CN107451605A - A kind of simple target recognition methods based on channel condition information and SVMs - Google Patents

Abstract

A kind of simple target recognition methods based on channel condition information and SVMs proposed by the present invention, it is not necessary to build special hardware facility, make full use of prior wireless network, the function of simple target identification is reached using common commercial router can.After CSI initial data is obtained, use density-based algorithms DBSCAN to cluster the subcarrier data in channel first with denoising, then the data after denoising are carried out using the moving average algorithm based on weights smooth.After data prediction, the present invention carries out characteristics extraction using Principal Component Analysis Algorithm to data.Data after pretreatment and feature extraction can more accurately reflect the Main change of signal and dimension substantially reduces, and are favorably improved target identification precision and reduce computation complexity.The present invention is by means of the SVM multi-classification algorithms based on one against one strategies, the statistical model of non-linear dependence between acquisition destination object and received signals fingerprint, so as to reach the purpose of simple target identification.

Description

A Simple Object Recognition Method Based on Channel State Information and Support Vector Machine

technical field

The invention relates to the field of target recognition, in particular to a method for simple target recognition based on channel state information and using support vector machine technology.

Background technique

Wi-Fi-based wireless local area networks are widely deployed indoors, and can provide simple target recognition services while providing data transmission services. The proportion of water in the human body is 70%, and water has a strong ability to absorb radio frequency signals, so the human body will reflect, scatter, diffract, and attenuate the surrounding Wi-Fi signals. By monitoring the special fingerprint characteristics formed by the interference caused by the human body to the Wi-Fi signal, it is possible to perform simple target identification on whether the target is a human.

Received Signal Strength Indicator (Received Signal Strength Indicator, RSSI) and Channel State Information (Channel State Information, CSI) can be obtained from the Wi-Fi signal. RSSI is currently the most widely used energy characteristic, but its coarse-grainedness and variability are not suitable for accurate perception in multipath indoor environments, and the accuracy of target recognition is poor. CSI is a physical layer feature that describes the attenuation factor of the signal propagating between the transmitter and receiver, including information such as scattering, environmental attenuation, and distance attenuation. It can resist interference from narrow-band signals in the frequency band and is stable enough in a static environment. It can respond immediately when it is interfered, and can distinguish signals from multiple paths, and the multipath effect has little influence. With the application of Orthogonal Frequency Division Multiplexing (Orthogonal Frequency Division Multiplexing, OFDM) technology in wireless local area networks, CSI can be obtained more conveniently at present. CSI contains the amplitude and phase information of each subcarrier, which can provide rich frequency domain information, thereby improving the accuracy of target recognition.

The process of simple target recognition is essentially a classification process. The interference of CSI signals on the human body is obvious and has a good degree of recognition. Therefore, the idea of classification can be used to solve the problem of simple target recognition. Support Vector Machines (Support Vector Machines, SVM) is a machine learning method based on statistical learning theory, which shows many advantages in solving high-dimensional and nonlinear problems. Statistical models of linear dependencies.

Contents of the invention

The present invention is a kind of simple target recognition method based on channel state information (CSI) and support vector machine (SVM) classification, comprises the following steps:

Step 1: Environmental deployment, Wi-Fi-based simple target recognition requires indoor coverage of Wi-Fi signals, select the 5G frequency band with less signal interference, and the equipment is two laptops, both of which are equipped with Intel link 5300agn wireless commercial network cards;

Step 2: CSI raw data collection, collecting several CSI raw data of different target objects, including: number of transmitting antennas, number of receiving antennas, sending frequency, channel state information CSI matrix;

Step 3: CSI data preprocessing, including: (1) remove the first dimension of the CSI matrix in the original data, convert the generated two-dimensional matrix from linear (level) space to logarithmic (power) space, and convert the matrix Each complex number is converted into a magnitude; (2) There are 30 subcarriers in each pair of transmitting and receiving antenna channels, and the density-based clustering algorithm Density-Based Spatial Clustering of Applications with Noise (DBSCAN) is applied to each channel Clustering, denoising by deleting outliers; (3) smoothing the denoised CSI data using a weight-based sliding average algorithm;

Step 4: CSI eigenvalue extraction, using the Principal Component Analysis (PCA) algorithm to perform dimensionality reduction and eigenvalue extraction on the preprocessed CSI data to generate CSI fingerprint samples;

Step 5: SVM model training; the steps are as follows: (1) normalize the CSI fingerprint samples; (2) establish a SVM multi-classification model based on the CSI fingerprint samples, each classification corresponds to a target;

Step 6: Simple target recognition, the steps are as follows: (1) Follow Step 2 to collect CSI raw data; (2) Follow Step 3 to perform CSI data preprocessing; (3) Follow Step 4 to perform CSI data dimensionality reduction and feature value extraction to obtain Real-time signal fingerprint; (4) Determine the target represented by the real-time measured CSI fingerprint according to the voting results of the SVM multi-classification model, and realize simple target recognition;

Description of drawings

Fig. 1 is a flow chart of a simple target recognition method based on CSI and SVM;

Figure 2 is an environmental deployment diagram of simple target recognition based on CSI;

detailed description

Specific embodiments of the present invention will be described below in conjunction with the accompanying drawings, so that those skilled in the art can better understand the present invention. It should be noted that in the following description, the introduction of known functions and designs related to the present invention may be downplayed and omitted.

In this embodiment, the present invention mainly includes the following links to the simple target recognition method: data collection, data preprocessing, feature extraction of data, and simple target classification and recognition. The flow chart is shown in Figure 1, and the specific implementation steps are as follows:

Step 1: Environmental deployment. Simple target recognition based on Wi-Fi requires indoor coverage of Wi-Fi signals. The layout of the experimental scene is 6 meters long and 7 meters wide. The system selects the 5G frequency band with less signal interference, and the equipment is two Laptops, the models are Lenovo THINKPAD T400 and X201i, both of them are equipped with intel link 5300agn wireless business network card, which has 3 antennas. T400 is set to send broadcast packet data in a directional way, as a device for sending signals, and X201i is set to monitor mode, as a device for receiving signals. Refer to Figure 2 for a schematic diagram of the layout.

Step 2: CSI raw data collection, training phase, will be divided into 4 scenarios to collect training data samples, namely no one, foam dummy, male experimenter and female experimenter. Let the target stay at the position shown in Figure 2, X201i collects 20 pieces of CSI raw data from T400 per second, including: number of transmitting antennas N _tx , number of receiving antennas N _rx , data packet sending frequency f, channel state information matrix H . The channel state information H is a three-dimensional matrix of N _tx ×N _rx ×30, the third dimension is the 30 subcarrier information h=|h|e ^jsinθ in the OFDM channel, |h| is the subcarrier amplitude, θ is the subcarrier phase .

Step 3: CSI data generation. For the collected CSI raw data, first remove the first dimension of the CSI matrix H to obtain N _tx two-dimensional matrices of N _rx ×30, and transform the two-dimensional matrix from the linear (level) space Convert to logarithmic (power) space and convert each complex number in the matrix to a magnitude.

Step 4: CSI data denoising, each pair of transmitting and receiving antennas forms a channel, so a pair of AP-MPs contains N _tx ×N _rx channels; each channel contains 30 subcarriers, so a pair of AP-MPs contains N _tx ×N _rx ×30 subcarriers. The CSI data set is divided into N _tx ×N _rx sub-data sets according to the channel, and each sub-data set contains 30 sub-carrier information. Let the serial number of each sub-carrier in the sub-dataset be index (1-30), and the amplitude be value, and apply the density-based clustering algorithm DBSCAN on each sub-dataset for clustering. The two parameters in DBSCAN are domain radius e and the minimum number of included points minOpt, and its clustering steps include: (1) mark all objects in the sub-dataset as unvisited, that is, "unvisited"; (2) randomly select an unvisited Visit object o(index,value), marked as "visited"; check if o's neighborhood contains at least minOpt objects: if not, mark o as an outlier; if yes, create a new cluster C for o and a candidate set N, put all objects in the neighborhood of o in the candidate set N; (3) DBSCAN iteratively adds objects in N that do not belong to other clusters to C until N is empty, and cluster C is completed; (4) Go to step (2) to process the next object; (5) Delete the sample data corresponding to the object marked as an outlier from the training set to achieve the purpose of data denoising.

Step 5: CSI data smoothing, using a weight-based moving average algorithm to smooth the denoised CSI data to reduce data fluctuations. Assuming that a CSI subcarrier sequence at time t is (v ₁ , v ₂ ,...,v _t ), then the CSI smoothing value at time t is the weighted average of the first m values, where m is the sliding window size:

Step 6: CSI eigenvalue extraction, (1) Assume that there are N _ap pairs of AP-MP combinations, each pair of AP-MPs contains N _tx × N _rx channels, and each channel has 30 subcarriers, then the dimension of each piece of CSI data The number is N _ap ×N _tx ×N _rx ×30, and the dimension is very high. The contribution of each value in the CSI data to the classification is different, so the principal component analysis (PCA) algorithm is used to reduce the dimensionality of the CSI data and extract the most effective features. The goal of PCA is to find r(r<N _ap ×N _tx ×N _rx ×30) new features, each new feature is a linear combination of the original features, the new features can reflect the main features of the original data, and can compress The size of the original data. (2) Assume that the training sample set matrix is X, and the number of samples is N. Through PCA, the conversion matrix C, the converted matrix S of the sample matrix X, and the feature weights L=(L ₁ , L ₂ , . . . , L _n ) arranged in descending order can be obtained, where n is the feature dimension. Calculate the cumulative contribution rate of the feature _Lr :

If L _r >95%, that is, the cumulative contribution rate of the first r features is greater _than the preset threshold of 95%, then take the _{first r} features F=(L ₁ ,L ₂ ,…,L _r ) as the extracted features. The first r columns in the matrix S form the principal component matrix R. R is used as sample data for subsequent model training.

Step 7: CSI fingerprint sample generation and normalization, assuming that the CSI sample set in the detection area is X={c ₁ ,c ₂ ,...,c _N }, N is the number of samples; c _i ={c _i1 ,c _i2 ,...,c _ir } represents the sample after feature extraction, r is the sample dimension, and normalized to get:

c _ij represents the jth eigenvalue of sample c _i , c _min represents the minimum value of all eigenvalues, and c _max represents the maximum value of all eigenvalues;

Step 8: SVM classification model training, perform SVM classification model training for each target. Since there are multiple recognition targets, multi-target classification is a multi-classification process. The one-against-one strategy is used to extend SVM from binary classification to multi-classification, and finally vote to obtain the final classification result. Assuming that the number of targets is K and each target corresponds to a class, then

The number of classes is K, and K(K-1)/2 classifiers need to be trained, and each classifier performs binary classification on two targets. Assuming that ω _i and ω _j represent the classes corresponding to target d _i and target d _j respectively, the training of the SVM classifier for class ω _i and class ω _j is to use the CSI fingerprint samples from ω _i and ω _j to solve the following problem :

Where C is a constant, ξ is a set of slack variables, c _t is a CSI fingerprint sample, and N is the number of samples.

Step 9: Online target recognition. During actual target recognition, the algorithm determines the classification of the target through the SVM classifier obtained through training based on the CSI fingerprints collected in real time. Including the following steps: (1) collecting CSI raw data in the manner of step 2; (2) performing preprocessing and feature extraction on the CSI data according to the manner of step 3 to step 6; (3) generating CSI fingerprints and merging them according to the manner of step 7 (4) Determine the target represented by the real-time measured CSI fingerprint based on the voting results of multiple SVM classification models.

The beneficial effects of the present invention are: the CSI-based simple target recognition solution does not need to build special hardware facilities, fully utilizes the existing wireless network, and can achieve the simple target recognition function by using ordinary commercial routers. After obtaining the original CSI data, the present invention preprocesses the data, first adopts the density-based clustering algorithm DBSCAN to cluster the subcarrier data in the channel to denoise, and then adopts the weight-based sliding average algorithm to denoise The subsequent data are smoothed. After data preprocessing, the present invention uses a principal component analysis algorithm to extract eigenvalues from the data. The data after preprocessing and feature extraction can more accurately reflect the main changes of the signal and greatly reduce the dimensionality, which helps to improve the accuracy of target recognition and reduce the computational complexity. The invention obtains the statistical model of the nonlinear dependence relationship between the target object and the signal fingerprint by means of the SVM multi-classification algorithm based on the one-against-one strategy, thereby achieving the purpose of simple target recognition. The invention can achieve target recognition accuracy of more than 98.7%.

Although the illustrative specific embodiments of the present invention have been described above, so that those skilled in the art can understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, As long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are included in the protection list.

Claims (2)

1. the present invention is a kind of simple target identification side classified based on channel condition information (CSI) and SVMs (SVM) Method, mainly including herein below：The collection and pretreatment of CSI data, CSI data characteristicses are extracted, to extracting feature off-line training Multiple SVM models, simple target ONLINE RECOGNITION stage.

Technical scheme is as follows：

Step 1：Simple target identification based on Wi-Fi requires in-door covering Wi-Fi signal, and selection signal disturbs less 5G Frequency range, equipment are two notebook computers, are respectively arranged with Intel link 5300agn wireless business network interface cards；

Step 2：Some CSI initial data of different target object are gathered, including：Transmission antenna number, reception antenna number, Transmission frequency, channel condition information CSI matrixes；

Step 3：CSI data predictions, including：(1) dimension of CSI matrixes first in initial data is removed, two by caused by Dimension matrix is transformed into logarithm (power) space from linear (level) space, by each complex conversion in matrix into value；(2) it is every There are 30 subcarriers in a pair of channels for sending and receiving antenna composition, using density-based algorithms Density- Based Spatial Clustering of Applications with Noise (DBSCAN) gather to each channel Class, by deleting outlier denoising；(3) the CSI data after denoising are carried out using the moving average algorithm based on weights smooth；

Step 4：CSI characteristics extractions, using principal component analysis (PCA) algorithm to pretreated CSI data carry out dimensionality reduction and Characteristics extraction, produce CSI sample fingerprints；

Step 5：SVM model trainings, step are as follows：(1) CSI sample fingerprints are normalized；(2) CSI sample fingerprints are based on, are built The more disaggregated models of vertical SVM, the corresponding target of each classification；

Step 6：Simple target identifies that step is as follows：(1) CSI raw data acquisitions are carried out according to step 2；(2) according to step Three carry out CSI data predictions；(3) CSI Data Dimensionality Reductions and characteristics extraction are carried out according to step 4, obtains live signal and refer to Line；(4) determine the CSI fingerprints measured in real time according to the voting results of the more disaggregated models of SVM representated by target.

A kind of 2. simple target classified based on channel condition information CSI and support vector machines according to claim 1 Recognition methods, its feature are (1) data prediction, and each channel is entered using density-based algorithms (DBSCAN) Row cluster, deletes outlier denoising；(2) CSI Data Dimensionality Reductions and characteristics extraction, using principal component analysis (PCA) algorithm to pre- CSI data after processing carry out dimensionality reduction and obtain CSI sample fingerprints；(3) by means of the SVM based on one-against-one strategies Multi-classification algorithm, the statistical model of non-linear dependence between target and received signals fingerprint is obtained, reach the purpose of target identification.