A personnel positioning system based on WIFI device

Abstract

The application discloses a personnel positioning system based on WIFI equipment, comprising a data enhancer, a feature extractor, a position predictor and a domain discriminator, wherein: the position mark data based on the WIFI equipment is input into the data enhancer to generate extended position mark data respectively, and the extended position mark data and unmarked data are input into the feature extractor to extract a feature vector, and then based on the feature vector, the position prediction feature irrelevant to user motion is learned through the position predictor and the domain discriminator to realize personnel positioning independent of user motion. The application can realize robustness for user action, and also can realize high-precision personnel positioning measurement. Specifically, the experimental results in a real scene show that, compared with other prior arts, the position prediction accuracy in the personnel positioning process can be increased by about 35%.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of WIFI sensing, and particularly relates to a personnel positioning system based on WIFI equipment. BACKGROUND

[0002] With the rapid development of network technology, stable and accurate indoor positioning service (ILBS) can provide new opportunities for many network applications. For example, by pushing location updates to users, functions such as geographic social networks in Internet of Things (IoT) applications, point of interest (POI) recommendations, and smart home automation can be realized. In addition, more accurate location awareness can also provide better solutions for applications such as unmanned supermarkets and augmented reality (AR).

[0003] In recent years, many indoor wireless positioning technologies have been widely studied, such as WIFI, RFID, acoustic, and visual-based positioning technologies. Among them, WIFI-based positioning technology has gradually become one of the popular wireless positioning technology solutions due to its ubiquitous infrastructure and low-cost setup.

[0004] WIFI-based positioning systems are mainly divided into device-oriented and device-free solutions. Device-oriented methods achieve decimeter-level positioning accuracy by measuring the angle of arrival (AoA), time of flight (ToF), or received signal strength indicator (RSSI) between a user's smart phone or other WIFI equipment and WIFI access points (APs). In contrast, device-free positioning systems have the advantage of requiring less smart equipment to be carried by users, making them applicable to a wider range of scenarios. Typically, device-free positioning systems associate the collected RSSI or channel state information (CSI) with user locations, then extract location fingerprints and train neural networks to achieve decimeter-level positioning accuracy. Moreover, CSI has rich frequency and spatial diversity information, and its amplitude is more stable in time series, so CSI fingerprints are more representative in terms of location features.

[0005] Although device-free fingerprint positioning technology provides some convenience for users, existing technologies still have many deficiencies in real-world applications, for the following reasons:

[0006] CSI in the system is fine-grainedly reflected the channel characteristics, and thus it is extremely sensitive to the surrounding environment. That is, in the application process, the changes of the environment and the user will affect the CSI location fingerprint. More importantly, when the positioning system is running, the user may also be accompanied by various actions, such as stretching the arm, bending the waist, swinging the arm, etc. Changes in these actions will also change the propagation characteristics of the WIFI signal, even for the same user at the same location. Moreover, the inconsistency caused by such movements is even more frequent than the changes in the environment and the user, which greatly limits the practicality and convenience of the positioning system.

[0007] In order to solve the above-mentioned problems existing in the prior art, in theory, if the CSI fingerprint of all possible movements of each location can be collected, a positioning system that is not affected by movement can be trained. However, considering the diversity of actions and the difference between different user body types, this will be very time-consuming and difficult, so this solution is not very feasible in practical applications. Another possible solution is to tag the CSI when each user performs a new action, and update the fingerprint library of the positioning system in real time to include all known actions. Unfortunately, there is currently no ready-made action recognition system that can accurately identify each action, especially for different locations and different people.

[0008] Therefore, there is no feasible technical solution in the prior art to solve the above-mentioned problems existing in the prior art. SUMMARY

[0009] The purpose of the present application is to provide a personnel positioning system based on WIFI devices to achieve accurate non-contact personnel positioning and solve the problems existing in the prior art.

[0010] The purpose of the present application is achieved by the following technical solutions:

[0011] A personnel positioning system based on WIFI devices, comprising a data enhancer, a feature extractor, a location predictor and a domain discriminator, wherein:

[0012] The location marked data based on WIFI devices is input into the data enhancer, and the data enhancer generates extended location marked data according to each location marked data;

[0013] The extended location marked data and unmarked data are input into a feature extractor based on a convolutional neural network (CNN), and a feature vector is obtained by the feature extractor, wherein the feature vector contains the features of the location and the features of the domain, and the domain refers to a pair of user and user movement action.

[0014] learning, by the position predictor and the domain discriminator, user position features irrelevant to user motion based on the feature vectors, to obtain position prediction features irrelevant to user motion; wherein the domain discriminator is configured to eliminate features in the feature vectors that are relevant to user motion;

[0015] Based on the position prediction features, the user within the range of the WIFI device can be identified in position, and personnel positioning independent of user motion is realized.

[0016] The data enhancer is implemented based on an adversarial autoencoder (AAE), and includes an encoder, a decoder and a discriminator.

[0017] The position-labeled data of the kth AAE is input into the encoder of the AAE numbered k, to generate a latent vector z;

[0018] The latent vector z is sent to the decoder, and a vector is generated for reconstructing the position-labeled data, and a reconstruction loss is defined using a mean square error (MSE) in the process of reconstructing the position-labeled data, and a corresponding loss function is:

[0019]

[0020] wherein, is the number of samples;

[0021] and,

[0022] The encoder is used as output data x of a generative adversarial network (GAN) generator. k is sent to the discriminator together with a vector z' subject to a prior distribution, and the discriminator is trained to standardize the reconstructed labeled data; wherein a cross-entropy is used as a loss function of the discriminator, and the loss function of the discriminator is:

[0023]

[0024] When x k is used as input, the label a k is 0, and when z' is used as input, the label a k is 1, is a predicted value of the ith sample, is a true value of the ith sample.

[0025] In the process of reconstructing the labeled data of each position, a position label is also assigned to the reconstructed labeled data, to create reconstructed labeled data with a position label.

[0026] The feature extractor includes the following in the process of extracting the feature vectors:

[0027] At each layer of the CNN, a convolutional layer with 2D convolutional kernels is used, a rectified linear unit (ReLU) is used to introduce nonlinearity, and a batch normalization layer is inserted to speed up the process, and a dropout layer is used to control the fitting process;

[0028] A max-pooling layer is also used to reduce the size of the feature vector.

[0029] The position prediction process performed by the position predictor includes:

[0030] The position predictor uses three fully connected layers and an activation function ReLU to learn the representation of the input feature vector, and maps the learned representation to a new latent space;

[0031] The learned representation is passed through an output layer with an activation function softmax to obtain a prediction probability vector of the position And the position labels of the labeled data and unlabeled data are and For labeled data and unlabeled data, cross-entropy is used as the loss function, i.e. the loss function of the position predictor is:

[0032]

[0033] Where N l and N u are the number of labeled data and unlabeled data used for training, respectively; La is the loss function of the labeled data, Lu is the loss function of the unlabeled data, and C is the total number of positions to be predicted.

[0034] The process performed by the domain discriminator includes:

[0035] According to the feature vector and the predicted position label output by the position predictor, a first domain label is predicted And, directly according to the feature vector output by the feature extractor, a second domain label is predicted

[0036] Cross-entropy is used as the loss function of the domain discriminator, i.e.:

[0037]

[0038] Where N is the number of training data, D is the number of domains, Ld is the loss function when aligning the posterior distribution of the source domain and the target domain, and Lo is the loss function when aligning the marginal distribution between the source domain and the target domain.

[0039] The process of identifying the location of the user within the range of the WIFI device also includes:

[0040] minimizing the loss function of the position predictor, maximizing the loss function of the domain discriminator, and determining the corresponding overall loss function as follows:

[0041] L all = L a + γL u - αL d - βL o ;

[0042] wherein, γ, α and β are hyperparameters, and the parameter values thereof are adjusted and set according to training results;

[0043] based on the overall loss function, identifying the positions of users located within the range of the WIFI device.

[0044] The loss function of the position predictor takes a positive value during back propagation, while the loss function of the domain discriminator takes a negative value during back propagation.

[0045] Compared with the prior art, the personnel positioning method based on a WIFI device provided by the present application can effectively overcome some limitations of the existing WIFI-based device-free personnel positioning scheme, and can be applied to personnel positioning in an indoor scene. In the specific personnel positioning process, the robustness of user actions can be realized, that is, the present application still achieves high positioning accuracy when facing different users and different actions by introducing domain self-adaptive technology and data enhancement technology. The implementation of the present application can achieve high-precision personnel positioning measurement. Specifically, the experimental results in a real scene show that, compared with other prior art, the present application can improve the position prediction accuracy in the personnel positioning process by about 35%. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 The system structure schematic diagram provided by the embodiments of the present application;

[0048] Figure 2 The data enhancer structure schematic diagram provided by the embodiments of the present application;

[0049] Figure 3 The comparative experimental result schematic diagram provided by the embodiments of the present application. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the specific contents of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application, which do not constitute a limitation of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative work belong to the protection scope of the present application.

[0051] Firstly, the terms possibly used in the present application are explained as follows:

[0052] The term "and / or" means either of the two or both, for example, X and / or Y means three cases including "X" or "Y" or "X and Y".

[0053] The terms "include", "contain", "have", "possess" or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw materials, components, ingredients, carriers, dosage forms, materials, sizes, parts, components, mechanisms, devices, steps, processes, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products or articles, etc.) should be interpreted as not only including the explicitly listed technical feature element, but also including other technical feature elements not explicitly listed in the art.

[0054] The term "consisting of" means excluding any technical feature element not explicitly listed. If this term is used in the claims, the term will make the claim closed, so that it does not contain technical feature elements other than the explicitly listed technical feature elements, except for conventional impurities related thereto. If the term only appears in a certain clause of the claim, it is only limited to the elements explicitly listed in the clause, and the elements described in other clauses are not excluded from the overall claim.

[0055] Unless otherwise explicitly specified or limited, the terms "mount", "connect", "connect", "fix", and the like should be understood broadly, for example: it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be mechanically connected, or it can be electrically connected; it can be directly connected, or it can be indirectly connected through an intermediate medium, or it can be connected inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0056] When concentrations, temperatures, pressures, sizes, or other parameters are expressed as numerical ranges, the numerical ranges should be interpreted as specifically disclosing all possible ranges of the upper value, lower value, and preferred values of the numerical range, whether or not the range is expressly recited; for example, if a numerical range of "2-8" is recited, the numerical range should be interpreted as including ranges of "2-7", "2-6", "5-7", "3-4 and 6-7", "3-5 and 7", "2 and 5-7", etc. Unless otherwise indicated, numerical ranges recited herein are inclusive of the recited whole integer and fraction values within the numerical range.

[0057] The terms "central", "longitudinal", "lateral", "length", "width", "thickness", "upper", "lower", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", "clockwise", "counterclockwise", and the like indicate the orientation or positional relationship shown in the drawings, and are used only for the convenience of description and simplification of description, and do not mean that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as limiting herein.

[0058] The purpose of the present application is to design a novel WIFI device-free personnel positioning method, which can position different users accompanying different actions, and can also ensure that the corresponding positioning processing can be realized by collecting and marking only a small amount of CSI fingerprints.

[0059] The technical solution provided by the present application mainly reduces the coupling of location fingerprint features (i.e., CSI based on location features) and user actions (i.e., user motion features accompanying the personnel positioning process) through domain adaptation technology and data enhancement technology, and generates rich data for training network models; based on the network model trained in this way, personnel accurate positioning based on WIFI devices can be realized.

[0060] For the convenience of understanding, the implementation mode of the present application will be described in detail below in combination with the drawings.

[0061] As Figure 1 shown, the technical solution provided by the present application can mainly include four processing modules, namely: a data enhancer, a feature extractor, a location predictor, and a domain discriminator; the working process of the personnel positioning system based on WIFI devices realized based on each component module can include:

[0062] (1) The position marked data based on the WIFI device enters the data enhancer based on the adversarial autoencoder (AAE), which can generate corresponding expanded position marked data, i.e. expansion data, according to the existing marked data of each position, to expand more potential user and action data;

[0063] (2) The enhanced data (i.e. expansion data) and unmarked data are input into the feature extractor based on the convolutional neural network (CNN), and the low-dimensional representation of the potential feature is extracted by the feature extractor, i.e. the corresponding feature vector is extracted, which contains the features of the position and the domain, and the domain refers to a pair of user and action, and the action refers to the action of user movement;

[0064] (3) Based on the generated potential feature (i.e. the feature vector), the user position feature irrelevant to user movement is learned by the position predictor and the domain discriminator, and the position prediction feature irrelevant to user movement is obtained; wherein the domain discriminator is used to eliminate the features in the feature vector related to user movement; i.e. the prediction position that maximizes the personnel positioning accuracy can be obtained by the position predictor;

[0065] Specifically, in this process, in order to eliminate the influence of different movements (i.e. domain-specific features), a corresponding domain discriminator is arranged in the system to predict each domain, i.e. to identify the action of the person to be positioned; i.e. the goal of the domain discriminator is to maximize the accuracy of domain marking (i.e. user action marked data), and the design purpose of the feature extractor is to try to deceive the domain discriminator, i.e. to minimize its prediction accuracy, while improving the position prediction accuracy, to reduce the influence of user action on position prediction in the personnel positioning process as much as possible.

[0066] In summary, by the above-mentioned manner, after learning the common features irrelevant to movement (i.e. user action) for defining position, the training model is further optimized by introducing two constraints of position predictor and domain discriminator to improve the accuracy of position prediction based on the training model.

[0067] For the convenience of understanding the present application, the specific implementation modes of the data enhancer, the feature extractor, the position predictor and the domain discriminator included in the present application will be described in detail respectively below with reference to the accompanying drawings.

[0068] (I) Data enhancer

[0069] The data enhancer uses data enhancement to generate more fingerprints (i.e. more position marked data). Specifically, as shown in FIG. 1, the data enhancer includes an autoencoder (AE) and an adversarial autoencoder (AAE). Figure 2As shown, the data enhancer is implemented based on AAE, which uses an AAE to process each position mark data after fingerprint position marking. That is, during the training process, the data enhancer is used to divide the position mark data into K subsets, each subset corresponds to the fingerprint x of a position k .

[0070] Reference Figure 2 As shown, the data enhancer may include an encoder, a decoder, and a discriminator. The data enhancement process performed by the data enhancer may include:

[0071] First, the position mark data corresponding to the kth AAE is input into the encoder of the kth AAE (i.e., AAE#k) to generate a latent vector z, which is sent to the decoder and generates a vector To reconstruct the position marker data; in the process of reconstructing the position marker data, the mean square error (MSE) can be used to define the reconstruction loss, and the corresponding reconstruction loss function can be:

[0072]

[0073] in, is the number of samples.

[0074] Then, the discriminator is trained to normalize the reconstructed position mark data; at this time, the encoder of AAE is the generator of GAN (Generative Adversarial Network), which converts the output data x of the generator into k It is sent to the discriminator together with the vector z' that obeys the prior distribution; for the discriminator, when it uses the data x k As input, label a k is 0, when it uses z' as input, label a k is 1; and cross entropy can be used as the loss function of the discriminator, that is:

[0075]

[0076] in, is the predicted value of the i-th sample, is the true value of the i-th sample;

[0077] Since data synthesis is performed on the labeled data of each position, position labels can be assigned to the labeled data after synthesis to create synthetic labeled data with position labels, that is, the reconstructed data and enhanced data obtained by processing the data enhancer are expanded data with position labels. The expanded data (that is, expanded data) can be used as more fingerprints (that is, more position labeled data, that is, expanded data of more potential users and actions) generated in subsequent learning and training processes.

[0078] (ii) feature extractor

[0079] The feature extractor is used to extract location-related features from the augmented data and unlabeled data output by the data augmenter to maximize the location prediction accuracy while minimizing the domain prediction accuracy. That is, all the augmented labeled data and unlabeled data output by the data augmenter are input into the feature extractor, and the feature vectors, i.e., location-related features, of the corresponding data are output through the feature extractor.

[0080] Specifically, the feature extractor uses a three-layer stacked CNN to extract features; that is, at each layer of the CNN, a convolutional layer with a 2D convolution kernel is used, a rectified linear unit (ReLU) is used to introduce nonlinearity, a batch normalization layer is inserted to speed up the training process, and a dropout layer is used to avoid overfitting; then, a max-pooling layer is used to reduce the size of the representation, i.e., the extracted feature vector.

[0081] As a result of using the above implementation, the feature extractor can obtain the desired feature matrix, i.e., the location-related feature vector, given the input data.

[0082] (iii) location predictor

[0083] After the feature vector is extracted by the above-mentioned feature extractor, it can be input into the location predictor to predict the location; the implementation process of the corresponding prediction location can include:

[0084] First, the location predictor uses three fully connected layers and an activation function ReLU to learn the representation of the input feature vector, i.e., to learn the feature vector;

[0085] Then, in order to predict the location, the learned representation is also mapped to a new latent space;

[0086] Finally, the learned representation is passed through an output layer with an activation function softmax to obtain the prediction probability vector of a certain location .

[0087] In order to improve the accuracy of location prediction, the location labels of labeled data and unlabeled data need to be predicted and recorded as and ; for labeled data and unlabeled data, cross-entropy can be used as the loss function, i.e., as the loss function of the corresponding location predictor, as follows:

[0088]

[0089] where N land N u is the number of labeled data and unlabeled data for training, La is the loss function of the labeled data, Lu is the loss function of the unlabeled data, and C is the total number of positions to be predicted.

[0090] The loss function of the labeled data and the loss function of the unlabeled data can also be collectively referred to as the loss function of the position predictor.

[0091] (Four) Domain discriminator

[0092] In the present application, the field adversarial deep learning technology is also used to eliminate the features related to user actions by using unlabeled data. In the adversarial network, the domain is defined as a pair of user and action (i.e. movement); the reason for such definition is that different movements have different effects on the CSI position fingerprint of WIFI, and different users have different body sizes. Specifically, the goal of the present application is to design a domain discriminator to identify different users and actions, and to make the feature extractor deceive the domain discriminator, i.e. to maximize the loss of the domain discriminator, so as to generate position features independent of user actions. Further, the specific processing process required to be performed by the domain discriminator can include:

[0093] Firstly, the output of the feature extractor contains not only position-specific features, but also domain-specific features; in order to match the conditional distribution, the feature matrix (i.e. feature vector) needs to be input into the domain discriminator together with the predicted label output by the position predictor to predict the first domain label And in order to improve the performance of the domain discriminator, not only the posterior distribution of the source domain (labeled data domain) and the target domain (unlabeled data domain) is aligned, but also the marginal distribution between the source domain and the target domain is aligned; for this purpose, the output of the feature extractor can be directly input into the domain discriminator, and the second domain label is predicted through the same process In the prediction process, cross-entropy can be used as the loss function of the domain discriminator, i.e.

[0094]

[0095] Where N is the number of training data, and D is the number of domains; correspondingly, Ld and Lo are the loss functions of the data predicted for all domains, Ld is the case of aligning the posterior distribution of the source domain and the target domain (i.e. Ld is the loss function when the posterior distribution of the source domain and the target domain is aligned), and Lo is the case of aligning the marginal distribution between the source domain and the target domain (i.e. Lo is the loss function when the marginal distribution between the source domain and the target domain is aligned).

[0096] Based on the outputs of the above position predictor and domain discriminator, the corresponding model training process can include:

[0097] In the model trained by the present application, the loss of the position predictor needs to be minimized to maximize the position prediction accuracy, and the loss of the domain discriminator needs to be maximized to learn the domain-independent position features, so that the overall loss function can be obtained as follows:

[0098] L all =L a +γL u -αL d -βL o ;

[0099] Wherein, gamma, alpha and beta are hyperparameters, wherein alpha, beta, gamma and lambda can be set to 1 in a specific training model, and for actual application, the parameters can also be appropriately fine-tuned according to the results of model training.

[0100] As can be seen from the above, the present application can use enhanced labeled data and unlabeled data to train the training model of the present application, and the parameters of the training model can also be iteratively updated during model training. It is worth noting that in order to achieve the goal of model training of the present application, the loss function of the position predictor takes a positive value during backpropagation, while the loss function of the domain discriminator takes a negative value during backpropagation.

[0101] As described above, the personnel positioning system based on WIFI provided by the present application includes a complete set of data processing and positioning algorithms, which can predict the position of unknown users and actions through careful processing of WIFI channel state. In the present application, a deep learning framework based on domain adaptation and data enhancement is also used, specifically using domain adaptation technology to reduce the coupling of position fingerprints with users and actions, i.e. different users and actions are considered as different domains, and the system provided by the present application can learn domain-independent position features. Then, the present application also uses data enhancement processing technology, so that only a small amount of position labeled data can be used to process various rich, position unlabeled data of different users and actions.

[0102] Therefore, the present application can effectively overcome the limitations of existing WIFI-based device-free personnel positioning schemes, and can produce the following significant technical effects in specific application process, including:

[0103] (1) The present application can realize non-contact personnel positioning through the setting of WIFI devices, avoiding the device carrying requirement based on the active positioning scheme, which is especially suitable for home care, continuous office environment;

[0104] (2) Since WIFI infrastructure exists widely and has low setting cost, compared with the implementation scheme of vision and millimeter wave radar, it can greatly reduce the cost of equipment and installation, so that large-scale deployment becomes economically feasible;

[0105] (3) The domain adaptation technology and the data enhancement technology are introduced in the application, so that higher positioning accuracy can be achieved when facing different users and different actions.

[0106] (4) The application is tested in a real scene, and the test results show that only three actions of three users are labeled, and accurate positioning (the nearest position distance is about one meter) of a total of 36 domains including 6 users and 6 actions can be achieved; compared with other prior arts, the system provided by the application can improve the position prediction accuracy by about 35%.

[0107] The application is tested in a real scene, and the test results show that only three actions of three users are labeled, and accurate positioning (the nearest position distance is about one meter) of a total of 36 domains including 6 users and 6 actions can be achieved; compared with other prior arts, the system provided by the application can improve the position prediction accuracy by about 35%. Figure 3 Figure 3 The application is tested in a real scene, and the test results show that only three actions of three users are labeled, and accurate positioning (the nearest position distance is about one meter) of a total of 36 domains including 6 users and 6 actions can be achieved; compared with other prior arts, the system provided by the application can improve the position prediction accuracy by about 35%. Figure 3 Figure 3 As shown in the experimental results in Table 1, the technical solution provided by the application can significantly improve the accuracy, and each position can be accurately predicted.

[0108] The above is only a preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the application, which should be covered in the protection scope of the application. Therefore, the protection scope of the application should be subject to the protection scope of the claims. The information disclosed in the background section of the present application is only intended to deepen the understanding of the overall background of the application, and should not be regarded as acknowledging or implying in any form that the information constitutes prior art known to those skilled in the art.​​

Claims

1. A personnel positioning system based on WIFI equipment, characterized in that: It includes a data enhancer, a feature extractor, a location predictor, and a domain discriminator, where: Inputting the location marker data based on the WIFI device into a data enhancer, wherein the data enhancer generates expanded location marker data according to each location marker data; Inputting the expanded position-marked data and the unmarked data into a feature extractor based on a convolutional neural network (CNN), extracting a feature vector using the feature extractor, wherein the feature vector includes a feature of the position and a feature of a domain, where the domain refers to an action between a pair of users and user motions; Based on the feature vector, learning user location features that are unrelated to user motion by the location predictor and the domain discriminator to obtain location prediction features that are unrelated to user motion; wherein the domain discriminator is used to eliminate features related to user motion in the feature vector; Based on the position prediction feature, the position of the user within the range of the WIFI device can be identified, thereby realizing personnel positioning independent of the user's movement.

2. The system according to claim 1, wherein: The data enhancer is implemented based on the adversarial autoencoder AAE and includes an encoder, a decoder and a discriminator, wherein: Input the position tag data of the kth AAE into the encoder of the AAE numbered k to generate the latent vector z; The latent vector z is sent to the decoder and generates a vector for reconstructing the position marker data In the process of reconstructing the position mark data, the mean square error MSE is used to define the reconstruction loss. The corresponding loss function is: in, is the number of samples; as well as, The encoder is used as the output data x of the GAN generator k Together with the vector z' that obeys the prior distribution, the reconstructed labeled data is sent to the discriminator to normalize the reconstructed labeled data by training the discriminator; wherein the cross entropy is used as the loss function of the discriminator, and the loss function of the discriminator is: When using x k As input, label a k is 0, when z' is used as input, label a k is 1, is the predicted value of the i-th sample, is the true value of the i-th sample; In the process of reconstructing the labeled data at each position, a position label is also assigned to the reconstructed labeled data to create reconstructed labeled data with the position label.

3. The system according to claim 1, wherein: The process of extracting and obtaining the feature vector by the feature extractor includes: In each layer of CNN, a convolutional layer with a 2D convolution kernel is used, nonlinearity is introduced using the rectified linear unit (ReLU), a batch normalization layer is inserted to accelerate the processing process, and a dropout layer is used to control the fitting process; A max pooling layer is also used to reduce the size of the feature vector.

4. The system according to any one of claims 1 to 3, characterized in that The process of predicting the position performed by the position predictor includes: The position predictor uses three fully connected layers and an activation function ReLU to learn representations for the input feature vector and maps the learned representations to a new latent space; The learned representation is passed through an output layer with a softmax activation function to obtain the predicted probability vector of the position And the position labels of the predicted labeled data and unlabeled data are and For labeled data and unlabeled data, cross entropy is used as the loss function, that is, the loss function of the position predictor is: Among them, N l and N u are the number of labeled and unlabeled data used for training, respectively; L a is the loss function for labeled data, L u is the loss function for unlabeled data, C is the total number of positions to be predicted, is the true label value of the cth position of the i-th labeled data, is the probability value of the cth position obtained after the position prediction of the i-th labeled data, is the probability value of the cth position obtained after the position prediction of the i-th unlabeled data.

5. The system according to claim 4, characterized in that The processing performed by the domain identifier includes: Predicting a first domain label based on the feature vector and the predicted location label output by the location predictor And, directly predicting the second domain label based on the feature vector output by the feature extractor We use cross entropy as the loss function for the domain discriminator, namely: Where N is the number of training data, D is the number of domains, and L d is the loss function when aligning the posterior distribution of the source domain and the target domain, L o is the loss function when aligning the marginal distribution between the source and target domains, d ij is the first domain label value of the jth domain of the i-th training data, is the first domain label prediction value of the jth domain of the i-th training data, Predict the second domain label for the jth domain of the i-th training data.

6. The system according to claim 5, characterized in that The process of identifying the location of a user within the range of the WIFI device further includes: Minimize the loss function of the position predictor, maximize the loss function of the domain discriminator, and determine the corresponding overall loss function as follows: L all =L a +γL u -αL d -βL o ; Among them, γ, α, and β are hyperparameters, and their parameter values ​​are adjusted according to the training results; The location of users within the range of the WIFI device is identified based on the overall loss function.

7. The system according to claim 6, characterized in that The loss function of the position predictor takes positive values ​​during back-propagation, while the loss function of the domain discriminator takes negative values ​​during back-propagation.

Images