A WIFI fingerprint positioning method and system for a multi-variable indoor environment

Abstract

The application relates to a WIFI fingerprint positioning method and system for a multi-variable indoor environment, and the method comprises the following steps: collecting online data of a user; inputting the online data of the user into an indoor positioning model to output a positioning result; wherein the indoor positioning model is constructed through a feature extraction network and a classifier and is obtained through training of a plurality of multi-domain fingerprint data; the feature extraction network respectively extracts domain-invariant features and domain-inter-invariant features; and the domain-invariant features and the domain-inter-invariant features are fused and input into the classifier. The application realizes general positioning in a multi-variable environment by splicing the two types of features, i.e. the domain-invariant features and the domain-inter-invariant features, to form domain-invariant features, and taking the domain-invariant features as position features for position estimation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of indoor fingerprint positioning technology, in particular to a WIFI fingerprint positioning method and system for a variable indoor environment. BACKGROUND

[0002] The advent of large models has brought profound changes to embodied intelligence, smart factories, intelligent transportation, and emergency rescue. Among them, indoor positioning provides key technical support for location-based services such as robot autonomous obstacle avoidance and navigation in embodied intelligence, equipment and material tracking in smart factories, vehicle positioning in intelligent transportation, and trapped personnel positioning and optimized rescue path in emergency rescue. However, most indoor positioning scenarios have large personnel flow, complex environmental structure changes, and frequent item movement, which severely affect the accuracy and stability of positioning. To address the challenges of complex and unpredictable environmental changes on indoor positioning, there is an urgent need to research high-precision, low-cost, and sustainable updating positioning technology in variable environments.

[0003] In complex and variable indoor environments, satellite signals are difficult to achieve high-precision positioning due to severe attenuation or interruption. Researchers have developed positioning systems based on indoor wireless signals. Existing technologies mainly fall into two categories: geometric positioning methods, which rely on distance or angle measurements, but are susceptible to non-line-of-sight interference affecting positioning stability; and fingerprint positioning methods, which achieve positioning by matching pre-stored signal characteristics, without the need for direct geometric parameter measurement. This approach has stronger robustness to multipath effects and non-line-of-sight environments, and has the advantages of flexible deployment, algorithm scalability, and sub-meter-level positioning accuracy in dynamic scenarios such as shopping malls and hospitals. Current mainstream fingerprint positioning technologies rely on WiFi, Bluetooth, radio frequency identification, ultra-wideband, and vision, among others. WiFi technology has become the most widely used solution due to its wide infrastructure coverage, low cost, and easy data acquisition.

[0004] Data-driven fingerprint positioning requires the pre-establishment of a fingerprint database, and the database samples only correspond to the indoor environment at the sampling time. In general, only a complete initial fingerprint database is collected in a fixed scenario. However, indoor environments are complex and variable, and wireless signal characteristics as fingerprint data will change with personnel movement, wall, furniture, and other obstacles, leading to significant distribution differences between offline fingerprint libraries and online data in practical applications, causing a sharp decline in fingerprint positioning performance.

[0005] And the current positioning method based on domain adversarial training mainly has the following limitations: first, it relies on the collection of target domain data after environmental changes (usually 50-100 groups of samples), which is difficult to meet the real-time positioning demand; second, it only supports linear knowledge transfer from a single source domain to a single target domain, and cannot cope with the fingerprint data distribution drift caused by continuous dynamic changes in indoor environments, resulting in limited cross-scene generalization ability; third, only by learning the multi-domain distribution alignment mechanism, it is difficult to extract enough domain-invariant features to achieve generalization positioning in unknown environment distribution. Therefore, it is urgent to study the cross-domain generalization positioning method to fully learn the characteristics of existing fingerprint data distribution to improve the positioning generalization ability in unknown distribution environment. SUMMARY

[0006] The purpose of the present application is to provide a WIFI fingerprint positioning method and system for a variable indoor environment, which extracts intra-domain invariant features of each multi-domain fingerprint data through Fourier phase transformation and knowledge distillation, learns inter-domain mutual invariant features using distribution-related alignment, and then forms domain-invariant features to achieve positioning in unknown environment changes through a position estimator. By maximizing the distance between intra-domain invariant features and inter-domain mutual invariant features, the overlap of invariant information between the two is reduced. By splicing the intra-domain invariant features and the inter-domain mutual invariant features to form domain-invariant features, the position estimation is performed as a position feature, and the generalization positioning in a variable environment is achieved.

[0007] To achieve the above purpose, the present application provides the following scheme:

[0008] A WIFI fingerprint positioning method for a variable indoor environment, comprising:

[0009] Collecting online data of a user;

[0010] Inputting the online data of the user into an indoor positioning model to output a positioning result; wherein the indoor positioning model is constructed by a feature extraction network and a classifier and is obtained by training with a plurality of multi-domain fingerprint data, the feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, and the intra-domain invariant features and the inter-domain mutual invariant features are fused and input into the classifier.

[0011] Optionally, the feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, comprising:

[0012] Training a teacher network with a plurality of multi-domain fingerprint data to obtain a trained teacher network;

[0013] Using a knowledge distillation technology to enable a student network to learn the feature extraction of the trained teacher network, and training the student network with a plurality of multi-domain fingerprint data to obtain an intra-domain invariant feature extractor;

[0014] The intra-domain invariant feature extractor and the inter-domain mutual invariant feature extractor are used to extract the intra-domain invariant features and the inter-domain mutual invariant features, respectively. The intra-domain invariant feature extractor and the inter-domain mutual invariant feature extractor are composed of CNN, ReLU and BatchNorm layers.

[0015] Optionally, the teacher network is trained using several of the multi-domain fingerprint data to obtain the trained teacher network, including:

[0016] Perform Fourier phase transform on several of the multi-domain fingerprint data to obtain Fourier phase information;

[0017] The Fourier phase information is input into the teacher network to extract Fourier phase fingerprint features;

[0018] The teacher network is trained by combining the Fourier phase fingerprint features with reference point labels.

[0019] Optionally, training the teacher network includes:

[0020]

[0021] in, and These represent the feature extractors of the teacher network. and classifier Learnable parameters, P tr Let E represent the distribution of the training data, and E be the expected value. Let x be the cross-entropy loss function for classification, and x be the multi-domain fingerprint data. y represents the Fourier phase fingerprint feature, and y represents the fingerprint data label.

[0022] Optionally, the features extracted by the teacher network after training can be learned by the student network through knowledge distillation techniques, including:

[0023]

[0024] in, These represent the feature extractors of the student network. and classifier The learnable parameters, where λ1 is a hyperparameter. For MSE loss, The loss was classified into student network fingerprints.

[0025] Optionally, the objective function of the indoor positioning model is:

[0026]

[0027] Among them, G c and G fThese represent the classifier and the feature extraction network, respectively, with corresponding parameters θ. c and θ f λ1, λ2, and λ3 are hyperparameters, d(.,.) represents the distance function, z1 and z2 are the intra-domain invariant and inter-domain mutually invariant features, respectively, and C i and C j These are the covariance matrix and ||.|| F Representing the Frobenius norm form of a matrix, The loss is the mutual invariant feature learning loss, where N is the dimension of the covariance matrix. This represents the loss of feature diversity.

[0028] Optionally, the positioning result is:

[0029] y = arg min G c (G f (x)).

[0030] The present invention also provides a system for implementing a WIFI fingerprint positioning method for variable indoor environments, comprising:

[0031] The data acquisition module is used to collect users' online data;

[0032] The positioning module is used to input the user's online data into an indoor positioning model and output the positioning result. The indoor positioning model is constructed through a feature extraction network and a classifier and trained with several multi-domain fingerprint data. The feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, and fuses the intra-domain invariant features and inter-domain mutual invariant features into the classifier.

[0033] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a WIFI fingerprint positioning method for variable indoor environments.

[0034] The beneficial effects of this invention are as follows: when the indoor environment changes, it is usually necessary to re-collect data and retrain the positioning model in order to maintain the accuracy of positioning. However, this invention can adapt to certain environmental changes without updating, thus reducing the cost of data collection and model training.

[0035] The actual operating environment will inevitably have differences in data distribution compared to the fingerprint database. Based on its generalization mechanism, this invention can effectively cope with dynamic changes in the environment and improve the accuracy and robustness of the positioning system.

[0036] This invention utilizes a knowledge distillation architecture to reduce the computational load of Fourier transforms, thereby accelerating the inference speed of the localization model while maintaining the same level of accuracy. It also reduces the consumption of hardware resources, which is beneficial for the application of fingerprint localization on lightweight devices. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a WIFI fingerprint positioning method for variable indoor environments according to an embodiment of the present invention;

[0039] Figure 2 This is the domain-invariant feature distillation learning architecture of this invention.

[0040] Figure 3 This is the architecture of the domain generalization localization method based on Fourier phase information in this embodiment of the invention;

[0041] Figure 4 The laboratory structure is shown in this embodiment of the invention;

[0042] Figure 5 The present invention relates to a WIFI signal transceiver device, wherein (a) is a receiver and (b) is a transmitter. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Example 1:

[0046] like Figure 1 As shown, this embodiment provides a WIFI fingerprint positioning method for variable indoor environments, including:

[0047] Collect users' online data;

[0048] The user's online data is input into the indoor positioning model, and the positioning result is output. The indoor positioning model is constructed through a feature extraction network and a classifier and trained with several multi-domain fingerprint data. The feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, and then fuses the intra-domain invariant features and inter-domain mutual invariant features into the classifier.

[0049] Specifically, this embodiment trains on multiple multi-domain fingerprint datasets, employing a dual feature extractor to learn intra-domain invariant features and inter-domain mutually invariant features respectively. These two types of features are then fused to form a unified domain-invariant feature, which is applied to the fingerprint classification task. To enhance the information diversity of the localization model, this embodiment introduces an exploratory loss mechanism, which reduces feature information overlap by maximizing the differences between the two types of features.

[0050] Furthermore, the feature extraction network extracts intra-domain invariant features and inter-domain mutually invariant features, including:

[0051] The teacher network was trained using several multi-domain fingerprint data sets to obtain the trained teacher network.

[0052] The knowledge distillation technique is used to extract features from the teacher network after the student network is trained, and the student network is trained with several multi-domain fingerprint data to obtain an invariant feature extractor within the domain.

[0053] Intra-domain invariant feature extractor and inter-domain mutual invariant feature extractor are used to extract intra-domain invariant features and inter-domain mutual invariant features, respectively. The intra-domain invariant feature extractor and inter-domain mutual invariant feature extractor are composed of CNN, ReLU and BatchNorm layers.

[0054] Specifically, such as Figure 3 As shown, the extraction of domain-invariant features is achieved using a distillation framework. In the teacher network, Fourier phase transform and feature extraction are performed on the fingerprint data to obtain domain-invariant features suitable for classification. Then, the feature extraction capability of the teacher network is learned through knowledge distillation. Knowledge distillation is a simple learning framework that enables different networks to extract common feature information.

[0055] The teacher network takes Fourier phase information as input and generates output based on reference point labels, thereby extracting Fourier phase fingerprint features suitable for classification tasks. This involves the Fourier transform of single-channel two-dimensional CSI data \mathbit{x}. It can be represented as:

[0056]

[0057] u and v are exponents, and H and W are the height and width of the CSI feature map, respectively. The Fourier transform can be efficiently computed using the FFT algorithm. The phase components are then represented as:

[0058]

[0059] Where R(x) and I(x) represent respectively The real and imaginary parts. For multi-channel CSI feature maps, Fourier transform can be performed on each channel separately to extract its independent phase information.

[0060] Furthermore, the teacher network is trained using several multi-domain fingerprint data sets, and the trained teacher network includes:

[0061] Perform Fourier phase transform on several multi-domain fingerprint data to obtain Fourier phase information;

[0062] Fourier phase information is input into the teacher network to extract Fourier phase fingerprint features;

[0063] The teacher network is trained by combining Fourier phase fingerprint features with reference point labels.

[0064] Furthermore, after obtaining the Fourier phase fingerprint features x of the original data x, it is combined with the label (x,y) to train the teacher network, which includes:

[0065]

[0066] in, and These represent the feature extractors of the teacher network. and classifier Learnable parameters, P tr Let E represent the distribution of the training data, and E be the expected value. Let x be the cross-entropy loss function for classification, and x be the multi-domain fingerprint data. y represents the Fourier phase fingerprint feature, and y represents the fingerprint classification label.

[0067] Furthermore, such as Figure 2 As shown, the features extracted by teachers from the network after students' online learning and training through knowledge distillation technology include:

[0068] After the teacher network completes training, the student network can be guided to learn Fourier phase information through feature knowledge distillation techniques. This process can be represented as:

[0069]

[0070] in, These represent the feature extractors of the student network. and classifier The learnable parameters, where λ1 is a hyperparameter. For MSE loss, The loss was classified into student network fingerprints.

[0071] Furthermore, the objective function of the indoor positioning model is:

[0072]

[0073] Among them, G c and G f These represent the classifier and the feature extraction network, respectively, with corresponding parameters θ. c and θ f λ1, λ2 and λ3 are hyperparameters.

[0074] Specifically, Fourier phase transform alone is insufficient to obtain enough domain-invariant positional information in fingerprint data for discrimination and classification. Therefore, this embodiment explores mutually invariant features by utilizing cross-domain knowledge contained in multiple fingerprint training domains. Given two domains... This embodiment uses a correlation alignment method to align their second-order statistics (correlation):

[0075]

[0076] in, and These are the covariance matrix and ||.||, respectively. F Representing the Frobenius norm form of a matrix, The loss is the mutual invariant feature learning loss, where N is the dimension of the covariance matrix, and n i For S i Number of domain samples, X i For S i Domain sample, i is S i Domain tag, n i For S j Number of domain samples, X i For S j Domain sample, j is S j Domain tags.

[0077] To avoid redundancy and duplication between internally invariant and mutually invariant features, this method designs a dual-fingerprint feature extractor, aiming to extract as many differentially invariant features as possible. Diverse invariant features are beneficial for achieving better domain generalization localization. To achieve this goal, this embodiment uses an additional regularization term to ensure that a certain distance is maintained between the internally invariant (z1) and mutually invariant features (z2):

[0078]

[0079] Where d(.,.) represents the distance function, and z1 and z2 are the intra-domain invariant feature and the inter-domain mutual invariant feature, respectively.

[0080] In general, this embodiment involves two steps. First, the teacher network is trained, and then the following objective function is optimized:

[0081]

[0082] The above formula integrates four types of loss functions: classification loss Internal invariant feature learning loss Mutually invariant feature learning loss and feature diversity loss The first and third loss terms are two general objective losses for invariant feature learning in domain generalization. Existing work shows that these two losses alone are insufficient for accurate localization in variable environments. By utilizing internally invariant features and feature diversity, this embodiment can mitigate the aforementioned problems and achieve better performance. Currently, the hyperparameters in the above equation are mainly adjusted empirically; future plans include designing more heuristic methods to automatically determine the hyperparameters.

[0083] Furthermore, for determining the user's location, the reference point can be predicted based on online data, and the positioning result is as follows:

[0084] y = arg minG c (G f (x)).

[0085] In summary, the localization method in this embodiment extracts intra-domain invariant features from multi-domain fingerprint data through Fourier phase transform and knowledge distillation, learns inter-domain mutual invariant features using distributional correlation alignment, and then constructs domain invariant features to achieve localization under unknown environmental changes through a location estimator. Simultaneously, by maximizing the distance between intra-domain and inter-domain mutual invariant features, the overlap of invariant information between them is reduced. By concatenating these two types of features to form domain invariant features, which are used as location features for location estimation, generalized localization under changing environments is achieved.

[0086] Let's illustrate the overall technical solution in a specific application environment with an example:

[0087] Example hardware and environment configuration:

[0088] This embodiment collected a fingerprint database of a small laboratory environment and used online fingerprints before and after environmental changes for location tracking. The laboratory contained various obstacles, including desks, sofas, display cases, chairs, and other facilities. Figure 4As shown, the laboratory area is approximately 4.0m × 8.0m. The entire laboratory space is equipped with 24 training fingerprint reference points and 15 test fingerprint reference points (the distance between adjacent reference points of the same type is 0.8 meters). Figure 5 As shown in (a)-(b), the WiFi positioning signal data transceiver includes a transmitter with one omnidirectional antenna and a receiver with three omnidirectional antennas. Both the receiver and the transmitter are industrial control computers with built-in Intel 5300 wireless network cards. The transmitter, acting as a mobile terminal, continuously transmits WiFi signal data packets with a 20MHz bandwidth. The receiver is placed in a fixed indoor location and stores the received data packets.

[0089] This embodiment sets up five spatial conditions with different environmental variations as shown in Table 1 in the experimental environment. In spatial condition 1, each reference point is sampled 1200 times consecutively. Then, using the method described in section 3.2, every 30 consecutive CSI data points are grouped into a single sample image, resulting in 40 samples per reference point, thus establishing a complete initial fingerprint database. Simultaneously, to simulate the limited data in the scene spatial conditions to be expanded, only 20 fingerprint samples are collected for each training reference point in scene spatial conditions 2-4, establishing corresponding real fingerprint databases. Scene spatial condition 1 corresponds to the initial fingerprint database, while scene spatial conditions 2-4 are the real fingerprint databases for the scene spatial conditions to be expanded. The CSI data is then sent to DLEI-DAM to expand the datasets in scene spatial conditions 2-4. The DLEI-DAM network is based on the PyTorch framework and uses a GPU (GTX 2080Ti) for accelerated training.

[0090] Table 1

[0091]

[0092] The training process for the generalized localization system is as follows:

[0093] 1. After collecting the CSI fingerprint dataset with various environmental variations, for each of the three independent antenna links, a CSI amplitude matrix containing 30 subcarriers is extracted. After min-max normalization, a single-channel feature map is formed. With a time-domain sampling number of 30, the time-frequency feature map dimension of each antenna link is 30×30 (time-subcarrier dimension). Finally, the normalized amplitude matrices of the three links are mapped to RGB three channels respectively to construct a 3×30×30 three-dimensional feature tensor map, thus constructing a fingerprint dataset with various environmental variations.

[0094] 2. Using the datasets from environments 1-4, apply the formula to the teacher network. The teacher network is trained using a multi-layer CNN structure for its feature extractor. After training, the teacher network parameters are fixed.

[0095] 3. The domain generalization localization model based on Fourier phase information is trained using the datasets from Environments 1-4. The entire model is trained by optimizing the objective function of DGLM-FPI, and the model parameters are fixed after training. The two feature extractors consist of CNN, ReLU, and BatchNorm layers, and the classifier consists of fully connected layers.

[0096] 4. After training the generalized localization model, fix the parameters as the location inference model. When there is a location requirement, collect the user's WIFI CSI signal, convert it into a fingerprint feature map, and feed it into the location inference model to obtain the user's predicted location.

[0097] Example 2:

[0098] A system for implementing a WIFI fingerprint positioning method for variable indoor environments includes:

[0099] The data acquisition module is used to collect users' online data;

[0100] The positioning module is used to input the user's online data into an indoor positioning model and output the positioning result. The indoor positioning model is constructed through a feature extraction network and a classifier and trained with several multi-domain fingerprint data. The feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, and fuses the intra-domain invariant features and inter-domain mutual invariant features into the classifier.

[0101] Example 3:

[0102] This embodiment also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements a WIFI fingerprint positioning method for variable indoor environments.

[0103] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A WIFI fingerprint positioning method for variable indoor environments, characterized in that, include: Collect users' online data; The user's online data is input into the indoor positioning model, and the positioning result is output. The indoor positioning model is constructed through a feature extraction network and a classifier and trained with several multi-domain fingerprint data. The feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, and the intra-domain invariant features and inter-domain mutual invariant features are fused and input into the classifier. The feature extraction network extracts intra-domain invariant features and inter-domain mutually invariant features, including: The teacher network is trained using several of the aforementioned multi-domain fingerprint data to obtain the trained teacher network; The student network learns the features extracted by the trained teacher network through knowledge distillation technology, and the student network is trained using several multi-domain fingerprint data to obtain an intra-domain invariant feature extractor. The intra-domain invariant feature extractor and the inter-domain mutual invariant feature extractor are used to extract the intra-domain invariant features and the inter-domain mutual invariant features, respectively. The intra-domain invariant feature extractor and the inter-domain mutual invariant feature extractor are composed of CNN, ReLU and BatchNorm layers. The teacher network is trained using several of the aforementioned multi-domain fingerprint data to obtain the trained teacher network, which includes: Perform Fourier phase transform on several of the multi-domain fingerprint data to obtain Fourier phase information; The Fourier phase information is input into the teacher network to extract Fourier phase fingerprint features; The teacher network is trained by combining the Fourier phase fingerprint features with reference point labels; Training the teacher network includes: in, and These represent the feature extractors of the teacher network. and classifier Learnable parameters Let E represent the distribution of the training data, and E be the expected value. For classification, cross-entropy loss function, For multi-domain fingerprint data, It is a Fourier phase fingerprint feature. It is a fingerprint classification label; The features extracted by the teacher's network after training, through knowledge distillation techniques, enable students to learn online from the following: in, These represent the feature extractors of the student network. and classifier Learnable parameters For hyperparameters, For MSE loss, Losses due to the classification of student network fingerprints; The objective function of the indoor positioning model is: in, and These represent the classifier and the feature extraction network, respectively, with corresponding parameters as follows: and E represents the expected value. For classification, cross-entropy loss function, Represents the distribution of the training data. For multi-domain fingerprint data, It is a Fourier phase fingerprint feature. It is a fingerprint classification label. For MSE loss, , and For hyperparameters, Represents the distance function. These are the intra-domain invariant features and inter-domain mutually invariant features, C i and C j These are the covariance matrix and Representing the Frobenius norm form of a matrix, The loss is learned from the mutual invariant features. Let be the dimension of the covariance matrix. This represents the loss of feature diversity.

2. The WIFI fingerprint positioning method for variable indoor environments according to claim 1, characterized in that, The location result is as follows: .

3. A system for implementing the method described in any one of claims 1-2, for a WIFI fingerprint positioning method in a variable indoor environment, characterized in that, include: The data acquisition module is used to collect users' online data; The positioning module is used to input the user's online data into an indoor positioning model and output the positioning result. The indoor positioning model is constructed through a feature extraction network and a classifier and trained with several multi-domain fingerprint data. The feature extraction network extracts intra-domain invariant features and inter-domain mutual invariant features respectively, and fuses the intra-domain invariant features and inter-domain mutual invariant features into the classifier.

4. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a WIFI fingerprint positioning method for variable indoor environments as described in any one of claims 1 to 2.

Images