CN116170874A

CN116170874A - Robust WiFi fingerprint indoor positioning method and system

Info

Publication number: CN116170874A
Application number: CN202310165399.8A
Authority: CN
Inventors: 张登银; 庄昌胜; 陈小星
Original assignee: JIANGSU YITONG HIGH-TECH CO LTD
Current assignee: JIANGSU YITONG HIGH-TECH CO LTD
Priority date: 2023-02-24
Filing date: 2023-02-24
Publication date: 2023-05-26

Abstract

The invention discloses a robust WiFi fingerprint indoor positioning method and a robust WiFi fingerprint indoor positioning system, wherein the method comprises the following steps: acquiring an RSS value of a WiFi access point of a reference point, and carrying out data enhancement on original data after acquisition is completed to construct a WiFi fingerprint database; training a feature extraction module by utilizing a fingerprint database, wherein the neural network structure of the feature extraction module is an improved stacked denoising self-encoder and is used for extracting robust features; the output of the feature extraction module is used as the input of a floor classification model and a coordinate regression model to carry out neural network training; preprocessing RRS data of the to-be-positioned point, inputting the RRS data into a feature extraction module after training is completed to obtain a stable feature, and inputting the feature as a floor classification model and a coordinate regression model to obtain the position estimation coordinate of the to-be-positioned point. The method effectively reduces the influence of WiFi signal fluctuation on the fingerprint positioning precision.

Description

Robust WiFi fingerprint indoor positioning method and system

Technical Field

The invention relates to a robust WiFi fingerprint indoor positioning method, and belongs to the field of wireless positioning and deep learning.

Background

Indoor positioning is widely applied to indoor positioning-based services such as wireless advertisement push, information retrieval, pedestrian navigation, and the like. At present, the fingerprint identification technology based on WiFi has the advantages of no invasiveness, higher positioning precision, easiness in deployment, strong universality and no need of additional equipment, and is the most popular positioning technology at present. The vector of received signal strengths of a plurality of WiFi Access Points (APs) measured at a certain location is a fingerprint that can mark its location, and the location of the user/device can be estimated directly by finding the closest match between the received signal strength (Receiving signal strength, RSS) measurements of WiFi and the fingerprint of known locations in the database, without having to consider the channel attenuation model in that environment.

The traditional indoor positioning method based on Wi-Fi fingerprint mainly adopts deterministic or probabilistic technology, the deterministic method estimates the target position according to the criterion of shortest distance (such as Euclidean distance), such as KNN, linear discriminant analysis and support vector machines (Support vector machine, SVM), RADAR is the most representative positioning system of the algorithm, the probabilistic method estimates the target position by using posterior probability calculated by the probabilistic inference method, and the Horus system is the representative of the algorithm.

Unlike traditional methods including probability, KNN, SVM, etc., which are computationally intensive and time consuming, and require complex filtering and parameter adjustments, deep learning can efficiently process large amounts of fingerprint data, extract reliable features, and construct internal representations from dynamic signals without additional human intervention, and in addition, due to the parametric nature of deep learning, the computational complexity after neural network training is complete is not dependent on the size of the data set, which can reduce the positioning delay to milliseconds, making the deep learning algorithm an ideal choice for real-time positioning applications.

The RSS signal has a certain mapping relation with the distance under ideal conditions, but due to the complexity of indoor scenes, especially in markets, train stations and teaching buildings where people gather more, interference of other wireless devices such as bluetooth, zigBee and the like causes fluctuation of WiFi signals.

The fluctuation of the WiFi signals enables the neural network model to be incapable of accurately fitting the mapping relation between the RSS fluctuation interval and the position coordinates, and has a large negative effect on the positioning accuracy, so that in order to keep good positioning performance, new samples need to be periodically acquired and the positioning model needs to be trained, but the data acquisition process is time-consuming and labor-consuming.

Disclosure of Invention

The technical problems to be solved by the invention are as follows: on the premise of ensuring positioning accuracy and low time delay, how to improve the robustness of a positioning system and reduce the negative influence of fluctuation of WiFi signals.

The principle of the invention is as follows: sampling of reference points at different moments of different APs (access points) generates new fingerprint data according to a certain rule, then filters the generated data, aims at a class of classification problems, utilizes the countermeasure idea of generating a countermeasure network (Generating adversarial network, GAN), uses the generator to provide a large number of negative samples, trains a discriminator, uses the trained discriminator to filter data, retains data higher than a threshold of the discriminator, and realizes supplementation of a fingerprint database, thereby improving positioning accuracy and reducing time-consuming manual acquisition processes; in addition, in order to accurately fit the mapping relation between the RSS fluctuation interval and the position, the SDAE is used for obtaining a robust time-independent feature from the dynamic WiFi signal as the input of a floor classification model and a regression model, and the noise adding mode of the SDAE is optimized at the same time, so that the distribution of the SDAE can approximate the distribution of new fingerprint data which fluctuates due to time, and the integrity of the original information is ensured; in order to enable the deep learning model to accurately fit the mapping relation between the RSS and the position coordinates, a one-dimensional CNN network is adopted to train the floor classification model, the robustness of the system is further improved by utilizing the characteristics of a convolution layer, and a coordinate regression model is built by using a multi-layer perceptron.

In order to solve the technical problems, the invention provides the following technical scheme:

a robust WiFi fingerprint indoor positioning method comprises the following steps:

1) Dividing the indoor area, and collecting RSS fingerprint data at a reference point;

2) According to the distribution characteristics of the RSS data, carrying out data enhancement on the acquired original data;

3) Filtering the expanded data by a discriminator, combining the filtered data with the original data, and constructing a fingerprint database;

4) Preprocessing data in a fingerprint database, and dividing the fingerprint database into a training set, a verification set and a test set;

5) The training feature extraction module is used for mapping the high-dimensional dynamic fingerprint data into a robust feature vector, and the feature extraction model is obtained after training;

6) Taking the output of the feature extraction module as a sample to finish the training of the floor classification model and the coordinate regression model;

7) The RRS data of the to-be-positioned point is input to a feature extraction module after being preprocessed, the output of the feature extraction module is input to the floor classification model and the coordinate regression model after training is finished again, and the position estimation coordinate of the to-be-positioned point is obtained through output.

A robust WiFi fingerprint indoor positioning system, comprising the following modules:

and a data acquisition module: dividing the indoor area, and collecting RSS fingerprint data at a reference point;

and a data enhancement module: according to the distribution characteristics of the RSS data, carrying out data enhancement on the acquired original data;

a database module: filtering the expanded data by a discriminator, combining the filtered data with the original data, and constructing a fingerprint database;

and a pretreatment module: preprocessing data in a fingerprint database, and dividing the fingerprint database into a training set, a verification set and a test set;

training module of the feature extraction module: the training feature extraction module is used for mapping the high-dimensional dynamic fingerprint data into a robust feature vector, and the feature extraction model is obtained after training;

model training module: taking the output of the feature extraction module as a sample to finish the training of the floor classification model and the coordinate regression model;

a position estimation module: the RRS data of the to-be-positioned point is input to a feature extraction module after being preprocessed, the output of the feature extraction module is input to the floor classification model and the coordinate regression model after training is finished again, and the position estimation coordinate of the to-be-positioned point is obtained through output.

The invention has the beneficial effects that: the invention uses a reliable data enhancement method, reduces the data acquisition burden, expands the combination range of fingerprints and the distribution of single-point RSS aiming at the problem of generating covariate offset, enhances the positioning accuracy of a neural network model through data set enhancement, fully uses the existing WiFi equipment, has no extra hardware requirement, is convenient to realize, does not need to carry a special sensor, can realize indoor positioning by using intelligent equipment such as a mobile phone and the like, and has universality; the invention provides a reliable data enhancement method, and aims at the problem of negative sample deletion of a class of classifier, a similar sample is generated by using a GAN generator to train a discriminator, and the generated data is filtered by using the discriminator; the invention improves SDAE aiming at the property of RSS, extracts stable characteristics from fingerprint data while avoiding key information loss, maps dynamic high-dimensional fingerprint data into low-dimensional stable characteristic vectors, and then takes the extracted characteristics as the input of a floor classification model and a coordinate regression model to improve the robustness of a neural network, thereby improving the positioning accuracy.

Drawings

FIG. 1 is a flow chart of a robust WiFi fingerprint indoor positioning method of the invention;

FIG. 2 is a schematic diagram of a data enhancement method;

FIG. 3 is a graph showing GAN results;

FIG. 4 is a schematic diagram of a DAE structure including a hidden layer;

FIG. 5 is a schematic diagram of an SDAE training procedure;

fig. 6 is a schematic diagram of a one-dimensional convolution process.

Detailed Description

The invention is further described below with reference to the accompanying drawings. The following examples are only for more clearly illustrating the technical aspects of the present invention, and are not intended to limit the scope of the present invention.

Example 1

Fig. 1 is a flowchart of a robust WiFi fingerprint indoor positioning method according to the present invention;

4) Preprocessing data in a fingerprint database, dividing the fingerprint database into a training set, a verification set and a test set, wherein the number of each floor is balanced by sampling the verification set;

7) The RRS data of the to-be-positioned point is input to the feature extraction module after being preprocessed, the output of the RRS data is input to the floor classification model and the coordinate regression model after training is finished again, and the position estimation coordinates of the to-be-positioned point are obtained through output.

In step 1), a coordinate origin is selected in a positioning area, a coordinate system is established in a positive north direction as a y-axis forward direction and in a positive east direction as an x-axis forward direction, reference points are selected at set intervals, the density of RSS fingerprint data acquisition is positively correlated with positioning accuracy, the data acquisition density can be determined according to requirements, and all indoor WiFi access points are numbered and marked as AP _i And mapping is performed through the hardware IDs of the APs, so as to determine which AP the signal of the reference point comes from, and there are M reference points and N APs.

Taking a reference point as a unit to collect fingerprint data, wherein the fingerprint data is obtained by RSS measured values RSS of N APs _i Composition ranging from-110 dBm to 0dBm, acquisition of data from AP at divided reference points _i RSS measurement RSS of (2) _i Then according to the hardware ID of the AP, mapping to the corresponding subscript, and storingRSS _i The missing value is set to be-110 dBm, after the RSS signal is collected, the fingerprint data is added with a label according to the actual position of the reference point, wherein the label comprises a building IDBuildID, a floor IDFloorID, longitude and latitude and a reference point ID, the preparation is carried out for training the neural network, and the fingerprint data is in the following format:

{RSS ₀ ,RSS ₁ ,RSS ₂ ,…,RSS _N-1 ,BuildID,FloorID,x,y,RPID}

RPID is a reference point ID from AP _i The RSS signal value is denoted as RSS _i Longitude and latitude are expressed using relative coordinates (x, y).

The more samples that are taken at the reference point, the better, but this is a time consuming process, which can be supplemented by data enhancement, the number of single point samples should be no less than 20 for the accuracy of the data enhancement method.

In step 2), the flow of the data enhancement method is shown in fig. 2, in the data enhancement method, the RSS value received by a certain AP corresponding to the reference point is not affected by other APs, and is only related to own propagation paths, namely, the APs are mutually independent, and the specific fingerprint data enhancement method comprises the following steps:

21 Classifying fingerprint data according to the ID tag of the reference point, classifying the RSS signals again by the reference point l according to the AP, forming a set of N RSS measured values by the reference point after classification, wherein the set is recorded as RSS_AP _i The reference point l is based on the different fingerprint numbers K that the original measured data can generate:

K＝k ₀ ×k ₁ ×…k _i …×k _N-1

wherein k is _i For RSS_AP _i The number of different RSS values in the set, the value of the fingerprint number K is usually far greater than the sampling number of the reference point l;

22 From rss_ap _i Is selected randomly as reference point/from the AP _i Extracting N measured values from the N sets to form new fingerprint data; unchanged aggregate RSS_AP _i The signal distribution of the reference point from a single AP can be ensured to be unchanged;

23 The RSS distribution presents normal distribution of left bias, and data expansion is carried out according to a distribution model, and the specific steps are as follows:

a) Computing aggregate RSS_AP _i Missing value (RSS) _i = -110) ratio r, mean μ after deletion value removal, standard deviation σ; the RSS distribution exhibits a normal distribution with left bias, and the normal distribution N (μ, σ) is fitted according to the mean and standard deviation ² ) From AP as reference point l _i An RSS signal distribution model of (a);

b) First in [0, 1]]Generating a random number rand in range, if rand<r, RSS is then applied _i Set to-110, otherwise according to normal distribution N (μ, σ) ² ) Generating random numbers as RSS _i Finally, generating RSS values of N APs in the reference point l according to the method to form new fingerprint data;

c) In view of the limited number of samples and the simpler sampling environment, the true mean is generally smaller than the mathematical expectation of the fitted model, thus preserving the normal distribution N (μ, σ) of the fit ² ) Besides, RSS_AP is added again _i The mean value mu minus 3 of the normal distribution is used as the mathematical expectation of the normal distribution, the standard deviation is unchanged, and a normal distribution N (mu-3, sigma) is regenerated ² ) For generating fingerprint data; the two fitting models generate equivalent data, and the signal distribution of a single AP is expanded on the basis of keeping the original data distribution unchanged, so that the possible data migration phenomenon in reality is simulated, and the robustness of the system is enhanced.

In step 3), the classifier is trained by using the GAN model as a filter, the data generated by the data enhancement method is filtered by using the generator, the data generated by the generator is judged by the classifier, the model generated by the generator in the training process of the GAN model tends to be real, meanwhile, the classifier is trained by obtaining a high-quality negative sample, after the training is finished, the classifier is used as a classifier for filtering the data, the problem of negative sample missing in a class of classifier is solved, the generator and the classifier are both composed of 3 layers of linear neural networks, the GAN structure is shown as a graph in fig. 3, and the training process of the GAN model is as follows:

a) A random input is given to the generator network model, and a generated sample set is obtained through corresponding output;

b) Training a network model of a discriminator, wherein the training process is a supervised two-classification problem, mixing original data with a generated sample set, and judging whether the original data is a true sample or a sample generated by using a generated network by the discriminator;

c) Concatenating the generator network model and the discriminator network model, wherein the loss function is a binary cross entropy function of the discriminator network, and updating parameters of the generator network model according to the loss function;

d) After the training of the generated network is completed, giving random input again to obtain a new sample set, and inputting the newly generated sample set into the discriminator network model for training, so that the further training of the discriminator network model is completed;

e) Repeating steps a) -d) until the loss functions of the generator and the discriminator are converged.

And after the training of the GAN model is completed, a discriminator of the GAN is taken as a filter, data generated by the data enhancement method is filtered, the default threshold of the discriminator is 0.5, and the data generated by enhancing the original data and the filtered data are enhanced.

In the step 4), normalization processing is carried out on the data set after data enhancement, so that gradient explosion of the neural network is avoided, the loss function is difficult to converge, and the normalization process is shown in the following formula:

wherein max_rss is the largest RSS measure in the data set, min_rss is the smallest RSS measure, RSS _i ' is the updated measured value rss _i ；

Then according to 9:1: the 1 scale divides the data set into three parts, a training set, a validation set and a test set. Furthermore, to test the robustness of the system, the test set data may be additionally sampled 7 days after the data acquisition.

For convenience of neural network construction, modeling was performed using a public WiFi fingerprint dataset ujiindorloc, which contained 21049 fingerprint samples, covering 3 buildings of 4 to 5 floors, 933 reference points, and coverage of 108703 square meters. Each sample contains 520 RSS measurements ranging from-110 dBm to 0dBm, and further contains reference point location information, such as building and floor numbers, latitude and longitude, time stamps, and user and device tags, each reference point averages about 23 samples, each RP averages 18 APs detected, 1111 samples were re-acquired the fourth month after the dataset was built, and the acquisition points were not repeated with the reference points, serving as a test set.

In step 5), the feature extraction module is trained, the feature extraction module is realized by improved SDAE, the denoising self-encoder (Denoising autoencoder, DAE) is one of AE, AE is a neural network which enables an output value to be equal to an input value by using a back propagation algorithm, the back propagation algorithm continuously adjusts the connection weight of the neural network according to the error between the prediction of forward response propagation and a reference signal, and the learning process of the back propagation algorithm is as follows:

a) The sample set is propagated layer by layer from the input layer to the output layer through the hidden layer, namely, forward propagation;

b) The difference between the expected output and the actual output of the neural network, an error signal is propagated from the output layer to the input layer by layer through an implicit layer, and the connection weight is corrected, namely error counter propagation;

c) Repeatedly and alternately performing memory training by forward propagation and error counter propagation;

d) When the loss function tends to 0 and the fluctuation is smaller than the set value, learning convergence is completed;

the AE neural network structure consists of two symmetrical parts, namely an encoder and a decoder, the DAE adds noise to the original input based on the AE, then the original input data is reconstructed through the self-encoder, the encoder is forced to learn and extract important robust features in the input data, the noise adding mode of the DAE is similar to a Dropout layer, and a certain proportion of inputs are randomly set to 0 to destroy the data. The neural network structure of the single-layer DAE is shown in fig. 4, and includes a noise adding part, an encoder and a decoder.

The DAE adds noise to training data through a noise adding part, so that a self-encoder is forced to learn to remove the noise to obtain real input which is not polluted by the noise, the encoder is enabled to learn to extract the most important features and learn the more robust representation in the input data, the SDAE is a training mode of an automatic encoder, the SDAE adopts a layer-by-layer training mode, the reconstruction error can be furthest reduced, the SDAE neural network structure consists of three stacked DAEs, the training process is shown in figure 5, and the specific steps are as follows:

a) Given initial input, training a layer of automatic encoder in an unsupervised mode, and reducing reconstruction errors to reach a threshold value;

b) Taking the output of the hidden layer of the last automatic encoder as the input of the current automatic encoder, and training the automatic encoder;

c) Repeating step b) until training of all automatic encoders is completed;

d) The output of the hidden layer of the last auto-encoder is used as the feature of the original data extraction.

Neurons in layers of SDAE are fully connected, there is no connection between each neuron in a layer, y is denoted as a given target value, L (·) is denoted as an objective function, and given any set of input modes (x (i), y (i)), i=1, 2,..:

summing and averaging over N data yields a loss function:

wherein w is a weight value in a neural network neuron, b is a bias, and θ is other parameters; h is a _w,b (x (i) -y (i)) is an error metric function,

is a weight decay term;

the objective function value is obtained through feedforward calculation, then a gradient descent algorithm is adopted according to a criterion of minimizing the loss function, the gradient value is reversely propagated, and the weight learning rule is as follows:

w _t ^(l) the method is characterized in that the weight of the first layer of the neural network in the t-th training is that epsilon represents the learning rate and is used for defining the updating amplitude of each parameter, SDAE adopts an Adam algorithm, and the learning rate of each parameter is dynamically adjusted according to the gradient first moment estimation and the second moment estimation of each parameter in the neural network by a loss function.

Traditional SDAE, directly employing Dropout layer as a noise addition, forces the encoder to learn and extract robust features from corrupted data. But the fingerprint data is sparse, taking the ujidroololoc data set as an example, the fingerprint data length is 520, the average effective signals are 18, after the data preprocessing, only 18 non-0 values are averaged, if the Dropout layer is randomly set to 0, the training effect is unstable, and the setting of 0 means that the effective signals are directly changed into invalid signals, for example, -10dBm is set to-110 dBm. Such wide-range fluctuation does not conform to the actual situation, and partial information may be lost, affecting the positioning accuracy of the regression model.

Reasonable noise is added into fingerprint data, so that the distribution of the fingerprint data can approximate to the distribution of new fingerprint data which fluctuates due to time, then an automatic encoder is used for extracting stable characteristics irrelevant to data time in a mode of re-reconstructing after encoding, so that the encoder learns to map dynamic fingerprint data into low-dimensional stable characteristic vectors, and the influence of signal fluctuation on positioning accuracy is reduced.

The fluctuation range of the RSS signal is generally within 20dBm, values are randomly taken between [ -10,10], noise addition is carried out on the data, and the values are randomly taken between [ -0.91,0.91] after normalization because the RSS range is between [ -110,0 ]; adding [ -0.91,0.91] noise between effective signals when the first DAE adds noise to input data, and simulating signal fluctuation, wherein the second DAE and the third DAE normally use a Dropout layer as a noise adding mode; the activation functions of the three DAE are all ReLu, and the number of neurons of the encoder output layer are 256, 128, and 64, respectively. After training is completed, 64 time independent robust features are finally extracted using SDAE.

In the step 6), in the floor classification model, the combination of the building ID and the floor ID is used as a label of the floor classification model, 13 types are shared, 64 features extracted by SDAE are used as inputs of the classification model, a neural network structure of the floor classification model is composed of three one-dimensional convolution layers and three full-connection layers, the number of convolution kernels of the three one-dimensional convolution layers is 32, 64 and 32, the convolution kernels are 15, 9 and 5, the step sizes are 2, 1 and 1, and the filling is 7, 4 and 2; the one-dimensional convolution process is shown in fig. 6, the number of neurons is 520, 128 and 32 respectively, dropout layers with parameters of 0.5 are added between each layer, overfitting is prevented, reLu is used as an activation function and a cross entropy function is used as a loss function in the whole network.

In the training of the coordinate regression model, 64 features extracted by SDAE are used as input of the coordinate regression model, the coordinate regression model comprises an 8-layer neural network, the number of neurons is 128, 256, 128, 64, 32, 16 and 8 respectively, reLu is used as an activation function, a regularization layer is adopted to accelerate the convergence process of a loss function, meanwhile, the problem of overfitting is relieved, as the relative coordinate range is between 0 and 400 m, the horizontal and vertical coordinates are normalized to be between 0 and 1 for accelerating the convergence, the activation function of an output layer is correspondingly set as Sigmoid, the output is mapped to (0, 1), and the mean square error is used as the loss function.

In step 7), the RRS data of the to-be-positioned point is input to the feature extraction module after normalization, the output of the RRS data is input to the floor classification model and the coordinate regression model after training is finished again, and the position estimation coordinate of the to-be-positioned point is obtained through output.

It will be appreciated by those skilled in the art that embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present application is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the application. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The foregoing is merely a preferred embodiment of the present invention, and it should be noted that modifications and variations could be made by those skilled in the art without departing from the technical principles of the present invention, and such modifications and variations should also be regarded as being within the scope of the invention.

Claims

1. The robust WiFi fingerprint indoor positioning method is characterized by comprising the following steps of:

2. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein:

in step 1), selecting a coordinate origin in a positioning area, establishing a coordinate system by taking the north direction as the y-axis forward direction and the east direction as the x-axis forward direction, selecting reference points at set intervals, acquiring the density of RSS fingerprint data, positively correlating with positioning accuracy, numbering all indoor WiFi access points, and marking as AP _i And mapping is performed through the hardware IDs of the APs, so as to determine which AP the signal of the reference point comes from, and there are M reference points and N APs.

3. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein:

in step 1), fingerprint data are collected by taking a reference point as a unit, wherein the fingerprint data are obtained by RSS measured values RSS of N APs _i Composition ranging from-110 dBm to 0dBm, acquisition of data from AP at divided reference points _i RSS measurement RSS of (2) _i Then according to the hardware ID of the AP, mapping to a corresponding subscript, and storing a measured value RSS _i The missing value is set to be-110 dBm, and after the RSS signal is collected, the fingerprint data is added with a label according to the actual position of the reference point, including a buildingThe object ID buildID, floor ID FloorID, longitude and latitude, reference point ID, fingerprint data format is as follows:

{RSS ₀ ,RSS ₁ ,RSS ₂ ,…,RSS _N-1 ,BuildID,FloorID,x,y,RPID}

4. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein: in step 2), the fingerprint data enhancement method comprises the steps of:

21 Classifying fingerprint data according to the ID tag of the reference point, classifying the RSS signals again by the reference point l according to the AP, forming a set of N RSS measured values by the reference point after classification, and recording the set as RSS_AP _i The reference point l is based on the different fingerprint numbers K that the former measured data generated are:

K＝k ₀ ×k ₁ ×…k _i …×k _N-1

wherein k is _i For RSS_AP _i Different numbers of RSS values in the set;

22 From the aggregate rss_ap _i Is selected randomly as reference point/from the AP _i Extracting N measured values from the N sets to form new fingerprint data;

a) Computing aggregate RSS_AP _i The ratio r of the missing values in (1) is the mean value mu and standard deviation sigma after the missing values are removed; the RSS distribution exhibits a normal distribution with left bias, and the normal distribution N (μ, σ) is fitted according to the mean and standard deviation ² ) From AP as reference point l _i An RSS signal distribution model of (a);

b) First in [0, 1]]Generating a random number rand in range, if rand<r, RSS is then applied _i Set to-110, otherwise according to normal distribution N (μ, σ) ² ) Generating random numbers as RSS _i Finally, according to the method, a reference point l is generatedRSS values of N APs in the database are used for forming new fingerprint data;

c) Preserving the fitted normal distribution N (μ, σ) ² ) Besides, RSS_AP is added again _i The mean value mu minus 3 of the normal distribution is used as the mathematical expectation of the normal distribution, the standard deviation is unchanged, and a normal distribution N (mu-3, sigma) is regenerated ² ) For generating fingerprint data.

5. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein:

in step 3), the data generated by the data enhancement method is filtered by using a GAN model training discriminator as a filter, the GAN model comprises a generator and a discriminator, the generator generates data, the discriminator judges the data generated by the generator, and the training process of the GAN model is as follows:

6. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein,

in step 4), normalization processing is performed on the data set after data enhancement, wherein the normalization process is shown in the following formula:

Then according to 9:1: the 1 scale divides the data set into three parts, a training set, a validation set and a test set.

7. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein:

in step 5), the SDAE neural network structure consists of three stacked DAE, and the training process includes the following steps:

c) Repeating step b) until training of all automatic encoders is completed;

8. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein,

in step 5), neurons in layers of SDAE are in fully connected relation, there is no connection between each neuron in a layer, y is noted as a given target value, L (·) is noted as an objective function, any set of input modes (x (i), y (i)) is given, i=1, 2,..:

summing and averaging over N data yields a loss function:

is a weight decay term;

w _t ^(l) is the weight of the first layer of the neural network trained at the t time, and epsilon represents the learning rate.

9. The robust WiFi fingerprint indoor positioning method according to claim 1, wherein:

in the step 6), in the floor classification model, the combination of the building ID and the floor ID is used as a label of the floor classification model, 64 features extracted by the SDAE are used as the input of the classification model, the neural network structure of the floor classification model is composed of three one-dimensional convolution layers and three full connection layers, the number of convolution kernels of the three one-dimensional convolution layers is respectively 32, 64 and 32, the sizes of the convolution kernels are respectively 15, 9 and 5, the step sizes are respectively 2, 1 and 1, and the filling is respectively 7, 4 and 2; the one-dimensional convolution process comprises three full-connection layers, the number of neurons is 520, 128 and 32 respectively, a Dropout layer with a parameter of 0.5 is added between each layer, the whole network uses ReLu as an activation function and uses a cross entropy function as a loss function;

in the training coordinate regression model, 64 features extracted by SDAE are used as input of the coordinate regression model, the coordinate regression model comprises 8 layers of neural networks, the number of neurons is 128, 256, 128, 64, 32, 16 and 8 respectively, and ReLu is used as an activation function.

10. A robust WiFi fingerprint indoor positioning system, comprising the following modules: