CN110536257B - Indoor positioning method based on depth adaptive network - Google Patents

Indoor positioning method based on depth adaptive network Download PDF

Info

Publication number
CN110536257B
CN110536257B CN201910774488.6A CN201910774488A CN110536257B CN 110536257 B CN110536257 B CN 110536257B CN 201910774488 A CN201910774488 A CN 201910774488A CN 110536257 B CN110536257 B CN 110536257B
Authority
CN
China
Prior art keywords
rss
domain
network
target domain
grid point
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910774488.6A
Other languages
Chinese (zh)
Other versions
CN110536257A (en
Inventor
殷光强
郭贤生
李耶
王磊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chengdu Dianke Huian Technology Co ltd
Original Assignee
Chengdu Dianke Huian Technology Co ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Chengdu Dianke Huian Technology Co ltd filed Critical Chengdu Dianke Huian Technology Co ltd
Priority to CN201910774488.6A priority Critical patent/CN110536257B/en
Publication of CN110536257A publication Critical patent/CN110536257A/en
Application granted granted Critical
Publication of CN110536257B publication Critical patent/CN110536257B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/02—Services making use of location information
    • H04W4/023—Services making use of location information using mutual or relative location information between multiple location based services [LBS] targets or of distance thresholds
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W4/00—Services specially adapted for wireless communication networks; Facilities therefor
    • H04W4/30—Services specially adapted for particular environments, situations or purposes
    • H04W4/33—Services specially adapted for particular environments, situations or purposes for indoor environments, e.g. buildings
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04W—WIRELESS COMMUNICATION NETWORKS
    • H04W64/00—Locating users or terminals or network equipment for network management purposes, e.g. mobility management
    • H04W64/006—Locating users or terminals or network equipment for network management purposes, e.g. mobility management with additional information processing, e.g. for direction or speed determination

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Position Fixing By Use Of Radio Waves (AREA)

Abstract

The invention belongs to the field of solving RSS volatility problem by adopting a deep migration learning method in a complex indoor environment, and particularly relates to an indoor positioning method based on a deep adaptive network, which is characterized by comprising the following steps: step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information; step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain; step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain; step 4, knowledge migration; and 5, inputting the target domain data into the trained network to obtain the position.

Description

Indoor positioning method based on depth adaptive network
Technical Field
The invention belongs to a method for positioning by using received signal strength in an indoor environment, in particular to a technology for positioning wireless equipment in a complex indoor environment by using a deep migration learning method, and particularly relates to an indoor positioning method based on a deep adaptive network.
Background
In recent years, the demand for indoor location services has been increasing due to the widespread use of mobile devices and the rapid development of social networks. For example, navigation in large airports and stations, real-time monitoring of worker positions by factories, rescue of indoor personnel in emergency situations, and the like are not exhaustive. Therefore, under the action of great market traction, the search for a high-precision real-time positioning system suitable for a complex indoor positioning environment has become a research focus in the industry.
Generally, RSS indoor positioning methods based on WiFi are generally divided into two types: triangulation and location fingerprinting. The triangulation location method is based on the geometric attributes of angles, and the location position of a user is calculated by using the geometric principle. Due to the existence of multipath effect and non-line-of-sight effect, the method has low positioning accuracy. The location fingerprinting method does not need to assume the presence of direct waves and to know the layout of the indoor environment, and it usually consists of two phases: an offline phase and an online phase. In an off-line stage, the RSS values of all the access points are collected through the mobile equipment at each divided lattice point, and then an off-line fingerprint database is constructed by using the collected RSS values and the corresponding coordinate information. In the online stage, an RSS sample of an unknown position is given, the RSS sample is matched with an offline fingerprint database by using a matching algorithm, and the position corresponding to the RSS value with the highest similarity is used as a positioning result.
When the conventional matching algorithm is used for realizing position fingerprint positioning, the offline fingerprint data and the online fingerprint data are generally required to meet the same distribution, but due to the change of the environment and the influence of heterogeneous equipment, the RSS value fluctuates, so that the positioning result is deviated. Document [1] solves the RSS value volatility problem with a migration learning method that reduces inter-domain differences by learning to shallow migratable features in one subspace. In this subspace, the Maximum Mean variance criterion (MMD) is used to reduce the marginal probability distribution difference between the source and target domains, while maximizing the variance of the RSS samples to preserve the properties of the data. In the stretched subspace, the distribution difference between domains is reduced, the volatility problem of the RSS value is solved, and then the actual position of the online fingerprint data can be obtained by utilizing the traditional matching algorithm. Although the method can improve the positioning accuracy to a certain extent, the method has the obvious disadvantage that the method mainly maps the source domain and the target domain into the subspace by learning a shallow mapping, and the shallow mapping can only learn the representation characteristics of the shallow layer, which is obviously insufficient for reducing the domain difference. Therefore, it is difficult to form accurate, real-time and stable positioning in a complicated indoor environment due to the above problems.
Disclosure of Invention
The invention aims to research and design a high-precision indoor positioning method based on a depth adaptive network aiming at the defects of a shallow representation feature learning method in the technical background. The method learns the deep migratable characteristics capable of reflecting inter-domain invariant factors through a deep adaptive network. The depth adaptive network reduces domain differences due to RSS fluctuations by matching mean embedding between domains in the Hilbert space and matching second-order statistics of hidden representations between domains. The method overcomes the defect that the traditional transfer learning method can only reduce the domain difference by learning shallow features, effectively utilizes the deep self-adaptive network to learn the deep transferable features, reduces the performance reduction caused by RSS value fluctuation, and improves the positioning precision and the robustness.
Aiming at the problems, the invention creatively provides an indoor positioning method based on a depth self-adaptive network, and the practicability of the algorithm is explained through the verification of the measured data.
In order to solve the problems in the prior art, the invention provides an indoor positioning method based on a depth adaptive network.
An indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain;
step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain;
step 4, knowledge migration;
and 5, inputting the target domain data into the trained network to obtain the position.
The detailed steps of the step 2 are as follows: acquiring data and forming RSS fingerprint database, and sequentially arranging mobile devicesAt each grid point, recording the grid point number and the RSS value from each access point to form an RSS vector
Figure BDA0002174613380000021
Where n issFor all RSS sample numbers for which the position marker is known, the corresponding position marker is denoted as yiE {1,2, …, C }, where C ═ 48 is the number of lattice points. The RSS values and corresponding location tags form a fingerprint database, namely the source field Ds＝{Xs,ys}，
Wherein the content of the first and second substances,
Figure BDA0002174613380000031
the detailed steps of the step 3 are as follows: collecting RSS values of mobile devices requesting positioning to form a target domain Dt＝{XtHere, where
Figure BDA0002174613380000032
Target field data matrix formed for RSS samples without position markers, ntIs the number of RSS samples collected.
The detailed steps of the step 4 are as follows:
step 4-1, inputting source domain data with position coordinate information and target domain data with unknown coordinate information into a depth adaptive network;
step 4-2, embedding the mean values of the matching source domain and the target domain in the Hilbert space in the last layer of the network hidden layer to form a first loss function;
4-3, matching second-order statistics of the source domain and the target domain in the last layer of the network hidden layer to form a second loss function;
4-4, calculating the training loss of the source domain by using a cross entropy loss function at a network classifier layer to form a third loss function;
and 4-5, optimizing network parameters by using a gradient descent method, and minimizing three loss functions.
The detailed process involved in the step 4 is as follows:
completing knowledge migration through a deep adaptive network
4-1, inputting the source domain data with position coordinate information and the target domain data with unknown coordinate information into the depth self-adaptive network, and obtaining the depth migratable characteristics of the source domain and the target domain at the last layer of the network hidden layer
Figure BDA0002174613380000033
And
Figure BDA0002174613380000034
wherein, b is a batch size, and d is the number of the neurons of the hidden layer.
4-2. mapping the depth migratable features of the source domain and the target domain into a high-dimensional regeneration kernel Hilbert space, matching the mean embedding of the source domain and the target domain, i.e. minimizing the following loss function:
Figure BDA0002174613380000035
where φ is a kernel mapping, kernel matrix Kij＝φ(hi)Tφ(hj)＝k(hi,hj) M is an MMD matrix, and the calculation method is as follows:
Figure BDA0002174613380000041
4-3. matching InterDomain hidden representation Hs,HtSecond order statistics in between to reduce domain differences, i.e. minimize the following loss function:
Figure BDA0002174613380000042
wherein, Cov (H)s) And Cov (H)t) The calculation method of the feature covariance matrix which is respectively represented by source domain hiding and target domain hiding comprises the following steps:
Cov(Hs)＝(Hs)TJbHs (4)
Cov(Ht)＝(Ht)TJbHt (5)
Jbis a central matrix, and the central matrix is a central matrix,
Figure BDA0002174613380000043
is a matrix of the units,
Figure BDA0002174613380000044
a column vector of all 1's.
4-4, at the network classifier level, computing the training loss of the source domain using a cross entropy loss function, i.e. minimizing the following loss function:
Figure BDA0002174613380000045
wherein, c (θ | x)i,yi) For categorical softmax loss, θ is a parameter of the network.
4-5, optimizing the network parameters by using a gradient descent method, and simultaneously minimizing the above three loss functions, namely minimizing the following loss functions:
L＝Ls+Lm+Lc (7)。
the invention has the beneficial effects that:
1. the invention solves the RSS volatility problem caused by environment change and heterogeneous equipment by using a deep migration learning method, improves the positioning precision and robustness, and overcomes the defect that the conventional matching algorithm does not work if the source domain data and the target domain data are uniformly distributed in a complicated and changeable indoor environment.
2. The method is based on the deep adaptive network, learns deep migratable features to reduce domain differences, and overcomes the defect that a migration learning method in the background art learns shallow features to reduce the domain differences.
3. The indoor positioning method based on the depth self-adaptive network is a positioning method with high positioning accuracy and good robustness.
Drawings
FIG. 1 is a diagram of a network architecture of the present invention;
FIG. 2 is a diagram of the cumulative percentage of positioning errors for a positioning method and the method of the present invention used in the background of the art for heterogeneous devices;
fig. 3 shows the cumulative percentage of positioning errors of the positioning method used in the background of the invention and the method of the invention when the environment changes.
The specific implementation mode is as follows:
example 1:
an indoor positioning method based on a depth adaptive network.
An indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain;
step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain;
step 4, knowledge migration;
and 5, inputting the target domain data into the trained network to obtain the position.
Example 2:
an indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain;
step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain;
step 4, knowledge migration;
and 5, inputting the target domain data into the trained network to obtain the position.
The detailed steps of the step 2 are as follows: acquiring data and forming an RSS fingerprint database, sequentially placing the mobile equipment in each grid point, recording the number of the grid point and the RSS value from each access point, and forming an RSS vector
Figure BDA0002174613380000061
Where n issFor all RSS sample numbers for which the position marker is known, the corresponding position marker is denoted as yiE {1,2, …, C }, where C ═ 48 is the number of lattice points. The RSS values and corresponding location tags form a fingerprint database, namely the source field Ds＝{Xs,ys}，
Wherein the content of the first and second substances,
Figure BDA0002174613380000062
the detailed steps of the step 3 are as follows: collecting RSS values of mobile devices requesting positioning to form a target domain Dt＝{XtHere, where
Figure BDA0002174613380000063
Target field data matrix formed for RSS samples without position markers, ntIs the number of RSS samples collected.
The detailed steps of the step 4 are as follows:
step 4-1, inputting source domain data with position coordinate information and target domain data with unknown coordinate information into a depth adaptive network;
step 4-2, embedding the mean values of the matching source domain and the target domain in the Hilbert space in the last layer of the network hidden layer to form a first loss function;
4-3, matching second-order statistics of the source domain and the target domain in the last layer of the network hidden layer to form a second loss function;
4-4, calculating the training loss of the source domain by using a cross entropy loss function at a network classifier layer to form a third loss function;
and 4-5, optimizing network parameters by using a gradient descent method, and minimizing three loss functions.
Example 3:
an indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain;
step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain;
step 4, knowledge migration;
and 5, inputting the target domain data into the trained network to obtain the position.
The detailed steps of the step 2 are as follows: acquiring data and forming an RSS fingerprint database, sequentially placing the mobile equipment in each grid point, recording the number of the grid point and the RSS value from each access point, and forming an RSS vector
Figure BDA0002174613380000071
Where n issFor all RSS sample numbers for which the position marker is known, the corresponding position marker is denoted as yiE {1,2, …, C }, where C ═ 48 is the number of lattice points. The RSS values and corresponding location tags form a fingerprint database, namely the source field Ds＝{Xs,ys}，
Wherein the content of the first and second substances,
Figure BDA0002174613380000072
example 4:
an indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain;
step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain;
step 4, knowledge migration;
and 5, inputting the target domain data into the trained network to obtain the position.
The detailed steps of the step 3 are as follows: collecting RSS values of mobile devices requesting positioning to form a target domain Dt＝{XtHere, where
Figure BDA0002174613380000073
Target field data matrix formed for RSS samples without position markers, ntIs the number of RSS samples collected.
Example 5:
an indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially placed in each grid point in the positioning environment, and the RSS value from each access point and corresponding grid point coordinate information at the moment are recorded to form an RSS offline fingerprint database, namely a source domain;
step 3, collecting the RSS value of the mobile equipment to be positioned to form a target domain;
step 4, knowledge migration;
and 5, inputting the target domain data into the trained network to obtain the position.
The detailed steps of the step 4 are as follows:
step 4-1, inputting source domain data with position coordinate information and target domain data with unknown coordinate information into a depth adaptive network;
step 4-2, embedding the mean values of the matching source domain and the target domain in the Hilbert space in the last layer of the network hidden layer to form a first loss function;
4-3, matching second-order statistics of the source domain and the target domain in the last layer of the network hidden layer to form a second loss function;
4-4, calculating the training loss of the source domain by using a cross entropy loss function at a network classifier layer to form a third loss function;
and 4-5, optimizing network parameters by using a gradient descent method, and minimizing three loss functions.
The detailed process involved in the step 4 is as follows:
completing knowledge migration through a deep adaptive network
4-1, inputting the source domain data with position coordinate information and the target domain data with unknown coordinate information into the depth self-adaptive network, and obtaining the depth migratable characteristics of the source domain and the target domain at the last layer of the network hidden layer
Figure BDA0002174613380000081
And
Figure BDA0002174613380000082
wherein, b is a batch size, and d is the number of the neurons of the hidden layer.
4-2. mapping the depth migratable features of the source domain and the target domain into a high-dimensional regeneration kernel Hilbert space, matching the mean embedding of the source domain and the target domain, i.e. minimizing the following loss function:
Figure BDA0002174613380000091
where φ is a kernel mapping, kernel matrix Kij＝φ(hi)Tφ(hj)＝k(hi,hj) M is an MMD matrix, and the calculation method is as follows:
Figure BDA0002174613380000092
4-3. matching InterDomain hidden representation Hs,HtSecond order statistics in between to reduce domain differences, i.e. minimize the following loss function:
Figure BDA0002174613380000093
wherein, Cov (H)s) And Cov (H)t) The calculation method of the feature covariance matrix which is respectively represented by source domain hiding and target domain hiding comprises the following steps:
Cov(Hs)＝(Hs)TJbHs (4)
Cov(Ht)＝(Ht)TJbHt (5)
Jbis a central matrix, and the central matrix is a central matrix,
Figure BDA0002174613380000094
is a matrix of the units,
Figure BDA0002174613380000095
a column vector of all 1's.
4-4, at the network classifier level, computing the training loss of the source domain using a cross entropy loss function, i.e. minimizing the following loss function:
Figure BDA0002174613380000096
wherein, c (θ | x)i,yi) For categorical softmax loss, θ is a parameter of the network.
4-5, optimizing the network parameters by using a gradient descent method, and simultaneously minimizing the above three loss functions, namely minimizing the following loss functions:
L＝Ls+Lm+Lc (7)。
example 6:
1. arrangement of experimental sites
The experimental environment is 308.4m2The library environment of (1) has seat bench and bookshelf etc. in the room, and the number of the WiFi access point that can detect in the location area is 620, divides the place into 48 check points earlier.
2. Acquiring data and forming RSS fingerprint database
The mobile equipment is sequentially arranged in each grid point, the grid point number and the RSS value from each access point are recorded, and an RSS vector is formed
Figure BDA0002174613380000101
Where n issFor all RSS sample numbers for which the position marker is known, the corresponding position marker is denoted as yiE {1,2, …, C }, where C ═ 48 is the number of lattice points. The RSS values and corresponding location tags form a fingerprint database, namely the source field Ds＝{Xs,ysAnd (c) the step of (c) in which,
Figure BDA0002174613380000102
3. collecting RSS values for a device to be located
Collecting RSS values of mobile devices requesting positioning to form a target domain Dt＝{XtHere, where
Figure BDA0002174613380000103
Target field data matrix formed for RSS samples without position markers, ntIs the number of RSS samples collected.
4. Completing knowledge migration through a deep adaptive network
4-1, inputting the source domain data with position coordinate information and the target domain data with unknown coordinate information into the depth self-adaptive network, and obtaining the depth migratable characteristics of the source domain and the target domain at the last layer of the network hidden layer
Figure BDA0002174613380000104
And
Figure BDA0002174613380000105
wherein b isOne batch size, d is the number of neurons in the hidden layer.
4-2. mapping the depth migratable features of the source domain and the target domain into a high-dimensional regeneration kernel Hilbert space, matching the mean embedding of the source domain and the target domain, i.e. minimizing the following loss function:
Figure BDA0002174613380000111
where φ is a kernel mapping, kernel matrix Kij＝φ(hi)Tφ(hj)＝k(hi,hj) M is an MMD matrix, and the calculation method is as follows:
Figure BDA0002174613380000112
4-3. matching InterDomain hidden representation Hs,HtSecond order statistics in between to reduce domain differences, i.e. minimize the following loss function:
Figure BDA0002174613380000113
wherein, Cov (H)s) And Cov (H)t) The calculation method of the feature covariance matrix which is respectively represented by source domain hiding and target domain hiding comprises the following steps:
Cov(Hs)＝(Hs)TJbHs (4)
Cov(Ht)＝(Ht)TJbHt (5)
Jbis a central matrix, and the central matrix is a central matrix,
Figure BDA0002174613380000114
is a matrix of the units,
Figure BDA0002174613380000115
a column vector of all 1's.
4-4, at the network classifier level, computing the training loss of the source domain using a cross entropy loss function, i.e. minimizing the following loss function:
Figure BDA0002174613380000116
wherein, c (θ | x)i,yi) For categorical softmax loss, θ is a parameter of the network.
4-5, optimizing the network parameters by using a gradient descent method, and simultaneously minimizing the above three loss functions, namely minimizing the following loss functions:
L＝Ls+Lm+Lc (7)
5. target domain data XtInputting the location information into the adaptive network obtained in step 4, and obtaining the location mark of the target domain, namely the location result.
The invention respectively carries out actual measurement positioning on 2 target domains in an experimental environment, wherein one target domain is a data sample obtained by acquiring RSS values by different equipment at different time, namely positioning under the condition of heterogeneous equipment, and the other target domain is a data sample obtained by acquiring the RSS values by the same equipment at different time, namely positioning under the condition of environmental change. Each target domain carries 3120 test samples, with an average localization error of 2.67m under heterogeneous device conditions and an average localization error of 2.58m under ambient variation conditions. Fig. 2 is a percentage of accumulated positioning errors of the positioning method and the method of the present invention used in the technical background of heterogeneous devices, and fig. 3 is a percentage of accumulated positioning errors of the positioning method and the method of the present invention used in the technical background of environmental changes. The method in the technical background has a significantly lower localization effect on each target domain than the method of the present invention, because the method maps the source domain and the target domain into the subspace by learning a shallow map, which can only learn shallow representation features, which is significantly insufficient for reducing domain differences. The method provided by the invention overcomes the defect that the traditional transfer learning method can only reduce the domain difference by learning shallow features, effectively utilizes the deep self-adaptive network to learn the deep transferable features, and can still obtain better positioning effect even in the target domain with larger RSS value fluctuation.

Claims (1)

1. An indoor positioning method based on a depth adaptive network is characterized by comprising the following steps:
step 1, dividing an indoor environment to be positioned into grid point areas with equal size, and recording corresponding coordinate information;
step 2, the mobile equipment is sequentially arranged in each grid point in the positioning environment, the RSS value from each access point and corresponding grid point coordinate information are recorded at the moment, an RSS offline fingerprint database which is a source domain is formed, data are acquired and an RSS fingerprint database is formed, the mobile equipment is sequentially arranged in each grid point, the grid point number and the RSS value from each access point are recorded, and an RSS vector is formed
Figure FDA0003331694200000011
i＝1,2,3,…,nsWhere n issFor all RSS sample numbers for which the position marker is known, the corresponding position marker is denoted as yiE {1,2, …, C }, where C is 48 is the number of lattice points, and the RSS values and corresponding location markers constitute a fingerprint database, i.e., a source domain
Figure FDA0003331694200000012
Wherein the content of the first and second substances,
Figure FDA0003331694200000013
Figure FDA0003331694200000014
step 3, collecting RSS value of mobile equipment to be positioned to form target domain
Figure FDA0003331694200000015
Here, the
Figure FDA0003331694200000016
Target field data matrix formed for RSS samples without position markers, ntThe number of RSS samples collected;
step 4, knowledge migration is completed through a depth self-adaptive network
Step 4-1, inputting source domain data with position coordinate information and target domain data with unknown coordinate information into the depth self-adaptive network, and obtaining the depth migratable characteristics of the source domain and the target domain at the last layer of the network hidden layer
Figure FDA0003331694200000017
And
Figure FDA0003331694200000018
wherein, b is a batch size, d is the number of the neurons of the hidden layer;
step 4-2, mapping the depth migratable features of the source domain and the target domain into a high-dimensional regeneration kernel Hilbert space, and matching the mean value embedding of the source domain and the target domain, namely minimizing the following loss function:
Figure FDA0003331694200000019
where φ is a kernel mapping, kernel matrix Kij＝φ(hi)Tφ(hj)＝k(hi,hj)，MijFor the MMD matrix, the calculation is as follows:
Figure FDA0003331694200000021
step 4-3. matching inter-domain hidden representation Hs,HtSecond order statistics in between to reduce domain differences, i.e. minimize the following loss function:
Figure FDA0003331694200000022
wherein, Cov (H)s) And Cov (H)t) The calculation method of the feature covariance matrix which is respectively represented by source domain hiding and target domain hiding comprises the following steps:
Cov(Hs)＝(Hs)TJbHs (4)
Cov(Ht)＝(Ht)TJbHt (5)
Jbis a central matrix, and the central matrix is a central matrix,
Figure FDA0003331694200000023
is a matrix of the units,
Figure FDA0003331694200000024
a column vector of all 1's;
and 4, calculating the training loss of the source domain by using a cross entropy loss function at a network classifier layer, namely minimizing the following loss function:
Figure FDA0003331694200000025
wherein, c (θ | x)i,yi) For categorical softmax loss, θ is a parameter of the network;
and 4-5, optimizing network parameters by using a gradient descent method, and simultaneously minimizing the three loss functions, namely minimizing the following loss functions:
L＝Ls+Lm+Lc (7)；
and 5, inputting the target domain data into the trained network to obtain the position.
CN201910774488.6A 2019-08-21 2019-08-21 Indoor positioning method based on depth adaptive network Active CN110536257B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910774488.6A CN110536257B (en) 2019-08-21 2019-08-21 Indoor positioning method based on depth adaptive network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910774488.6A CN110536257B (en) 2019-08-21 2019-08-21 Indoor positioning method based on depth adaptive network

Publications (2)

Publication Number Publication Date
CN110536257A CN110536257A (en) 2019-12-03
CN110536257B true CN110536257B (en) 2022-02-08

Family

ID=68662425

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910774488.6A Active CN110536257B (en) 2019-08-21 2019-08-21 Indoor positioning method based on depth adaptive network

Country Status (1)

Country Link
CN (1) CN110536257B (en)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111885703B (en) * 2020-07-21 2021-04-23 电子科技大学 Indoor positioning method
CN111935629B (en) * 2020-07-30 2023-01-17 广东工业大学 Self-adaptive positioning method based on environmental feature migration
CN112954632B (en) * 2021-01-26 2022-01-28 电子科技大学 Indoor positioning method based on heterogeneous transfer learning
CN113132931B (en) * 2021-04-16 2022-01-28 电子科技大学 Depth migration indoor positioning method based on parameter prediction
CN115103441B (en) * 2022-06-20 2024-04-05 燕山大学 Positioning method based on priori knowledge and transfer learning
CN116233747B (en) * 2023-03-09 2023-12-22 江苏亨鑫科技有限公司 BLE positioning method and device based on transfer learning and fingerprint library updating

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105574859A (en) * 2015-12-14 2016-05-11 中国科学院深圳先进技术研究院 Liver tumor segmentation method and device based on CT (Computed Tomography) image
CN106874957A (en) * 2017-02-27 2017-06-20 苏州大学 A kind of Fault Diagnosis of Roller Bearings
CN107037399A (en) * 2017-05-10 2017-08-11 重庆大学 A kind of Wi Fi indoor orientation methods based on deep learning
CN108449295A (en) * 2018-02-05 2018-08-24 西安电子科技大学昆山创新研究院 Combined modulation recognition methods based on RBM networks and BP neural network
CN109102126A (en) * 2018-08-30 2018-12-28 燕山大学 One kind being based on depth migration learning theory line loss per unit prediction model

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9924490B2 (en) * 2013-10-09 2018-03-20 International Business Machines Corporation Scaling multi-core neurosynaptic networks across chip boundaries
CN108234370B (en) * 2017-12-22 2020-12-15 西安电子科技大学 Communication signal modulation mode identification method based on convolutional neural network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105574859A (en) * 2015-12-14 2016-05-11 中国科学院深圳先进技术研究院 Liver tumor segmentation method and device based on CT (Computed Tomography) image
CN106874957A (en) * 2017-02-27 2017-06-20 苏州大学 A kind of Fault Diagnosis of Roller Bearings
CN107037399A (en) * 2017-05-10 2017-08-11 重庆大学 A kind of Wi Fi indoor orientation methods based on deep learning
CN108449295A (en) * 2018-02-05 2018-08-24 西安电子科技大学昆山创新研究院 Combined modulation recognition methods based on RBM networks and BP neural network
CN109102126A (en) * 2018-08-30 2018-12-28 燕山大学 One kind being based on depth migration learning theory line loss per unit prediction model

Also Published As

Publication number Publication date
CN110536257A (en) 2019-12-03

Similar Documents

Publication Publication Date Title
CN110536257B (en) Indoor positioning method based on depth adaptive network
CN106793086B (en) Indoor positioning method
CN109444813B (en) RFID indoor positioning method based on BP and DNN double neural networks
WO2019233054A1 (en) Fingerprint-based positioning method and system for smart classroom
CN104853435B (en) A kind of indoor orientation method based on probability and device
CN104394588B (en) Indoor orientation method based on Wi Fi fingerprints and Multidimensional Scaling
CN110536245B (en) Deep learning-based indoor wireless positioning method and system
CN109348410B (en) Indoor positioning method based on global and local joint constraint transfer learning
CN104392414A (en) Establishing method of regional CORS coordinate time series noise model
CN108627798B (en) WLAN indoor positioning algorithm based on linear discriminant analysis and gradient lifting tree
CN110728411A (en) High-low altitude area combined rainfall prediction method based on convolutional neural network
CN107677989B (en) A kind of indoor location localization method carrying out RSSI removal noise based on RSSI maximum value
CN107360552B (en) Indoor positioning method for global dynamic fusion of multiple classifiers
CN111854732B (en) Indoor fingerprint positioning method based on data fusion and width learning
CN112653991A (en) WLAN indoor positioning method of TebNet neural network model based on deep learning
CN107290714B (en) Positioning method based on multi-identification fingerprint positioning
CN116008671A (en) Lightning positioning method based on time difference and clustering
Zhang et al. Dual-band wi-fi based indoor localization via stacked denosing autoencoder
CN111885703B (en) Indoor positioning method
CN109858517B (en) Track similarity measurement method taking motion direction as dominant
CN108731679A (en) Mobile robot environmental characteristic localization method
CN108696818B (en) Distance measurement method and system for Wi-Fi fingerprint positioning
CN114679683A (en) Indoor intelligent positioning method based on derivative fingerprint migration
CN110321528A (en) A kind of Hyperspectral imaging heavy metal-polluted soil concentration evaluation method based on semi-supervised geographical space regression analysis
WO2022242018A1 (en) Indoor target positioning method based on improved cnn model

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
GR01 Patent grant
GR01 Patent grant
PE01 Entry into force of the registration of the contract for pledge of patent right
PE01 Entry into force of the registration of the contract for pledge of patent right

Denomination of invention: An indoor location method based on depth adaptive network

Effective date of registration: 20220523

Granted publication date: 20220208

Pledgee: Jinniu sub branch of Bank of Chengdu Co.,Ltd.

Pledgor: CHENGDU DIANKE HUIAN TECHNOLOGY Co.,Ltd.

Registration number: Y2022510000134