CN110691319B - A method for realizing high-precision indoor positioning of heterogeneous devices using domain adaptation - Google Patents

Abstract

The present invention provides a method for realizing high-precision indoor positioning of heterogeneous equipment using domain adaptation, which minimizes the difference between the two domains by aligning the fingerprint vector collected by the online positioning terminal and the second-order statistical information of the fingerprint database fingerprint in the offline stage. offset, and does not require any information about the terminal label. Based on the transfer learning framework, domain adaptation and elimination of terminal differences are combined to improve the scalability of the positioning system. Before the classifier is trained, the fingerprint database collected by the fixed terminal in the offline phase is used as the target feature, and the source-target whitening alignment of the fingerprint of any terminal in the online phase can greatly reduce the damage to the localization performance caused by the heterogeneity. The method of the invention realizes on-line adjustment simply and quickly, and achieves ideal performance in actual multi-terminal positioning.

Description

A method for realizing high-precision indoor positioning of heterogeneous devices using domain adaptation

technical field

The invention belongs to the technical field of positioning, relates to an indoor positioning technology based on WLAN, and in particular relates to a method for realizing high-precision indoor positioning of heterogeneous equipment by using field adaptation.

Background technique

The accuracy of the terminal-based fingerprint positioning system depends on the communication packets received by the terminal from each access point AP, and the RSS value is the basis for positioning. Different data acquisition devices are equipped with different hardware chips and antennas, and may have different signal perception capabilities, thus resulting in different data distributions. When the signal distribution changes in the online phase, it is necessary to re-collect and update a new fingerprint database to maintain the positioning accuracy of the system. In large and complex buildings, collecting and maintaining the signal strength fingerprint database is extremely labor-intensive. The terminal version iteration speed is so fast that the collection device of the fingerprint database cannot cover all smart terminals on the market. In the actual positioning system, various mobile devices used by users are different from those used to construct radio maps, which seriously affects the promotion and use of the positioning system. In addition, during the positioning of heterogeneous devices, terminal domain offset is prone to occur, resulting in a decrease in positioning performance and insufficient positioning accuracy.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention proposes a method for realizing high-precision indoor positioning of heterogeneous devices using field adaptation, so as to achieve better positioning performance when positioning heterogeneous devices online. The present invention utilizes relative alignment to minimize the distance of the source and target data margins. According to the fingerprint database data as a reference, the operation of transforming and aligning the second-order statistical characteristics of the fingerprints in the offline phase and the online phase is performed. After confusing the source domain and the target domain, the classifier cannot distinguish the source domain and the target domain, which can naturally alleviate the impact of device differences.

In order to achieve the above object, the present invention provides the following technical solutions:

A method for realizing high-precision indoor positioning of heterogeneous devices using domain adaptation, comprising the following steps:

(1) Simulate the positioning process of a variety of equipment heterogeneous situations in actual scenarios through common terminals;

(2) For the indoor positioning system, it is necessary to divide the sampling points evenly in the test environment to collect and build a fingerprint database of the wireless signal strength of each surrounding access point; then conduct offline integrated learning and training. The integrated learning composed of perceptron regression and multi-layer perceptron classification, training to establish the mapping relationship between the position labels of each reference point and their corresponding fingerprints, and saving the trained position estimation network model; finally, the online position estimation test is carried out. When the fingerprint of the surrounding access point is reached, after the relevant alignment is adjusted online, the trained algorithm is used for real-time positioning;

(5) During the online adjustment stage, first perform a whitening operation on the test source domain data received by another mobile phone to eliminate the feature correlation of the source domain data; and then calculate the correlation between the training target domain data in the fingerprint database. The correlation acts on the whitened source data to reconstruct the features of the source data; through the continuous approximation between the second-order statistical information of the source domain and the target domain, the source domain and the target domain are fused; the second-order statistical characteristics preserve the data feature alignment At the same time, it also retains the distribution information, which generally includes the heterogeneity of the device caused by the difference in the chip, antenna direction and installation material;

(6) The fingerprints received in real time pass through the trained position estimation network, and the algorithm can predict the position coordinates.

Further, the step of building a fingerprint database in the step (2) specifically includes the following process:

(11) The area to be measured is evenly divided into grid reference points with a certain interval according to the area, as the sampling points of the positioning area;

(12) On each reference point, terminal A receives and records the wireless signal strengths sent by all access points APs in the vicinity for a period of time, and forms a form including the reference point position [x _i , y _i ], the MAC address of each AP and the corresponding reception to the fingerprint vector ([rss ₁ ,rss ₂ ,...,rss _n ],[x _i ,y _i ]) of signal strengths [rss ₁ ,rss ₂ ,...,rss _n ];

(13) The fingerprint vector of each reference point can be combined and stored as an offline fingerprint database of the area to be measured.

Further, the off-line integrated learning and training step in the step (2) includes the following process:

(11) After normalizing the RSS data [rss ₁ , rss ₂ ,..., rss _n ] in the fingerprint database to zero mean and standard deviation, divide the offline data set into the fingerprint database data according to the ten-fold cross-validation method; The database data is divided into training set, validation set and test set; the training process is performed on the training set, and the final positioning result is tested on the test set;

(12) Using random forest regression, multi-layer perceptron regression and multi-layer perceptron classification model as the base learner, the training set and the validation set are combined into an ensemble learning. The objective function of the training process is the encoded neurons and the The mean square error of the mapping relationship between the positions of the labels minimizes the objective function optimization parameters;

(13) The distance between the predicted position and the actual position is used as a critical criterion for accuracy, and it is recorded that there are L unknown unknown reference points in the area to be located, and the real geographical position of the i-th reference point is recorded as ( _xi , y _i ), The unknown estimated in the input positioning system is denoted as ( _xi ', y _i '), then ALE can be expressed as:

Further, the step (12) includes the following process:

The base class learner random forest regression learning method is composed of multiple decision trees; the n APs of the measured signals [rss ₁ , rss ₂ ,..., rss _n ] can be regarded as the characteristics of the division conditions, as the algorithm The input data of ; the reference point position label [x _i , y _i ] is equivalent to the leaf node of the decision tree, that is, the two-dimensional output result of the algorithm; using the heuristic node selection method C4.5 algorithm, first find out the information gain Ent(·) is higher than the average level of attributes, and then select the attributes with the largest information gain rate Gain(·) to build a decision tree; the construction of a decision tree requires grid coordination to adjust some parameters to effectively prevent overfitting and underfitting Fitting; random forest estimates the importance of each feature in the classification result, and takes the class that appears most in the output of the decision tree as the final output of the random forest classifier;

A recursive process consisting of a smaller subset of the input dataset trains each decision tree until all tree nodes achieve a similar output goal; the random forest classifier takes input-based weights as parameters similar to the number of decision trees; instead The classification algorithm, the RFR algorithm is selected to train the mapping of fingerprints and labels, and the feature division standard splitting standard is the minimum mean square error:

In the formula, y _L* represents the value of the left subtree, and y _R* represents the value of the right subtree;

The multilayer perceptron algorithm is a feedforward artificial neural network; the hidden layer, the number of neurons and the nonlinear activation function in this algorithm need to be carefully selected; every node except the input node is under the action of the nonlinear activation function. The neurons of the multi-layer perceptron are used, and two hidden layers with 100 neurons in each layer are used; when the loss function is minimized, the SGD method is used to update the parameters:

Among them, η is the learning rate of the control step size in the search parameter space, and Loss is the loss function of the entire network;

Ensemble learning comprehensively considers the results of multiple basic learner models; the average method is used to combine strategies:

Among them, the position (x _i , y _i ) is the prediction result of each basic learner _, and Wi is the weight of the prediction result of the basic learner to the final position estimation.

Further, the online position estimation step in the step (2) includes the following process:

利用已经训练好的位置指纹特征与位置标签的网络模型，针对新输入的在线特征估计其对应的位置。Fingerprints from n APs collected at the i-th test point in the test set

Using the network model of the trained location fingerprint feature and location label, the corresponding location of the newly input online feature is estimated.

Further, in the step (3), the online feature whitening and reconstruction process includes the following initialization stage, modeling stage and optimization stage:

希望其对齐的目标域的数据是指纹库中的训练数据R_T＝(r₁,…,r_k)，其中，

源域和目标域的特征都是n维向量；记这两个域特征向量的均值和协方差矩阵分别为μ_S、μ_T和Cov_S、Cov_T，经过中心化标准后，μ_S＝μ_T＝0，但是Cov_S≠Cov_T；(11) In the initialization phase, for the data in the given test set as the source domain data RSS _S =(rss ₁ ,...,rss _m ), where,

The data of the target domain whose alignment is desired is the training data R _T =(r ₁ ,...,r _k ) in the fingerprint library, where,

The features of the source domain and the target domain are all n-dimensional vectors; the mean and covariance matrices of the eigenvectors of these two domains are respectively μ _S , μ _T and Cov _S , Cov _T , after the centralization standard, μ _S = μ _T = 0, but Cov _S ≠ Cov _T ;

(12) In the modeling stage, in order to reduce the gap between the second-order statistical characteristics of the source domain and target domain features, that is, the distance between the covariance Cov _S and Cov _T , the source domain features are linearly changed to minimize the source domain after the change. The distance between the feature and the target feature covariance; then the objective function is:

表示线性变化后源域特征的协方差矩阵。in,

Represents the covariance matrix of the source domain features after linear variation.

的最优解为

其中，

由于协方差矩阵是对称矩阵，对Cov_S特征分解得Cov_S＝U_SΣ_SV_S：(13) In the optimization stage, perform singular value decomposition on the covariance matrix of the objective function Cov _T =U _T Σ _T V _T , where U _T[1:r] and Σ _T[1:r] V _{T[1:r ]} are the left singular vector of Cov _T , the largest r singular values and the corresponding right singular vector, then

The optimal solution of is

in,

Since the covariance matrix is a symmetric matrix, the eigendecomposition of Cov _S can obtain Cov _S =U _S Σ _S V _S :

得，

则线性变化重建过程推导为：Depend on

have to,

Then the linear change reconstruction process is deduced as:

为白化过程，第二部分

为对齐过程；In the above formula, the first part

for the whitening process, part 2

for the alignment process;

次幂，消除源域数据的相关性；再以目标函数的二阶统计特性对白化特征作相关性叠加变换；(14) For users who need online positioning, provide the fingerprint vector at the test point, and calculate the covariance matrix of the target domain fingerprint database and the source domain online test fingerprint respectively. inverse matrix

power to eliminate the correlation of the source domain data; then use the second-order statistical characteristics of the objective function to perform correlation stacking transformation on the whitening features;

(15) The online features after CORAL alignment use the data in the original training set for location estimation.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

The invention adopts the method of confusing training set and test set equipment features by related alignment, and aligns the real-time fingerprint online with the fingerprint in the training fingerprint database. The invention only needs to perform transformation and alignment operations on the second-order statistical characteristics of the fingerprints in the offline stage and the online stage, and does not need to know the device labels used in the two stages. The entire positioning system only needs to integrate regression learning in the offline stage, training to establish a high-dimensional manifold graph mapping relationship between position coordinates and multi-dimensional features, and CORAL alignment for device heterogeneity can be quickly adjusted in the online positioning stage, which minimizes the number of domains. offset, which well compensates for the degradation of localization performance due to terminal domain offset.

Description of drawings

FIG. 1 is a schematic diagram of the training flow of the present invention.

Figure 2 shows the CORAL domain adaptation process in the online adjustment stage of the present invention, wherein (a) is the initial centralization stage, (b) is the whitening source domain feature stage, and (c) is the reconstructed source domain feature stage.

Detailed ways

The technical solutions provided by the present invention will be described in detail below with reference to specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and not to limit the scope of the present invention.

The present invention proposes a domain adaptation method using correlation alignment, which can compensate for performance degradation due to terminal domain offset.

A method for realizing high-precision indoor positioning of heterogeneous devices provided by the present invention, as shown in FIG. 1, includes the following steps:

(1) First, consider the online positioning process of users. The training of machine learning greatly improves the positioning performance, so the generalization requirements for the positioning system are also improved. End markets with such rapid iterations require positioning systems to maintain high-precision positioning while being compatible with multiple models of devices. We use common terminals to simulate the positioning process of various equipment heterogeneous situations in actual scenarios. In the existing positioning experiment system, different devices are used to collect RSS fingerprint values in the offline and online stages (for example, the fingerprint database for online positioning and matching of Xiaomi mobile phones is collected and constructed by Huawei mobile phones), and the heterogeneous positioning of devices is constructed. Happening.

(2) For the indoor positioning system, we first need to evenly divide the sampling points in the test environment. In this example, the interval between sampling points is 1.6m*1.6m to collect and sort out the fingerprints of the wireless signal strength of the surrounding access points. library. In the training stage, through the ensemble learning consisting of random forest regression, multi-layer perceptron regression and multi-layer perceptron classification, the training establishes the mapping relationship between each reference point position label and its corresponding fingerprint, and saves the trained position estimation network model. When online users receive the fingerprints of surrounding access points, they can use the trained algorithm for real-time positioning after online adjustment of relevant alignment.

Specifically, step (2) includes building a fingerprint database, offline ensemble learning and training, and online position estimation testing.

The building of the fingerprint database includes the following steps:

(11) The area to be measured is evenly divided into grid reference points with an interval of 1-2m according to the area, as the sampling points of the positioning area.

(12) At each reference point, terminal A receives and records the wireless signal strengths sent by all access points APs in the vicinity for 1 minute, and forms a form including the reference point position [x _i , y _i ], the MAC address of each AP and the corresponding received Fingerprint vector ([rss ₁ ,rss ₂ ,...,rss _n ],[x _i ,y _i ]) of signal strengths [rss ₁ ,rss ₂ ,...,rss _n ].

(13) The fingerprint vector of each reference point can be combined and stored as an offline fingerprint database of the area to be measured.

Offline ensemble learning training includes the following steps:

(11) After normalizing the RSS data [rss ₁ , rss ₂ ,..., rss _n ] in the fingerprint database to zero mean and standard deviation, divide the offline data set into the fingerprint database data according to the ten-fold cross-validation method. The fingerprint database data is divided into training set, validation set and test set. During the training process, training is performed on the training set, and the positioning accuracy is verified on the validation set to improve the training ability. The final positioning result is tested on the test set, which prevents overfitting while ensuring high accuracy.

(12) Using random forest regression, multilayer perceptron regression and multilayer perceptron classification model as the base learner to combine the training set and the validation set into ensemble learning, the objective function of the training process is the encoded neurons (ie The mean square error of the mapping relationship between the reconstructed wireless signal fingerprint feature) and the labelled position minimizes the objective function optimization parameters.

The base class learner random forest regression learning method is composed of multiple decision trees. The _n APs of the measured signals [rss ₁ , rss ₂ , . The reference point position label [x _i , y _i ] is equivalent to the leaf node of the decision tree, that is, the two-dimensional output result of the algorithm. The heuristic node selection method C4.5 algorithm is adopted. First, find the attributes whose information gain Ent(·) is higher than the average level, and then select the attributes with the largest information gain rate Gain(·) to construct a decision tree. The construction of decision tree also requires grid coordination to adjust some parameters to effectively prevent overfitting and underfitting, such as feature selection criteria, maximum depth of decision tree, minimum impurity of node division, etc. The choice of parameters is also another factor that affects the positioning accuracy. Random forest estimates the importance of each feature in the classification result, and takes the class that appears most in the output of the decision tree as the final output of the random forest classifier.

A recursive process consisting of smaller subsets of the input dataset trains each decision tree until all tree nodes achieve a similar output goal. Random forest classifiers take input-based weights as parameters similar to the number of decision trees. Instead of the classification algorithm, RFR (Random Forest Regression, random forest regression learning algorithm) is selected to train the mapping of fingerprints and labels, and the feature division standard splitting standard is the minimum mean square error:

In the formula, y _L* represents the value of the left subtree, and y _R* represents the value of the right subtree.

The multilayer perceptron algorithm is a feedforward artificial neural network. The hidden layer, the number of neurons and the nonlinear activation function in the algorithm also need to be carefully selected. Every node except the input node is a neuron under the action of a non-linear activation function. We use a multilayer perceptron with two hidden layers of 100 neurons each. To speed up the learning process, SGD (Stochastic Gradient Descent) is used to update the parameters when minimizing the loss function.

where η is the learning rate that controls the step size in the search parameter space, and Loss is the loss function of the entire network.

Ensemble learning comprehensively considers the results of multiple base learner models. An important advantage of the averaging combined strategy does not require prior tuning of each underlying weak learner. The averaging method effectively combines the algorithms of a variety of basic learners to improve the accuracy, robustness and generality of the system, and is one of the most popular voting methods. It works as follows: Although each individual learner can estimate the position, the performance of the learner is not the same, and the average method combined with the strategy can prevent the basic learner from overfitting.

Among them, the position (x _i , y _i ) is the prediction result of each basic learner _, and Wi is the weight of the prediction result of the basic learner to the final position estimation. Here we simply set the weights to 1.

(13) Although each individual learner can estimate the position, the performance of the learner is not the same. Using the average method combined with the strategy can prevent the basic learner from overfitting. For the measurement of the positioning system, the distance between the predicted position and the actual position is used here as a critical criterion for accuracy. Remember that there are L unknown unknown reference points in the area to be located, and the real geographic location of the ith reference point is recorded as ( _xi , y _i ), and the unknown estimated in the input positioning system is recorded as ( _xi ′, y _i ′ ), then ALE (average localization error, average localization error) can be expressed as:

The online location estimation process includes the following steps:

利用已经训练好的位置指纹特征与位置标签的网络模型，针对新输入的在线特征估计其对应的位置。Fingerprints from n APs collected at the i-th test point in the test set

Using the network model of the trained location fingerprint feature and location label, the corresponding location of the newly input online feature is estimated.

(3) During the online adjustment stage, it is necessary to perform a whitening operation on the test source domain data received by another mobile phone to eliminate the feature correlation of the source domain data; then calculate the correlation between the training target domain data in the fingerprint database, The correlation is applied to the whitened source data to reconstruct the characteristics of the source data. Through the continuous approximation between the second-order statistical information of the source domain and the target domain, the source domain and the target domain are fused. The second-order statistical properties retain the data feature alignment and distribution information, and generally include the heterogeneity caused by differences in the chip, antenna orientation, and mounting materials of the device.

The steps of whitening and reconstructing the online features in step (3) include: an initialization phase, a modeling phase and an optimization phase, specifically including:

希望其对齐的目标域的数据是指纹库中的训练数据R_T＝(r₁,…,r_k)，其中，

源域和目标域的特征都是n维向量。记这两个域特征向量的均值和协方差矩阵分别为μ_S、μ_T和Cov_S、Cov_T，经过中心化标准后，μ_S＝μ_T＝0，但是Cov_S≠Cov_T。(11) In the initialization phase, for the data in the given test set as the source domain data RSS _S =(rss ₁ ,...,rss _m ), where,

The data of the target domain whose alignment is desired is the training data R _T =(r ₁ ,...,r _k ) in the fingerprint library, where,

The features of the source and target domains are both n-dimensional vectors. Denote the mean and covariance matrix of these two domain eigenvectors as μ _S , μ _T and Cov _S , Cov _T , respectively. After the centralization standard, μ _S = μ _T = 0, but Cov _S ≠ Cov _T .

(12) In the modeling stage, in order to reduce the difference between the second-order statistical characteristics of the source domain and target domain features, that is, the distance between the covariance Cov _S and Cov _T , the source domain features are linearly changed to minimize the source domain after the change. The distance between the feature and the target feature covariance is sufficient.

Then the objective function is:

表示线性变化后源域特征的协方差矩阵。in,

Represents the covariance matrix of the source domain features after linear variation.

的最优解为

其中，

由于协方差矩阵是对称矩阵，对Cov_S特征分解很容易得Cov_S＝U_SΣ_SV_S：(13) In the optimization stage, perform singular value decomposition on the covariance matrix of the objective function Cov _T =U _T Σ _T V _T , where U _T[1:r] and Σ _T[1:r] V _{T[1:r ]} are the left singular vector of Cov _T , the largest r singular values and the corresponding right singular vector, then

The optimal solution of is

in,

Since the covariance matrix is a symmetric matrix, it is easy to decompose the Cov _S eigendecomposition to Cov _S =U _S Σ _S V _S :

得，

则线性变化重建过程可以很容易地推导为Depend on

have to,

Then the linearly varying reconstruction process can be easily deduced as

为白化过程，第二部分

为对齐过程In the above formula, the first part

for the whitening process, part 2

for the alignment process

Instead of simply performing feature normalization, CORAL aligns two different distributions. Since correlation alignment only changes features, it can be applied to any basic classifier. The target domain is changed efficiently, avoiding the difficulty of subspace projection and dimension selection in manifold learning.

次幂，消除源域数据的相关性。再以目标函数的二阶统计特性对白化特征作相关性叠加变换。(14) For users who need online positioning, provide the fingerprint vector at the test point, and calculate the covariance matrix of the target domain fingerprint database and the source domain online test fingerprint respectively. For the source domain covariance matrix, whitening transformation is performed first, and the generalized inverse matrix of the covariance matrix is used.

power to remove the correlation of the source domain data. Then use the second-order statistical characteristics of the objective function to perform correlation stacking transformation on the whitening features.

(15) The online features after CORAL alignment can use the data in the original training set for location estimation.

(4) Relying on the new test data fused by the target domain, the influence of equipment heterogeneity is well compensated, and the estimated position coordinates can be obtained through the trained position estimation network. To prevent errors caused by randomness to the experimental results, we are simulating the online positioning of various heterogeneous devices respectively.

The present invention minimizes the offset between the two fields by aligning the fingerprint vector collected by the online positioning terminal with the second-order statistical information of the fingerprint database fingerprint in the offline phase, and does not require any information about the terminal label. Based on the transfer learning framework, domain adaptation and elimination of terminal differences are combined to improve the scalability of the positioning system. Before the classifier is trained, the fingerprint database collected by the fixed terminal in the offline phase is used as the target feature, and the source-target whitening alignment of the fingerprint of any terminal in the online phase can greatly reduce the damage to the localization performance caused by the heterogeneity. The algorithm is simple and fast to realize online adjustment, and achieves ideal performance in actual multi-terminal positioning.

The technical means disclosed in the solution of the present invention are not limited to the technical means disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be pointed out that for those skilled in the art, without departing from the principle of the present invention, several improvements and modifications can be made, and these improvements and modifications are also regarded as the protection scope of the present invention.

Claims (6)

1. A method for realizing high-precision indoor positioning of heterogeneous equipment in a self-adaptive mode in the field of use is characterized by comprising the following steps:

(1) simulating a positioning process of various equipment heterogeneous conditions in an actual scene through a common terminal;

(2) for an indoor positioning system, sampling points need to be uniformly divided for an environment to be tested so as to collect and build a fingerprint database of wireless signal intensity of each peripheral access point; then off-line ensemble learning training is carried out, in the training stage, the mapping relation between each reference point position label and the corresponding fingerprint is trained and established through ensemble learning formed by random forest regression, multilayer perceptron regression and multilayer perceptron classification, and the trained position estimation network model is stored; finally, carrying out an online position estimation test, and carrying out real-time positioning by using a trained algorithm after related alignment online adjustment when an online user receives fingerprints of surrounding access points;

(3) in the online adjustment stage, whitening operation is firstly carried out on test source domain data received by another mobile phone, and the characteristic correlation of the source domain data is eliminated; then calculating the correlation among the training target domain data in the fingerprint database, acting the correlation on the whitened source data, and reconstructing the characteristics of the source data; fusing the source domain and the target domain through continuous approximation between the second-order statistical information of the source domain and the target domain; the second-order statistical characteristic keeps the data feature alignment and also keeps the distribution information;

(4) the fingerprints received in real time are passed through a trained position estimation network, and the algorithm can predict position coordinates.

2. The method for realizing high-precision indoor positioning of heterogeneous equipment in a use field self-adaption mode according to claim 1, wherein the step of building the fingerprint database in the step (2) specifically comprises the following steps:

(11) uniformly dividing a region to be measured into grid reference points at certain intervals according to the area, and taking the grid reference points as sampling points of a positioning region;

(12) at each reference point, the wireless signal strength sent by all access points AP in a period of time nearby is recorded by the acceptance of the terminal A, and a reference point is formedPosition [ x ]_i,y_i]MAC address of each AP and corresponding received signal strength rss₁,rss₂,...,rss_n]Is given as a fingerprint vector ([ rss ]₁,rss₂,...,rss_n],[x_i,y_i])；

(13) The fingerprint vectors of each reference point can be combined and stored as an off-line fingerprint library of the area to be detected.

3. The method for realizing high-precision indoor positioning of heterogeneous equipment by using domain self-adaptation as claimed in claim 1, wherein the step of off-line integrated learning training in the step (2) comprises the following processes:

(11) for RSS data [ RSS ] in fingerprint database₁,rss₂,...,rss_n]After normalization to zero mean and standard variance, dividing the off-line data set into fingerprint database data according to a cross-over verification method, [ rss [ [ rss ]₁,rss₂,...,rss_n]Is the signal strength; the fingerprint database data is divided into a training set, a verification set and a test set; in the training process, training is carried out on a training set, and the final positioning result is tested on a test set;

(12) combining a random forest regression model, a multilayer perceptron regression model and a multilayer perceptron classification model as a base learner on a training set and a verification set to form integrated learning, wherein an objective function in the training process is the mean square error of a mapping relation between coded neurons and positions with labels, and objective function optimization parameters are minimized;

(13) taking the distance between the predicted position and the actual position as a critical standard of precision, and recording that the to-be-positioned area has L unknown reference points, wherein the real geographic position of the ith reference point is recorded as (x)_i,y_i) The estimated position in the input positioning system is noted as (x)_i′,y_i'), the average positioning error ALE can be expressed as:

4. the method for realizing high-precision indoor positioning of heterogeneous equipment by using field adaptive method according to claim 3, wherein the step (12) comprises the following steps:

the RFR algorithm of the base class learner random forest regression learning method consists of a plurality of decision trees; measured signal [ rss ]₁,rss₂,...,rss_n]Can be regarded as the characteristic of the division condition, as the input data of the algorithm; reference point location label [ x ]_i,y_i]The leaf nodes of the decision tree are equivalent, namely the two-dimensional output result of the algorithm; a heuristic node selection method C4.5 algorithm is adopted, the attribute that the information Gain Ent (-) is higher than the average level is found out, and then the attribute with the maximum information Gain rate Gain (-) is selected to construct a decision tree; the construction of the decision tree requires the coordination and adjustment of some parameters of the grids to effectively prevent over-fitting and under-fitting; estimating the importance of each feature in the classification result by the random forest, and taking the class with the most occurrence in the decision tree output as the final output of the random forest classifier;

training each decision tree by a recursive process consisting of smaller subsets in the input data set until all tree nodes reach similar output targets; the random forest classifier takes the weight value based on input as a parameter similar to the number of the decision trees; instead of a classification algorithm, the RFR algorithm is chosen to train the mapping of fingerprints to labels, and the feature classification criterion is the minimum mean square error:

in the formula, y_L*Value, y, representing the left sub-tree_R*The value of the right subtree represented;

the multi-layer perceptron algorithm is a feedforward type artificial neural network; hidden layers, the number of neurons and nonlinear activation functions in the algorithm need to be carefully selected; except for the input nodes, each node is a neuron under the action of a nonlinear activation function, and when a multilayer perceptron is used, two hidden layers with 100 neurons in each layer are adopted; the parameters are updated using the SGD method when minimizing the loss function:

wherein eta is the learning rate of the control step length in the search parameter space, and Loss is the Loss function of the whole network;

the integrated learning comprehensively considers the results of a plurality of basic learner models; the average method is adopted to combine the strategies:

wherein, the position (x)_i,y_i) Is the result of the prediction of the respective base learner, and W_iIs the weight of the base learner prediction result to the final position estimate.

5. The method for realizing high-precision indoor positioning of heterogeneous equipment by using domain self-adaptation as claimed in claim 1, wherein the online position estimation step in the step (2) comprises the following processes:

fingerprints from n APs collected at the ith test point in the test set

And estimating the corresponding position of the newly input online feature by using the trained network model of the position fingerprint feature and the position label.

6. The method for realizing high-precision indoor positioning of heterogeneous equipment by using domain adaptation as claimed in claim 1, wherein the whitening and rebuilding process of the online features in the step (3) comprises the following initialization stage, modeling stage and optimization stage:

(11) initialization phase, data RSS as source domain for data in given test set_S＝(rss₁,…,rss_m) Wherein the signal strength

The data of the target domain whose alignment is desired is the training data R in the fingerprint library_T＝(r₁,…,r_k) Wherein

the characteristics of the source domain and the target domain are n-dimensional vectors; the mean and covariance matrices of the two domain eigenvectors are recorded as mu respectively_S、μ_TAnd Cov_S、Cov_TAfter centering criterion, μ_S＝μ_T0, but Cov_S≠Cov_T；

(12) A modeling stage for reducing the difference between the second-order statistical properties of the source domain and target domain features, namely covariance Cov_SAnd Cov_TThe source domain features are linearly changed, and the distance between the covariance of the changed source domain features and the covariance of the changed target features is minimized; the objective function is then:

wherein,

a covariance matrix representing the source domain features after linear variation;

(13) in the optimization stage, singular value decomposition Cov is carried out on the covariance matrix of the objective function_T＝U_TΣ_TV_TWherein, U_T[1:r]、Σ_T[1:r]V_T[1:r]Are each Cov_TThe largest r singular values and the corresponding right singular vectors, then

Is optimally solved as

Wherein,

since the covariance matrix is a symmetric matrix, for Cov_SFeature decomposition to get Cov_S＝U_SΣ_SV_S：

By

So as to obtain the compound with the characteristics of,

the linear variation reconstruction process is derived as:

in the above formula, the first part

For the whitening process, the second part

An alignment process;

(14) providing fingerprint vectors at test points for users needing online positioning, and respectively calculating covariance matrixes of online test fingerprints of a target domain fingerprint library and a source domain; the source domain covariance matrix is firstly subjected to whitening transformation, and the covariance matrix is used as a generalized inverse matrix

The power of the next, eliminating the correlation of the source domain data; then using the second order of the objective functionThe statistical characteristics perform correlation superposition transformation on the whitened features;

(15) and carrying out position estimation on the online characteristics after CORAL alignment by using data in the original training set.

Images