A method and system for identifying ETC interference signal transmitters based on deep residual network

Through the identification method based on deep residual network, combined with ground sensing coil and principal component analysis, and introducing technologies such as identity mapping and batch normalization, the problem of individual identification of ETC interference signal transmitters is solved, and efficient identification and generalization are achieved Improve the ability to improve.

CN112163574BActive Publication Date: 2024-12-03NANJING AEROSPACE IND TECH CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202011320179.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-11-23
Publication Date
2024-12-03
Estimated Expiration
2040-11-23

AI Technical Summary

Technical Problem

The prior art cannot effectively identify and distinguish individual ETC interference signal transmitters, resulting in insufficient generalization capabilities and it is difficult to achieve accurate identification of interference signal transmitters.

Method used

Using the recognition method based on deep residual network, the ETC interference signal is received through ground sensing coil directionally, the data set is constructed and principal component analysis is performed, the deep residual network is designed, and the identity mapping, batch normalization and label smoothing strategies are introduced, and the training is carried out. Networks to extract deep abstract features and improve generalization capabilities.

Benefits of technology

It realizes accurate identification of ETC interference signal transmitters, enhances the generalization ability of the identification network, can effectively identify and determine the type of interference signal transmitters, and supports subsequent investigation and punishment measures.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112163574B_ABST
    Figure CN112163574B_ABST
Patent Text Reader

Abstract

The present invention proposes a method and system for identifying ETC interference signal transmitters based on a deep residual network. The identification method mainly includes the following steps: Step 1, by combining the ground sensing coil to receive the ETC interference signal in a directional manner, an ETC interference signal data set is constructed; Step 2, the ETC interference signal data set is divided into a training set, a validation set and a test set according to a predetermined ratio; Step 3, the interference signals of the training set, the validation set and the test set are subjected to principal component analysis to obtain a low-dimensional feature representation of the interference signal; Step 4, a deep residual network is designed, the deep residual network is trained by the training set, and the optimal deep residual network recognition model is selected according to the validation set, so as to identify the type of ETC interference signal. The present invention can automatically extract abstract features useful for ETC interference signal recognition from the training set data without manual participation in feature extraction. In addition, label smoothing, batch normalization and other methods are introduced to further improve the generalization performance of the recognition network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method and system for identifying an ETC interference signal transmitter based on a deep residual network, and relates to the field of individual identification of wireless signal transmitters. Background Art

[0002] ETC is an electronic toll collection system that automatically collects tolls on highways or bridges. Through dedicated short-range communication between the on-board electronic tag installed on the vehicle windshield and the microwave antenna on the ETC lane of the toll station, computer networking technology is used to conduct back-end settlement processing with the bank, so that the vehicle can pay the highway or bridge fee without stopping at the highway or bridge toll station. At present, some lawless elements use radio jammers to interfere with the ETC toll collection system to evade tolls, causing certain economic losses and having a greater negative impact on the development of highway ETC. Therefore, a method for identifying individual ETC jammers is needed to identify interference information and record the characteristics of the individual transmitting the interference source, so as to achieve the subsequent processing of retaining evidence, recovering tolls and even pursuing legal liability.

[0003] In the prior art, there are only methods for locating interference sources. Such positioning methods are unable to perform deep feature extraction, resulting in low generalization ability of network recognition, making it difficult to effectively identify individual interference signal transmitters. Summary of the invention

[0004] Purpose of the invention: One purpose is to propose a method for identifying ETC interference signal transmitters based on a deep residual network to solve the above problems existing in the prior art. A further purpose is to propose a system for implementing the above method.

[0005] Technical solution: A method for identifying ETC interference signal transmitters based on a deep residual network, comprising the following steps:

[0006] Step 1: Construct an ETC interference signal dataset by combining the ground sensing coil to receive the ETC interference signal in a directional manner;

[0007] Step 2: Divide the ETC interference signal data set into a training set, a validation set, and a test set according to a predetermined ratio;

[0008] Step 3: Perform principal component analysis on the interference signals of the training set, validation set, and test set to obtain low-dimensional feature representation of the interference signals;

[0009] Step 4: Design a deep residual network, train the deep residual network through the training set, select the optimal deep residual network recognition model based on the verification set, test the optimal deep residual network recognition model through the test set, and then identify the type of ETC interference signal.

[0010] In a further embodiment, the training set, validation set and test set in step 2 account for 60%, 20% and 20% of each type of interference signal data respectively;

[0011] The dimension of the interference sample feature vector obtained by principal component analysis in step 3 accounts for 30% of the initial interference sample feature dimension.

[0012] In a further embodiment, step 4 further comprises:

[0013] Step 4-1, introducing the identity mapping into the deep residual network and further extracting the deep abstract features of the ETC interference signal;

[0014] Step 4-2: Introduce batch normalization operation and label smoothing strategy in the deep residual network;

[0015] Step 4-3: Train the deep residual network and iteratively update the connection weights between layers through the back propagation algorithm;

[0016] Step 4-4: Use the trained deep residual recognition network to identify the type of test interference samples.

[0017] In a further embodiment, the deep residual network introduces the idea of ​​identity mapping, which, to a certain extent, solves the problem that the recognition performance of the recognition model decreases due to the increase in the depth of the convolutional neural network, gives full play to the advantages of the depth of the convolutional neural network, and can extract the deep abstract features of the ETC interference signal. Through the identity mapping, the conventional convention that the output of the n-1 layer of the traditional neural network can only be given to the n layer as input is broken. After such processing, the generalization ability of the network is stronger. Among them, the basic structure of the designed deep residual recognition network mainly includes a convolution layer, a pooling layer, a nonlinear mapping layer, a residual block and a fully connected layer. The characteristics of "local connection" and "weight sharing" of the convolution operation are used to effectively solve the training overhead. Through operations such as convolution, pooling and nonlinear mapping, the input interference signal is processed, and the mapping between the input target type and the output target type is realized in the connection layer. Each feature mapping extracts a local feature of the interference signal through a one-dimensional convolution filter. According to the principle of local correlation, the pooling layer downsamples the features obtained by the convolution layer, thereby reducing the amount of model training while retaining useful feature information.

[0018] In a further embodiment, in order to speed up the convergence speed of the deep residual recognition network, a batch normalization (BN) operation is introduced when designing the network. The definition of BN is as follows:

[0019]

[0020]

[0021] Here, it is assumed that one layer of the neural network has a d-dimensional input feature vector , represents the mean of the input feature vector x, represents the variance of the input feature vector x, represents the eigenvalue of the kth dimension of the input feature vector x, represents the eigenvalue of the kth dimension of the normalized input eigenvector x, represent The normalized eigenvalues ​​after scaling and translation, and Represents learnable translation and scaling parameters, and BN is inserted in any layer of the network;

[0022] Since deep network models are prone to overfitting, in order to enhance the generalization ability of the network, a label smoothing strategy is introduced when designing the model. Label smoothing is defined as follows:

[0023] In the study of neural networks, for classification and recognition problems, the labels of training samples are often converted into one-hot vector encoding. For the cross entropy loss function commonly used in classification and recognition, the use of one-hot encoding cannot guarantee the generalization ability of the model and is prone to overfitting. Label smoothing is a method of correcting the loss function, which allows the neural network to consider the loss contributed by incorrect labels to a certain extent.

[0024] Assume that the modulated signal sample is , taking binary classification as an example, is the training label, which is 0 or 1. In each iteration of training, Put it in the training set, but set an error rate ,by The probability will be Substitute into training, The probability will be Substitute it into the training. In this way, the model has both correct label input and wrong label input during training. The model trained in this way will not fully match every training label, but will match it to a certain extent. In this way, if an incorrect label appears, the model will be less affected.

[0025] When the neural network adopts the cross entropy loss function, for each sample , the loss function of the network is expressed as:

[0026]

[0027] In the formula Represents the network prediction sample The probability of class 1, Indicates the predicted sample category, and the meanings of the other symbols are the same as above;

[0028] When label smoothing is introduced, the loss function of the network becomes:

[0029]

[0030] in:

[0031]

[0032] When the label is 0, it is not put directly into the training, but replaced by a smaller number. Similarly, if the label is 1, replace it with a number closer to 1 .

[0033] In a further embodiment, during the training of the deep residual network, the connection weights between the layers are continuously updated iteratively through the back propagation algorithm, so that the entire ETC interference signal recognition network can be learned more accurately. Wherein, formula (1) reflects the relationship between the features extracted by the recognition network at the kth layer and the features at the k-1th layer.

[0034]

[0035]

[0036] In the formula Represents a nonlinear activation function, which can enhance the nonlinear expression ability of the network. After the last convolution operation, the fully connected layer is introduced. The fully connected layer plays the role of mapping the "distributed feature representation" learned by the convolution layer and the pooling layer to the sample label space. represents the nonlinear activation function, represents the eigenvalue, represents the jth feature of the kth layer, represents the i-th eigenvalue of the k-1th layer, represents the weight value connecting the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, It represents the bias between the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, and N represents the number of neurons in the k-1th layer.

[0037] In a further embodiment, step 4-4 further includes: using a trained deep residual recognition network to identify the type of test interference samples. First, the test sample is processed by PCA and then input into a deep residual recognition network. The deep residual recognition network extracts features from the test sample, and the features extracted by the last convolutional layer are input into the fully connected layer. The distributed features learned by the convolutional layer are further mapped to the interference sample label space, and finally the output of the fully connected layer is subjected to a softmax operation to obtain the interference signal category corresponding to the test sample. By comparing the category output by the deep residual recognition network with the category label of the test sample itself, a quantitative measure of the recognition accuracy of the recognition network can be calculated. Through this network, the type of ETC interference signal can be accurately identified, and then the type of interference transmitter can be determined, which is convenient for tracking down and punishing illegal vehicles in the later stage.

[0038] A system for identifying ETC interference signal transmitters based on a deep residual network is characterized by comprising a first module for directionally receiving ETC interference signals and constructing an ETC interference signal data set; a second module for dividing the ETC interference signal data set into a training set, a verification set and a test set according to a predetermined ratio; a third module for performing principal component analysis on the interference signals of the training set, the verification set and the test set to obtain a low-dimensional feature representation of the interference signal; and a fourth module for constructing a deep residual network, training the deep residual network through the training set, selecting the optimal deep residual network recognition model according to the verification set, testing the optimal deep residual network recognition model through the test set, and then identifying the type of ETC interference signal.

[0039] In a further embodiment, the second module further divides the training set, the validation set and the test set into 60%, 20% and 20% of the total sample set respectively;

[0040] The third module further uses principal component analysis to obtain interference sample feature vector dimensions that account for 30% of the initial interference sample feature dimensions;

[0041] The fourth module further introduces identity mapping into the deep residual network and extracts deep abstract features of the ETC interference signal; introduces batch normalization operation and label smoothing strategy into the deep residual network; trains the deep residual network, and iteratively updates the connection weights between layers through the back propagation algorithm; uses the trained deep residual recognition network to identify the type of test interference samples; inputs the test samples into the deep residual recognition network after PCA processing, extracts features of the test samples based on the deep residual recognition network, and inputs the features extracted by the last convolutional layer into the fully connected layer; further maps the distributed features learned by the convolutional layer to the interference sample label space, and finally performs a softmax operation on the output of the fully connected layer to obtain the interference signal category of the corresponding test sample; compares the category output by the deep residual recognition network with the category label of the test sample itself, and calculates a quantitative measure of the recognition accuracy of the recognition network; accurately identifies the type of ETC interference signal through the deep residual recognition network, and then determines the type of interference transmitter.

[0042] In a further embodiment, the deep residual network in the deep residual network constructed by the fourth module includes a convolution layer, a pooling layer, a nonlinear mapping layer, a residual block and a fully connected layer; through convolution, pooling and nonlinear mapping operations, the input interference signal is processed, and the mapping between the input target type and the output target type is realized in the connection layer. Each feature mapping extracts a local feature of the interference signal through a one-dimensional convolution filter. According to the principle of local correlation, the pooling layer downsamples the features obtained by the convolution layer;

[0043] The batch normalization operation is defined as follows:

[0044]

[0045]

[0046] Here, it is assumed that one layer of the neural network has a d-dimensional input feature vector , represents the mean of the input feature vector x, represents the variance of the input feature vector x, represents the eigenvalue of the kth dimension of the input feature vector x, represents the eigenvalue of the kth dimension of the normalized input eigenvector x, represent The normalized eigenvalues ​​after scaling and translation, and Representing learnable translation and scaling parameters, BN is inserted in any layer of the network.

[0047] The label smoothing strategy is as follows:

[0048] Define the modulation signal sample as ,in, is the training label, which is 0 or 1;

[0049] In each iteration of training, set an error rate ,by The probability will be Substitute into training, The probability will be Substitution training;

[0050] When the neural network adopts the cross entropy loss function, for each sample , the loss function of the network is expressed as:

[0051]

[0052] In the formula, It indicates the probability that the network predicts the sample category to be 1. Indicates the predicted sample category (the sample category is 1 or 0), and the meanings of the other symbols are the same as above;

[0053] When label smoothing is introduced, the loss function of the network becomes:

[0054]

[0055] In the formula, It represents the training label after label smoothing, and the meanings of other symbols are the same as above;

[0056] in:

[0057]

[0058] When the label is 0, it is not put directly into the training, but replaced by a smaller number. , when the label is 1, replace it with a number closer to 1 ;

[0059] The relationship between the features of the kth layer and the k-1th layer in the deep residual network is as follows:

[0060]

[0061]

[0062] In the formula, represents the nonlinear activation function, represents the eigenvalue, represents the jth feature of the kth layer, represents the i-th eigenvalue of the k-1th layer, represents the weight value connecting the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, It represents the bias between the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, and N represents the number of neurons in the k-1th layer.

[0063] Beneficial effects: The present invention designs a corresponding deep residual recognition network model based on the characteristics of ETC transmitter interference signals, which can accurately identify the individual type of transmitters that transmit interference signals. In the field of ETC identification of individual types of transmitters that transmit interference signals, a deep residual network is introduced for the first time for identification. The designed recognition network can accurately identify the individual type of transmitters that transmit interference signals. In order to enhance the generalization ability of the recognition network, label smoothing, batch normalization, and L2 regularization are introduced to enhance the generalization ability of the recognition network. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 This is a schematic diagram of the ETC interference signal recognition model based on the deep residual network.

[0065] Figure 2 It is a work flow chart of the present invention. DETAILED DESCRIPTION

[0066] In the following description, a large number of specific details are provided to provide a more thorough understanding of the present invention. However, it is apparent to those skilled in the art that the present invention can be implemented without one or more of these details. In other examples, in order to avoid confusion with the present invention, some technical features well known in the art are not described.

[0067] The applicant believes that in the prior art, there are only methods for locating interference sources, but no methods for individually identifying interference signal transmitters. To this end, the present invention aims to provide a method for identifying ETC interference signal transmitters based on a deep residual network, and further proposes a system that can implement the above method, effectively improve the recognition accuracy of ETC interference signals, and solve the current problem of evading highway tolls.

[0068] The basic idea of ​​this paper is to design a corresponding deep residual recognition network based on the characteristics of the ETC transmitter interference signal, and perform deep feature extraction on the interference signal to accurately identify the individual type of the transmitter that emits the interference signal. First, an ETC interference signal dataset is constructed by combining the ground sensing loop to directional receive the ETC interference signal; the obtained ETC interference signal dataset is divided into training set, validation set and test set according to a certain ratio; then, based on the characteristics of the ETC interference signal, an ETC interference signal recognition model based on a deep residual network is designed, which can automatically extract abstract features that are useful for ETC interference signal recognition from the training set data without manual participation in feature extraction. In addition, label smoothing and batch normalization are introduced to further improve the generalization performance of the recognition network. Its ETC interference signal transmitter recognition model based on a deep residual network is as follows: Figure 1 The above ETC interference recognition model based on deep residual network includes 2 convolutional layers, 2 pooling layers, 3 residual blocks (each residual block includes two convolution operations and 2 nonlinear mapping operations) and 3 fully connected layers.

[0069] The specific steps are as follows:

[0070] Step 1: Construct an ETC interference signal dataset by combining the ground sensing coil to receive the ETC interference signal in a directional manner;

[0071] Step 2: Divide the ETC interference signal data set into a training set, a validation set, and a test set according to a certain ratio, where the training set, validation set, and test set account for 60%, 20%, and 20% of the total sample set.

[0072] Step 3: Since the data dimension of the ETC interference signal is large, in order to eliminate the redundant features of the feature vector of the interference sample and reduce the impact of noise, a more efficient and lower-dimensional feature representation is obtained. Therefore, the interference signals of the training set, validation set, and test set are subjected to principal component analysis (PCA) to obtain a low-dimensional feature representation of the interference signal. Assuming that the initial feature dimension of the interference signal is A, the feature vector A is PCA-transformed. The principle is to represent most of the characteristics of the initial feature dimension A with fewer feature components, and to convert the components with higher correlation in the initial feature dimension A into feature components that are unrelated to each other. The new feature components are principal components B. Experiments show that when the feature dimension of B accounts for 30% of the feature dimension of A, the accuracy of interference signal recognition is optimally improved.

[0073] Step 4: Design a deep residual network, train the deep residual network through the training set, select the optimal deep residual network recognition model based on the verification set, test the optimal deep residual network recognition model through the test set, and then identify the type of ETC interference signal.

[0074] Step 4-1, the deep residual network introduces the idea of ​​identity mapping, which to a certain extent solves the problem that the recognition performance of the recognition model decreases due to the increase in the depth of the convolutional neural network, gives full play to the advantages of the depth of the convolutional neural network, and can extract the deep abstract features of the ETC interference signal. Through the identity mapping, the convention that the output of the n-1 layer of the traditional neural network can only be given to the n layer as input is broken. After this processing, the generalization ability of the network is stronger. Among them, the basic structure of the designed deep residual recognition network mainly includes convolution layer, pooling layer, nonlinear mapping layer, convolution block and residual block. The "local connection" and "weight sharing" characteristics of the convolution operation are used to effectively solve the training overhead. Among them, the number of convolution kernels in the first two convolutional layers is 32, 64, the number of convolution kernels in the residual block is 32, 64, and the number of neurons in the fully connected layer is 32, 64, 128. The nonlinear activation function adopted by the recognition network is the ReLU activation function, and the learning rate of the network adopts a strategy of decreasing with the number of training rounds, and the initial learning rate is set to 0.001. The input interference signal is processed through operations such as convolution, pooling and nonlinear mapping, and the mapping between the input interference signal and the output target type is realized in the connection layer. Each feature mapping extracts a local feature of the interference signal through a one-dimensional convolution filter. According to the principle of local correlation, the pooling layer downsamples the features obtained by the convolution layer, thereby reducing the amount of model training while retaining useful feature information.

[0075] Step 4-2: In the process of training the deep residual network, the connection weights between the layers are continuously updated through the back propagation algorithm, so that the entire ETC interference signal recognition network can learn more accurately. The following formula reflects the relationship between the features extracted by the recognition network in the kth layer and the features in the k-1th layer.

[0076]

[0077]

[0078] In the formula Represents a nonlinear activation function, which can enhance the nonlinear expression ability of the network. After the last convolution operation, the fully connected layer is introduced. The fully connected layer plays the role of mapping the "distributed feature representation" learned by the convolution layer and the pooling layer to the sample label space. represents the nonlinear activation function, represents the eigenvalue, represents the jth feature of the kth layer, represents the i-th eigenvalue of the k-1th layer, represents the weight value connecting the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, It represents the bias between the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, and N represents the number of neurons in the k-1th layer.

[0079] Step 4-3: In order to speed up the convergence of the deep residual recognition network, batch normalization (BN) is introduced when designing the network. In a deep neural network, if the data distribution of each layer is different, it will make the network very difficult to converge and train. The role of BN is to convert the data of each layer into a state with a mean of 0 and a variance of 1, which makes the data distribution of each layer consistent and the network easier to converge. The definition of BN is as follows:

[0080]

[0081]

[0082] Here, it is assumed that one layer of the neural network has a d-dimensional input feature vector , represents the mean of the input feature vector x, represents the variance of the input feature vector x, represents the eigenvalue of the kth dimension of the input feature vector x, represents the eigenvalue of the kth dimension of the normalized input eigenvector x, represent The normalized eigenvalues ​​after scaling and translation, and Representing learnable translation and scaling parameters, BN is inserted in any layer of the network.

[0083] Step 4-4: Since deep network models are prone to overfitting, in order to enhance the generalization ability of the network, a label smoothing strategy is introduced when designing the model. Label smoothing is defined as follows:

[0084] In the study of neural networks, for classification and recognition problems, the labels of training samples are often converted into one-hot vector encoding. For the cross entropy loss function commonly used in classification and recognition, the use of one-hot encoding cannot guarantee the generalization ability of the model and is prone to overfitting. Label smoothing is a method of correcting the loss function, which allows the neural network to consider the loss contributed by incorrect labels to a certain extent.

[0085] Assume that the modulated signal sample is , taking binary classification as an example, is the training label, which is 0 or 1. In each iteration of training, Put it in the training set, but set an error rate ,by The probability will be Substitute into training, The probability will be Substitute it into the training. In this way, the model has both correct label input and wrong label input during training. The model trained in this way will not fully match every training label, but will match it to a certain extent. In this way, if an incorrect label appears, the model will be less affected.

[0086] When the neural network adopts the cross entropy loss function, for each sample , the loss function of the network is expressed as:

[0087]

[0088] In the formula Represents the network prediction sample The probability of class 1, Indicates the predicted sample category, and the meanings of the other symbols are the same as above;

[0089] When label smoothing is introduced, the loss function of the network becomes:

[0090]

[0091] in:

[0092]

[0093] When the label is 0, it is not put directly into the training, but replaced by a smaller number. Similarly, if the label is 1, replace it with a number closer to 1 .

[0094] Step 4-5, use the trained deep residual recognition network to identify the type of test interference samples. First, the test samples are processed by PCA and then input into the deep residual recognition network. The deep residual recognition network extracts features from the test samples, inputs the features extracted by the last convolutional layer into the fully connected layer, and further maps the distributed features learned by the convolutional layer to the interference sample label space. Finally, the output of the fully connected layer is subjected to a softmax operation to obtain the interference signal category of the corresponding test sample. By comparing the category output by the deep residual recognition network with the category label of the test sample itself, a quantitative measure of the recognition accuracy of the recognition network can be calculated. Through this network, the type of ETC interference signal can be accurately identified, and then the type of interference transmitter can be determined, which is convenient for tracking and punishing illegal vehicles in the later stage.

[0095] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for identifying ETC interference signal transmitters based on deep residual network, Its characteristics are The following steps are involved: Step 1: Construct an ETC interference signal dataset by combining the ground sensing coil to receive the ETC interference signal in a directional manner; Step 2: Divide the ETC interference signal data set into a training set, a validation set, and a test set according to a predetermined ratio; Step 3: Perform principal component analysis on the interference signals of the training set, validation set, and test set to obtain low-dimensional feature representation of the interference signals; Step 4: Design a deep residual network, train the deep residual network through the training set, select the optimal deep residual network recognition model according to the verification set, test the optimal deep residual network recognition model through the test set, and then identify the type of ETC interference signal; Step 4-1, introducing the identity mapping into the deep residual network and further extracting the deep abstract features of the ETC interference signal; Step 4-2: Introduce batch normalization operation and label smoothing strategy in the deep residual network; The batch normalization operation is defined as follows: Here, it is assumed that a layer of the neural network has a d-dimensional input feature vector x = (x 1 ,…,x d ), E[x (k) ] represents the mean of the input feature vector x, Var[x (k) ] represents the variance of the input feature vector x, x (k) represents the eigenvalue of the kth dimension of the input feature vector x, Represents the eigenvalue of the kth dimension of the normalized input eigenvector x, y (k) represent The normalized eigenvalue after scaling and translation, γ (k) and β (k) Represents learnable translation and scaling parameters, and BN is inserted in any layer of the network; The label smoothing strategy is as follows: Define the modulation signal sample as (x i ,y i ), where y i is the training label, which is 0 or 1; In each iteration of training, an error rate ε is set, and (x i ,y i ) into the training, and (x i ,1-y i ) into training; When the neural network adopts the cross entropy loss function, for each sample x i , the loss function of the network is expressed as: In the formula Represents the network prediction sample x i The probability of class 1, Indicates the predicted sample category, and the meanings of the other symbols are the same as above; When label smoothing is introduced, the loss function of the network becomes: In the formula, y i ′ It represents the training label after label smoothing, and the meanings of other symbols are the same as above; in: When the label is 0, it is not put directly into the training, but replaced by ε. When the label is 1, it is replaced by 1-ε. Step 4-3: Train the deep residual network and iteratively update the connection weights between layers through the back propagation algorithm; The relationship between the features of the kth layer and the k-1th layer in the deep residual network is as follows: f(x)=max(0,x) In the formula, f(x) represents the nonlinear activation function, x represents the eigenvalue, represents the jth feature of the kth layer, represents the i-th eigenvalue of the k-1th layer, represents the weight value connecting the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, represents the bias between the jth eigenvalue of the kth layer and the ith eigenvalue of the k-1th layer, and N represents the number of neurons in the k-1th layer; Step 4-4: Use the trained deep residual recognition network to identify the type of test interference samples.

2. According to claim 1, a method for identifying an ETC interference signal transmitter based on a deep residual network, Features: The training set, validation set and test set described in step 2 account for 60%, 20% and 20% of each type of interference signal data respectively; The dimension of the interference sample feature vector obtained by principal component analysis in step 3 accounts for 30% of the initial interference sample feature dimension.

3. According to claim 1, a method for identifying an ETC interference signal transmitter based on a deep residual network, It is characterized in that The deep residual network described in step 4-1 includes a convolution layer, a pooling layer, a nonlinear mapping layer, a residual block and a fully connected layer; the input interference signal is processed through convolution, pooling and nonlinear mapping operations, and the mapping between the input and the output target type is realized in the connection layer. Each feature mapping extracts a local feature of the interference signal through a one-dimensional convolution filter. According to the principle of local correlation, the pooling layer downsamples the features obtained by the convolution layer.

4. According to claim 1, a method for identifying an ETC interference signal transmitter based on a deep residual network, It is characterized in that Step 4-4 further includes: Step 4-4a, after PCA processing, the test sample is input into the deep residual recognition network, the deep residual recognition network extracts features from the test sample, and the features extracted by the last convolutional layer are input into the fully connected layer; Step 4-4b, further map the distributed features learned by the convolutional layer to the interference sample label space, and finally perform a softmax operation on the output of the fully connected layer to obtain the interference signal category of the corresponding test sample; Step 4-4c: Compare the category output by the deep residual recognition network with the category label of the test sample itself, and calculate a quantitative measure of the recognition accuracy of the recognition network; accurately identify the type of ETC interference signal through the deep residual recognition network, and then determine the type of interference transmitter.

Citation Information

Patent Citations

  • Newborn pain expression recognition method and system based on deep 3D residual network

    CN108596069A

  • Image recognition method of multi-feature source residual network

    CN110197205A

  • Video action recognition method based on residual 3D CNN and multi-modal feature fusion strategy

    CN111325155A

  • 1D-CNN-Based Distributed Optical Fiber Sensing Signal Feature Learning and Classification Method

    US20180357542A1