CN113449259A - Channel state information feature extraction method and system based on deep learning - Google Patents

Channel state information feature extraction method and system based on deep learning Download PDF

Info

Publication number
CN113449259A
CN113449259A CN202110761547.3A CN202110761547A CN113449259A CN 113449259 A CN113449259 A CN 113449259A CN 202110761547 A CN202110761547 A CN 202110761547A CN 113449259 A CN113449259 A CN 113449259A
Authority
CN
China
Prior art keywords
matrix
channel state
state information
determining
amplitude
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.)
Pending
Application number
CN202110761547.3A
Other languages
Chinese (zh)
Inventor
杜留锋
宋长源
杨文强
王占奎
田熙燕
秦国庆
郭新
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.)
Henan Institute of Science and Technology
Original Assignee
Henan Institute of Science and Technology
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 Henan Institute of Science and Technology filed Critical Henan Institute of Science and Technology
Priority to CN202110761547.3A priority Critical patent/CN113449259A/en
Publication of CN113449259A publication Critical patent/CN113449259A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Molecular Biology (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention relates to a channel state information feature extraction method and system based on deep learning. The method comprises the steps of carrying out filtering processing on channel state information; converting the channel state information after filtering into amplitude matrixes of each spatial stream according to time sequence, and combining the amplitude matrixes of each spatial stream into a third-order tensor; obtaining a structured information matrix and a fluctuation data matrix according to the amplitude matrix in the third-order tensor, and determining a gradient matrix related to time diversity and a gradient matrix related to frequency diversity; determining model driving data of four orders according to the structured information matrix, the fluctuation data matrix and the two gradient matrices; according to the fourth-order model driving data, a trained channel state information feature extraction depth model is determined based on a 3D convolutional neural network and a gated cyclic unit network framework, the mapping relation between features and target situations can be accurately extracted, and the overall performance of the context awareness technology is improved.

Description

Channel state information feature extraction method and system based on deep learning
Technical Field
The invention relates to the field of wireless signal processing, in particular to a channel state information feature extraction method and system based on deep learning.
Background
The popularity of wireless terminals and the development of mobile computing technology have led to wireless local area networks based on Wi-Fi technology reaching deep into the spaces of human activities. The Wi-Fi signal transmission process is influenced by physical space change, and related parameters of signals and links are caused to change in real time, so that the Wi-Fi signal transmission process can be used for sensing scene information via paths while transmitting data. As a physical layer parameter of a Wi-Fi link, data representation of Channel State Information (CSI) has both frequency and space diversity of the MIMO-OFDM technology, and the signal passing change is more finely described. The received CSI is analyzed and processed by a specific means to acquire the contextual information of an interest target in a transmission space, so that the Wi-Fi sensing technology can be applied more deeply in the fields of smart home, medical treatment, man-machine interaction and the like.
In a complex indoor environment, the CSI can suffer from serious multipath effect, and the data has vivid non-stationary random statistical characteristics. Although the real-time situation of the perceived target is clearly mapped into the CSI, the characteristic difference caused by the small amplitude change of the target is often easily submerged in the overall data fluctuation of the multipath effect. Traditional methods, such as principal component analysis, clustering algorithm, shallow neural network and the like, are difficult to accurately capture the information which contributes little to the data pattern, so that the perception performance is improved and the bottleneck is encountered. In recent years, as a leading branch of the field of artificial intelligence, deep learning has been successfully applied in a plurality of technical fields. Deep learning spontaneously extracts internal rules and advanced features from sample data through a multi-level network structure, and has better feature learning and nonlinear mapping capability. In view of this, some successful depth models are also being used to achieve situational annotation of perceptual targets. However, due to the complexity of the spatial environment, the dynamics of the target, multipath fading, and other factors, the characterizations contained in the CSI are very sensitive, and although the existing deep learning-based scheme has a breakthrough in the sensing accuracy, the stability and the generalization thereof still need to be enhanced. In this context, a reliable CSI data feature extraction method is particularly important.
The 'channel state information-based active person number estimation method' (application number: 201710195058X, publication number: CN107154088B, publication date: 2019.03.26) applied by the university of Western electronic technology removes outliers based on the mean and variance of original CSI data, and forms a feature vector of the number of associated persons by using the data variance of all subcarriers of CSI; in a patent of Haerbin engineering university applied for 'speed self-adaptive indoor human body detection method based on subcarrier dynamic selection' (application number: 2018110868979, application publication number: CN109409216A, published date: 2019.03.01), CSI data is divided into designated lengths along an acquisition time axis, and expectation, standard deviation, average energy and maximum amplitude difference values of subcarriers in the lengths are fused into detection characteristics to realize dynamic and static target detection. The feature extraction method based on the traditional linear statistical strategy is poor in mapping accuracy of extracted features and target situations due to insufficient mining of CSI frequency and space diversity information.
A Minh Tu Hoang et al (A CNN-LSTM quantum for Single Access Point CSI Indor Localization, arXiv:200506394, 2020) proposes a combined deep learning method to learn the associated position characteristics in CSI for the target position perception problem. The method utilizes 2D CNN to search the space information of CSI, and a cascaded long-short term memory network LSTM is used for calibrating the time attribute of space characteristics. However, the method does not effectively process the original CSI data, and the extracted features and the target position lack stable matching degree, so that the robustness of the model and the execution effect of the whole scheme are influenced.
Therefore, a new extraction method is needed to solve the above problems.
Disclosure of Invention
The invention aims to provide a channel state information feature extraction method and system based on deep learning, which can accurately extract the mapping relation between features and target situations, and further improve the overall performance of a context awareness technology.
In order to achieve the purpose, the invention provides the following scheme:
a channel state information feature extraction method based on deep learning comprises the following steps:
acquiring channel state information in a perception scene, and filtering the channel state information;
converting the channel state information after filtering into amplitude matrixes of each spatial stream according to time sequence, and combining the amplitude matrixes of each spatial stream into a third-order tensor;
performing low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix;
determining a gradient matrix related to time diversity and a gradient matrix related to frequency diversity according to the amplitude matrix in the third-order tensor;
determining fourth-order model driving data according to the structural information matrix, the fluctuation data matrix, the gradient matrix related to time diversity and the gradient matrix related to frequency diversity;
determining a trained channel state information feature extraction depth model based on a 3D convolutional neural network and a gated cyclic unit network framework according to the model driving data of the fourth order; and the channel state information feature extraction depth model takes the fourth-order model driving data as input and takes the channel state information feature of the perception target situation as output.
Optionally, the converting the channel state information after the filtering process into amplitude matrices of each spatial stream according to a time sequence, and combining the amplitude matrices of each spatial stream into a third-order tensor specifically includes:
using formulas
Figure BDA0003149244840000031
Determining N magnitude vectors for any one spatial stream
Figure BDA0003149244840000032
Determining a magnitude matrix of the corresponding spatial stream in time sequence;
using formulas
Figure BDA0003149244840000033
Determining a third-order tensor of amplitude matrix combination of the nine spatial streams;
wherein the content of the first and second substances,
Figure BDA0003149244840000034
for each spatial stream amplitude vector, ht,rIs the complex vector of the spatial stream from the t-th transmitting antenna to the r-th receiving antenna in the channel state information after filtering processing,
Figure BDA0003149244840000038
the number of the dot product is the number of dots,
Figure BDA0003149244840000039
is a complex vector ht,rConjugation of (a) HiIs the amplitude matrix of the ith spatial stream, K is the total number of carriers of the spatial stream, and K is the index of the subchannel.
Optionally, the performing low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix specifically includes:
to be provided with
Figure BDA0003149244840000035
Taking s.t. I + F as H as an objective function; the sum of the I, the F,
Figure BDA0003149244840000036
determining a Lagrangian function as a constraint condition;
the Lagrangian function includes:
Figure BDA0003149244840000037
solving the Lagrangian function by using an alternating direction multiplier method to determine a structured information matrix and a fluctuation data matrix;
wherein I is structured information, F is fluctuation data, H is an amplitude matrix in a third-order tensor, | | | | | survival*Is the kernel norm, | | | | luminanceFIs the Frobenius norm,
Figure BDA0003149244840000041
is composed of
Figure BDA0003149244840000045
Norm, the purpose is sparse column vector, V is multiplier matrix, rho is penalty factor, and lambda is hyper-parameter for regulating fluctuation ratio.
Optionally, the determining a gradient matrix for time diversity and a gradient matrix for frequency diversity according to the amplitude matrix in the third order tensor specifically includes:
using formulas
Figure BDA0003149244840000042
Determining a gradient matrix with respect to time diversity;
using formulas
Figure BDA0003149244840000043
Determining a gradient matrix for frequency diversity;
wherein HnIs a gradient matrix with respect to time diversity, HkIs a gradient matrix with respect to frequency diversity.
Optionally, the determining the fourth-order model driving data according to the structural information matrix, the fluctuation data matrix, the gradient matrix related to time diversity, and the gradient matrix related to frequency diversity specifically includes:
using formulas
Figure BDA0003149244840000044
Determining model driving data of fourth order;
wherein [ H ]n；[0]1×K]Is HnBottom-added K-dimensional all-0 row vector, [ H ]k,[0]N×1]Is HkThe right end is supplemented with an N-dimensional all 0-column vector.
A deep learning based channel state information feature extraction system, comprising:
the information processing module is used for acquiring channel state information in a perception scene and filtering the channel state information;
the third-order tensor determining module is used for converting the channel state information after the filtering processing into amplitude matrixes of each spatial stream according to time sequence and combining the amplitude matrixes of each spatial stream into a third-order tensor;
the low-rank matrix decomposition module is used for performing low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix;
the gradient matrix determination module is used for determining a gradient matrix related to time diversity and a gradient matrix related to frequency diversity according to the amplitude matrix in the third-order tensor;
a fourth-order model driving data determination module, configured to determine fourth-order model driving data according to the structured information matrix, the fluctuation data matrix, the gradient matrix related to time diversity, and the gradient matrix related to frequency diversity;
the trained channel state information feature extraction depth model determining module is used for determining a trained channel state information feature extraction depth model based on a 3D convolutional neural network and a gated cyclic unit network framework according to the fourth-order model driving data; and the channel state information feature extraction depth model takes the fourth-order model driving data as input and takes the channel state information feature of the perception target situation as output.
Optionally, the third order tensor determining module specifically includes:
an amplitude matrix determination unit for using the formula
Figure BDA0003149244840000051
Determining N magnitude vectors for any one spatial stream
Figure BDA0003149244840000052
Determining a magnitude matrix of the corresponding spatial stream in time sequence;
a third order tensor determining unit for utilizing the formula
Figure BDA0003149244840000053
Determining a third-order tensor of amplitude matrix combination of the nine spatial streams;
wherein the content of the first and second substances,
Figure BDA0003149244840000054
for each spatial stream amplitude vector, ht,rIs the complex vector of the spatial stream from the t-th transmitting antenna to the r-th receiving antenna in the channel state information after filtering processing,
Figure BDA0003149244840000058
the number of the dot product is the number of dots,
Figure BDA0003149244840000055
is a complex vector ht,rConjugation of (a) HiIs the amplitude matrix of the ith spatial stream, K is the total number of carriers of the spatial stream, and K is the index of the subchannel.
Optionally, the low rank matrix decomposition module specifically includes:
a Lagrangian function determination unit for determining the function of the input signal
Figure BDA0003149244840000056
Taking s.t. I + F as H as an objective function; the sum of the I, the F,
Figure BDA0003149244840000057
determining a Lagrangian function as a constraint condition;
the Lagrangian function includes:
Figure BDA0003149244840000061
the structured information matrix and fluctuation data matrix determining unit is used for solving the Lagrangian function by using an alternating direction multiplier method to determine a structured information matrix and a fluctuation data matrix;
wherein I is structured information, F is fluctuation data, H is an amplitude matrix in a third-order tensor, | | | | | survival*Is the kernel norm, | | | | luminanceFIs the Frobenius norm,
Figure BDA0003149244840000062
is composed of
Figure BDA0003149244840000066
Norm, the purpose is sparse column vector, V is multiplier matrix, rho is penalty factor, and lambda is hyper-parameter for regulating fluctuation ratio.
Optionally, the gradient matrix determining module specifically includes:
gradient matrix determination unit for time diversity using a formula
Figure BDA0003149244840000063
Determining a gradient matrix with respect to time diversity;
gradient matrix determination unit for frequency diversity, for using the formula
Figure BDA0003149244840000064
Determining a gradient matrix for frequency diversity;
wherein HnIs a gradient matrix with respect to time diversity, HkIs a gradient matrix with respect to frequency diversity.
Optionally, the fourth-order model driving data determining module specifically includes:
a fourth order model driving data determination unit for using the formula
Figure BDA0003149244840000065
Determining model driving data of fourth order;
wherein [ H ]n；[0]1×K]Is HnBottom-added K-dimensional all-0 row vector, [ H ]k,[0]N×1]Is HkThe right end is supplemented with an N-dimensional all 0-column vector.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention provides a Channel State Information feature extraction method and system based on deep learning, and provides a model multidimensional driving data method based on the time-space correlation of Channel State Information (CSI) data. The method makes full use of the MIMO-OFDM technology of the Wi-Fi standard, effectively separates stable structural information and scene fluctuation in the CSI, considers time and frequency changes of real-time information, and improves the richness and targeting of the input data materials of the depth model. The adopted deep learning model is subjected to multi-dimensional information extraction by the 3D CNN, and GRUs with fewer network parameters are subjected to nonlinear feature regression. The 3D CNN can extract value data from three dimensions of time, space and frequency, so that the level and precision of characterization learning are improved; the GRU which is good at processing the time sequence can effectively map the time-varying characteristics of the perception target, and the stability and the generalization of the output characteristics are enhanced.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present invention, and it is obvious for those skilled in the art to obtain other drawings without inventive exercise.
Fig. 1 is a schematic flow chart of a deep learning-based channel state information feature extraction method provided in the present invention;
fig. 2 is a schematic general flow chart of a deep learning-based channel state information feature extraction method provided in the present invention;
FIG. 3 is a low-rank analysis diagram of an actual CSI amplitude matrix under different situations;
FIG. 4 is a schematic diagram of a single spatial stream CSI amplitude matrix generation;
FIG. 5 is a flow chart of a CSI amplitude matrix distributed decomposition algorithm;
FIG. 6 is a schematic diagram of a channel state information feature extraction depth model structure;
FIG. 7 is a schematic diagram of an experimental environment layout;
FIG. 8 is a diagram illustrating cumulative distribution of absolute errors of test results of a simulation experiment;
fig. 9 is a schematic structural diagram of a deep learning-based channel state information feature extraction system provided in the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The invention aims to provide a channel state information feature extraction method and system based on deep learning, which can accurately extract the mapping relation between features and target situations, and further improve the overall performance of a context awareness technology.
In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.
The CSI belongs to time sequence data of space propagation, and parameter distribution of close time and adjacent space has strong space-time correlation. Before the depth model executes learning, stable structural information in the CSI and fluctuation data caused by the target situation are effectively separated by means of a targeted decomposition algorithm, and feature extraction of the depth model is facilitated. The invention expands the CSI data into a two-dimensional array according to the carrier and the time sequence, and the time-space correlation of the CSI is reflected as the low-rank property of the matrix. The low rank of the matrix is proved by the singular value decomposition of the CSI amplitude matrix for different situations, see figure 1. Based on the method, firstly, model high-order driving data based on a CSI amplitude matrix is constructed, then, a deep feature learning model is designed, and CSI feature extraction facing to a typical scene perception problem is achieved.
Fig. 1 is a schematic flow chart of a channel state information feature extraction method based on deep learning according to the present invention, fig. 2 is a schematic general flow chart of a channel state information feature extraction method based on deep learning according to the present invention, as shown in fig. 1 and fig. 2, a channel state information feature extraction method based on deep learning according to the present invention includes:
s101, acquiring channel state information in a perception scene, and filtering the channel state information;
as shown in fig. 7, a sensing scene CSI collection facility is configured, and a 3-antenna wireless signal transmitting terminal TX and a signal receiving terminal RX under a Linux operating system and equipped with a 3-antenna wireless network card are arranged in the invention, and the wireless network card needs to modify firmware by using an open source software package Atheros CSI tool.
Firmware modification is carried out on the wireless network card by utilizing an open source software package Atheros CSI tool, CSI data package field acquisition is carried out aiming at specific perception problems, and the Ping speed of the data package is set to be 100/second.
Extracting CSI in the collected data packets, wherein each packet comprises a 3 multiplied by 56 original CSI numerical matrix; and carrying out filtering processing on the extracted data, wherein the filtering processing comprises abnormal value elimination and burst noise filtering in the data.
The raw data for the CSI in each valid received packet may be represented as:
Figure BDA0003149244840000091
c represents the CSI for all subcarriers of 9 spatial streams (3 × 3 TX/RX antenna pairs), where
Figure BDA0003149244840000092
Represents the t-th transmit antenna to the r-th receive antenna spatial stream complex vector,
Figure BDA0003149244840000093
is a center frequency fkK is the total number of carriers of the spatial stream. The invention adopts the wireless network card Atheors 9K as the RX end, and analyzes all the sub-carrier CSI under the 2.4G/20MHz mode, namely K is 56.
The raw data C is filtered. Abnormal acquisition values caused by noise, transmission protocol standards and the like are removed by a Hampel identification method, and burst noise in the original CSI is suppressed by a low-pass filtering method.
S102, converting the channel state information after filtering into amplitude matrixes of each spatial stream according to time sequence, and combining the amplitude matrixes of each spatial stream into a third-order tensor;
s102 specifically comprises the following steps:
using formulas
Figure BDA0003149244840000094
Determining N magnitude vectors for any one spatial stream
Figure BDA0003149244840000095
Determining a magnitude matrix of the corresponding spatial stream in time sequence;
using formulas
Figure BDA0003149244840000096
Determining a third-order tensor of amplitude matrix combination of the nine spatial streams;
wherein the content of the first and second substances,
Figure BDA0003149244840000097
for each spatial stream amplitude vector, ht,rIs the complex vector of the spatial stream from the t-th transmitting antenna to the r-th receiving antenna in the channel state information after filtering processing,
Figure BDA00031492448400000910
the number of the dot product is the number of dots,
Figure BDA0003149244840000098
is a complex vector ht,rConjugation of (a) HiIs the amplitude matrix of the ith spatial stream, K is the total number of carriers of the spatial stream, and K is the index of the subchannel.
The adjacent amplitude matrix is represented as:
Figure BDA0003149244840000099
information coverage and perception considering extracted featuresThe method has the advantages that the real-time performance of the task and the Ping speed of the set data packet are realized, the time sequence length of each CSI amplitude array is set to be 0.5 second, namely N is 50; the adjacent amplitude matrix time span is 0.1 seconds, i.e., Δ N ═ 10. The generation of either spatial stream amplitude matrix is shown in fig. 4.
S103, performing low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix;
s103 specifically comprises the following steps:
to be provided with
Figure BDA0003149244840000101
Taking s.t. I + F as H as an objective function; the sum of the I, the F,
Figure BDA0003149244840000102
determining a Lagrangian function as a constraint condition;
the Lagrangian function includes:
Figure BDA0003149244840000103
solving the Lagrangian function by using an alternating direction multiplier method to determine a structured information matrix and a fluctuation data matrix;
wherein I is structured information, F is fluctuation data, H is an amplitude matrix in a third-order tensor, | | | | | survival*Is the kernel norm, | | | | luminanceFIs the Frobenius norm,
Figure BDA0003149244840000104
is composed of
Figure BDA0003149244840000107
Norm, the purpose is sparse column vector, V is multiplier matrix, rho is penalty factor, and lambda is hyper-parameter for regulating fluctuation ratio. The invention sets the lambda range to [0.2, 0.3 ] for a specific context aware task]。
As shown in fig. 3 and 5, the specific method of updating the parameters is as follows:
1) fixing other parameters, updating F: Fi←Sλ/ρ(H-Ιi-1-ρ-1Vi-1)。
Figure BDA0003149244840000105
For the contraction factor, operator [ x ]]+＝max(x,0)。
2) Fixing other parameters, updating I:
Figure BDA0003149244840000106
is a singular threshold function.
3) Fixing other parameters, and updating multiplier matrix Vi←Vi-1-ρ(H-Fi-Ii-1)。
4) Update penalty coefficient rhoi←[αρi-1,ρmax]-. Alpha > 1 is an amplification factor, operator [ x, y]-＝min(x,y)。
S104, determining a gradient matrix related to time diversity and a gradient matrix related to frequency diversity according to the amplitude matrix in the third-order tensor;
s104 specifically comprises the following steps:
using formulas
Figure BDA0003149244840000111
Determining a gradient matrix with respect to time diversity;
using formulas
Figure BDA0003149244840000112
Determining a gradient matrix for frequency diversity;
wherein HnIs a gradient matrix with respect to time diversity, HkIs a gradient matrix with respect to frequency diversity.
The process of calculating the gradient of the amplitude matrix H with respect to time, frequency diversity is as follows:
Figure BDA0003149244840000113
Figure BDA0003149244840000114
s105, determining fourth-order model driving data according to the structural information matrix, the fluctuation data matrix, the gradient matrix related to time diversity and the gradient matrix related to frequency diversity;
s105 specifically comprises the following steps:
using formulas
Figure BDA0003149244840000115
Determining model driving data of fourth order;
wherein [ H ]n；[0]1×K]Is HnBottom-added K-dimensional all-0 row vector, [ H ]k,[0]N×1]Is HkThe right end is supplemented with an N-dimensional all 0-column vector.
S106, determining a trained channel state information characteristic extraction depth model based on a 3D convolutional neural network and a gating cycle unit network framework according to the fourth-order model driving data; and the channel state information feature extraction depth model takes the fourth-order model driving data as input and takes the channel state information feature of the perception target situation as output.
And (3) constructing a channel state information feature extraction depth model based on a 3D CNN frame by utilizing open source TensorFlow 2.1, wherein the structure is shown as an attached figure 6. In the specific setting, the convolution step lengths (strides) in the frame direction (1 st dimension direction) of the four groups of 3D convolution kernels are all set to be 1, and the convolution step lengths in the height direction and the width direction (2 nd dimension and 3 rd dimension directions) are all set to be 2; the boundary padding is set to "SAME". After the convolution is finished, array dimensionality reduction is carried out, and three dimensions of 2, 3 and 4 are combined into 1 dimension. The module is finally a fully connected layer that performs a random deactivation with probability 0.5 when trained to prevent overfitting. The module is activated using a ReLU linear function for each layer.
The structure of the characteristic regression module is shown in figure 6. The input end of the module is a full connection layer, and data reduction is carried out on the output of the 3D CNN. The number of GRU units is consistent with the batch capacity of the input CSI amplitude matrix, 128 hidden layers are arranged in each unit, and the initial input array of the first unit is set to be 0. The output layer is a fully-connected network, a Softmax function is used for activation, and the network scale M can be set according to specific perception problem requirements.
The depth model for CSI feature extraction executes supervised learning. The cross entropy is adopted as a loss function during training, and is expressed as:
Figure BDA0003149244840000121
wherein the tag data ylabelAccording to specific problem making, B represents the total number of the input matrix batch at one time. If the target quantity detection problem is solved, a designed characteristic regression module output layer is combined, M-bit One-Hot coding is used for formulating training labels, and the position of 1 in the One-Hot code corresponds to the target quantity, such as labels
Figure BDA0003149244840000122
Representing the true number of objects in the scene as M-1.
The drive tensor to be constructed is dependent on the specific perception problem
Figure BDA0003149244840000123
Sending the model into a depth model, and training the model until the convergence of the loss function is stable; and verifying by using a ten-fold intersection method, adjusting super parameters such as a convolution kernel and the like according to results, repeatedly executing the processes, and finally establishing reliable mapping of the target situation and the extracted CSI characteristics.
The effects of the present invention can be illustrated by the following simulation experiments.
1. And simulating CSI data acquisition and experimental setting.
The actual measurement scene layout of the simulation experiment is shown in the attached figure 7; the distributed low-rank matrix decomposition is based on a CVXPY 1.1 module under Python 3.7; the depth model building is based on a TensorFlow 2.1 depth learning framework, and the main operation hardware is double NVIDIARTX2060 GPU.
2. Simulation experiment content and test results.
Experiment 1, personnel number detection, and the layout of the experimental scene is shown in the attached figure 7. In the off-line training stage, CSI under 6 situations of normal activities of no person (0 person) and 1 to 5 persons in the environment is respectively collected; the training process refers to the attached figure 1, and the corresponding labels are 6-bit full-experience One-Hot codes and respectively correspond to 0 to 5 persons. Online evaluation is divided into two modes: 1) the test object of the mode I is a data acquisition participant. The groups were divided into 5 groups of 1-5 persons, each group had 2 minutes of activity, and the results of 50 evaluations were randomly selected, as shown in Table 1. 2) In the mode II, non-trainers form a test group with the scale of 1-5 persons at random, 200 estimation results are randomly selected for evaluation, and the cumulative distribution CDF of absolute errors between the estimation values and the actual values is used as an evaluation standard, and the result is shown in figure 8. And the two modes are evaluated in an off-line mode, namely, CSI of a test scene is collected firstly, and then characteristic extraction and result comparison are performed.
TABLE 1
Figure BDA0003149244840000131
The evaluation results in table 1 show that, based on the feature extraction method provided by the invention, the detection result shows higher estimation accuracy, especially under the situation of relatively few people; the estimation of more people also achieves the recognition rate of about 95 percent. Meanwhile, the method is helpful for improving the generalization of the system, and as shown in fig. 8, the method also presents a high recognition rate for the detection of any person.
Experiment 2, target situation recognition, and the layout of the experimental scene is shown in the attached figure 7. The offline phase collected 4 situational CSI for 3 testers, i.e., fast moving, slow moving, standing still and sitting still, respectively. The corresponding label is 4-bit full-history One-Hot code and respectively corresponds to four situations. The online test object is a data acquisition participant, and the evaluation adopts an offline mode. Each person in each situation was randomly selected 100 times, the estimated result was the average of 3 persons, and the accuracy is shown in table 2.
TABLE 2
Figure BDA0003149244840000141
As can be seen from table 2, the method provided by the present invention also provides a higher recognition rate for four common target situations. For target actions with larger dominance, the average accuracy rate is close to 98%; for the static situation which is difficult to judge, a higher recognition rate is finally obtained based on the proposed feature extraction method.
Fig. 9 is a schematic structural diagram of a deep learning-based channel state information feature extraction system provided by the present invention, and as shown in fig. 9, the deep learning-based channel state information feature extraction system provided by the present invention includes:
an information processing module 901, configured to acquire channel state information in a sensing scene, and perform filtering processing on the channel state information;
a third-order tensor determining module 902, configured to convert the filtered channel state information into amplitude matrices of each spatial stream according to a time sequence, and combine the amplitude matrices of each spatial stream into a third-order tensor;
a low-rank matrix decomposition module 903, configured to perform low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix;
a gradient matrix determining module 904, configured to determine a gradient matrix for time diversity and a gradient matrix for frequency diversity according to the magnitude matrix in the third order tensor;
a fourth-order model driving data determining module 905, configured to determine fourth-order model driving data according to the structured information matrix, the fluctuation data matrix, the gradient matrix related to time diversity, and the gradient matrix related to frequency diversity;
a trained channel state information feature extraction depth model determining module 906, configured to determine a trained channel state information feature extraction depth model based on a 3D convolutional neural network and a gated cyclic unit network framework according to the fourth-order model driving data; and the channel state information feature extraction depth model takes the fourth-order model driving data as input and takes the channel state information feature of the perception target situation as output.
The third order tensor determining module 902 specifically includes:
an amplitude matrix determination unit for using the formula
Figure BDA0003149244840000151
Determining N magnitude vectors for any one spatial stream
Figure BDA0003149244840000152
Determining a magnitude matrix of the corresponding spatial stream in time sequence;
a third order tensor determining unit for utilizing the formula
Figure BDA0003149244840000153
Determining a third-order tensor of amplitude matrix combination of the nine spatial streams;
wherein the content of the first and second substances,
Figure BDA0003149244840000154
for each spatial stream amplitude vector, ht,rIs the complex vector of the spatial stream from the t-th transmitting antenna to the r-th receiving antenna in the channel state information after filtering processing, wherein omicron is the dot product,
Figure BDA0003149244840000155
is a complex vector ht,rConjugation of (a) HiIs the amplitude matrix of the ith spatial stream, K is the total number of carriers of the spatial stream, and K is the index of the subchannel.
The low-rank matrix decomposition module 903 specifically includes:
a Lagrangian function determination unit for determining the function of the input signal
Figure BDA0003149244840000156
Taking s.t. I + F as H as an objective function; the sum of the I, the F,
Figure BDA0003149244840000157
determining a Lagrangian function as a constraint condition;
the Lagrangian function includes:
Figure BDA0003149244840000158
the structured information matrix and fluctuation data matrix determining unit is used for solving the Lagrangian function by using an alternating direction multiplier method to determine a structured information matrix and a fluctuation data matrix;
wherein I is structured information, F is fluctuation data, H is an amplitude matrix in a third-order tensor, | | | | | survival*Is the kernel norm, | | | | luminanceFIs the Frobenius norm,
Figure BDA0003149244840000159
is composed of
Figure BDA00031492448400001510
Norm, the purpose is sparse column vector, V is multiplier matrix, rho is penalty factor, and lambda is hyper-parameter for regulating fluctuation ratio.
The gradient matrix determining module 904 specifically includes:
gradient matrix determination unit for time diversity using a formula
Figure BDA0003149244840000161
Determining a gradient matrix with respect to time diversity;
gradient matrix determination unit for frequency diversity, for using the formula
Figure BDA0003149244840000162
Determining a gradient matrix for frequency diversity;
wherein HnIs a gradient matrix with respect to time diversity, HkIs a gradient matrix with respect to frequency diversity.
The fourth-order model driving data determining module 905 specifically includes:
a fourth order model driving data determination unit for using the formula
Figure BDA0003149244840000163
Determining model driving data of fourth order;
wherein [ H ]n；[0]1×K]Is HnBottom-added K-dimensional all-0 row vector, [ H ]k,[0]N×1]Is HkThe right end is supplemented with an N-dimensional all 0-column vector.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. For the system disclosed by the embodiment, the description is relatively simple because the system corresponds to the method disclosed by the embodiment, and the relevant points can be referred to the method part for description.
The principles and embodiments of the present invention have been described herein using specific examples, which are provided only to help understand the method and the core concept of the present invention; meanwhile, for a person skilled in the art, according to the idea of the present invention, the specific embodiments and the application range may be changed. In view of the above, the present disclosure should not be construed as limiting the invention.

Claims (10)

1. A channel state information feature extraction method based on deep learning is characterized by comprising the following steps:
acquiring channel state information in a perception scene, and filtering the channel state information;
converting the channel state information after filtering into amplitude matrixes of each spatial stream according to time sequence, and combining the amplitude matrixes of each spatial stream into a third-order tensor;
performing low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix;
determining a gradient matrix related to time diversity and a gradient matrix related to frequency diversity according to the amplitude matrix in the third-order tensor;
determining fourth-order model driving data according to the structural information matrix, the fluctuation data matrix, the gradient matrix related to time diversity and the gradient matrix related to frequency diversity;
determining a trained channel state information feature extraction depth model based on a 3D convolutional neural network and a gated cyclic unit network framework according to the model driving data of the fourth order; and the channel state information feature extraction depth model takes the fourth-order model driving data as input and takes the channel state information feature of the perception target situation as output.
2. The method according to claim 1, wherein the step of converting the filtered channel state information into amplitude matrices of each spatial stream in time sequence and combining the amplitude matrices of each spatial stream into a third-order tensor specifically comprises:
using formulas
Figure FDA0003149244830000011
Determining N magnitude vectors for any one spatial stream
Figure FDA0003149244830000012
Determining a magnitude matrix of the corresponding spatial stream in time sequence;
using formulas
Figure FDA0003149244830000013
Determining a third-order tensor of amplitude matrix combination of the nine spatial streams;
wherein the content of the first and second substances,
Figure FDA0003149244830000014
for each spatial stream amplitude vector, ht,rIs the complex vector of the spatial stream from the t-th transmitting antenna to the r-th receiving antenna in the channel state information after filtering processing,
Figure FDA0003149244830000015
the number of the dot product is the number of dots,
Figure FDA0003149244830000016
is a complex vector ht,rConjugation of (a) HiIs the amplitude matrix of the ith spatial stream, K is the total number of carriers of the spatial stream, and K is the index of the subchannel.
3. The method for extracting channel state information features based on deep learning according to claim 2, wherein the performing low rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix specifically comprises:
to be provided with
Figure FDA0003149244830000021
Taking s.t. I + F as H as an objective function;
Figure FDA0003149244830000022
determining a Lagrangian function as a constraint condition;
the Lagrangian function includes:
Figure FDA0003149244830000023
solving the Lagrangian function by using an alternating direction multiplier method to determine a structured information matrix and a fluctuation data matrix;
wherein I is structured information, F is fluctuation data, H is an amplitude matrix in a third-order tensor, | | | | | survival*Is the kernel norm, | | | | luminanceFIs the Frobenius norm,
Figure FDA0003149244830000024
is 12,1Norm, the purpose is sparse column vector, V is multiplier matrix, rho is penalty factor, and lambda is hyper-parameter for regulating fluctuation ratio.
4. The method for extracting channel state information features based on deep learning according to claim 3, wherein the determining a gradient matrix for time diversity and a gradient matrix for frequency diversity according to the magnitude matrix in the third-order tensor specifically includes:
using formulas
Figure FDA0003149244830000025
Determining a gradient matrix with respect to time diversity;
using formulas
Figure FDA0003149244830000026
Determining a gradient matrix for frequency diversity;
wherein HnIs a gradient matrix with respect to time diversity, HkIs a gradient matrix with respect to frequency diversity.
5. The method for extracting channel state information features based on deep learning according to claim 4, wherein the determining of the model driving data of fourth order according to the structured information matrix, the fluctuation data matrix, the gradient matrix with respect to time diversity, and the gradient matrix with respect to frequency diversity specifically comprises:
using formulas
Figure FDA0003149244830000031
Determining model driving data of fourth order;
wherein [ H ]n；[0]1×K]Is HnBottom-added K-dimensional all-0 row vector, [ H ]k,[0]N×1]Is HkThe right end is supplemented with an N-dimensional all 0-column vector.
6. A deep learning-based channel state information feature extraction system, comprising:
the information processing module is used for acquiring channel state information in a perception scene and filtering the channel state information;
the third-order tensor determining module is used for converting the channel state information after the filtering processing into amplitude matrixes of each spatial stream according to time sequence and combining the amplitude matrixes of each spatial stream into a third-order tensor;
the low-rank matrix decomposition module is used for performing low-rank matrix decomposition on the amplitude matrix in the third-order tensor to obtain a structured information matrix and a fluctuation data matrix;
the gradient matrix determination module is used for determining a gradient matrix related to time diversity and a gradient matrix related to frequency diversity according to the amplitude matrix in the third-order tensor;
a fourth-order model driving data determination module, configured to determine fourth-order model driving data according to the structured information matrix, the fluctuation data matrix, the gradient matrix related to time diversity, and the gradient matrix related to frequency diversity;
the trained channel state information feature extraction depth model determining module is used for determining a trained channel state information feature extraction depth model based on a 3D convolutional neural network and a gated cyclic unit network framework according to the fourth-order model driving data; and the channel state information feature extraction depth model takes the fourth-order model driving data as input and takes the channel state information feature of the perception target situation as output.
7. The system according to claim 6, wherein the third order tensor determining module specifically includes:
an amplitude matrix determination unit for using the formula
Figure FDA0003149244830000041
Determining N magnitude vectors for any one spatial stream
Figure FDA0003149244830000042
Determining a magnitude matrix of the corresponding spatial stream in time sequence;
a third order tensor determining unit for utilizing the formula
Figure FDA0003149244830000043
Determining a third-order tensor of amplitude matrix combination of the nine spatial streams;
wherein the content of the first and second substances,
Figure FDA0003149244830000044
for each spatial stream amplitude vector, ht,rIs the complex vector of the spatial stream from the t-th transmitting antenna to the r-th receiving antenna in the channel state information after filtering processing,
Figure FDA0003149244830000045
the number of the dot product is the number of dots,
Figure FDA0003149244830000046
is a complex vector ht,rConjugation of (a) HiIs the amplitude matrix of the ith spatial stream, K is the total number of carriers of the spatial stream, and K is the index of the subchannel.
8. The deep learning-based channel state information feature extraction system according to claim 7, wherein the low-rank matrix decomposition module specifically includes:
a Lagrangian function determination unit for determining the function of the input signal
Figure FDA0003149244830000047
Taking s.t. I + F as H as an objective function;
Figure FDA0003149244830000048
determining a Lagrangian function as a constraint condition;
the Lagrangian function includes:
Figure FDA0003149244830000049
the structured information matrix and fluctuation data matrix determining unit is used for solving the Lagrangian function by using an alternating direction multiplier method to determine a structured information matrix and a fluctuation data matrix;
wherein I is structured information, F is fluctuation data, H is an amplitude matrix in a third-order tensor, | | | | | survival*Is the kernel norm, | | | | luminanceFIs the Frobenius norm,
Figure FDA00031492448300000410
is 12,1Norm, the purpose is sparse column vector, V is multiplier matrix, rho is penalty factor, and lambda is hyper-parameter for regulating fluctuation ratio.
9. The system according to claim 8, wherein the gradient matrix determining module specifically includes:
gradient matrix determination unit for time diversity using a formula
Figure FDA0003149244830000051
Determining a gradient matrix with respect to time diversity;
gradient matrix determination unit for frequency diversity, for using the formula
Figure FDA0003149244830000052
Determining a gradient matrix for frequency diversity;
wherein HnIs a gradient matrix with respect to time diversity, HkIs a gradient matrix with respect to frequency diversity.
10. The system according to claim 9, wherein the fourth-order model-driven data determination module specifically includes:
a fourth order model driving data determination unit for using the formula
Figure FDA0003149244830000053
Determining model driving data of fourth order;
wherein [ H ]n；[0]1×K]Is HnBottom-added K-dimensional all-0 row vector, [ H ]k,[0]N×1]Is HkThe right end is supplemented with an N-dimensional all 0-column vector.
CN202110761547.3A 2021-07-06 2021-07-06 Channel state information feature extraction method and system based on deep learning Pending CN113449259A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110761547.3A CN113449259A (en) 2021-07-06 2021-07-06 Channel state information feature extraction method and system based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110761547.3A CN113449259A (en) 2021-07-06 2021-07-06 Channel state information feature extraction method and system based on deep learning

Publications (1)

Publication Number Publication Date
CN113449259A true CN113449259A (en) 2021-09-28

Family

ID=77815571

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110761547.3A Pending CN113449259A (en) 2021-07-06 2021-07-06 Channel state information feature extraction method and system based on deep learning

Country Status (1)

Country Link
CN (1) CN113449259A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114997228A (en) * 2022-05-30 2022-09-02 平安科技（深圳）有限公司 Action detection method and device based on artificial intelligence, computer equipment and medium

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114997228A (en) * 2022-05-30 2022-09-02 平安科技（深圳）有限公司 Action detection method and device based on artificial intelligence, computer equipment and medium
CN114997228B (en) * 2022-05-30 2024-05-03 平安科技（深圳）有限公司 Action detection method and device based on artificial intelligence, computer equipment and medium

Similar Documents

Publication Publication Date Title
CN112036433B (en) CNN-based Wi-Move behavior sensing method
Kulikajevas et al. Detection of sitting posture using hierarchical image composition and deep learning
CN108182456B (en) Target detection model based on deep learning and training method thereof
CN106295694B (en) Face recognition method for iterative re-constrained group sparse representation classification
Arshad et al. Leveraging transfer learning in multiple human activity recognition using WiFi signal
KR102318775B1 (en) Method for Adaptive EEG signal processing using reinforcement learning and System Using the same
Chang et al. Device-free indoor localization using Wi-Fi channel state information for Internet of Things
CN110522412B (en) Method for classifying electroencephalogram signals based on multi-scale brain function network
CN109299647B (en) Vehicle control-oriented multitask motor imagery electroencephalogram feature extraction and mode recognition method
Zhang et al. Few-shot learning for fine-grained signal modulation recognition based on foreground segmentation
CN113466852A (en) Millimeter wave radar dynamic gesture recognition method applied to random interference scene
CN116953642A (en) Millimeter wave radar gesture recognition method based on adaptive coding Vision Transformer network
CN113449259A (en) Channel state information feature extraction method and system based on deep learning
CN114861740A (en) Self-adaptive mechanical fault diagnosis method and system based on multi-head attention mechanism
CN106407975B (en) Multiple dimensioned layering object detection method based on space-optical spectrum structural constraint
Sheng et al. CDFi: Cross-Domain Action Recognition using WiFi Signals
CN116306780B (en) Dynamic graph link generation method
CN116311349A (en) Human body key point detection method based on lightweight neural network
CN108021873B (en) Electroencephalogram signal epilepsy classification method and system for clustering asymmetric mutual information
Hammouch et al. A two-stage deep convolutional generative adversarial network-based data augmentation scheme for agriculture image regression tasks
CN115937977A (en) Few-sample human body action recognition method based on multi-dimensional feature fusion
Hsieh et al. End-to-end deep learning-based human activity recognition using channel state information
CN112488238B (en) Hybrid anomaly detection method based on countermeasure self-encoder
Auge et al. Hand gesture recognition in range-Doppler images using binary activated spiking neural networks
CN114724245A (en) CSI-based incremental learning human body action identification method

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