A cross-domain WiFi personnel identification method based on multi-branch domain alignment model

Abstract

The present invention is applicable to the technical field of personnel identification. The present invention discloses a cross-domain WiFi personnel identification method based on a multi-branch domain alignment model. The cross-domain WiFi personnel identification method includes establishing a multi-branch domain alignment DDA model, and using a loss function to supervise the multi-branch domain alignment model, so that the domain-specific classifier accurately identifies different gait features, divides the overall source domain into two or more groups, and aligns each group of source domains with the target domain respectively, and shortens the distance between the source domain and the target domain of each branch and the distance between multiple branches through a series of loss functions, so as to achieve simultaneous migration and adaptation of multiple interference domains. It can simultaneously eliminate the negative effects brought about by multiple environmental factors that appear simultaneously, improve the accuracy of WiFi human identity recognition and the reliability of the system, and reduce the difficulty of migration and adaptation to multiple interference domains.

Description

Technical Field

[0001] The present invention relates to network personnel identification, and in particular to a cross-domain WiFi personnel identification method based on a multi-branch domain alignment model Background Art

[0002] When WiFi devices transmit radio frequency signals in the form of WiFi spectrum during communication, these signals will be reflected and diffracted by the human body in the area, and the fluctuations of the signals can be used to identify various human movements, such as hands, body or breathing. In real life scenarios, this recognition method can be applied to multiple fields, such as security monitoring systems, elderly motor function status monitoring, and smart home devices. Since everyone's walking posture (referred to as "gait") is theoretically uniquely bound to biometrics. Some existing technologies attempt to use the channel state information (CSI, Channel State Information) corresponding to the human gait collected by WiFi devices for human identification (HI, Human Identification), which is suitable for personalized deployment in various sampling locations such as smart homes and offices. However, the real multi-path environment contains multiple interference factors, also known as domains. For example, the walking trajectory direction, walking speed, WiFi device location, etc. will cause changes in the human gait signal, resulting in a decrease in the accuracy of human identity recognition. The existing technology attempts to use transfer learning to eliminate the negative impact of a single interference domain, but it is impossible to remove multiple types of interference domains at the same time. Summary of the invention

[0003] The main technical problem solved by the present invention is to provide a cross-domain WiFi personnel identification method based on a multi-branch domain alignment model. The cross-domain WiFi personnel identification method based on a multi-branch domain alignment model can avoid the reduction in human identity recognition accuracy caused by only eliminating a single interference domain, eliminate the negative impact of multiple interference domains, and improve the accuracy of human identity recognition.

[0004] In order to solve the above problems, the present invention provides a cross-domain WiFi personnel identity recognition method based on a multi-branch domain alignment model. The cross-domain WiFi personnel identity recognition method based on a multi-branch domain alignment model recognizes the identity of a person by analyzing the gait characteristics of the person, including:

[0005] A series of loss functions are used to supervise the multi-branch domain alignment model to align the distribution between the source domain and the target domain, so as to encourage the domain-specific classifier to accurately identify different gait features. The domain adaptation loss function in the series of loss functions shortens the distance between the source domain and the target domain of each branch and the distance between multiple branches, so as to enable simultaneous migration and adaptation of multiple interference domains.

[0006] The multi-branch domain alignment DDA model includes dividing the entire source domain into at least two groups, aligning each source domain with the target domain respectively, and the network model includes a common feature extractor E and a feature extractor H for each source domain of each group. i , learn the gait features in each group respectively, and align the embedding features of the target domain obtained by passing the features of the target domain through the network models corresponding to the sub-domains of the two source domains with the embedding features of the two sub-source domains.

[0007] In one embodiment, an MMD domain adaptation loss function is used to minimize the distance between two sets of source domains and target domains. The MMD domain adaptation loss function is formally expressed as:

[0008]

[0009] in, and Respectively represent the expectations of two sets of data distributions p and q, represents the reproducing kernel Hilbert space, the function φ(x s ) and φ(x t ) represents the feature map that converts the original sample into the reproducing kernel Hilbert space, and the kernel function k(x s ,x t ) are defined as the feature representations φ(x s ) and the feature representation φ(x t ), x s and x t Represent the WiFi data samples of the source domain and the target domain respectively.

[0010] In one embodiment, the series of loss functions includes a domain adaptation loss function Convergence loss function and the cross entropy loss function

[0011] In one embodiment, the domain adaptation loss function The formal expression is:

[0012]

[0013] Among them, x s and x t represent the WiFi data samples of the source domain and the target domain respectively, N represents the number of network branches, Indicates A simplified estimate of The MMD loss applied in the two branch networks of the DDA model helps to align the distribution between the source and target domains in each branch, E(x t )and Indicates that the source domain and target domain samples are extracted by the common feature extractor E, and H i (E(x t )) represents the source domain grouping feature extractor H i The extracted features, s represents the source domain, t represents the target area, and i ranges from 1 to N.

[0014] In one embodiment, the number of the network branches is 2, 3 or 4.

[0015] In one embodiment, the cross entropy loss function shown in the DDA model is For supervised classification learning, the cross entropy loss function The formal expression is:

[0016] Among them, K represents the number of gait labels, l y represents the true category label of the sample, σ refers to the SoftMax function operation, C i Represents the classifier of branch i, which is used to convert the network output results into probability values. The range of i is 1 to N. The denotation parameter is the expectation of the source domain samples.

[0017] In one embodiment, the DDA model uses unlabeled samples of the target domain to close the distance between multiple branches and helps the model achieve domain alignment of two specific branches through a convergence loss function. The convergence loss function The formal expression is:

[0018]

[0019] In one embodiment, the domain alignment loss function The formal expression is:

[0020]

[0021] Among them, λ 4 and λ 5 Respectively and The importance weight of .

[0022] In one embodiment, in the early stage of training, the two weights are dynamically adjusted using a progressive strategy, according to the formula λ 4 =λ 5=2 / (1+exp(-θp′))-1, which gradually changes from 0 to 1 as the training iterations proceed, where p′ changes linearly during the training process, θ is empirically taken as 10 and remains unchanged throughout the experiment, and exp(-θp′) means the -θp′ power function with exponent e as the base.

[0023] The present invention discloses a cross-domain WiFi personnel identification method based on a multi-branch domain alignment model. The cross-domain WiFi personnel identification method adopts a multi-branch domain alignment (DDA, Divided Domain Adaptation) model and a series of loss functions, divides the overall source domain into two or more groups, and aligns each source domain with a target domain respectively, shortens the distance between the source domain and the target domain of each branch and the distance between multiple branches through a series of series of loss functions, and realizes simultaneous migration adaptation of multiple interference domains. It can simultaneously eliminate the negative effects brought by multiple environmental factors that appear simultaneously, improve the accuracy of WiFi human identity recognition and the reliability of the system, and reduce the difficulty of migration adaptation to multiple interference domains. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the description only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 It is a schematic diagram of the multi-branch domain alignment model of the present invention.

[0026] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0027] The claims of the present invention are further described in detail below in conjunction with specific embodiments and drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments proposed by ordinary technicians in this field before making creative work also fall within the scope of protection of the present invention.

[0028] It should be understood that, in the description of the embodiments of the present invention, all directional indication terms, such as "up", "down", "left", "right", "front", "back", etc., indicate the orientation or position relationship based on the orientation, position relationship shown in the drawings or the orientation or position relationship in which the invented product is usually placed when in use, which is only for the convenience of simplifying the description of the present invention, and does not expressly or imply that the device, element or component referred to must have a specific orientation and a specific orientation structure, and should not be understood as a limitation on the present invention. It is only used to explain the relative position relationship, movement, etc. between the components shown in the drawings. When the specific posture changes, the directional indication may also change accordingly.

[0029] In addition, ordinal numbers such as "first" and "second" in the present invention are only used for the purpose of distinction, and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. The features defined as "first" and "second" can expressly or implicitly indicate at least one of the technical features. In the present invention, for the convenience of explanation, there may be quantifiers such as "one", "two", "multiple", "at least one" or "several" to define the technical (structural) features, where "multiple" means at least two, that is, two or more; "at least one" and "several" mean one or one and the corresponding technical features; quantifiers and other limitations are only for illustrating the specific embodiment or the best implementation method, and they should be included as long as they are not contrary to the overall concept of the invention and the purpose of the invention, unless it is clearly stated that they must be defined by the quantifiers to be implemented or to achieve the purpose of the invention.

[0030] In the present invention, unless otherwise clearly stipulated and limited, the terms such as "installation", "setting", "connection", "fixation" and "screw-on" should be understood in a broad sense. For example, the positional relationship between components can be relatively fixed, or the components can be physically fixedly connected; they can be detachably connected or integrated; they can be mechanically connected or electrically connected; they can be directly connected or indirectly connected through an intermediate medium or component; they can be internally connected between two elements or interact with each other. Unless otherwise clearly defined in the specification, other interpretations may not achieve the corresponding functions or effects. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0031] The controller and control circuit involved in the present invention are conventional control technologies or units of those skilled in the art. For example, the control circuit of the controller can be realized by ordinary technicians in the field using existing technologies, such as simple programming. If the software or program involved in the control result is not described in detail, it belongs to the use of existing technologies or conventional technologies of ordinary technicians in the field. The power supply also uses the existing technologies in the field, and the main technical point of the present invention is to improve the mechanical device, so the present invention will not describe the specific circuit control relationship and circuit connection in detail.

[0032] The disclosure of the present invention provides many different embodiments or examples to realize different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described in the present invention. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or reference letters in different examples, and this repetition is for the purpose of simplicity and clarity, which itself does not indicate the relationship between the various embodiments and / or settings discussed. In addition, the present invention provides various specific examples of processes and materials, but those of ordinary skill in the art can be aware of the application of other processes and / or the use of other materials.

[0033] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0034] like Figure 1 As shown, the present invention provides an embodiment of a cross-domain WiFi personnel identity recognition method based on a multi-branch domain alignment model.

[0035] WiFi CSI (Channel State Information) human gait data can be divided into source domain data and target domain data. The source domain contains multiple interference domains, such as walking trajectory direction, walking speed, and WiFi device location; each interference domain contains multiple subdomains. The target domain contains multiple interference domains and a small number of subdomains. First, an unsupervised generative adversarial network model is used to learn the migration correspondence between the source domain and the target domain, and a feature extractor E of WiFi CSI data is obtained. The generative adversarial network model enables the feature extractor E to have preliminary domain adaptation capabilities, but the generated adversarial network model is not enough to alleviate the negative impact of the coexistence of multiple heterogeneous domains, which include source domains or target domains. The trained feature extractor E is used as the backbone network, and the multi-branch domain alignment (DDA, Divided Domain Adaptation) model is used to segment the source domain data, and the source domain data is converted into multiple subsets, namely sub-source domains. For example, it can be converted into more subdivided subdomains such as two subsets, and the two subdomains can be combined to form the entire source domain. For example, five different source domain receivers are laid out and the data corresponding to the five different source domain receivers are divided into two clusters of data, one cluster of data includes two subsets, and the other cluster of data includes three subsets.

[0036] The source domain CSI (Channel State Information) data is split and converted into sub-source domains, namely source domain 1 and source domain 2, and the target domain CSI data is converted into target domain 1. Subsequently, the multi-branch domain alignment model, namely DDA (Divided Domain Adaptation), uses the pre-designed network model with the same structure H 1 and the mesh model H 2 , mesh model H 1 and the mesh model H 2 The feature extractor E of a certain source domain branch extracts features, extracts data features of a specific branch, and learns gait features in each sub-source domain respectively. 1 and the mesh model H 2The gait features in source domain 1 and source domain 2 are learned respectively, and the features of the target domain are respectively passed through the sub-domains of the two source domains, that is, the embedded features obtained by the network model corresponding to the sub-source domain are aligned and learned with the embedded features of the two sub-source domains. In this embodiment, different arrows represent the data processing process of different domains. In order to achieve domain adaptation, the MMD (Maximum Mean Discrepancy) domain adaptation loss is introduced to minimize the distance between the two sets of source domains and the target domain. The MMD domain adaptation loss is also referred to as MMD loss. The MMD domain adaptation loss function can be formally expressed as shown in formula (1):

[0037]

[0038] in, and denote the expectations relative to distributions p and q, respectively. represents the reproducing kernel Hilbert space, which is a mathematical basis for analyzing and manipulating data. The space is equipped with a characteristic kernel function k, which is used to measure the distance between sample pairs. The function φ(x s ) and φ(x t ) represents the feature map that transforms the original sample into the reproducing kernel Hilbert space. The function k(x s ,x t ) are defined as the feature representations φ(x s ) and the feature representation φ(x t ), x s and x t Represent the WiFi data samples of the source domain and the target domain respectively.

[0039] Specifically, the MMD loss applied in the two branch networks of the DDA model can be formally expressed as follows:

[0040]

[0041] In formula (2), N represents the number of network branches (the actual value is 2, 3 or 4), Indicates A simplified estimate of The MMD loss applied in the two branch networks of the DDA model helps to align the distribution between the source and target domains in each branch, E(x t )and Indicates that the source domain and target domain samples are extracted by the common feature extractor E, and H i (E(x t )) represents the source domain grouping feature extractor H iThe extracted features, s represents the source domain, t represents the target area, and i ranges from 1 to N.

[0042] In order to make full use of the known label characteristics of source domain data, the DDA model uses the cross entropy loss function shown in formula (3) for supervised classification learning. Specifically, the cross entropy loss is formalized as follows:

[0043]

[0044] In the formula, K represents the number of gait labels, that is, the category of gait features, l y represents the true category label of the sample, σ refers to the SoftMax function operation, C i represents the classifier of branch i, which is used to convert the network output into a probability value. K represents the number of gait labels, and the range of i is 1 to K. The denotation parameter is the expectation of the source domain samples.

[0045] Since the two domains, i.e., source and target domain specific classifiers are trained on different sets of source domains, they may have disagreements in their predictions of target domain samples. The consistency between classifiers helps to eliminate the domain differences between branches and further improves the domain adaptation performance. To achieve this goal, a convergence loss is designed. To help the model achieve domain alignment of two specific branches. Specifically, the domain alignment model designed in this patent will calculate the absolute value between the probability outputs of two specific classifiers, using The loss is used to calculate the distance between different branches and the output results of the target domain data prediction, so as to narrow the distance between multiple branches and eliminate the differences between them. From the perspective of the overall network training logic, the DDA model uses unlabeled samples in the target domain to narrow the distance between multiple branches as a supervised convergence loss, and its definition is formally expressed as shown in (4).

[0046]

[0047] The DDA model uses a series of loss functions to supervise the loss, including the domain adaptation loss function Convergence loss function and the cross entropy loss function To align the distribution between the source domain and the target domain, such as Figure 1 As shown, and the domain-specific classifier is prompted to accurately identify different gait features. The domain-specific classifier is referred to as the domain classifier or classifier for short, as shown in Figure 1 Classifier 1 and Classifier 2. In general, the domain alignment loss function It can be summarized as formula (5):

[0048]

[0049] Among them, λ 4 and λ 5 Respectively and The importance weight of .

[0050] In the early stages of training, these two weights are dynamically adjusted using a progressive strategy, according to the formula λ 4 =λ 5 =2 / (1+exp(-θp′))-1, that is, it gradually changes from 0 to 1 as the training iteration proceeds, where p′ changes linearly during the training process, θ is empirically taken as 10 and remains unchanged throughout the experiment, and exp(-θp′) means the -θp′ power function with exponent e as the base. After the training is completed, during the test of the target domain data samples, the model uses the average prediction results of two or more classifiers, which are shown in the box "Average (Test)" in the figure, to predict the data samples in the target domain.

[0051] The present invention provides a cross-domain WiFi personnel identification method based on a multi-branch domain alignment model, including establishing a multi-branch domain alignment model, i.e., a DDA model, and using a loss function to supervise the DDA model, so that a domain-specific classifier can accurately identify different gait features, divide the overall source domain into two or more groups, and align each source domain with the target domain respectively, and shorten the distance between the source domain and the target domain of each branch and the distance between multiple branches through a series of loss functions, so as to achieve simultaneous migration and adaptation of multiple interference domains. It can simultaneously eliminate the negative effects brought about by multiple environmental factors that appear simultaneously, improve the accuracy of WiFi human identity identification and the reliability of the system, and reduce the difficulty of migration and adaptation of multiple interference domains.

[0052] The multi-branch domain alignment model divides the overall source domain into two or more groups, and aligns each group of source domains with the target domain respectively, shortens the distance between the source domain and the target domain of each branch through a series of loss functions, and shortens the distance between multiple branches at the same time, thereby achieving simultaneous migration and adaptation of multiple interference domains.

[0053] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents, and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A cross-domain WiFi personnel identification method based on a multi-branch domain alignment model, which recognizes the identity of a person through his or her gait characteristics, is characterized by: include, A series of loss functions are used to align the distribution between the source domain and the target domain in the multi-branch domain alignment DDA model, so as to encourage domain-specific classifiers to accurately identify different gait features. The domain adaptation loss function in the series of loss functions shortens the distance between the source domain and the target domain of each branch and the distance between multiple branches, so as to enable simultaneous migration and adaptation of multiple interference domains. The multi-branch domain alignment DDA model includes dividing the entire source domain into at least two groups, aligning each source domain with the target domain respectively, and the network model includes a common feature extractor and a feature extractor for the source domain of each group. The gait features in each group are learned respectively, and the embedded features of the target domain obtained by the network models corresponding to the sub-domains of the two source domains are aligned with the embedded features of the two sub-source domains respectively. The MMD domain adaptation loss function is used to minimize the distance between the two groups of source domains and the target domain. The MMD domain adaptation loss function is formally expressed as: in, and Respectively represent the expectations of two sets of data distributions p and q, represents the reproducing kernel Hilbert space, the function φ(x s ) and φ(x t ) represents the feature map that converts the original sample into the reproducing kernel Hilbert space, and the kernel function k(x s ,x t ) are defined as the feature representations φ(x s ) and the feature representation φ(x t ), x s and x t Represent the WiFi data samples of the source domain and the target domain respectively.

2. According to claim 1, the cross-domain WiFi personnel identification method based on the multi-branch domain alignment model is characterized in that: The series of loss functions include a domain adaptation loss function Convergence loss function and the cross entropy loss function 3. The cross-domain WiFi personnel identification method based on the multi-branch domain alignment model according to claim 2 is characterized in that: The domain adaptation loss function The formal expression is: Among them, x s and x t represent the WiFi data samples of the source domain and the target domain respectively, N represents the number of network branches, Indicates A simplified estimate of The MMD loss applied in the two branch networks of the DDA model helps to align the distribution between the source and target domains in each branch, E(x t )and Indicates that the source domain and target domain samples are extracted by the common feature extractor E, and H i (E(x t )) represents the source domain grouping feature extractor H i The extracted features, s represents the source domain, t represents the target area, and i ranges from 1 to N.

4. According to claim 3, the cross-domain WiFi personnel identification method based on the multi-branch domain alignment model is characterized in that: The number of the network branches is 2, 3 or 4.

5. The cross-domain WiFi personnel identification method based on the multi-branch domain alignment model according to claim 3 is characterized in that: Cross entropy loss function shown for the DDA model For supervised classification learning, the cross entropy loss function The formal expression is: Among them, K represents the number of gait labels, l y represents the true category label of the sample, σ refers to the SoftMax function operation, C i Represents the classifier of branch i, which is used to convert the network output results into probability values. The range of i is 1 to N. The denotation parameter is the expectation of the source domain samples.

6. The cross-domain WiFi personnel identification method based on the multi-branch domain alignment model according to claim 5 is characterized in that: The DDA model uses unlabeled samples in the target domain to shorten the distance between multiple branches and converges the loss function Help the model achieve a specific domain alignment of the two branches, the convergence loss function The formal expression is:

7. The cross-domain WiFi personnel identification method based on the multi-branch domain alignment model according to claim 6 is characterized in that: Domain Alignment Loss Function The formal expression is: Among them, λ4 and λ5 represent and The importance weight of .

8. The cross-domain WiFi personnel identification method based on the multi-branch domain alignment model according to claim 7 is characterized in that: In the early stage of training, the two weights are dynamically adjusted using a progressive strategy. According to the formula λ4=λ5=2 / (1+exp(-θp′))-1, they gradually change from 0 to 1 as the training iterations proceed, where p′ changes linearly during the training process, θ is empirically taken as 10 and remains unchanged throughout the experiment, and exp(-θp′) means the -θp′ power function with exponent e as the base.

Images