CN111460901B - Wi-Fi signal and transfer learning-based activity classification model construction method and system - Google Patents

Wi-Fi signal and transfer learning-based activity classification model construction method and system Download PDF

Info

Publication number
CN111460901B
CN111460901B CN202010144596.8A CN202010144596A CN111460901B CN 111460901 B CN111460901 B CN 111460901B CN 202010144596 A CN202010144596 A CN 202010144596A CN 111460901 B CN111460901 B CN 111460901B
Authority
CN
China
Prior art keywords
csi
layer
user activity
network
active
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202010144596.8A
Other languages
Chinese (zh)
Other versions
CN111460901A (en
Inventor
冯宏伟
明星霞
卜起荣
冯筠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
NORTHWEST UNIVERSITY
Original Assignee
NORTHWEST UNIVERSITY
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by NORTHWEST UNIVERSITY filed Critical NORTHWEST UNIVERSITY
Priority to CN202010144596.8A priority Critical patent/CN111460901B/en
Publication of CN111460901A publication Critical patent/CN111460901A/en
Application granted granted Critical
Publication of CN111460901B publication Critical patent/CN111460901B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computational Linguistics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Evolutionary Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention discloses an activity classification model construction method and system based on Wi-Fi signals and transfer learning, comprising the following steps: preprocessing Wi-Fi channel state information corresponding to user activities through Butterworth filtering, singular value decomposition, phase correction and other operations; training a deep convolution feature extraction network in a training mode through data triplet combination in a source domain, adopting transfer learning, and extracting active features in the Wi-FiCSI fragment after preprocessing in a target domain by using a feature extraction network model; and utilizing the extracted activity features to combine with an SVM classifier to realize a cross-domain activity classification task. According to the method, a novel Wi-Fi signal-based cross-domain activity classification method is established by adopting a migration learning mode, and a novel cross-domain activity classification framework is provided for solving the problem that a model application is invalid after an environment is changed in the conventional activity classification method.

Description

Wi-Fi signal and transfer learning-based activity classification model construction method and system
Technical Field
The invention relates to the technical field of activity classification and neural networks, in particular to an activity classification method based on Wi-Fi signals and transfer learning, which is mainly used for realizing user activity classification under the cross-domain condition.
Background
The human activity classification has important research value and significance in the fields of intelligent home, man-machine interaction, intelligent security and the like. In a smart home, we can intelligently adjust indoor temperature, light, humidity and the like for a user through a perception result of user activities. For families with the old, the intelligent nursing for the old can be provided. In the field of human-computer interaction, the perception of the behavior of a user is the basis for understanding and realizing the demands of the user. In the field of intelligent security, the identity of a user can be effectively judged through mining the behavior characteristics of the user, and the property and personal safety problems caused by a single identity authentication mode are avoided. Under such a development trend, the human activity classification technology (Human Activity Recognition, HAR) has great application value and potential.
In recent years, the rapid development of wireless communication technology is not only convenient for people's daily life, but also creates possibility for realizing passive human activity perception. The low cost and wide application of Wi-Fi make Wi-Fi-based activity classification technology rapidly one of the popular technologies in the wireless sensing field. The core of the Wi-Fi-based activity classification technology is to realize an activity classification task by mining the influence of user activities on Wi-Fi signal propagation. Because the signal of the Wi-Fi receiving end is the superposition effect of the multipath signals sent by the sending end through reflection, refraction, diffraction and the like. Thus, when there is no user in the environment, the signal fading in Wi-Fi propagation is only related to the deployment of the environment, which has a certain impact on Wi-Fi signal propagation. On the other hand, these characteristics of Wi-Fi signal propagation also lead to problems of failure of existing activity classification model applications after environmental deployment changes. A large amount of marker data needs to be retrieved in a new scene and the activity classification model needs to be re-built.
With the continuous research of researchers in recent years, some methods of cross-domain activity classification have been proposed, and although the task of cross-domain activity classification based on Wi-Fi signals is achieved to some extent, one disadvantage common to these methods is that a large amount of training data needs to be collected in the target domain. This deficiency limits the widespread deployment and application of Wi-Fi based active technologies, e.g., if one wants to apply such technologies to private spaces, we have difficulty collecting enough tagged samples in these environments. Therefore, a more robust, low cost Wi-Fi cross-domain activity classification algorithm remains to be studied.
Disclosure of Invention
Aiming at the defects existing in the prior art, the invention aims to provide an activity classification method based on Wi-Fi signals and transfer learning, solves the problems that the identification accuracy is low and a large number of labeled samples are required in a cross-domain activity classification task in the prior art, and provides an effective model training mode based on junction data aiming at the problem that the existing deep learning model training method is not suitable for cross-domain activity classification.
In order to solve the technical problems, the invention adopts the following technical scheme:
The invention provides a user activity classification model construction method based on Wi-Fi signals and transfer learning, which is used for cross-domain user activity classification from a source domain to a target domain in Wi-Fi signal environment and comprises the following steps,
step 1, respectively acquiring CSI samples corresponding to each user activity sample in a source domain and a target domain, and obtaining a CSI image corresponding to each user activity sample through the CSI sample corresponding to each user activity sample;
step 2, an active feature extraction model based on a VGG16 network is established by adopting a transfer learning technology, simultaneously, the CSI images corresponding to all the user active samples in the source domain obtained in the step 1 are established into a triplet CSI image sample pair by adopting a triplet data pair structure, and the triplet CSI image sample pair is used as input to train the active feature extraction model; then, the CSI images corresponding to all the user activity samples in the target domain obtained in the step 1 are constructed into a triplet CSI image sample pair, and the triplet CSI image sample pair is sent into the trained activity feature extraction model to perform feature coding, so that feature vectors corresponding to each user activity sample in the target domain are obtained;
and 3, establishing a linear SVM user activity classification model, and training by utilizing the feature vector obtained in the step 2 to obtain a trained user activity classification model.
Specifically, obtaining the CSI image corresponding to each user activity through the CSI sample corresponding to each user activity sample refers to: and preprocessing the CSI samples corresponding to each user activity acquired in the source domain and the target domain by sequentially adopting data preprocessing methods of Butterworth filtering, singular value decomposition, phase correction, amplitude phase splicing and data normalization, and finally obtaining the active CSI images corresponding to each user activity in the source domain and the target domain respectively.
More specifically, step 1.1, performing butterworth filtering on original CSI amplitude data in CSI samples corresponding to each user active sample to obtain CSI amplitudes after removing high-frequency noise;
step 1.2, singular value decomposition is carried out on the CSI amplitude data preprocessed in the step 1.1;
step 1.3, carrying out phase correction on the original CSI phase in the CSI sample corresponding to each user active sample to obtain the CSI phase after randomness and uncertainty are eliminated;
step 1.4, singular value decomposition is carried out on the CSI phase data preprocessed in the step 1.3;
and step 1.5, organizing the CSI amplitude and the CSI phase obtained in the step 1.2 and the step 1.4 after singular value decomposition into an image data format and splicing left and right to obtain a movable CSI image.
In addition, in step 1.5, the left and right splicing is performed using the data format of the following formula:
Figure BDA0002400295970000031
wherein G is sample A sample of the activity is represented and,
Figure BDA0002400295970000032
CSI amplitude representing the current active sample after preprocessing, +.>
Figure BDA0002400295970000033
Indicating the CSI phase of the corrected current active sample, t indicating the duration of the user activity, P rate Representing Wi-Fi connectionsThe data sampling frequency of the receiving end;
after splicing, obtaining a CSI feature matrix, normalizing floating point data in the CSI feature matrix to a [0,255] interval, copying three layers, and organizing the three layers into a three-channel image data format, namely, a movable CSI image.
In addition, in step 2, an active feature extraction model is built and trained, and the active feature extraction model is an improved deep active feature extraction convolutional network (Activityfeatureextraction neural network, AFEN), specifically comprising:
firstly, establishing an active feature extraction model by adopting a transfer learning technology, and constructing a triplet CSI image sample pair by grouping three CSI images into a pair;
secondly, constructing a deep active feature extraction convolutional network AFEN which is input as a triplet CSI image sample pair;
and finally, training the AFEN by adopting a triplet CSI image sample pair and the feature coding distances of three CSI image samples in the sample pair to obtain the trained AFEN.
The above-described construction of a deep active feature extraction convolutional network with inputs as triplet CSI image sample pairs includes constructing a deep convolutional network for extracting environmental deployment independent active features that improves upon VGG16 pre-trained on ImageNet datasets: layer 1 output data format of AFEN
Figure BDA0002400295970000034
Wherein->
Figure BDA0002400295970000035
Representing anchor activity samples,/->
Figure BDA0002400295970000036
Representing activity samples like anchor activity samples, < >>
Figure BDA0002400295970000037
Representing activity samples of different classes from the anchor activity sample;
the network receives one triplet data, and then the 21 st layer of the network can obtain 3 active feature vectors corresponding to the triplet data;
the specific structure of the network is as follows: the network comprises 21 layers, wherein the 1 st layer is an input layer, namely an input triplet CSI image sample pair; the 2 nd layer and the 3 rd layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 64; the 4 th layer is a pooling layer, and the size of the pooling core is 2x2; the 5 th layer and the 6 th layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 128; the 7 th layer is a pooling layer, and the size of a pooling core is 2x2; the 8 th, 9 th and 10 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 256; the 11 th layer is a pooling layer, and the size of a pooling core is 2x2; the 12 th, 13 th and 14 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 512; the 15 th layer is a pooling layer, and the size of a pooling core is 2x2; the 16 th, 17 th and 18 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 1024; the 19 th layer is a pooling layer, and the size of a pooling core is 2x2; layer 20 is the global pooling layer; layer 21 is the feature stitching layer.
The above-mentioned adoption triplet CSI image sample pair, and the characteristic code of three CSI image samples in sample pair is apart from AFEN that training step obtained, obtain the AFEN after training, specifically include: the method of transfer learning is adopted, an AFEN network is initialized by pre-trained VGG16 network parameters in the training process, then network parameters of layers 2-11 of the AFEN network are fixed, the rest network parameters are trained, the loss function of the AFEN network in training is the absolute value square error between active feature codes between a triplet active sample pair, and the adjustable parameters in the AFEN network are optimized by a random gradient descent method in the training process; after training is completed, the AFEN network parameters are saved, and the trained AFEN network is obtained;
the feature vector extraction section includes: organizing the movable CSI image of the target domain obtained in the step 1 into a triplet CSI image sample pair; parameters of layers 2-20 of the trained AFEN network are kept unchanged, and layer 21 outputs three active CSI images to input corresponding feature vectors, so that feature extraction is carried out on the triplet CSI image sample pairs by utilizing the trained AFEN network to obtain the feature vector of each active CSI image.
The invention also provides a user activity classification method based on Wi-Fi signals and transfer learning, which comprises four parts of data preprocessing, feature extraction model training, activity feature extraction and activity classification of a target domain, and comprises the following steps:
step 1, preprocessing CSI samples corresponding to each user activity acquired in a source domain and a target domain by adopting data preprocessing methods such as Butterworth filtering, singular value decomposition, phase correction, amplitude phase splicing, data normalization and the like. And finally, obtaining the active CSI image corresponding to each user activity in the source domain and the target domain.
And 2, constructing the active CSI image in the source domain obtained in the step 1 into a triplet CSI image sample pair by adopting a triplet data pair structure. The training method is characterized in that the paired triple CSI image sample pairs are used as input, and an active feature extraction model (the model is a convolutional neural network based on VGG16 improvement) designed by the training method is trained, so that a trained active feature extraction model of a user is obtained.
And 3, organizing the movable CSI image of the target domain obtained in the step 1 into a triplet CSI image sample pair, sending the triplet CSI image sample pair into the movable feature extraction model trained in the step 2, and carrying out feature coding on the movable feature extraction model by adopting the feature extraction model to finally obtain a feature vector corresponding to each user movable sample in the target domain.
And 4, setting model parameters of a support vector machine (Support Vector Machine, SVM) to obtain an activity classification model of the target domain activity classification. A small number of labeled activity samples are selected from each class as a training set, and the rest is used as a test set. And training the activity classification model by adopting a training set to obtain the activity classification model of the target domain. And then testing the test set by using the trained activity classification model, and finally performing activity classification by using the activity classification model.
Further, the overall structure of the pretreatment part in step 1 is as shown in fig. 2, and the establishment process includes:
step 1.1, performing Butterworth filtering on original CSI amplitude data to obtain CSI amplitude after high-frequency noise is removed;
step 1.2, singular value decomposition is carried out on the CSI amplitude data preprocessed in the step 1.1;
step 1.3, performing phase correction on the original CSI phase to obtain the CSI phase after randomness and uncertainty are eliminated, and making
Figure BDA0002400295970000051
Representing the approximate true phase, can be expressed as follows:
Figure BDA0002400295970000052
wherein, is less than H x, (f i The method comprises the steps of carrying out a first treatment on the surface of the ) Representing the true phase, i representing the current subcarrier, t being the subcarrier f i At the current point in time, α and β are the slopes and offsets representing the phase changes over all subcarriers. The α and β on each pair of transmit and receive antennas can be calculated by the following formula:
Figure BDA0002400295970000053
Wherein, the liquid crystal display device comprises a liquid crystal display device,
Figure BDA0002400295970000054
representing the phase value of the 30 th CSI subcarrier acquired.
And step 1.4, performing singular value decomposition on the CSI phase data preprocessed in the step 1.3.
Step 1.5, splicing the CSI amplitude and the CSI phase obtained in the steps 1.2 and 1.4 by adopting the following data format:
Figure BDA0002400295970000055
wherein G is sample Representing a current active sample of the sample,
Figure BDA0002400295970000061
CSI amplitude representing the current active sample after preprocessing, +.>
Figure BDA0002400295970000062
Representing the CSI phase of the corrected current active sample. N=t×p rate T represents the duration of the user activity, P rate Representing the data sampling frequency of the Wi-Fi receiver.
After splicing, floating point data in the CSI feature matrix are normalized to the [0,255] interval, and three layers are copied to be organized into a three-channel image data format, which is called as a movable CSI image for short.
And step 1.6, adjusting the size of the active CSI image obtained in the step 1.5.
Further, the feature extraction model training section described in step 2 includes:
step 2.1, forming three groups of CSI images processed by the pretreatment part of the activity classification model into a pair to construct a triplet CSI image sample pair;
step 2.2, designing a deep active feature extraction convolutional network (Activityfeatureextraction neural network, AFEN) input as a triplet CSI image sample pair.
And 2.3, training the AFEN obtained in the step 2.2 by adopting a triplet CSI image sample pair and the characteristic coding distances of three CSI image samples in the sample pair to obtain the AFEN after parameter adjustment.
Further, the design input described in step 2.2 is a deep active feature extraction convolutional network of triplet CSI image sample pairs, and specifically includes:
first, a deep convolutional network AFEN for extracting environmental deployment-independent activity features is constructed, which is improved based on VGG16 pre-trained on ImageNet dataset. Layer 1 output data format of AFEN
Figure BDA0002400295970000063
Wherein s is a Representing anchor activity samples,/->
Figure BDA0002400295970000064
Representing activity samples like anchor activity samples, < >>
Figure BDA0002400295970000065
Representing activity samples of a different class than the anchor activity sample. Because the AFEN network can accept one triplet data, the 21 st layer of the AFEN network can obtain 3 active feature vectors corresponding to the triplet data samples;
the network structure comprises 21 layers, wherein the 1 st layer is an input layer, and the input of the network is a triplet CSI image sample pair obtained in the step 2.1; the 2 nd layer and the 3 rd layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 64; the 4 th layer is a pooling layer, and the size of the pooling core is 2x2; the 5 th layer and the 6 th layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 128; the 7 th layer is a pooling layer, and the size of a pooling core is 2x2; the 8 th, 9 th and 10 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 256; the 11 th layer is a pooling layer, and the size of a pooling core is 2x2; the 12 th, 13 th and 14 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 512; the 15 th layer is a pooling layer, and the size of a pooling core is 2x2; the 16 th, 17 th and 18 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 1024; the 19 th layer is a pooling layer, and the size of a pooling core is 2x2; layer 20 is the global pooling layer; layer 21 is the feature stitching layer.
Further, the training of the deep convolutional network AFEN of the feature extraction part through the triplet CSI image sample pair in step 2.3 specifically includes:
the AFEN network is initialized by pre-trained VGG16 network parameters in the training process, then the network parameters of layers 2-11 are fixed, and the rest network parameters are trained.
The loss function of the AFEN network during training is the absolute square error between active feature codes between triplet active sample pairs, which can be calculated by:
Figure BDA0002400295970000071
where a represents the minimum difference between the positive and negative pair of feature distances,
Figure BDA0002400295970000072
representing anchor sample, ++>
Figure BDA0002400295970000073
Representation and->
Figure BDA0002400295970000074
Activity samples belonging to the same category, +.>
Figure BDA0002400295970000075
Representation and->
Figure BDA0002400295970000076
Active samples that do not belong to the same category.
Optimizing adjustable parameters in the AFEN network by a random gradient descent method in the training process; after training, the parameters of the AFEN network are saved, and the AFEN network with the parameters adjusted is obtained.
Further, the activity feature extraction section of step 3 includes:
and 3.1, keeping parameters of the 2 nd layer to the 20 th layer of the AFEN network trained in the step 2.3 unchanged, wherein the output of the 21 st layer of the network is the feature vector corresponding to the input of the active CSI image. And organizing the movable CSI image of the target domain obtained in the step 1 into a triplet CSI image sample pair
And 3.2, performing feature extraction on the triplet CSI image sample pairs of the target domain obtained in the step 3.1 by adopting the AFEN network with the fixed parameters obtained in the step 3.1 to obtain feature vectors of each movable CSI image.
Further, the activity classification section of the target domain in step 4 includes:
and 4.1, designing model parameters of a support vector machine (Support Vector Machine, SVM) to obtain an activity classification model of the target domain activity classification.
And 4.2, selecting a small number of labeled active samples from the coded active samples obtained in the step 3.2 as a training set, and taking the rest as a test set. And obtaining a training set and a testing set of the target domain.
And 4.3, training the activity classification model by adopting the training set divided in the step 4.2 to obtain the activity classification model of the target domain.
And 4.4, testing the test set by using the trained activity classification model to obtain an activity class corresponding to each sample in the target domain test data set, and finally realizing a target domain user activity classification task with low cost based on few samples.
On the basis, the invention also provides a user activity classification model construction device based on Wi-Fi signals and transfer learning, which is characterized by comprising a preprocessing module, a feature extraction model training and user activity feature extraction module and a target domain activity classification module;
The preprocessing module is used for preprocessing CSI sample data corresponding to each user activity acquired in the source domain and the target domain to obtain an active CSI image corresponding to each user activity in the source domain and the target domain;
the feature extraction model training and user activity feature extraction module adopts a migration learning technology to establish an activity feature extraction model based on a VGG16 network, adopts a triple data pair structure to construct the CSI images corresponding to all user activity samples in a source domain obtained by the preprocessing module into a triple CSI image sample pair, takes the triple CSI image sample pair as input, and trains the activity feature extraction model; then, the CSI images corresponding to all the user activity samples in the target domain obtained by the preprocessing module are constructed into a triplet CSI image sample pair, and the triplet CSI image sample pair is sent into the trained activity feature extraction model to perform feature coding, so that feature vectors corresponding to each user activity sample in the target domain are obtained;
the target domain activity classification module is used for establishing a linear SVM user activity classification model, training by utilizing the feature extraction model and the feature vector obtained by the user activity feature extraction module, and obtaining a trained user activity classification model.
Finally, the invention provides a user activity classification system based on Wi-Fi signals and transfer learning, which comprises an image input device and the user activity classification model building device based on Wi-Fi signals and transfer learning;
the image input module is used for inputting CSI images corresponding to user classification activities of a source domain and a target domain to be classified;
the Wi-Fi signal and transfer learning-based user activity classification model construction device is used for classifying cross-domain user activities by using the constructed user activity classification model to obtain a user activity classification result.
Compared with the prior art, the invention has the following technical effects:
1. according to the construction method, when the active feature extraction model is designed, the transfer learning technology is adopted, the network parameters of the VGG16 pre-training model are utilized, and the multi-level active features are fused, so that the existing feature extraction model can be finely adjusted by using less active sample data aiming at the Wi-Fi-based cross-domain active classification problem, and the same feature extraction effect is achieved with lower calculation cost.
2. The invention provides a cross-domain activity classification model aiming at the problem that the prior cross-domain activity classification method needs to collect a large number of labeled samples in a source domain and a target domain to remove environmental influence, namely: the feature extraction network is trained by constructing the triplet active sample pairs, so that the network can extract active feature vectors with close similar distances and far different similar distances. The training mode can enable the activity feature extraction model to avoid being influenced by environmental deployment, and achieves good feature coding effect on the cross-domain activity classification problem.
Drawings
FIG. 1 is an overall block diagram of a model building method of the present invention;
FIG. 2 is a schematic diagram of a portion of the overall structure of the CSI preprocessing for activity in the present invention
FIG. 3 is a schematic diagram of the result of the CSI amplitude preprocessing (horizontal axis: data sampling point, vertical axis: amplitude) according to the present invention
FIG. 4 is a diagram of a network model of a training feature extraction portion of the present invention;
FIG. 5 is a diagram of a network model of a test feature extraction section of the present invention;
fig. 6 is a schematic diagram of wireless signal propagation affected by environmental deployment.
Detailed Description
Specific examples of the present invention are given below, and it should be noted that: 1. the invention is not limited to the following specific embodiments, and all equivalent changes made on the basis of the technical proposal of the application fall within the protection scope of the invention. 2. The data set employed in the embodiment contains 1760 user activity samples in total, with the data of class 34 activities in the source domain as the training set for the user activity feature extractor, training the AFEN model. Three samples of each type of activity in 10 types of user activities in the target domain are randomly selected as training sets of the target domain activity classification model, and the rest data in the target domain are used as test sets.
The terms involved in the present invention are explained as follows:
Wi-FiCSI: known collectively as WiFi channel state information (Channel state information, CSI), which reflects signal attenuation characteristics during WiFi signal propagation.
In the field of cross-domain activity classification based on wireless signals, the propagation characteristics of wireless signals are greatly affected by environmental deployment. As shown in fig. 6, the propagation of a wireless signal includes both signal propagation on a dynamic path and signal propagation on a static path. When the environmental deployment changes, the collected WiFi CSI data will also change accordingly, which may result in that the collected WiFi CSI is different when the same user does the same activity in different scenarios. If a researcher collects WiFi CSI data of user activities in the environment 1, an activity classification model is established, and the model can have the problem of application failure in the environment 2. This is because environment 2 changes static path C compared to environment 1 and will eventually react on the acquired WiFi CSI.
An environment deployment is generally referred to as a domain, and is further specifically divided into a source domain and a target domain in the cross-domain activity classification problem. Wherein:
source domain: generally refers to an experimental scenario, such as a scenario in which a researcher is collecting data when building a model.
Target domain: generally referred to as the scenario in which the activity classification model is actually applied. Such as the user's home.
As shown in fig. 6, if a researcher collects WiFi CSI data of a user activity in an environment 1 and expects to apply the WiFi CSI data to an environment 2, the user activity is identified, the environment 1 is called a source domain, the environment 2 is called a target domain, this is a cross-domain process, and mainly means that a feature extractor is trained by using data of the source domain, feature extraction of a target domain sample is achieved, and finally an activity classification task of the target domain is achieved.
Example 1:
the implementation provides an activity classification model construction method based on Wi-Fi signals and transfer learning, wherein the classification model constructed by the method is used for carrying out activity classification by applying Wi-Fi signals, and comprises three parts which are respectively: the system comprises a preprocessing part, a feature extraction model training part, an activity feature extraction part and a target domain activity classification part, wherein a specific model framework diagram is shown in figure 1.
The method comprises the following steps:
step 1, respectively acquiring CSI samples corresponding to each user activity sample in a source domain and a target domain, and preprocessing the CSI samples corresponding to each user activity sample to obtain a CSI image corresponding to each user activity sample;
The establishing process of the preprocessing part in the step 1 comprises the following steps:
step 1.1, performing Butterworth filtering on original CSI amplitude data to obtain CSI amplitude after high-frequency noise is removed; in this embodiment, the cut-off frequency of the filter used is 0.9425Hz. The result of CSI amplitude preprocessing is shown in fig. 3.
Step 1.2, further noise reduction treatment is carried out on the pre-processed CSI amplitude data, and common methods include principal component analysis (Principal Component Analysis, PCA), singular value decomposition and the like; in this embodiment, singular value decomposition is adopted, and after the original CSI amplitude is subjected to singular value decomposition, the weight corresponding to the first principal component is set to 0.
Step 1.3, carrying out phase correction on the original CSI phase to obtain the CSI phase after randomness and uncertainty are eliminated;
step 1.4, further denoising the preprocessed CSI phase data, wherein common methods include PCA, singular value decomposition and the like; in this embodiment, singular value decomposition is adopted, and after the original CSI phase is subjected to singular value decomposition, the weight corresponding to the first principal component is set to 0.
And step 1.5, organizing the CSI amplitude and the CSI phase obtained in the steps 1.2 and 1.4 into image data, splicing, and adjusting the size of the image after splicing. The data organization method which can be generally adopted comprises the steps of overlapping the phase data and the amplitude data, and splicing the amplitude data and the phase data left and right. In this embodiment, the active CSI amplitude data and the phase data are spliced right and left.
Step 1.6, the size of the movable CSI image obtained in the step 1.5 is adjusted, so that the subsequent training process is facilitated; in order to keep consistent with the input of the feature extraction network, the active CSI image size is adjusted to a size of 90×180 in this embodiment; if the input to the feature extraction network changes, the image size may be adjusted accordingly.
Consider that the method solves the problem of activity recognition based on application of Wi-Fi signals. The activity data sample based on Wi-Fi signals is small, and a deep neural network with a huge parameter amount is difficult to train. Therefore, in the method, a migration learning method is adopted, a VGG16 network model pre-trained on tens of thousands of natural image data sets ImageNet is finely tuned, and trainable parameters of an active feature extraction model are reduced.
Step 2, an active feature extraction model is established by adopting a transfer learning technology, simultaneously, the CSI images corresponding to all the user active samples in the source domain obtained in the step 1 are established into a triplet CSI image sample pair by adopting a triplet data pair structure, and the triplet CSI image sample pair is used as input to train the active feature extraction model, so that a trained user active feature extraction model is obtained; then, the CSI images corresponding to all the user activity samples in the target domain obtained in the step 1 are constructed into triplet CSI image sample pairs, and the triplet CSI image sample pairs are sent into the user activity feature extraction model to perform feature coding, so that feature vectors corresponding to each user activity sample in the target domain are obtained;
The establishing process of the feature extraction part in the step 2 comprises the following steps:
and 2.1, forming three groups of CSI images processed by the pretreatment part of the activity classification model into a pair to construct a triplet CSI image sample pair. In this embodiment, the data format adopted is
Figure BDA0002400295970000111
Wherein->
Figure BDA0002400295970000121
Representing anchor sample, ++>
Figure BDA0002400295970000122
Representation and->
Figure BDA0002400295970000123
Activity samples belonging to the same category, +.>
Figure BDA0002400295970000124
Representation and->
Figure BDA0002400295970000125
Active samples that do not belong to the same category.
Step 2.2, designing a deep active feature extraction convolutional network (AFEN) with input as triplet CSI image sample pairs.
First, a deep convolutional network AFEN is built for extracting environmental deployment-independent activity features, which is based on VGG16 pre-trained on ImageNet datasetImprovement. As shown in FIG. 4, the AFEN layer 1 output data format is
Figure BDA0002400295970000126
Wherein->
Figure BDA0002400295970000127
Representing anchor activity samples,/->
Figure BDA0002400295970000128
Representing activity samples like anchor activity samples, < >>
Figure BDA0002400295970000129
Representing activity samples of a different class than the anchor activity sample. Because the AFEN network can accept one triplet data, the 21 st layer of the AFEN network can obtain 3 active feature vectors corresponding to the triplet data samples;
a deep convolutional network (AFEN) is constructed for extracting active features. The network structure includes 21 layers in total, as shown in table 1: the layer 1 is an input layer, and the network input is the active CSI image with the size of 90-180 obtained after the adjustment in the step 1.6; the 2 nd layer and the 3 rd layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 64; the 4 th layer is a pooling layer, and the size of the pooling core is 2x2; the 5 th layer and the 6 th layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 128; the 7 th layer is a pooling layer, and the size of a pooling core is 2x2; the 8 th, 9 th and 10 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 256; the 11 th layer is a pooling layer, and the size of a pooling core is 2x2; the 12 th, 13 th and 14 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 512; the 15 th layer is a pooling layer, and the size of a pooling core is 2x2; the 16 th, 17 th and 18 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 1024; the 19 th layer is a pooling layer, and the size of a pooling core is 2x2; layer 20 is the global pooling layer; layer 21 is the feature stitching layer.
TABLE 1 structural composition of AFEN network
Figure BDA00024002959700001210
Figure BDA0002400295970000131
The 2 nd to 19 th layers of the AFEN network are used for extracting active features of the input active CSI image, the 20 th layer is used for carrying out global pooling and splicing fusion on active features of different layers and finally outputting active feature vectors extracted by the network. The network parameters of layers 2-11 in the network come from the VGG16 network model trained on ImageNet.
And 2.3, training the AFEN obtained in the step 2.2 by adopting a triplet CSI image sample pair and characteristic coding distances of three CSI image sample pairs in the sample pair to obtain the AFEN after parameter adjustment.
Firstly, setting the layer 1 of the deep convolutional network AFEN as three inputs, so that the AFEN network can accept the three inputs; the regulated AFEN network can accept three inputs at the same time, so that the 21 st layer of the AFEN network can obtain three active feature vectors corresponding to the inputs.
Then, the loss function of the AFEN network during training is the absolute value square error of the feature vectors of the three input active CSI images, and the trainable parameters in the AFEN network are optimized by a random gradient descent method. After training is completed, only the parameters of the AFEN network need to be saved, and the network can be used for extracting the characteristics of the active CSI image. In this embodiment, a batch training mode is adopted, the batch size adopted in the training process is 10, the iteration number epoch is 100, the selected optimizer is Adam, and the learning rate is 0.0001.
The process of activity feature extraction in step 2 includes:
and 2.4, as shown in fig. 5, parameters of the 2 nd to 19 th layers of the AFEN network trained in the step 2.3 are kept unchanged, and at the moment, the output of the 21 st layer of the network is the feature vector corresponding to the input of 3 active CSI images. And organizing the active CSI image of the target domain obtained in the step 1 into a triplet CSI image sample pair.In this embodiment, the data format adopted is< i ,j,k>Wherein x is i ，x j ，x k Representing any three active data samples in the target domain, respectively.
And 2.5, performing feature extraction on the triplet CSI image sample pairs of the target domain obtained in the step 2.1 by adopting the AFEN network with the fixed parameters obtained in the step 2.4 to obtain feature vectors corresponding to each movable CSI image in the target domain. In this embodiment, when the number of active samples in the target domain is not a multiple of 3, the last sample is duplicated, and the number of samples is complemented to be a multiple of 3, and the last sample is duplicated at most 2 times. After feature encoding, discarding feature vectors corresponding to the redundant samples.
And step 3, establishing a user activity classification model, and training by utilizing the feature vector obtained in the step 2 to obtain a trained user activity classification model for realizing user activity classification.
The process for establishing the target domain activity classification model comprises the following steps:
and 3.1, designing model parameters of a support vector machine (Support Vector Machine, SVM) to obtain an activity classification model of the target domain activity classification. In this embodiment, the penalty coefficient is 10, the kernel function is a linear kernel function, and the kernel function coefficient is 0.01.
And 3.2, selecting a small number of labeled active samples from the coded active samples obtained in the step 2.5 as a training set, and taking the rest as a test set. And obtaining a training set and a testing set of the target domain. In this embodiment, the number of labeled data samples selected by each class in the target domain is 3.
And 3.3, training the activity classification model by adopting the training set divided in the step 3.2 to obtain the activity classification model of the target domain.
And 3.4, testing the test set by using the trained activity classification model to obtain an activity class corresponding to each sample in the target domain test data set, and finally realizing a cross-domain activity classification task with low cost based on few samples.
And (3) effect verification:
the embodiment compares the effect of activity classification on the target domain by 2 different methods through experiments, and the specific method is as follows:
And (3) SVM: firstly, amplitude and phase preprocessing is carried out on the active CSI data of the target domain, then the CSI amplitude and phase data are organized into an image format, and the preprocessing flow is shown in figure 2. And then, training an SVM classification model by using few label samples in a target domain in an end-to-end mode, and using the trained SVM classification model to realize classification on the active CSI image to be classified.
AFEN: firstly, amplitude and phase preprocessing is carried out on the active CSI data in the target domain, and then the CSI amplitude and phase data are organized into an image format. The active samples in the source domain are organized into a format of triplet data sample pairs, training the feature extraction model part proposed in the present invention. Next, all active data samples in the target domain will be encoded using the active feature extraction model. And finally, training an SVM classification model by adopting the few label samples in the target domain after coding, and for the activity samples to be classified in the target domain, using the trained SVM classification model to realize activity classification, so as to obtain the activity types of the samples to be classified.
The experimental results of this embodiment are shown in the following table, and the evaluation index of the experimental results is the classification accuracy, the value range of the evaluation index is [0,1], and the higher the value is, the more accurate the representing method is.
TABLE 2 comparison of effects between different methods
Method Accuracy rate of
SVM 0.308108108
AFEN 0.875675676
As can be seen from Table 2, the method for identifying by using Wi-Fi signals provided by the invention can obtain a better result under the cross-domain condition. The problem of application failure of the activity classification model due to environmental changes is alleviated.
The previous description of the embodiments is provided to facilitate a person of ordinary skill in the art in order to make and use the present invention. It will be apparent to those skilled in the art that various modifications can be readily made to these embodiments and the generic principles described herein may be applied to other embodiments without the use of the inventive faculty. Therefore, the present invention is not limited to the above-described embodiments, and those skilled in the art, based on the present disclosure, should make improvements and modifications without departing from the scope of the present invention.

Claims (8)

1. A method for constructing a user activity classification model based on Wi-Fi signals and transfer learning is characterized in that the model constructed by the method is used for cross-domain user activity classification from a source domain to a target domain in a Wi-Fi signal environment,
step 1, respectively acquiring CSI samples corresponding to each user activity sample in a source domain and a target domain, and obtaining a CSI image corresponding to each user activity sample through the CSI sample corresponding to each user activity sample;
Step 2, an active feature extraction model based on a VGG16 network is established by adopting a transfer learning technology, simultaneously, the CSI images corresponding to all the user active samples in the source domain obtained in the step 1 are established into a triplet CSI image sample pair by adopting a triplet data pair structure, and the triplet CSI image sample pair is used as input to train the active feature extraction model; then, the CSI images corresponding to all the user activity samples in the target domain obtained in the step 1 are constructed into a triplet CSI image sample pair, and the triplet CSI image sample pair is sent into the trained activity feature extraction model to perform feature coding, so that feature vectors corresponding to each user activity sample in the target domain are obtained;
establishing and training an active feature extraction model, wherein the active feature extraction model is an improved deep active feature extraction convolutional network AFEN, and specifically comprises the following steps:
firstly, establishing an active feature extraction model by adopting a transfer learning technology, and constructing a triplet CSI image sample pair by grouping three CSI images into a pair;
secondly, constructing a deep active feature extraction convolutional network AFEN which is input as a triplet CSI image sample pair;
constructing a deep active feature extraction convolutional network input as triplet CSI image sample pairs, comprising constructing a deep convolutional network for extracting environmental deployment independent active features, the network being improved based on VGG16 pre-trained on ImageNet dataset: layer 1 output data format of AFEN
Figure FDA0004096983700000011
Wherein->
Figure FDA0004096983700000012
Representing anchor activity samples,/->
Figure FDA0004096983700000013
Representing activity samples like anchor activity samples, < >>
Figure FDA0004096983700000014
Representing activity samples of different classes from the anchor activity sample;
the network receives one triplet data, and then the 21 st layer of the network can obtain 3 active feature vectors corresponding to the triplet data;
the specific structure of the network is as follows: the network comprises 21 layers, wherein the 1 st layer is an input layer, namely an input triplet CSI image sample pair; the 2 nd layer and the 3 rd layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 64; the 4 th layer is a pooling layer, and the size of the pooling core is 2x2; the 5 th layer and the 6 th layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 128; the 7 th layer is a pooling layer, and the size of a pooling core is 2x2; the 8 th, 9 th and 10 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 256; the 11 th layer is a pooling layer, and the size of a pooling core is 2x2; the 12 th, 13 th and 14 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 512; the 15 th layer is a pooling layer, and the size of a pooling core is 2x2; the 16 th, 17 th and 18 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 1024; the 19 th layer is a pooling layer, and the size of a pooling core is 2x2; layer 20 is the global pooling layer; layer 21 is the output layer;
Finally, training the AFEN by adopting a triplet CSI image sample pair and the feature coding distances of three CSI image samples in the sample pair to obtain the trained AFEN;
and 3, establishing a linear SVM user activity classification model, and training by utilizing the feature vector obtained in the step 2 to obtain a trained user activity classification model.
2. The method for constructing a classification model of user activity based on Wi-Fi signals and transfer learning as claimed in claim 1, wherein obtaining CSI images corresponding to each user activity from CSI samples corresponding to each user activity sample means: and preprocessing the CSI samples corresponding to each user activity acquired in the source domain and the target domain by sequentially adopting data preprocessing methods of Butterworth filtering, singular value decomposition, phase correction, amplitude phase splicing and data normalization, and finally obtaining the active CSI images corresponding to each user activity in the source domain and the target domain respectively.
3. The method for constructing the user activity classification model based on Wi-Fi signals and transfer learning according to claim 2, which is characterized by comprising the following steps:
step 1.1, performing Butterworth filtering on original CSI amplitude data in the CSI samples corresponding to each user active sample to obtain CSI amplitudes after high-frequency noise is removed;
Step 1.2, singular value decomposition is carried out on the CSI amplitude data preprocessed in the step 1.1;
step 1.3, carrying out phase correction on the original CSI phase in the CSI sample corresponding to each user active sample to obtain the CSI phase after randomness and uncertainty are eliminated;
step 1.4, singular value decomposition is carried out on the CSI phase data preprocessed in the step 1.3;
and step 1.5, organizing the CSI amplitude and the CSI phase obtained in the step 1.2 and the step 1.4 after singular value decomposition into an image data format and splicing left and right to obtain a movable CSI image.
4. The method for constructing the user activity classification model based on Wi-Fi signals and transfer learning as claimed in claim 2, wherein in step 1.5, the left and right concatenation is performed by adopting a data format of the following formula:
Figure FDA0004096983700000021
wherein G is sample Representing a current active sample of the sample,
Figure FDA0004096983700000031
CSI amplitude representing the current active sample after preprocessing, +.>
Figure FDA0004096983700000032
Indicating the CSI phase of the corrected current active sample, t indicating the duration of the user activity, P rate Representing the data sampling frequency of a Wi-Fi receiving end;
after splicing, obtaining a CSI feature matrix, normalizing floating point data in the CSI feature matrix to a [0,255] interval, copying three layers, and organizing the three layers into a three-channel image data format, namely, a movable CSI image.
5. The method for constructing the user activity classification model based on Wi-Fi signals and transfer learning according to claim 1, wherein the method is characterized in that the method comprises the steps of obtaining the trained AFEN by adopting a triplet CSI image sample pair and feature coding distance training of three CSI image samples in the sample pair, and specifically comprises the following steps: the method of transfer learning is adopted, an AFEN network is initialized by pre-trained VGG16 network parameters in the training process, then network parameters of layers 2-11 of the AFEN network are fixed, the rest network parameters are trained, the loss function of the AFEN network in training is the absolute value square error between active feature codes between a triplet active sample pair, and the adjustable parameters in the AFEN network are optimized by a random gradient descent method in the training process; after training is completed, the AFEN network parameters are saved, and the trained AFEN network is obtained;
the feature vector extraction section includes: organizing the movable CSI image of the target domain obtained in the step 1 into a triplet CSI image sample pair; parameters of layers 2-20 of the trained AFEN network are kept unchanged, and layer 21 inputs the output active CSI images into corresponding feature vectors, so that feature extraction is carried out on the triplet CSI image sample pairs by utilizing the trained AFEN network, and the feature vectors of each active CSI image are obtained.
6. A user activity classification method based on Wi-Fi signals and transfer learning, which is characterized in that the cross-domain user activity data to be classified is input into a user activity classification model constructed by the user activity classification model construction method according to any one of claims 1 to 5, and a user activity classification result is obtained.
7. The device is characterized by comprising a preprocessing module, a feature extraction model training and user activity feature extraction module and a target domain activity classification module;
the preprocessing module is used for preprocessing CSI sample data corresponding to each user activity acquired in the source domain and the target domain to obtain an active CSI image corresponding to each user activity in the source domain and the target domain;
the feature extraction model training and user activity feature extraction module adopts a migration learning technology to establish an activity feature extraction model based on a VGG16 network, adopts a triple data pair structure to construct the CSI images corresponding to all user activity samples in a source domain obtained by the preprocessing module into a triple CSI image sample pair, takes the triple CSI image sample pair as input, and trains the activity feature extraction model; then, the CSI images corresponding to all the user activity samples in the target domain obtained by the preprocessing module are constructed into a triplet CSI image sample pair, and the triplet CSI image sample pair is sent into the trained activity feature extraction model to perform feature coding, so that feature vectors corresponding to each user activity sample in the target domain are obtained;
Establishing and training an active feature extraction model, wherein the active feature extraction model is an improved deep active feature extraction convolutional network AFEN, and specifically comprises the following steps:
firstly, establishing an active feature extraction model by adopting a transfer learning technology, and constructing a triplet CSI image sample pair by grouping three CSI images into a pair;
secondly, constructing a deep active feature extraction convolutional network AFEN which is input as a triplet CSI image sample pair;
constructing a deep active feature extraction convolutional network input as triplet CSI image sample pairs, comprising constructing a deep convolutional network for extracting environmental deployment independent active features, the network being improved based on VGG16 pre-trained on ImageNet dataset: layer 1 output data format of AFEN
Figure FDA0004096983700000041
Wherein->
Figure FDA0004096983700000042
Representing anchor activity samples,/->
Figure FDA0004096983700000043
Representing activity samples like anchor activity samples, < >>
Figure FDA0004096983700000044
Representing activity samples of different classes from the anchor activity sample;
the network receives one triplet data, and then the 21 st layer of the network can obtain 3 active feature vectors corresponding to the triplet data;
the specific structure of the network is as follows: the network comprises 21 layers, wherein the 1 st layer is an input layer, namely an input triplet CSI image sample pair; the 2 nd layer and the 3 rd layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 64; the 4 th layer is a pooling layer, and the size of the pooling core is 2x2; the 5 th layer and the 6 th layer are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 128; the 7 th layer is a pooling layer, and the size of a pooling core is 2x2; the 8 th, 9 th and 10 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 256; the 11 th layer is a pooling layer, and the size of a pooling core is 2x2; the 12 th, 13 th and 14 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 512; the 15 th layer is a pooling layer, and the size of a pooling core is 2x2; the 16 th, 17 th and 18 th layers are convolution layers, the convolution kernel size is 3x3, and the number of the convolution kernels is 1024; the 19 th layer is a pooling layer, and the size of a pooling core is 2x2; layer 20 is the global pooling layer; layer 21 is the output layer;
Finally, training the AFEN by adopting a triplet CSI image sample pair and the feature coding distances of three CSI image samples in the sample pair to obtain the trained AFEN;
the target domain activity classification module is used for establishing a linear SVM user activity classification model and training by utilizing the activity feature vectors obtained by the user activity feature extraction module to obtain a trained user activity classification model.
8. A Wi-Fi signal and transfer learning based user activity classification system, comprising an image input device and a Wi-Fi signal and transfer learning based user activity classification model construction device according to claim 7;
the image input module is used for inputting CSI images corresponding to user classification activities of a source domain and a target domain to be classified;
the Wi-Fi signal and transfer learning-based user activity classification model construction device is used for classifying cross-domain user activities by using the constructed user activity classification model to obtain a user activity classification result.
CN202010144596.8A 2020-03-04 2020-03-04 Wi-Fi signal and transfer learning-based activity classification model construction method and system Active CN111460901B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010144596.8A CN111460901B (en) 2020-03-04 2020-03-04 Wi-Fi signal and transfer learning-based activity classification model construction method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010144596.8A CN111460901B (en) 2020-03-04 2020-03-04 Wi-Fi signal and transfer learning-based activity classification model construction method and system

Publications (2)

Publication Number Publication Date
CN111460901A CN111460901A (en) 2020-07-28
CN111460901B true CN111460901B (en) 2023-05-26

Family

ID=71684174

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010144596.8A Active CN111460901B (en) 2020-03-04 2020-03-04 Wi-Fi signal and transfer learning-based activity classification model construction method and system

Country Status (1)

Country Link
CN (1) CN111460901B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113920577A (en) * 2021-08-06 2022-01-11 南京邮电大学 Indoor non-contact human body activity identification method and system
CN114444293B (en) * 2022-01-21 2023-06-23 南京森林警察学院 Interference-like signal phase decomposition method based on transfer learning

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108062753A (en) * 2017-12-29 2018-05-22 重庆理工大学 The adaptive brain tumor semantic segmentation method in unsupervised domain based on depth confrontation study
CN108629380A (en) * 2018-05-11 2018-10-09 西北大学 A kind of across scene wireless signal cognitive method based on transfer learning
CN110555467A (en) * 2019-08-13 2019-12-10 深圳创新奇智科技有限公司 industrial data classification method based on model migration

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108932480B (en) * 2018-06-08 2022-03-15 电子科技大学 Distributed optical fiber sensing signal feature learning and classifying method based on 1D-CNN

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108062753A (en) * 2017-12-29 2018-05-22 重庆理工大学 The adaptive brain tumor semantic segmentation method in unsupervised domain based on depth confrontation study
CN108629380A (en) * 2018-05-11 2018-10-09 西北大学 A kind of across scene wireless signal cognitive method based on transfer learning
CN110555467A (en) * 2019-08-13 2019-12-10 深圳创新奇智科技有限公司 industrial data classification method based on model migration

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
基于迁移学习的全连接神经网络舌象分类方法;杨晶东等;《第二军医大学学报》(第08期);全文 *

Also Published As

Publication number Publication date
CN111460901A (en) 2020-07-28

Similar Documents

Publication Publication Date Title
CN107633513B (en) 3D image quality measuring method based on deep learning
CN108921051B (en) Pedestrian attribute identification network and technology based on cyclic neural network attention model
CN109447099B (en) PCA (principal component analysis) dimension reduction-based multi-classifier fusion method
CN108960299B (en) Method for identifying multi-class motor imagery electroencephalogram signals
US11263435B2 (en) Method for recognizing face from monitoring video data
CN106548159A (en) Reticulate pattern facial image recognition method and device based on full convolutional neural networks
CN107967458A (en) A kind of face identification method
CN111460901B (en) Wi-Fi signal and transfer learning-based activity classification model construction method and system
CN106096535A (en) A kind of face verification method based on bilinearity associating CNN
CN109446991A (en) Gait recognition method based on global and local Fusion Features
CN102024145A (en) Layered recognition method and system for disguised face
CN106156767A (en) Driving license effect duration extraction method, server and terminal
CN110717481A (en) Method for realizing face detection by using cascaded convolutional neural network
CN110222592A (en) A kind of construction method of the timing behavioral value network model generated based on complementary timing behavior motion
CN113095370A (en) Image recognition method and device, electronic equipment and storage medium
CN110062379A (en) Identity identifying method based on channel state information under a kind of human body behavior scene
CN110730473A (en) Signal feature extraction method for WiFi activity recognition
CN112801054A (en) Face recognition model processing method, face recognition method and device
CN104978569A (en) Sparse representation based incremental face recognition method
CN112395941A (en) Radio signal radiation source individual identification method, computer device and storage medium
CN111242028A (en) Remote sensing image ground object segmentation method based on U-Net
CN106295574A (en) Face characteristic based on neutral net extracts modeling, face identification method and device
Jeny et al. PassNet-Country Identification by Classifying Passport Cover Using Deep Convolutional Neural Networks
Sasipriyaa et al. SFGDO: Smart flower gradient descent optimization enabled generative adversarial network for recognition of Tamil handwritten character
CN111859000A (en) Method for constructing and updating human face feature database under deep learning model

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant