A source domain selection method for WiFi cross-scenario perception

Abstract

The present invention belongs to the technical field of cross-domain behavior perception of wireless signals, discloses a source domain selection method suitable for WiFi cross-scene perception, fully considers the global data distribution of the source domain and the target domain and the local distribution of labeled data, designs an evaluation strategy based on a linear combination of global and local distances, selects the source domain closest to the target domain to construct a cross-domain perception model, and ensures the reliability and accuracy of cross-scene passive perception; it also provides a new research idea for the cross-scene perception problem based on WiFi.

Description

Technical Field

[0001] The present invention belongs to the technical field of cross-domain behavior perception of wireless signals, and in particular relates to a source domain selection method suitable for WiFi cross-scenario perception. Background Art

[0002] With the popularization of Internet of Things technology and the widespread application of indoor WiFi devices, passive personnel perception technology based on WiFi signals has shown great potential. Considering that the model trained in one environment cannot be directly applied to another environment, cross-domain perception based on WiFi signals has become a research hotspot. For the cross-domain perception problem in actual scenarios, the difference between the source domain and the target domain may lead to the phenomenon of negative transfer. Therefore, selecting a source domain dataset suitable for a given target domain from multiple available source domains is of great significance to improve the accuracy of cross-domain perception.

[0003] Existing cross-domain perception methods based on WiFi signals often ignore the source domain selection problem and directly build complex network models (such as adversarial training networks) to learn domain adaptive features. If the data of an original scene is randomly selected from multiple source domains to build a cross-domain model, the blind selection of the source domain will greatly reduce the reliability of cross-scene perception. In other research fields, such as the document "Mental workload identification based on multi-source domain deep domain adaptation" (from "Journal of East China University of Science and Technology (Natural Science Edition)"), patent applications CN112364994A, CN114330451A, CN115034296A, the data is directly projected into a reproducible Hilbert space or Euclidean space, and the maximum mean difference (MMD distance or improved MMD distance) is used to directly calculate the difference between the two data sources. However, the data projection process ignores the correlation between data and labels, resulting in poor recognition ability of the projection space; in addition, for behavior perception tasks, it contains multiple different action categories, and it is necessary to pay attention to the similarity between samples of the same category in multiple domains at the same time. CN114254676A trains multiple source domain models and counts their test results in the target domain, generates a multi-source classifier by voting, and further uses an iterative method to select the source domain. However, this method is complicated to operate. Summary of the invention

[0004] To solve the above technical problems, the present invention provides a source domain selection method suitable for WiFi cross-scenario perception. The method fully considers the global data distribution of the source domain and the target domain and the local distribution of labeled data, designs an evaluation strategy based on a linear combination of global and local distances, and selects the source domain closest to the target domain to construct a cross-domain perception model.

[0005] The present invention provides a source domain selection method suitable for WiFi cross-scenario perception, comprising the following steps:

[0006] Step 1: Collect WiFi signals of different actions in multiple indoor scenes and extract channel signal status information CSI. The data of one scene is used as the target domain. , data from other scenarios as multiple source domains , Label the data samples in the target domain and all source domains;

[0007] Step 2: Use a deep convolutional neural network to train the i-th source domain data;

[0008] Step 3: Use the trained deep convolutional neural network as a feature extractor to extract the source domain Characteristics of all data, as well as the target domain A small amount of features with labeled data;

[0009] Step 4: In the source domain and the target domain with labels In the data, class prototypes of different categories are calculated, and the cosine similarity of the same class prototype is used as the local similarity of the class. The reciprocal of the sum of the local similarities of all classes is counted as the local distance;

[0010] Step 5: Use JS divergence to measure the source domain and target domain The similarity of the global data probability distribution is used as the global distance;

[0011] Step 6: Linearly combine the local distance and the global distance as the distance measure between the two domains;

[0012] Step 7: Repeat steps 2-6 to traverse all other source domains, and select the source domain corresponding to the minimum distance for subsequent cross-scene perception.

[0013] Furthermore, in the target domain In the training, only a small amount of data has category label information for training; multiple source domains In the above example, all data are marked with category label information.

[0014] Furthermore, a deep convolutional neural network AlexNet is used to train a certain source domain data.

[0015] Furthermore, step 3 is specifically to use the trained neural network as a feature extractor, take the last fully connected layer as the sample feature, extract the features of all the data in the source domain, and the features of a small amount of labeled data in the target domain.

[0016] Furthermore, step 4 is specifically to use the average value of each category feature in the source domain and the labeled target domain data as the class prototype, and the formula is:

[0017] ;

[0018] in, is the class prototype of the k-th action, is the number of samples of the k-th action, is the i-th sample of the k-th action, is a feature extractor;

[0019] Then, the cosine similarity is used to measure the similarity between the prototypes of the same category in the source domain and the target domain, and the inverse of the sum of the similarities of the prototypes of each category is taken as the local distance; and target domain The local distance It is expressed as:

[0020] ,

[0021] in, , are the class prototypes of the k-th action in the source domain and the target domain respectively, K refers to the number of action categories, is the cosine similarity.

[0022] Furthermore, in step 5, JS divergence is used to measure the source domain and target domain The similarity of the global probability distribution of all sample data in is used as the global distance :

[0023] ;

[0024] in, , are the probability distributions of the two domains respectively.

[0025] Furthermore, in step 6, the local distance and the global distance are linearly combined as the distance measure between the two domains:

[0026] ;

[0027] Among them, u is the equilibrium parameter.

[0028] Furthermore, in step 7, the distance from each source domain to the target domain is calculated for all source domains, that is, , and select the source domain corresponding to the minimum distance :

[0029] ;

[0030] in, Represents the source domain data corresponding to the smallest element in the set; using the selected source domain and a given target domain Proceed to build subsequent cross-domain perception models.

[0031] The beneficial effects described in the present invention are as follows: the present invention proposes a source domain selection method suitable for WiFi cross-scenario perception for the problem of cross-domain perception based on WiFi signals, and innovatively proposes a local distance based on class prototypes in consideration of the sparsity of label data in the target scene; considers the source domain samples and label data at the same time, trains the network model to extract sample features with strong recognition ability, and uses the model to project labeled samples into the feature space; calculates class prototypes of different categories in all data in the source domain and labeled data in the target domain; evaluates the class prototype similarity of each behavior category in the two domains; and finally calculates the local similarity. Since the local distance fully considers the differences in the distribution of samples of different categories, it can effectively evaluate the local similarity of the two domains; in addition, the present invention further calculates the global distance based on all data, comprehensively considers the global and local distances, and more comprehensively and effectively evaluates the similarity between the source domain and the target domain; by selecting the source domain closest to the target domain for cross-domain perception, the occurrence of negative transfer can be avoided. In view of the fact that most of the current cross-domain perception problems based on WiFi ignore the importance of source domain selection, the present invention also provides a new research idea for the cross-scene perception problem based on WiFi. After the source domain data is selected, this method can be applied before all existing cross-domain perception algorithms, thereby ensuring the reliability and accuracy of cross-scene passive perception. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a flow chart of a source domain selection method applicable to WiFi cross-scenario perception in an embodiment of the present invention;

[0033] Figure 2 This is a rendering of the source domain selection solution described in the present invention. DETAILED DESCRIPTION

[0034] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments in conjunction with the accompanying drawings.

[0035] A source domain selection method suitable for WiFi cross-scenario perception includes the following steps:

[0036] Step 1: Collect WiFi signals of different actions in multiple indoor scenes and extract channel signal state information CSI. Take the data of one scene as the target domain and the data of other scenes as multiple source domains, and label the samples with categories.

[0037] Step 2: Use a deep convolutional neural network to train a source domain data;

[0038] Step 3: Use the trained neural network as a feature extractor to extract the features of all the data in the source domain and the features of a small amount of labeled data in the target domain;

[0039] Step 4: In the source domain and the labeled target domain data, class prototypes of different categories are calculated, the cosine similarity of the same class prototype is used as the local similarity of the class, and the reciprocal of the sum of the local similarities of all classes is counted as the local distance;

[0040] Step 5: Use JS divergence to measure the similarity of the global data probability distribution of the source domain and the target domain, and use it as the global distance;

[0041] Step 6: Linearly combine the local distance and the global distance as the distance measure between the two domains;

[0042] Step 7: Use steps 2-6 to traverse all source domains and select the source domain corresponding to the minimum distance for subsequent cross-scene perception.

[0043] The following is a detailed description of a source domain selection method suitable for WiFi cross-scenario perception provided by the present invention.

[0044] See also Figure 1 , which is a flowchart of a source domain selection method applicable to WiFi cross-scenario perception according to an embodiment of the present invention.

[0045] In step 1, a wireless router is used as a WiFi transmitter and a PC equipped with an Intel 5300 is used as a receiver to collect CSI data of wireless signals in multiple indoor scenes. The collected data is pre-processed by filtering, etc. The data of one of the scenes is used as the target domain. In the target domain, only a small amount of data has category label information as the training set, and all the rest of the data is used as the test set; the data of other multiple scenarios are used as multiple source domains. , and all are marked with category label information.

[0046] In step 2, select data from a source domain , use CSI samples and corresponding labels to train the deep convolutional neural network AlexNet.

[0047] In step 3, the trained neural network is used as a feature extractor , taking the last fully connected layer as the sample feature; assuming the sample is , then the corresponding feature ;use Extract all data from the source domain and features of a small amount of labeled data in the target domain.

[0048] In step 4, the average value of each category feature in the source domain and the labeled target domain data is used as the class prototype; specifically, assuming , , are the class prototype, sample number and i-th sample of the k-th action respectively, then:

[0049]

[0050] The cosine similarity is used to measure the similarity between prototypes of the same category in two domains, and the inverse of the sum of similarities of prototypes of each category is taken as the local distance. and target domain The local distance It is expressed as:

[0051]

[0052] In the above formula, , are the class prototypes of the k-th action in the source domain and the target domain, respectively. is the cosine similarity.

[0053] In step 5, JS divergence is used to measure the source domain and target domain The similarity of the global probability distribution of all sample data in is used as the global distance :

[0054]

[0055] In the above formula, , are the probability distributions of the two domains respectively.

[0056] In step 6, the local distance and the global distance are linearly combined as the distance measure between the two domains:

[0057]

[0058] In the above formula, u is the equilibrium parameter.

[0059] In step 7, steps 2-6 are used for all source domains to calculate the distance from each source domain to the target domain, that is, , and select the source domain corresponding to the minimum distance :

[0060]

[0061] In the above formula, Represents the source domain data corresponding to the smallest element in the set; using the selected source domain and a given target domain Proceed to build subsequent cross-domain perception models.

[0062] Figure 2 This is the effect diagram of different source domain selections. Figure 2 In the figure, S1, S2, and S3 are data collected from three different scenes. When S1, S2, and S3 are set as the target domains respectively, different perception accuracies can be obtained by selecting different source domain data for cross-scene perception.

[0063] Table 1 shows the results of the source domain selection scheme mentioned in the embodiment. Figure 2 As can be seen from Table 1, the embodiment method can select a more reasonable source domain model and achieve high-precision recognition of various actions in the cross-domain perception problem based on WiFi. Figure 2 It can be seen that when S1 is used as the target domain, selecting S2 as the source domain can obtain higher perception accuracy; and from Table 1, it can be seen that when S1 is used as the target domain, the embodiment method selects S2 as the source domain. Similarly, if S2 or S3 is used as the target domain, similar conclusions can be obtained;

[0064] .

[0065] The above description is only a preferred embodiment of the present invention and is not intended to be a further limitation of the present invention. All equivalent changes made using the contents of the present specification and drawings are within the protection scope of the present invention.

Claims

1. A source domain selection method suitable for WiFi cross-scenario perception, characterized in that: The steps of the method are: Step 1: Collect WiFi signals of different actions in multiple indoor scenes and extract channel signal status information CSI. The data of one scene is used as the target domain. , data from other scenarios as multiple source domains , Label the data samples in the target domain and all source domains; Step 2: Use a deep convolutional neural network to train the i-th source domain data; Step 3: Use the trained deep convolutional neural network as a feature extractor to extract the source domain Characteristics of all data, as well as the target domain A small amount of features with labeled data; Step 4: In the source domain and the target domain with labels In the data, class prototypes of different categories are calculated, and the cosine similarity of the same class prototype is used as the local similarity of the class. The reciprocal of the sum of the local similarities of all classes is counted as the local distance; Step 5: Use JS divergence to measure the source domain and target domain The similarity of the global data probability distribution is used as the global distance; Step 6: Linearly combine the local distance and the global distance as the distance measure between the two domains; Step 7: Repeat steps 2-6 to traverse all other source domains, and select the source domain corresponding to the minimum distance for subsequent cross-scene perception.

2. According to claim 1, a source domain selection method suitable for WiFi cross-scenario perception is characterized in that: In the target domain In the training, only a small amount of data has category label information for training; multiple source domains In the above example, all data are marked with category label information.

3. According to a source domain selection method suitable for WiFi cross-scenario perception according to claim 1, it is characterized in that: The deep convolutional neural network AlexNet is used to train a certain source domain data.

4. According to claim 1, a source domain selection method suitable for WiFi cross-scenario perception is characterized in that: Specifically, step 3 uses the trained neural network as a feature extractor and the last fully connected layer as the sample feature to extract the features of all the data in the source domain and the features of a small amount of labeled data in the target domain.

5. According to claim 1, a source domain selection method suitable for WiFi cross-scenario perception is characterized in that: Step 4 is to use the average value of each category feature in the source domain and the labeled target domain data as the class prototype, and the formula is: ， in, is the class prototype of the k-th action, is the number of samples of the k-th action, is the i-th sample of the k-th action, is a feature extractor; Then, the cosine similarity is used to measure the similarity between the prototypes of the same category in the source domain and the target domain, and the inverse of the sum of the similarities of the prototypes of each category is taken as the local distance; and target domain The local distance It is expressed as: ， in, , are the class prototypes of the k-th action in the source domain and the target domain respectively, K refers to the number of action categories, is the cosine similarity.

6. According to claim 5, a source domain selection method suitable for WiFi cross-scenario perception is characterized in that: In step 5, JS divergence is used to measure the source domain and target domain The similarity of the global probability distribution of all sample data in is used as the global distance : ， in, , are the probability distributions of the two domains respectively.

7. According to claim 6, a source domain selection method suitable for WiFi cross-scenario perception is characterized in that: In step 6, the local distance and the global distance are linearly combined as the distance measure between the two domains: ， Among them, u is the equilibrium parameter.

8. According to claim 7, a source domain selection method suitable for WiFi cross-scenario perception is characterized in that: In step 7, the distance from each source domain to the target domain is calculated for all source domains, that is, , and select the source domain corresponding to the minimum distance : ， in, Represents the source domain data corresponding to the smallest element in the set; using the selected source domain and a given target domain Proceed to build subsequent cross-domain perception models.

Images