time output weight matrix, λ are the constraint percentage contribution for reflecting Laplce's manifold
Weight coefficient.
Another aspect, the present invention also provides a kind of adaptive tuning on-line device of terminal, described device includes one
Or multiple processing units；And
Internal storage location, wherein being stored with the computer instruction that can call and carry out operation by processor unit；
The computer instruction executes the adaptive tuning on-line method of terminal as described above.
Compared with prior art, technical solution of the present invention not only inherits the advantages of ELM algorithm, while using without calibration
The iteration that signal data carries out model updates, to solve the imeliness problem and data problem of calibrating of model, and with mould
The otherness feature of the continuous renewal of type, device end hardware will be protruded constantly, so that model will increasingly adapt to the terminal,
Positioning accuracy will also be continuously improved.
Detailed description of the invention
In order to more clearly explain the embodiment of the invention or the technical proposal in the existing technology, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is only this
Some embodiments of invention for those of ordinary skill in the art without creative efforts, can be with
Other attached drawings are obtained according to these attached drawings.
Fig. 1 is single hidden layer feedforward network SLFN of the embodiment of the present invention；
Fig. 2 is the increment type localization method schematic diagram for equipment variability issues of the embodiment of the present invention；
Fig. 3 is the SOS-ELM algorithm flow chart of the embodiment of the present invention；
Fig. 4 is that positioning accuracy of each algorithm of the embodiment of the present invention under different terminals location data collection compares；
Fig. 5 a be the embodiment of the present invention over time, SOS-ELM positioning accuracy situation of change；
Fig. 5 b is the error distance of the embodiment of the present invention when being 5m, and the SOS-ELM positioning accuracy of different time compares；
Fig. 6 a is that the red X mobile phone test data set of the embodiment of the present invention compares under different terminals model；
Fig. 6 b is that the Vi X mobile phone test data set of the embodiment of the present invention compares under different terminals model.
Specific embodiment
The embodiment of the present invention is described in detail with reference to the accompanying drawing.It will be appreciated that described embodiment is only this
Invention a part of the embodiment, instead of all the embodiments.Based on the embodiments of the present invention, those of ordinary skill in the art exist
All other embodiment obtained under the premise of creative work is not made, shall fall within the protection scope of the present invention.
Those skilled in the art should know it is further that following specific embodiments or specific embodiment, which are the present invention,
The set-up mode of series of optimum explaining specific summary of the invention and enumerating, and being between those set-up modes can be mutual
In conjunction with or it is interrelated use, unless clearly proposing some of them or a certain specific embodiment or embodiment party in the present invention
Formula can not be associated setting or is used in conjunction with other embodiments or embodiment.Meanwhile following specific embodiment or
Embodiment is only as the set-up mode optimized, and not as the understanding limited the scope of protection of the present invention.
Embodiment 1:
Below with a specific embodiment to half utilized in adaptive tuning on-line method proposed in the present invention
The principle and specific example for supervising the very fast learning machine of increment type (SOS-ELM) are described in detail.
In conventional mode, the method based on very fast learning machine (ELM) can be completely suitable for indoor positioning field, basic
Principle is referring to Fig. 1.The basic principle of ELM is as follows:
ELM (Extreme Learning Machine) is quick by one kind of Nanyang Technolohy University Huang Guangbin et al. proposition
Efficient machine learning method.ELM belongs to the scope of artificial neural network, is before one kind to neural networks with single hidden layer (Single
Hidden Layer Feed-forward Neural Network,SLFN).It is short with the training time, network structure is simple etc.
Feature.For an input vector x ∈ RnWith the SLFN with L hidden node, output form is defined as following formula:
Wherein ai=[ai1,ai2,L,ain]TIt indicates the weight of input node and i-th of hidden node, indicates i-th of hidden layer
The offset parameter of node, G (ai,bi, x) be i-th of hidden node output, β=[βi1,βi2,L,βim]TIndicate i-th of connection
The weight of hidden node and output node.
The activation primitive (Activation Function) for choosing hidden node is g (x), then:
G(ai,bi, x) and=g (ai·x+bi),ai∈Rn,bi∈R (2)
For N number of training sample data collection of inputWherein, xj
∈RnRepresent the signal characteristic vector of input, tj∈RnThe corresponding physical coordinates of representative sample point, then have:
G(ai,bi,xj)=g (ai·xj+bi),ai∈Rn,bi∈ R, j=1,2, L, N (3)
To sum up, it is expressed in matrix as
According to formula (1), can obtain
H β=T (5)
This formula indicates that the SLFN of building will be with zero error training pattern parameter.
According to document, parameter ai,biIt, only need to be in initial stage random assignment without adjusting in the training process.Therefore, H and
Known to T, therefore β is parameter to be solved.Above-mentioned equation can be considered as a linear system, and solving above-mentioned equation can be by solving this
The least mean-square error of system is realized, that is, is equivalent to solution formula (6):
It solves:
Wherein,It is the generalized inverse matrix of H.
For ELM algorithm, need in advance using the terminals such as mobile phone acquire a collection of location position signal data sample set (X,
), T then off-line training obtains hidden layer output weight matrix β, and then in the on-line prediction stage, terminal is adopted in position to be positioned
Collect signal characteristic data X*, the positioning result of prediction can be calculated using T=H β.And semi-supervised very fast learning machine (SS-ELM) and
The very fast learning machine of increment type (OS-ELM) is respectively from different perspectives optimized ELM.On the one hand, in actual environment, pass through
The cost of the signal data sample set of artificial acquisition calibration is very high, and the signal data that do not demarcate is readily available, Yong Hu
In localization region, as long as being positioned using mobile phone positioning software, so that it may to collect the signal data of no calibration simultaneously.For
This, SS-ELM makes full use of the structural smooth feature of training data, and the data of no calibration are added to training sample and are concentrated, rich
Rich training sample set, improves the precision of model.On the other hand, since wireless signal has high dynamic on a timeline
Feature, therefore the timeliness in RSSI sample fingerprint library will affect the positioning accuracy of model.For this reason, OS-ELM passes through newly-increased training sample
Notebook data is persistently iterated update to model, and the timeliness of model has been effectively ensured, to improve positioning accuracy.On however
Three kinds of face model still has the following problems.
SS-ELM algorithm does not solve the imeliness problem of model, and over time, positioning accuracy will constantly decline；
When using terminal positionings such as mobile phones, there are the variability issues of equipment, and SS-ELM algorithm does not consider this point, model instruction
After white silk, different terminals use identical model, therefore even if in same position, since different terminal equipment acquires
Wireless signal data it is inconsistent, it is larger to also result in positioning result otherness, and positioning confidence level is poor；Although OS-ELM being capable of benefit
The data more new model acquired with terminal makes model gradually adapt to the terminal, but the acquisition of the incremental data of this calibration at
This is excessively high, does not have actual availability.
To solve the above-mentioned problems, the invention proposes the very fast learning machine of semi-supervised increment type (SOS-ELM), the learning machines
The advantages of not only inheriting ELM algorithm, while being updated using the iteration that the signal data of no calibration carries out model, to solve mould
The imeliness problem and data problem of calibrating of type, and with the continuous renewal of model, the otherness feature of device end hardware
It will constantly protrude, so that model will increasingly adapt to the terminal, positioning accuracy will be also continuously improved.
It is as shown in Figure 2 for the increment type indoor orientation method principle of equipment variability issues.Specifically, this method mentions
The adaptive tuning on-line process of terminal out are as follows: first according to a batch calibration finger print data training initial model, then by model
It being applied in different positioning terminals, each terminal can collect the signal data of no calibration during positioning service in real time,
Terminal model will carry out online incrementally updating to itself without demarcation signal data using this, and over time, model will
Present terminal is increasingly adapted to, to have terminal adaptive performance.
This method proposes the semi-supervised very fast learning machine (SOS- of increment type by modification SS-ELM algorithm combination increment type frame
ELM).The specific algorithm of SS-ELM is as follows:
SS-ELM is a kind of extended method based on ELM, it mentions calibration sample in conjunction with not demarcating sample and train together
Forecast result of model is risen, main target is to make up marker samples deficiency bring shadow by excavating the information of unmarked sample
It rings.In order to make SS-ELM obtain better precision of prediction compared to ELM, while over-fitting is avoided, generalization ability is improved, according to structure
Risk minimization is theoretical, and model needs weighing apparatus of making even between empiric risk and the complexity of learning function f.This method introduces figure
(Graph) Laplce (Laplacian) operator does manifold constraint, according to document, is come using the smoothness function S (f) of figure
Indicate model complexity, is defined as:
Wherein, LαRepresent the Laplace operator of figure.According to document,
Wherein l is the data volume of calibration, and u is the data volume that do not demarcate.In order to consider empiric risk and learning function simultaneously
Complexity, loss function may be expressed as:
It is easy to calculate, it enables J=diag (1,1, L, 0,0), 1 number is the data volume l of calibration, and 0 number is not mark
Determine data volume u, following formula can be obtained by H β=f:
To β derivationOptimal solution, which can be obtained, is
β=(HTJH+λHTLH)-1HTJT (12)
λ is set to 0 in formula (12), (10), it is meant that nonstandard random sample notebook data is ignored, then β degenerates for the knot of formula (7)
Fruit.
And incremental learning method (OS-ELM) is updated prediction model using most newly arrived data, it can more preferable body
The structural adjustment of existing prediction model is in order to obtain increment type real-time learning effect, can be with for the training data X* newly increased
It is combined into a new training set with old training sample, again training according to conventional location fingerprint calculation method
The problem of journey acquisition model, this way, which is to compute repeatedly legacy data, will cause the Space-time Complexity of model to be continuously increased,
The low of Model Practical has gone amendment for this reason, this method introduces a kind of increment type ELM learning method, with the Δ β that new data X* is contributed
Some training pattern parameter betas0, so that new model parameter β * is obtained, such as following formula:
β *=β0+Δβ(X*) (13)
Specifically, according to formula (7), it is assumed that existing learning systemWhereinWhen new
Increase N1A sample point data setThen β is solved by following formula:
Solution can obtain:
Wherein
Accordingly, β1Iterative formula can indicate as follows:
The structure of the positive meeting formula of above formula (13), β1Calculating be based on β0, it is not necessary to all initial data are re-started into meter again
It calculates, so that calculation amount is greatly reduced, in addition,Part embodies newly-increased sample information (H1,T1) to instruction
Practice the amendment of model
In solution procedure,For figure Laplce commonly used in the art, Jk+1、Hk+1Then calculated with J, H above-mentioned
Mode is identical, and calculation conventional in the art can also be used certainly and obtain.
It is repeatedly updated in calculating to enable above formula to be used in, needs to be extended to recursion:
The very fast learning machine of semi-supervised increment type (SOS-ELM) concrete methods of realizing in the present embodiment is as follows:
For convenience of calculation, K=(H is enabledTJH+λHTLH), then formula (12) is converted into
β=K-1HTJT (20)
It is now assumed that initial a batch training sample set isIncluding calibration sample (xi,ti) with
Sample set x ' is not demarcatediIn order to minimize training error, output weight matrix trained for the first time can be obtained by formula (12) are as follows:
Wherein,
Assuming that obtaining the new sample data set of a batch nowIt wherein also include calibration sample
Sample is not demarcated according to SS-ELM, and in the case where merging the data set of old sample and new samples, new output weight matrix can be indicated
Are as follows:
Wherein,
For formula (23), by above it will be appreciated that, J0,J1It is the diagonal matrix that element is 0 or 1, is easy to get；T0,T1It is sample
The position coordinates matrix of collection, if it is non-nominal data collection, position coordinates are unknown, in order to meet the dimension requirement of matrix operation,
It needs arbitrarily to fill coordinate value so thatMatrix is (N0+N1) m dimension, due to the effect of matrix J, the coordinate value of filling is simultaneously
It will not influence prediction result；H0,H1And can directly calculate solve and for formula (24), occur the old sample data of fusion and
The parameter of new samples dataThese require us to need to calculate old sample data again, this and we
Analytical mathematics in OS-ELM are disagreed, and will cause the waste of time and computing resource to this, from formula (8)s, by most
Smallization Mean square error loss function is punished plus smoothness to calculate output weight matrix, therefore is ignoredItem is right
Final result will generate faint influence but will greatly reduce calculation amount, at this time, formula (24) conversion are as follows:
Formula (25) is brought into formula (23), K is solved to obtain1Expression formula:
In addition,
By formula (26), (27) substitute into formula (22), obtain β1It is iterative:
So far, it is similar with OS-ELM to have obtained the increment type model based on semi-supervised sample shaped like formula (13) for we, on
Formula can be extended to recursive expression:
By after original training set carries out model training, not marking for newly-increased calibration every time or in above-mentioned model
Random sample notebook data carries out the real-time amendment of model by SOS-ELM method, because only calculating newly-increased sample data, institute every time
It is greatly improved with training speed, meanwhile, this method both ensure that the timeliness of model, and greatly reduce the work of data scaling
Make
To sum up, learning algorithm SOS-ELM very fast to semi-supervised increment type can be as shown in Figure 3, and training process is summarized as follows: right
In a collection of training dataset, including nominal data collection { (xi,ti)|xi∈Rn,ti∈Rm, i=1,2 ..., Nl, non-nominal data
Collect { x 'i|x′i∈Rn, i=1,2 ..., Nu, system parameter is determined, such as hidden node number L, smoothness penalty coefficient λ, activation
Functional form g (x) (as select " rbf ") and to practical problem, the number of hidden nodes L of ELM generally take fully it is more than enough can (such as
1000), this method experiment takes L=1000.Next, training process is classified into two steps:
(1) initial stage: a small amount of initial training sample set is utilizedTo calculate output weight
Matrix β0And K0, it is specific as follows:
A) input weight a is given by random fashioniWith biasing biAssignment, i=1,2 ..., L；
B) initial matrix J is calculated0, the Laplacian Matrix of figureAnd hidden layer output matrix H0；
C) calculating matrix K0With initial output weight matrix β0；
(2) incremental stages: when newly-increased a collection of sample dataWhen, by iterative calculation output power
Weight βk+1, it is specific as follows:
A) nominal data number l and non-nominal data number u in newly-increased sample is counted, then calculating matrix Jk+1,With
And Hk+1；
B) the output weight matrix β updated is calculated using formula (29)k+1.
Plan-validation:
In order to verify superiority and inferiority of the different Fingerprint Model algorithms on actual location data set, choose ELM, BP, SVM algorithm with
SOS-ELM algorithm compares experiment using 1810 valid data of acquisition in first week as the initial training of SOS-ELM model
Sample, while whole training samples as non-incremental model, wherein nominal data amount and without nominal data amount ratio be 1:1.
Model parameter λ value 0.45. chooses data of 4 kinds of terminals from the 2nd week to the 6th week in the incremental stages of SOS-ELM model respectively
Incremental learning is carried out, 4 kinds of different SOS-ELM models are finally trained.Test data set chooses the 7th week data, wherein
Magnificent X mobile phone totally 338 valid data, red x mobile phone 285, evil spirit X mobile phone totally 339, Vi X mobile phone totally 340.Different terminals
The test data set of SOS-ELM model selection counterpart terminal carries out test experimental result as shown in Figure 4 and is analyzed from Fig. 4,
SOS-ELM algorithm whole positioning accuracy in each terminal is better than other algorithms, and especially when error distance is 3~5m, advantage is more
Add obvious
There is timeliness sexual clorminance in order to verify the SOS-ELM algorithm proposed in the present invention, by comparing different time SOS-
The positioning accuracy of ELM is tested.Choose red X mobile phone terminal, to the data that acquire weekly generate a SOS-ELM model its
In, initial training sample is identical as above-mentioned sample.Incremental portion chooses the 2nd week to currently week collected data, such as the 1st week
The incremental data of SOS-ELM model is sky, and the incremental data of the 2nd week model is the data of acquisition in the 2nd week, the increasing of the 3rd week model
Measure data be the 2nd, 3 week acquire data, and so on test data experimental result identical as previous experiment such as Fig. 5 a, 5b institute
Show.Known to analysis, over time, the whole positioning accuracy of SOS-ELM model is constantly promoted；It is 5m in error distance
When, the model orientation precision improvement from the 1st week to the 6th week is particularly evident, and the timeliness for effectively demonstrating SOS-ELM algorithm is excellent
Gesture.
Equipment variability issues are efficiently solved in order to verify SOS-ELM algorithm proposed by the present invention, by different terminals
SOS-ELM model using the test data set of non-terminal carry out test experiment have chosen Vi x terminal SOS-ELM model and
The SOS-ELM model of red X mobile phone terminal, by test data set cross-beta experimental result such as Fig. 6 a, 6b institute of two terminals
Show.Known to analysis, locating effect of the particular terminal model on other terminal test data sets is obviously not so good as in this terminal test
Locating effect on data set, this is because the SOS-ELM model of different terminals has otherness, particular terminal model can more reflect
The characteristic distributions of the data of particular terminal acquisition.In practical application, terminal can will be real in positioning software when using positioning service
When acquire carry out online incremental learning without nominal data, as time goes by, the location model of terminal will be increasingly able to
The characteristics of reflecting terminal wireless communication hardware, to solve the problems, such as equipment otherness.
Embodiment 2:
In a further embodiment, technical solution of the present invention can also be realized by way of system or device.It needs
It is noted that those systems or device are only a kind of preferable embodiments, the conventional structure that carries out on this basis or
Deformation, dismantling or the combination of person's module, should all be contemplated as falling within protection scope of the present invention.Also, following embodiment
In system perhaps device can be realized or carry and runs method as described in example 1 above, to realize the adaptive of terminal
Tuning on-line.
On the one hand, the present invention also provides a kind of adaptive tuning on-line system of terminal, the system be can be set fixed
In the terminal of position, the system comprises:
Signal acquisition module acquires the signal data without calibration in the positioning terminal use process；
Model memory module, at the beginning of the system use, storage is trained by the data set of a certain amount of calibration fingerprint
The initial model established of mode, and, the initial model is updated by the signal data of the no calibration, according to new
Increasing sample data, model after being updated, the master mould before model substitution updates after the update realize constantly changing to model
In generation, updates.
Preferably, the initial model are as follows:
β=K-1HTJT
In formula,
H indicates hidden layer output matrix, and β is expressed as output weight matrix, and T indicates output classification matrix；J=diag (1,
1 ..., 0,0) it is interim intermediate parameters matrix that, 1 number, which is flag data number l, K,.
Preferably, the initial model obtains the initial ginseng of model by the data set of a certain amount of calibration fingerprint
Number, the initial parameter includes β0、K0。
Preferably, the initial parameter battle array β0And K0It obtains in the following manner:
In formula, λ is the weight coefficient for reflecting the constraint percentage contribution of Laplce's manifold.
Preferably, acquisition output weight matrix β trained for the first time0And K0It is accomplished by the following way:
Input weight a is provided by random fashioniWith biasing biAssignment, i=1,2 ..., L；
Calculate initial matrix J0Laplacian MatrixAnd hidden layer output matrix H0；
Calculating matrix K0With initial output weight matrix β0。
Preferably, the input weight aiWith biasing biFor the forward direction neural networks with single hidden layer with L hidden node,
It is exported are as follows:
In formula, G (ai,bi, x) be i-th of hidden node output, β=[βi1,βi2,L,βim]TI-th of expression connection hidden
The weight of node layer and output node.
Preferably, the model memory module is in the following manner updated model:
Nominal data number l and non-nominal data number u in newly-increased sample is counted, then calculating matrix Jk+1,And
Hk+1；Then the output weight matrix β of update is calculatedk+1。
Preferably, the βk+1It is iterated in the following manner:
In formula, βk+1Indicate that kth+1 time output weight matrix, λ are the constraint percentage contribution for reflecting Laplce's manifold
Weight coefficient.
Another aspect, the present invention also provides a kind of adaptive tuning on-line device of terminal, described device includes one
Or multiple processing units；And
Internal storage location, wherein being stored with the computer instruction that can call and carry out operation by processor unit；
The computer instruction executes the adaptive tuning on-line method of terminal as described above.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage medium
In, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be magnetic
Dish, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random Access
Memory, RAM) etc..
The above description is merely a specific embodiment, but scope of protection of the present invention is not limited thereto, any
In the technical scope disclosed by the present invention, any changes or substitutions that can be easily thought of by those familiar with the art, all answers
It is included within the scope of the present invention.Therefore, protection scope of the present invention should be subject to the protection scope in claims.
Claims (11)
1. a kind of adaptive tuning on-line method of terminal, which is characterized in that the described method includes:
Step 1, according to it is a certain amount of calibration fingerprint data set, initial model is established by trained mode；
The initial model is applied in different positioning terminals by step 2；
Step 3, the positioning terminal in use, acquire the signal data without calibration；
The initial model in step 4, the positioning terminal is updated by the signal data of the no calibration, according to new
Increase sample data, model after being updated, the master mould before model substitution updates after the update；
Step 5 repeats step 3, step 4, so that model increasingly adapts to the positioning terminal after the update.
2. the method according to claim 1, wherein the initial model of the step 1 tracking are as follows:
β=K-1HTJT
In formula,
H indicates hidden layer output matrix, and β is expressed as output weight matrix, and T indicates output classification matrix；J=diag (1,1 ..., 0,
0) it is interim intermediate parameters matrix that, 1 number, which is flag data number l, K,.
3. according to the method described in claim 2, it is characterized in that, the step 1 further include: pass through a certain amount of calibration
The data set of fingerprint, obtains the initial parameter of model, and the initial parameter includes β0、K0。
4. according to the method described in claim 3, it is characterized in that, obtaining the initial parameter of model, comprising:
Step 101 provides input weight a by random fashioniWith biasing biAssignment, i=1,2 ..., L；
Step 102 calculates initial matrix J0Laplacian MatrixAnd hidden layer output matrix H0；
Step 103, calculating matrix K0With initial output weight matrix β0。
5. the method according to claim 1, wherein the step 4 further comprises:
Step 401 is based on nominal data number l and non-nominal data number u in newly-increased sample, the parameter of more new model；It is described
Parameter includes hidden layer output matrix, hidden node quantity, matrix J, wherein J=diag (1,1 ..., 0,0), 1 number are mark
Remember data amount check l；
Step 402, the calculated result based on the step 401 updates output weight matrix β.
6. according to the method described in claim 5, it is characterized in that, in the step 402, update output weight matrix by with
Under type carries out:
In formula, βk+1Indicate that kth+1 time output weight matrix, λ are the weight for reflecting the constraint percentage contribution of Laplce's manifold
Coefficient.
7. a kind of adaptive tuning on-line system of terminal, which is characterized in that the system can be set in positioning terminal, described
System includes:
Signal acquisition module acquires the signal data without calibration in the positioning terminal use process；
Model memory module, at the beginning of the system use, storage passes through the side of the data set training of a certain amount of calibration fingerprint
The initial model that formula is established, and, the initial model is updated by the signal data of the no calibration, according to newly-increased sample
Notebook data, model after being updated, the master mould before model substitution updates after the update are realized to the continuous iteration of model more
Newly.
8. system according to claim 7, which is characterized in that the initial model are as follows:
β=K-1HTJT
In formula,
H indicates hidden layer output matrix, and β is expressed as output weight matrix, and T indicates output classification matrix；J=diag (1,1 ..., 0,
0) it is interim intermediate parameters matrix that, 1 number, which is flag data number l, K,.
9. system according to claim 8, which is characterized in that the initial model is referred to by a certain amount of calibration
The data set of line, obtains the initial parameter of model, and the initial parameter includes β0、K0。
10. system according to claim 9, which is characterized in that the initial parameter battle array β0And K0It obtains in the following manner
:
In formula, λ is the weight coefficient for reflecting the constraint percentage contribution of Laplce's manifold.
11. a kind of adaptive tuning on-line device of terminal, which is characterized in that described device includes one or more processing units；
And
Internal storage location, wherein being stored with the computer instruction that can call and carry out operation by processor unit；
The computer instruction perform claim requires any adaptive tuning on-line method of terminal of 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811161901.3A CN109581280A (en) | 2018-09-30 | 2018-09-30 | The adaptive tuning on-line method, system and device of terminal |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811161901.3A CN109581280A (en) | 2018-09-30 | 2018-09-30 | The adaptive tuning on-line method, system and device of terminal |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109581280A true CN109581280A (en) | 2019-04-05 |
Family
ID=65919978
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811161901.3A Pending CN109581280A (en) | 2018-09-30 | 2018-09-30 | The adaptive tuning on-line method, system and device of terminal |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109581280A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110049442A (en) * | 2019-05-23 | 2019-07-23 | 江西师范大学 | Indoor WiFi fingerprint location automatic calibration method and system based on smart phone |
| CN112543470A (en) * | 2019-09-23 | 2021-03-23 | 中国移动通信集团重庆有限公司 | Terminal positioning method and system based on machine learning |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107423547A (en) * | 2017-04-19 | 2017-12-01 | 江南大学 | Increment type location algorithm based on the semi-supervised learning machine that transfinites |
-
2018
- 2018-09-30 CN CN201811161901.3A patent/CN109581280A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107423547A (en) * | 2017-04-19 | 2017-12-01 | 江南大学 | Increment type location algorithm based on the semi-supervised learning machine that transfinites |
Non-Patent Citations (3)
| Title |
|---|
| 刘军发等: "具有时效机制的增量式无线定位方法", 《计算机学报》 * |
| 朱顺涛等: "基于半监督极限学习机的增量式定位算法", 《传感技术学报》 * |
| 黄正宇等: "基于众包数据的室内定位方法和平台", 《地球信息科学学报》 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110049442A (en) * | 2019-05-23 | 2019-07-23 | 江西师范大学 | Indoor WiFi fingerprint location automatic calibration method and system based on smart phone |
| CN112543470A (en) * | 2019-09-23 | 2021-03-23 | 中国移动通信集团重庆有限公司 | Terminal positioning method and system based on machine learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tang et al. | A joint tensor completion and prediction scheme for multi-dimensional spectrum map construction | |
| CN106714336A (en) | Wireless sensor network temperature monitoring method based on improved Kriging algorithm | |
| Yu et al. | Improvement of positioning technology based on RSSI in ZigBee networks | |
| CN106912105A (en) | 3-D positioning method based on PSO_BP neutral nets | |
| CN109379711B (en) | positioning method | |
| Liu et al. | Geomagnetism-based indoor navigation by offloading strategy in NB-IoT | |
| CN112866904B (en) | Channel-training-free large-dimension communication beam alignment method based on beam index map | |
| CN109581280A (en) | The adaptive tuning on-line method, system and device of terminal | |
| CN103209102A (en) | Web quality of service distributed measurement system and method | |
| CN113543026B (en) | Multi-floor indoor positioning method based on radial basis function network | |
| Qin et al. | A wireless sensor network location algorithm based on insufficient fingerprint information | |
| Li et al. | UAV trajectory optimization for spectrum cartography: a PPO approach | |
| Etiabi et al. | Federated learning based hierarchical 3D indoor localization | |
| Liu et al. | Low-cost indoor wireless fingerprint location database construction methods: A review | |
| CN114513805A (en) | Wireless modeling method and system | |
| CN111854745A (en) | Clock prediction method based on Internet of things indoor positioning | |
| CN108445445B (en) | Distributed passive positioning method and device | |
| Wang et al. | Radio map construction based on BERT for fingerprint-based indoor positioning system | |
| Wu et al. | Research on RSS based indoor location method | |
| CN108495262B (en) | Sparse representation and matching positioning method for indoor space ubiquitous positioning signal fingerprint database | |
| CN115942231A (en) | RSS-based 5G outdoor positioning method | |
| Mo et al. | A novel technique for human traffic based radio map updating in Wi-Fi indoor positioning systems | |
| CN109951874B (en) | Method for tracking moving unknown node in real time in sensor network | |
| CN112929818A (en) | Indoor positioning method based on Kalman filtering and pan-kriging interpolation | |
| Ma et al. | A Method for Establishing Tropospheric Atmospheric Refractivity Profile Model Based on Multiquadric RBF and k-means Clustering |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190405 |
|
| RJ01 | Rejection of invention patent application after publication |