CN114866166B - Wi-Fi subcarrier cross-protocol interference identification method based on CNN - Google Patents

Abstract

The invention discloses a CNN-based Wi-Fi sub-carrier interference identification method, which specifically includes the following steps: Step 1, collecting and classifying channel characteristic information of Wi-Fi sub-carriers in a heterogeneous wireless network; Step 2, preprocessing and The data set that obtains is divided into training set and test set; Step 3, the training set that step 2 obtains is learned and trained by CNN model, obtains the subcarrier classification model after training; Described CNN model comprises input layer, intermediate layer L, output layer; the middle layer includes two layers of convolutional layers, nonlinear fitting function, fully connected layer, and activation function; step 4, according to the classification model trained in step 3, the subcarriers in the test set are subjected to interference identification, and the location is interfered subcarriers and undisturbed subcarriers. Experimental results and performance analysis show that, compared with the prior art, the method provided by the present invention has improved accuracy, generalization ability and the like.

Description

Wi-Fi subcarrier cross-protocol interference identification method based on CNN

technical field

The present invention relates to the field of heterogeneous wireless network communication, in particular to a convolutional neural network-based Wi-Fi sub-carrier cross-protocol interference identification method.

Background technique

A heterogeneous wireless network is a complex communication unit formed by fusing multiple existing communication technologies in the same ISM open frequency band. The phenomenon that multiple incompatible wireless communication technologies share the same spectrum resource in the same time and space is called a heterogeneous wireless network. Network coexistence. In the era of rapid development of artificial intelligence, application scenarios where heterogeneous wireless networks coexist (intelligent transportation, Internet of Vehicles, etc.) This is an important challenge for efficient data transmission in wireless communications. Generally, wireless devices sharing spectrum resources can use the Carrier Sense Multiple Access with Collision Avoidance (CSMA/CA) protocol to implement data transmission, monitor channel status through CCA, and use binary exponential backoff algorithm to avoid conflicts, Obtain a greater share of available bandwidth from other devices at the expense of one device. However, CSMA/CA relies on random backtracking, significantly reducing network throughput as more devices attempt to access the channel. And it cannot be applied to the "high-power interference" scenario, because high-power devices cannot detect the existence of low-power devices, and the signal gain of low-power devices is much smaller than that of high-power signals, which eventually leads to high-power devices monopolizing the shared spectrum. Low-power devices "starve to death". In the coexistence environment of heterogeneous wireless networks, if heterogeneous devices using different communication protocols cannot share limited spectrum resources, it will cause interference between signals, resulting in packet loss, signal distortion, and increased communication delay, which will lead to successful data transmission. rate and spectrum utilization are reduced.

The interference of data transmission in heterogeneous wireless networks mainly lies in the incompatibility of the underlying architectures of different communication protocols. Since various wireless communications were not designed to share spectrum resources in the same time domain across protocols, devices following different protocol standards compete for resources in the shared frequency band, resulting in co-channel interference. So far, the interference problem caused by multi-protocol simultaneous communication in heterogeneous wireless networks has been widely concerned by many researchers and industrial circles in the field of communication at home and abroad, and a series of solutions have been proposed, which can be mainly divided into interference avoidance schemes, There are three types of interference elimination schemes and cross-protocol data transmission schemes. No matter which solution is used, the first prerequisite is the accurate identification of cross-protocol interference, and then different interference solutions can be selected. For example: the 2.4GHz frequency band is a globally shared ISM frequency band, and wireless communication technologies such as Wi-Fi, Bluetooth, and ZigBee can all work in this frequency band. However, because different communication methods follow different MAC layer protocols and standards, such as Wi-Fi follows the IEEE802.11b/g/n protocol, and ZigBee follows the IEEE802.15.4 protocol, due to the incompatibility between the two in the physical layer modulation/demodulation , multiple devices working in the same frequency band under the same heterogeneous wireless network will generate cross-protocol interference, which will affect the overall performance of the heterogeneous wireless network. And in the actual communication process, narrowband ZigBee signal interference has different effects on different Wi-Fi subcarriers. There are two types of interfered subcarriers and non-interfered subcarriers. The difference in subcarrier performance can bring more Wi-Fi communication. Much flexibility.

However, the existing channel estimation technology is to make coarse-grained channel estimation and analysis on the Wi-Fi channel as a whole, which has not been refined to the sub-carrier level, and it is impossible to perform channel feature information on all sub-carriers during the communication process. After the analysis, the subcarriers are classified to determine the transmission scheme.

At present, a variety of commonly used channel estimation methods have been proposed one after another:

(1) Channel estimation based on training sequence. By sending known training sequences, initial channel estimation is performed at the receiving end. When useful information data is sent, a judgment update is performed using the initial channel estimation results to complete real-time channel estimation.

(2) Blind channel estimation. The channel estimation is carried out by using the channel structure information and the statistical characteristics of the transmitted information symbols. This method does not need a training sequence, and only obtains the channel state information by correlating the received signals.

(3) Semi-blind channel estimation. No or only very short training sequences are required. A channel estimation method that combines the advantages of blind estimation and training sequence-based estimation.

The above channel estimation and analysis methods have their own advantages and disadvantages, which are suitable for different specific application scenarios, and may have different disadvantages for different situations, but none of them pay attention to the channel state information of a single subcarrier of WiFi. In a heterogeneous wireless network, after cross-protocol interference, the channel status of different sub-carriers is different. If all sub-carriers are treated equally, the communication performance may be degraded. Therefore, it is first necessary to perform channel estimation from the sub-carrier level, and use an effective classification model to identify the interference of Wi-Fi sub-carriers, and quickly locate the interfered sub-carriers and non-interference sub-carriers.

Contents of the invention

In order to solve the above problems, the present invention proposes a Wi-Fi subcarrier interference identification method based on a convolutional neural network. The method first collects and visually analyzes the channel characteristic information of the subcarriers in a heterogeneous wireless network in a real scene. It is divided into two types: disturbed subcarriers and non-interfered subcarriers. After preprocessing the channel feature information, the training set and test set are constructed, and then the proposed Convolutional Neural Networks (CNN) is used to generate a classification model for subcarrier interference recognition through training. Interference identification of subcarriers in structured wireless networks. Here, the subcarrier interference identification problem is equivalent to learning the conditional probability distribution of subcarrier classification, and using the maximum conditional probability to classify subcarriers.

In order to achieve the above tasks, the present invention adopts the following technical solutions:

A CNN-based Wi-Fi subcarrier interference identification method, specifically comprising the following steps:

Step 1, collect and classify the channel characteristic information of Wi-Fi subcarriers in the heterogeneous wireless network;

Step 2. Perform data preprocessing on the channel characteristic information collected in step 1, including filling in missing values and outliers, and normalize the channel characteristic information, and then divide the preprocessed data set into a training set and the test set;

Step 3, the training set obtained in step 2 is trained through CNN model learning and training to obtain a trained subcarrier classification model; the CNN model includes an input layer, an intermediate layer L, and an output layer; the intermediate layer includes two layers of convolutional layers , nonlinear fitting function, fully connected layer, activation function;

Step 4, according to the classification model trained in step 3, carry out interference identification on the subcarriers in the test set, and locate the interfered subcarriers and non-interfered subcarriers.

Further, the specific operation of the step 1 is as follows:

In the time slot T, the sending end uses the subcarrier S _i to send K data packets to the receiving node. For the subcarrier S _i , the receiving end feeds back the channel information of the subcarrier to the sending end, and the sending end device sets the subcarrier S The channel characteristic information of _i , after obtaining the channel characteristic information of all subcarriers, the sending end divides the categories of n subcarriers.

Further, the step 2 specifically includes the following sub-steps:

Step 21, data filling is performed on the channel characteristic information to obtain a data set after data filling; the specific operation is as follows:

Adopt the box diagram analysis method to carry out outlier detection to the value that deviates from the normal range in the k sampling of the subcarrier, and treat the outlier as a missing value, and use the K nearest neighbor algorithm to fill in the missing value;

In step 22, normalize the filled data set obtained in step 21 using a linear function to obtain a preprocessed data set.

Step 23, divide the preprocessed data set into training set and test set.

Further, in the step 3, the function of the middle layer is as follows:

a) Two convolutional layers are used for deep-level feature perception and extraction, and local perception of sub-carrier channel features;

b) Connect all the features through the fully connected layer, and send the output value to the classifier;

c) Introduce the dropout operation in the fully connected layer, randomly delete some neurons in the neural network in one cycle, and then perform the training and optimization process of the network in this cycle, repeat this process in the next cycle, and use the loss function To measure the quality of the model prediction, as the training progresses, the loss value generally shows a downward trend until the end of the training;

Further, in the step 3, the two convolution layers use a 3x3 convolution kernel.

Further, in the step 3, the BP algorithm is used for update optimization, the learning and training of the network is completed, and the prediction error is minimized; wherein, the optimization process has two stages:

1) Forward calculation calculates the final output value according to the input from front to back, and calculates the gap between the current output and the target, that is, calculates the loss function;

2) The reverse update uses the stochastic gradient descent method to minimize the loss function, and updates the weight by passing the error value;

Further, in the step 3, for the middle layer L and L+i, the output of the middle layer L is nonlinearly mapped by a nonlinear function; the input of the L layer is expressed as x ^(L) , and the weight is expressed as u ^{( L)} ;

Then the input x ^(L+1) of the L+1 layer is expressed as:

x ^(L+1) = f(u ^(L) x ^(L) )

The weight of the middle layer L = (u ⁽¹⁾ , u ⁽²⁾ , ...u ^(L) ), input = (x ⁽¹⁾ , x ⁽²⁾ , ... x ^(L) ), f( ) means not linear function.

Further, in the step 3, the fully connected layer obtains the conditional probability of realizing subcarrier classification through the activation function softmax, and the sum of the conditional probabilities of different categories is 1; the conditional probability distribution is as follows:

The softmax activation function expression is as follows:

Among them, x ⁽⁰⁾ represents the input layer data, u represents the weight, j=0 or 1 represents the possible output subcarrier category, exp(y(j|u, x ⁽⁰⁾ )) indicates that the output layer output subcarrier category is The value given at j.

Further, in the step 3, the expression of the cross-entropy loss function used by the CNN model is:

Among them, n represents the number of input samples, k represents the number of categories, S _i represents the true category of input sample i, j represents the predicted category, 1{ } means that the function is equal to 1 only when the expression in the brackets is true, otherwise it is equal to 0 .

Further, in the step 3, the objective function of the CNN model:

G(u, x ⁽⁰⁾ )=-klog(P(k=1|u,x ⁽⁰⁾ ))-(1-k)log(P(k=0|u,x ⁽⁰⁾ ))

＝-log(P _K (k))

Among them, k represents the number of categories, x ⁽⁰⁾ represents the input layer data, u represents the weight, P(k=1|u, x ⁽⁰⁾ ) represents the probability of outputting category 1, and P _K (k) represents the condition of the predicted category Probability distributions.

Compared with the prior art, the present invention has the following technical characteristics:

(1) The present invention refines the research on the Wi-Fi channel state to the sub-carrier level, and conducts fine-grained analysis and classification after the sub-carrier channel characteristic information is collected, which is conducive to more efficient use of spectrum resources in the frequency domain.

(2) The present invention applies CNN technology to the scene of subcarrier interference identification for the first time. In the process of model training, the channel feature information of all subcarriers is comprehensively considered, realizing the interference identification of subcarriers, and quickly locating the interfered subcarriers and It is free from interference subcarriers, and has good accuracy and generalization ability.

(3) The present invention collects subcarrier channel feature information based on real scenarios, constructs training sets and test sets, and proves the feasibility of the proposed CNN model for subcarrier identification. It is important and valuable to use the data collected in the real scene for training and testing.

Description of drawings

Figure 1 is a flow chart of the method of the present invention.

Figure 2 is an example diagram of subcarrier EVM data;

Figure 3 is the Wi-Fi subcarrier energy monitoring situation before and after being interfered, where (a) is before being interfered, and (b) is after being interfered;

Figure 4 is the classification accuracy rate of the training data set under the subcarrier interference recognition CNN model;

Figure 5 is the classification accuracy rate of the test data set under the subcarrier interference recognition CNN model;

Figure 6 shows the classification accuracy of the test data set under different configurations of the subcarrier interference recognition CNN model;

Figure 7 is a comparison of the classification accuracy between the CNN model and the SVM model.

The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments.

detailed description

Relevant technical terms involved in the present invention:

1. CNN: Convolutional Neural Networks.

2. Subcarrier channel characteristic information: refers to when Wi-Fi devices in a heterogeneous wireless network are interfered by heterogeneous devices (such as ZigBee), in a fixed time slot T, use subcarriers S _i (0<i<n , where i represents the subcarrier index, n represents the number of subcarriers) during data transmission, all relevant information reflecting the channel state of the subcarriers, such as bit error rate, energy detection, amplitude, EVM (Error Vector Magnitude), etc.

In order to quickly locate interfered subcarriers and undisturbed subcarriers, the present invention proposes a CNN-based subcarrier interference identification method.

The CNN-based subcarrier interference identification method comprehensively considers the real subcarrier channel feature information, and obtains the subcarrier classification model through CNN training. First, collect the channel characteristic information of the Wi-Fi subcarrier in the heterogeneous wireless network, determine the interference situation of the subcarrier according to the subcarrier channel characteristic information, and generate a training set and a test set after data preprocessing; second, the The training set of the training set generates a classification model through CNN autonomous learning training including the input layer, the middle layer L (convolution layer, nonlinear fitting function, fully connected layer, activation function), and the output layer; finally, the subcarrier classification model generated by training is used Interference identification is performed on the subcarriers, that is, the maximum conditional probability decision is made according to the conditional probability values output by the subcarriers into different categories, and the subcarriers are divided into different categories.

The CNN-based Wi-Fi subcarrier interference identification method provided by the present invention comprises the following steps:

Step 1, collecting and classifying channel characteristic information of Wi-Fi subcarriers in a heterogeneous wireless network. The specific operation is as follows:

In the time slot T, the sending end uses the subcarrier S _i to send K data packets to the receiving node. For the subcarrier S _i , the sending process of each data packet adopts a closed-loop feedback idea, and the receiving end feeds back the channel information of the subcarrier to the sending end. The sending end device collects the channel characteristic information of the subcarrier S _i fed back by the receiving end through the host computer, and after obtaining the channel characteristic information of all subcarriers, the sending end visually analyzes the channel characteristic information and adds label parameters to the corresponding subcarriers, that is, Classify n subcarriers.

Step 2. Perform data preprocessing on the channel feature information collected in step 1, including filling in missing values and abnormal values to ensure that the data dimension is complete, and normalize the channel feature information, and preprocess The data set is divided into training set and test set; specifically, the following sub-steps are included:

Step 21, data filling is performed on the channel feature information to obtain a data set after data filling.

This is due to the existence of cross-protocol interference in heterogeneous wireless networks. When the closed-loop approach is used to collect channel characteristic information of certain subcarriers, some data is often abnormal or missing. For example: in the information collection process of the i time (i<k, k is the number of samples), some data distributions are: 0.5, 160, 12, 4, 3, 25...; such a data sequence cannot be directly used as a data set use.

Specifically, outlier detection is performed on values deviating from the normal range in the k samples of subcarriers. In this embodiment, the box diagram analysis method is used, that is, a standard for identifying outliers is provided, beyond the upper bound (UB) and lower bound (LB) The values and garbled values within the defined range are identified as outliers, and the outliers are regarded as missing values, which are handed over to the missing value processing method for processing.

Before processing missing values, it is necessary to determine the mechanism and form of missing data, and whether outliers (missing values) are completely random missing, random missing or non-random missing. In this embodiment, the missing data belongs to random missing. In the k samples of subcarriers, the missing data is related to the interference of the subcarriers. The disturbed subcarriers in the closed-loop feedback process are more likely to have outliers and missing values. If missing records are discarded, a large amount of information will be lost, resulting in systematic differences between incomplete and complete observations. For the processing of missing values, a missing value should be filled according to the distribution of characteristic data in the remaining samples in the sample set.

Specifically, the K-nearest neighbor algorithm is used to process the missing value, and the missing value is filled according to the average value of the missing value Di in other k channel feature information;

Step 22, performing normalization processing on the data set filled with data obtained in step 21 to obtain a preprocessed data set.

For the singular sample data in the collected channel feature information, for example, there are discrete values 0.5 and 150 in the collected subcarrier channel feature information, such discrete data will slow down the network convergence speed, so the data needs to be processed according to certain rules Normalize discrete data between [0,1]. In this embodiment, the linear function normalization method is used to perform proportional scaling processing, as shown in the following formula:

Wherein, D is the channel feature information, D _max is the maximum value of the channel feature information, D _min is the minimum value of the channel feature information, and D _norm is the value after normalization processing.

Step 23, divide the preprocessed data set into training set and test set.

Step 3, the training set obtained in step 2 is trained through CNN model learning and training, and the subcarrier classification model after training is obtained; the network model includes an input layer, an intermediate layer (including two layers of convolutional layers, a nonlinear fitting function, a fully connected layer, activation function), output layer. Among them, the role of the middle layer is as follows:

a) Two convolutional layers are used for deep-level feature perception and extraction, and local perception of sub-carrier channel features;

b) Connect all the features through the fully connected layer, and send the output value to the classifier (ie, the activation function).

c) Introduce the dropout operation in the fully connected layer (normally set to 0.5), randomly delete some neurons in the neural network in one cycle, and then perform the training and optimization process of the network in this cycle, and repeat this in the next cycle A process, and use the loss function to measure the quality of the model prediction. As the training progresses, ideally, the loss value generally shows a downward trend until the end of the training. Through the Dropout operation, a neural unit can be ordered to work together with randomly selected neurons, which weakens the joint adaptability between neuron nodes, effectively prevents the network from over-fitting, and enhances the network generalization ability.

Specifically, the two convolutional layers in step 3 can locally perceive the subcarrier channel characteristics, the key lies in the determination of the convolution kernel size and the weight optimization process, as follows:

Combined with the known channel characteristic information and the characteristics of Wi-Fi subcarriers being interfered by cross-protocols, when building a CNN model, it is necessary to design a convolution kernel that conforms to the characteristics of sample data and cross-protocol interference on the basis of conventional experience, so as to speed up the CNN model. In the process of identifying subcarriers, a 3x3 convolution kernel is used here. In the problem of subcarrier interference identification in heterogeneous networks, in terms of the role of convolution itself, compared with large convolution kernels (5x5, 7x7, ...), 3x3 convolution kernels are sufficient to capture changes in subcarrier channel characteristics . Assuming that the number of convolution kernels is n, the parameter quantity of one 5x5 convolution kernel is 25n, and the parameter quantity of two 3x3 convolution kernels is 18n, and the receptive fields of two 3x3 convolution kernels and one 5x5 convolution kernel are the same. The former can effectively reduce the computational complexity and the amount of parameters; and the stack of two convolutional layers with a convolution kernel size of 3x3 has more nonlinear transformations than one convolutional layer with a convolution kernel size of 5x5. The former can use two The sub-nonlinear activation function, while the latter has only one time, (the former feature diversity increases) makes the network capacity larger, CNN's ability to learn features is stronger, and the ability to distinguish different types of subcarriers is stronger; from the perspective of model compression In other words, under the premise of the same receptive field, small convolution kernel stacking is used here to obtain a deeper network with fewer parameters and increase the fitting ability of the network.

For the weight optimization process, the present invention adopts the BP algorithm to update, complete the learning and training of the network, and minimize the prediction error. The optimization process mainly has two stages:

1) Forward calculation Calculate the final output value according to the input from front to back, and calculate the gap between the current output and the target, that is, calculate the loss function.

2) The reverse update uses the stochastic gradient descent method to minimize the loss function, and updates parameters such as weights by passing the error value.

Specifically, for the intermediate layer L and L+1 in step 3, the output of the intermediate layer L must be nonlinearly mapped using a nonlinear function to enhance the network's ability to classify complex features;

Non-linear processing is required between the intermediate layer L and L+1 of the CNN network, where the input of the Lth layer can be expressed as x ^(L) and the weight can be expressed as u ^(L) . In this embodiment, the activation function softsign is used.

Then the input x ^(L+1) of the L+1 layer can be expressed as:

x ^(L+1) = f(u ^(L) x ^(L) )

Here, the weights of each layer in the middle layer can be expressed as u=(u ⁽¹⁾ , u ⁽²⁾ , ...u ^(L) ), and the input can be expressed as x=(x ⁽¹⁾ , x ⁽²⁾ , ... x ^(L) ), f(·) represents a nonlinear function.

Assuming that the output features are not processed nonlinearly in the middle layer of CNN, then the relationship between the output layer and the input layer can be expressed as a simple linear relationship. CNN is equivalent to being simplified to a network model with only one layer. At this time, the relationship between the two can be expressed as Expressed as follows:

x ^(O) = u ^(L) u ^(L-1) …u ⁽¹⁾ x ⁽⁰⁾

Among them, x ⁽⁰⁾ represents the input data, and x ^(O) represents the output data. If the nonlinear processing process is not added in the middle layer, it is impossible to learn complex features for classification.

Specifically, the fully connected layer in step 3 needs to obtain the conditional probability of subcarrier classification through the activation function softmax, and the sum of the conditional probabilities of different categories is 1;

Because the subcarrier category is divided into 2 categories: interfered/undisturbed, each vector length in the softmax output data is 2, each of which has a value between 0 and |, describing the input data Likelihood of belonging to the class it represents relative to other classes. Assuming that the subcarrier category S∈{0, 1} predicted by the classifier here is a discrete random variable, s=0 and s=1 represent interfered subcarriers and undisturbed subcarriers, and vice versa. In this case, the predicted probability P conforms to the Bernoulli distribution, its probability mass function can be expressed as P _S (s) = P ^s (1-P) ^1-s , and the conditional probability distribution can be written as the following formula:

Further, the expression of the softmax activation function used in step 3 is as follows:

Among them, x ⁽⁰⁾ represents the input layer data, u represents the weight, j=0 or 1 represents the possible output subcarrier category, exp(y(j|u, x ⁽⁰⁾ )) indicates that the output layer output subcarrier category is The value given at j. Furthermore, the conditional probability P of the k-classification problem can be derived as:

Specifically, the expression of the cross-entropy loss function used by the CNN model in step 3 is:

Among them, n represents the number of input samples, k represents the number of categories, S _i represents the true category of the input sample i, j represents the predicted category, 1{ } means that the function is equal to 1 only when the expression in the brackets is true, otherwise it is equal to 0 .

On the basis of the above loss function, the objective function of the CNN model is established. The objective function describes the relationship between the conditional probability distribution and the loss function:

G(u, x ⁽⁰⁾ )=-klog(P(k=1|u,x ⁽⁰⁾ ))-(1-k)log(P(k=0|u,x ⁽⁰⁾ ))

＝-log(P _K (k))

Among them, k represents the number of categories (k=2 in this embodiment), x ⁽⁰⁾ represents (input layer data), u represents the weight, P(k=1|u, x ⁽⁰⁾ ) represents the output of category 1 Probability, P _K (k) represents the conditional probability distribution of the predicted class.

Step 4, according to the classification model trained in step 3, carry out interference identification on the sub-carriers to be tested, and locate the interfered sub-carriers and non-interference sub-carriers.

Specifically, the accuracy rate of predicting the subcarrier category described in step 4 can be expressed as:

Where n represents the number of samples, S _i represents the true category of the input sample i, and j represents the predicted category.

Step 4 Use the classification model trained in step 3 to predict the subcarrier category S∈{0, 1} on the test set obtained after data preprocessing in step 2, and use the maximum conditional probability to make classification decisions, and locate the interfered subcarriers and unaffected subcarriers. interfering subcarriers.

In order to prove the feasibility and effectiveness of the method of the present invention for subcarrier interference identification, the present invention carried out the following experimental evaluation on the data sample set collected by WARP equipment in a real heterogeneous wireless network scenario.

1. Data set description

The data set used in this experiment is collected by the experimental equipment. We measured the required sub-carrier information on the Wi-Fi device side in a heterogeneous network environment. In this experiment, the collected sub-carrier information is divided into training data sets. and test data sets, in which the training data set and the test data set include the throughput of subcarriers, SER (Symbol ErrorRate), CSI (Channel State Information) amplitude, subcarrier energy information, EVM (Error VectorMagnitude), Table 1 shows the specifics of the data set. As shown in Figures 1 and 2, part of the captured channel information is displayed. Figure 1 shows the EVM value and interference data distribution of the Wi-Fi subcarriers after cross-protocol interference. Figure 2 shows the Wi-Fi subcarriers The channel energy changes before and after cross-protocol interference, the information needs to be able to truly and effectively reflect the subcarrier channel information.

Statistics of the dataset in Table 1

2. Analyze the experimental results

Figure 3 shows the classification performance of the CNN model on the subcarriers of the training set as the number of training epochs increases. Compared with the high SINR and low SINR environments, the medium SINR environment has fewer training epochs to achieve 100% classification accuracy. This is because in the high-interference environment of heterogeneous networks, when the Wi-Fi subcarriers are subject to cross-protocol interference, there will be avoidance and fallback in the time domain, resulting in a decrease in network convergence speed. In a low-interference environment, at this time The cross-protocol interference feature is weak, and the environmental noise will cause the network convergence speed to decrease.

Figure 4 shows the classification performance of the CNN model on the test set under different SINRs. It can be seen that under the medium SINR, the classification accuracy of the test data set is the highest, which is caused by the fact that the subcarriers are subject to cross-protocol interference in the real environment, and under different SINRs, the CNN model has achieved convergence. The effect proves the robustness of the CNN model for interference identification on Wi-Fi subcarriers.

Figure 5 shows the performance change of the CNN model when increasing the number of convolutional layers. By increasing the number of convolutional layers, the classification performance of the test set under different SINRs is compared. The experimental results show that the classification performance after increasing the number of convolutional layers is basically the same as that of the original configuration, which indicates that the increase in the number of convolutional layers in this network model will not significantly improve the classification performance.

Figure 6 shows the performance comparison of the CNN model and the SVM model for subcarrier interference identification under different SINRs. Support Vector Machine (SVM) requires subcarrier features to be selected from the time domain or frequency domain. This method is simple to train and fast to identify. However, its performance depends heavily on the selection of features and cannot adapt to changes in the environment. This is because SVM relies on feature extraction, and manually extracted features cannot fully represent the channel characteristics of Wi-Fi subcarriers in a heterogeneous network coexistence environment. Through the CNN model to identify the interference of the subcarriers, it can be found that the classification accuracy of the subcarriers under different SINRs is significantly higher than that of the SVM, indicating that the model has better generalization ability.

Claims (10)

1. A Wi-Fi subcarrier interference identification method based on CNN is characterized by comprising the following steps:

step 1, collecting channel characteristic information of Wi-Fi subcarriers in a heterogeneous wireless network and classifying the channel characteristic information;

step 2, performing data preprocessing on the channel characteristic information acquired in the step 1, wherein the data preprocessing comprises filling missing values and abnormal values, performing normalization processing on the channel characteristic information, and dividing a preprocessed data set into a training set and a test set;

step 3, learning and training the training set obtained in the step 2 through a CNN model to obtain a trained subcarrier classification model; the CNN model comprises an input layer, an intermediate layer and an output layer; the middle layer comprises two convolutional layers, a nonlinear fitting function, a full connection layer and an activation function;

and 4, carrying out interference identification on the subcarriers in the test set according to the classification model obtained by training in the step 3, and positioning the interfered subcarriers and the non-interfered subcarriers.

2. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein the specific operation of step 1 is as follows:

in time slot T, the transmitting end adopts subcarrier S _i Sending K data packets to a receiving node, for subcarrier S _i The receiving end feeds back the sub-carrier channel information to the sending end, and the sending end equipment acquires the sub-carrier S fed back by the receiving end _i After acquiring the channel characteristic information of all the subcarriers, the sending end divides the categories of the n subcarriers.

3. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein the step 2 specifically includes the following substeps:

step 21, performing data filling on the channel characteristic information to obtain a data set after the data filling; the specific operation is as follows:

abnormal value detection is carried out on values deviating from a normal range in K times of sampling of subcarriers by adopting a boxed graph analysis method, the abnormal values are regarded as missing values, and filling processing is carried out on the missing values by adopting a K-nearest neighbor algorithm;

step 22, performing normalization processing on the data set filled with the data obtained in the step 21 by adopting a linear function to obtain a preprocessed data set;

and step 23, dividing the preprocessed data set into a training set and a test set.

4. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein in the step 3, the role of the intermediate layer is as follows:

a) The two layers of convolution layers are used for carrying out deep level feature sensing and extraction and locally sensing the channel features of the sub-carriers;

b) Connecting all the characteristics through the full-connection layer, and sending an output value to a classifier;

c) And introducing dropout operation in a full connection layer, randomly deleting part of neurons in the neural network in one cycle, then performing the training and optimizing process of the network in the cycle, repeating the process in the next cycle, measuring the quality predicted by the model by using a loss function, and leading the loss value to be in a descending trend along with the training until the training is finished.

5. The CNN-based Wi-Fi subcarrier interference identification method of claim 1, wherein in the step 3, the two convolutional layers employ a 3x3 convolutional kernel.

6. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein in step 3, updating optimization is performed by using a BP algorithm to complete learning and training of a network and minimize a prediction error; the process of weight optimization has two stages:

1) Forward calculation is carried out to obtain a final output value by sequentially calculating from front to back according to input, and the difference between the current output and a target is calculated, namely a loss function is calculated;

2) And the reverse updating utilizes a random gradient descent method to minimize a loss function, and updates the weight value through a transfer error value.

7. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein in the step 3, for the intermediate layers L and L +1, the output of the intermediate layer L is nonlinearly mapped by using a nonlinear function; the input at L < th > level is denoted as x ^(L) The weight is represented as u ^(L) ；

Then input x of layer L +1 ^(L+1) Expressed as:

x ^(L+1) ＝f(u ^(L) x ^(L) )

weight u = (u) of each layer of the intermediate layer ⁽¹⁾ ,u ⁽²⁾ ,…u ^(L) ) Input x = (x) ⁽¹⁾ ,x ⁽²⁾ ,…x ^(L) ) And f (·) denotes a nonlinear function.

8. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein in the step 3, the full connectivity layer obtains conditional probabilities for implementing subcarrier classification by activating function softmax, and the sum of the conditional probabilities of different classes is 1; the conditional probability distribution is as follows:

the softmax activation function is expressed as follows:

A(s＝j|u，x ⁽⁰⁾ )∈[0，1]

wherein x is ⁽⁰⁾ Represents input layer data, u represents a weight, j =0 or 1 represents a subcarrier class that may be output, y (j | u, x) ⁽⁰⁾ ) Which represents the value given when the output layer outputs a subcarrier class of j.

9. The method for identifying Wi-Fi subcarrier interference based on CNN according to claim 1, wherein in the step 3, the cross entropy loss function expression used by the CNN model is:

where n represents the number of input samples, k represents the number of classes, S _i Representing the true class of the input sample i, j representing the prediction class, 1 {. Cndot.) represents that the function is equal to 1 only if the expression in brackets is true, and 0 otherwise.

10. The CNN-based Wi-Fi subcarrier interference identification method of claim 1, wherein in the step 3, an objective function of the CNN model:

G(u，x ⁽⁰⁾ )＝-klog(P(k＝1|u，x ⁽⁰⁾ ))-(1-k)log(P(k＝0|u，x ⁽⁰⁾ ))

＝-log(P _K (k))

where k denotes the number of classes, x ⁽⁰⁾ Denotes input layer data, u denotes a weight, P (k =1 calc u ⁽⁰⁾ ) Representing the probability, P, of outputting class 1 _K (k) Representing a conditional probability distribution of the prediction classes.

Images