A Method for Identifying the MFR Working Mode for an Open Environment

Abstract

The present invention relates to a method for identifying the working mode of an MFR for an open environment, including: obtaining the PDW parameters of the MFR to be measured, performing maximum-minimum normalization processing on each dimension of the PDW parameters respectively to obtain normalized PDW parameters; inputting the normalized PDW parameters into the trained working mode recognition network to obtain the MFR working mode recognition result; wherein, the working mode recognition network includes a parallel attention-temporal feature perception neural network and a classifier. The parallel attention-temporal feature perception neural network is used to extract features from the normalized PDW parameters to obtain the feature vectors corresponding to the PDW parameters. According to the comparison result between the shortest Euclidean distance from the feature vectors to each prototype and the rejection threshold, the classifier classifies the feature vectors to obtain the MFR working mode recognition result. The method of the present invention can better process the PDW parameters with temporal characteristics, making the recognition accuracy of the working mode recognition network higher and the rejection ability stronger.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar operating mode recognition, and particularly relates to a method for recognizing the operating mode of an MFR (multifunction radar) for an open environment. Background Art

[0002] In a complex electromagnetic environment, a reconnaissance aircraft inevitably receives an unknown operating mode in an existing database during actual operation. However, traditional algorithms based on parameter range matching and deep neural network algorithms are both models constructed for a closed set and do not have the ability to reject unknown classes, resulting in a great impact on their generalization ability and practicality.

[0003] Currently, Zhang Shuangtuo et al. proposed in the literature "Research on Radar Function Judgment Method and Software Design" to use the radar waveform as a medium to first determine the waveform attributes of the pulse train through the correspondence between the parameter range and the radar waveform, and then determine its operating mode. This method solves the problem of ambiguity in the inference result to a certain extent, but there is a problem of poor recognition effect for multifunction radars with complex modulation. Chen H Y et al. proposed in the literature "Function Recognition of Multi-function Radar Via CNN-GRU Neural Network" a method for recognizing the operating mode of a multifunction phased array radar based on the CNN-GRU network. This method improves the accuracy of recognizing the operating mode of a multifunction phased array radar with complex modulation to a certain extent, but it can only recognize existing operating modes and will misclassify into existing operating modes when facing unknown modes. Wang Chunsheng et al. proposed in the literature "Radiation Source Individual Recognition for Open Set Scenarios" a method based on the PN-SE prototype network model to handle open set recognition problems. This method can achieve the recognition of existing operating modes and the rejection of unknown modes to a certain extent, but it is mainly based on a convolutional neural network and has poor effects on processing PDW parameters with time series characteristics.

[0004] Hong-Ming Yang et al. proposed the Convolutional Prototype Network (CPN) in the literature "Convolutional prototype network for open set recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 44(5): 2358-2370" to address the open set recognition problem. Wang Chunsheng et al. proposed the PN-SE prototype network model in the literature "Radiation source individual recognition for open set scenarios[J]. Journal of Xi'an Jiaotong University, 2022" to handle the open set recognition problem.

[0005] Most of the above open set recognition methods are based on convolutional neural networks. However, this network has difficulty capturing high-dimensional features such as sequence changes in PDW parameters with temporal characteristics, and there is also the problem of a small receptive field, resulting in low recognition accuracy and rejection ability of the above open set recognition methods. Applying the above recognition methods to reconnaissance aircraft will lead to a decrease in the recognition accuracy of reconnaissance aircraft. Summary of the Invention

[0006] To solve the above problems existing in the prior art, the present invention provides a method for recognizing the MFR working mode for an open environment. The technical problems to be solved by the present invention are achieved through the following technical solutions:

[0007] The present invention provides a method for recognizing the MFR working mode for an open environment, including:

[0008] Step 1: Obtain the PDW parameters of the MFR to be measured, and perform maximum-minimum normalization processing on each dimension of the PDW parameters to obtain normalized PDW parameters;

[0009] Step 2: Input the normalized PDW parameters into the trained working mode recognition network to obtain the MFR working mode recognition result; wherein,

[0010] The working mode recognition network includes a parallel attention-temporal feature perception neural network and a classifier. The parallel attention-temporal feature perception neural network is used to extract features from the normalized PDW parameters to obtain feature vectors corresponding to the PDW parameters, and the classifier classifies the feature vectors to obtain the MFR working mode recognition result.

[0011] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0012] 1. The method for identifying the MFR working mode for an open environment of the present invention utilizes the designed parallel attention-temporal feature perception network with a larger receptive field, which can better process the PDW parameters with temporal characteristics, enabling the identification accuracy of the working mode recognition network to be higher and the rejection ability to be stronger. Applying it to a reconnaissance aircraft can improve the recognition accuracy of the reconnaissance aircraft.

[0013] 2. The method for identifying the MFR working mode for an open environment of the present invention utilizes the joint loss function to make the working mode recognition network pay more attention to difficult-to-train samples during the training process, thereby improving the recognition accuracy of the network.

[0014] The above description is only an overview of the technical solution of the present invention. In order to be able to more clearly understand the technical means of the present invention, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of the present invention more obvious and understandable, the following specifically gives preferred embodiments and, in conjunction with the drawings, details are described as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a schematic diagram of a method for identifying the MFR working mode for an open environment provided by an embodiment of the present invention;

[0016] Figure 2 is a schematic structural diagram of a parallel attention-temporal feature perception neural network provided by an embodiment of the present invention;

[0017] Figure 3 is an implementation framework diagram for the training and testing of a working mode recognition network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] In order to further elaborate on the technical means and effects adopted by the present invention to achieve the intended invention purpose, the following, in conjunction with the drawings and specific embodiments, details a method for identifying the MFR working mode for an open environment proposed according to the present invention.

[0019] The foregoing and other technical contents, features and effects of the present invention can be clearly presented in the following detailed description in conjunction with the drawings. Through the description of the specific embodiments, a more in-depth and specific understanding of the technical means and effects adopted by the present invention to achieve the intended purpose can be obtained. However, the attached drawings are only for reference and illustration, and are not used to limit the technical solution of the present invention.

[0020] Please refer to Figure 1 , Figure 1 is a schematic diagram of a method for identifying the MFR working mode for an open environment provided by an embodiment of the present invention. As shown in the figure, the method for identifying the MFR working mode for an open environment of this embodiment includes:

[0021] Step 1: Obtain the PDW parameters of the radar to be measured, perform maximum-minimum normalization processing on each dimension of the PDW parameters respectively, and obtain the normalized PDW parameters;

[0022] In this embodiment, the PDW parameters of the radar to be measured include five parameters: carrier frequency, PRI, bandwidth, amplitude, and pulse width. Maximum-minimum normalization processing is performed on each parameter. The formula for maximum-minimum normalization processing is as follows:

[0023] (1);

[0024] Wherein, represents the PDW parameters of the radar to be measured obtained.

[0025] Step 2: Input the normalized PDW parameters into the trained working mode recognition network to obtain the MFR working mode recognition result.

[0026] In an optional embodiment, the working mode recognition network includes a parallel attention-temporal feature perception neural network and a classifier. The parallel attention-temporal feature perception neural network is used to extract features from the normalized PDW parameters to obtain feature vectors corresponding to the PDW parameters, and the classifier classifies the feature vectors to obtain the MFR working mode recognition result.

[0027] Please refer to Figure 2 for the structural schematic diagram of the parallel attention-temporal feature perception neural network, as Figure 2 shown, the parallel attention-temporal feature perception neural network includes: a first linear layer, an activation function layer, a position encoding layer, a plurality of encoder modules, a DANet module, and a second linear layer.

[0028] Wherein, the normalized PDW parameters are sequentially passed through the first linear layer and the activation function layer to obtain non-linear transformation features. The position encoding layer adds position information to the non-linear transformation features. The plurality of encoder modules sequentially extract features from the non-linear transformation features with added position information to obtain a feature matrix corresponding to the PDW parameters. The DANet module extracts the channel-position features of the feature matrix, and the channel-position features pass through the second linear layer to obtain the feature vectors corresponding to the PDW parameters.

[0029] In this embodiment, the first linear layer contains 5 nodes, linearly connected to 16 nodes, and the activation function of the activation function layer is the ReLU activation function. The position encoding layer is to add a learnable special character and a learnable position encoding , that is, , Represents the output of the position encoding layer. The second linear layer contains 16 nodes and is linearly connected to 5 nodes.

[0030] Optionally, in this embodiment, 6 encoder modules are provided. Each encoder module has the same structure and includes: a first normalization layer, a multi-head attention layer, a first Dropout layer, a second normalization layer, a first dilated convolution layer, a first ReLU activation layer, a second Dropout layer, a second dilated convolution layer, a second ReLU activation layer, a third Dropout layer, a third dilated convolution layer, a third ReLU activation layer, and a fourth Dropout layer.

[0031] Among them, the first normalization layer, the multi-head attention layer, and the first Dropout layer are connected in sequence; the output of the first Dropout layer is added to the input of the first normalization layer and then used as the input of the second normalization layer; the second normalization layer, the first dilated convolution layer, the first ReLU activation layer, the second Dropout layer, the second dilated convolution layer, the second ReLU activation layer, the third Dropout layer, the third dilated convolution layer, the third ReLU activation layer, and the fourth Dropout layer are connected in sequence; the input of the second normalization layer is added to the output of the fourth Dropout layer and then used as the output of the encoder module.

[0032] In this embodiment, the normalization layer is Layer Normalization (LN) layer normalization, the multi-head attention layer has 5 heads, and the Dropout layer is used to avoid overfitting. Among them, the first dilated convolution layer has a convolution kernel size of 3, a stride of 1, a padding of 1, and a dilation factor of 1; the second dilated convolution layer has a convolution kernel size of 3, a stride of 1, a padding of 2, and a dilation factor of 2; the third dilated convolution layer has a convolution kernel size of 3, a stride of 1, a padding of 4, and a dilation factor of 4.

[0033] In an optional embodiment, the DANet module includes a position attention module and a channel attention module. Among them, the position attention module extracts position features from the input feature matrix to obtain corresponding position features; the channel attention module extracts channel features from the input feature matrix to obtain corresponding channel features; the channel features and the position features are added to obtain channel-position features.

[0034] In this embodiment, the channel attention module includes 3 convolutional layers, a Reshape layer, a Reshape-Transpose layer, and a Softmax layer; the position attention module includes 3 Reshape layers, a Reshape-Transpose layer, and a Softmax layer. The specific connection relationship is as Figure 2 shown.

[0035] In this embodiment, the three convolutional layer structure models in the channel attention module are exactly the same, and their convolutional kernel size is 1.

[0036] Furthermore, in combination with Figure 3 the implementation framework diagram of the training and testing of a working mode recognition network provided by the embodiment of the present invention as shown. The training process of the working mode recognition network in this embodiment and the process of using the trained working mode recognition network to perform MFR working mode recognition will be specifically described.

[0037] In this embodiment, the training process of the working mode recognition network includes:

[0038] Step a: Obtain a dataset of PDW parameters, and perform maximum-minimum normalization processing on each dimension of the PDW parameters in the dataset to obtain a training sample set. The dataset includes PDW parameters corresponding to multiple categories of MFR working modes, and each PDW parameter is attached with a corresponding working mode category label;

[0039] Step b: Input the training sample set into the constructed working mode recognition network, and extract the feature vectors corresponding to the training samples;

[0040] In this embodiment, the feature vector is a mapping expression with a dimension of 5.

[0041] Step c: Calculate the class prototype according to the feature vector;

[0042] Among them, the formula for the

[0043] th class prototype is as follows:

[0044] Among them, represents the th class prototype of the base class, represents the set composed of the training samples of the th class of the base class, is the total number of samples of the th class of training samples, represents the training sample, represents the training sample corresponding feature vector.

[0045] Step d: Calculate the Euclidean distance from the feature vector of each training sample in the training sample set to each class prototype, and use the category of the class prototype corresponding to the shortest Euclidean distance as the prediction result of the MFR working mode; among them, the Euclidean distance .

[0046] Step e: According to the prediction results of the MFR working mode corresponding to the training samples and the working mode class labels, use the joint loss function to backpropagate and update the network parameters of the working mode recognition network;

[0047] Optionally, the joint loss function is composed of an improved MCE loss function and an improved PL loss function, and the joint loss function is expressed as:

[0048] (3);

[0049] Wherein, represents the joint loss function, represents the improved MCE loss function, represents the improved PL loss function, represents the weight value of the improved PL loss function.

[0050] In this embodiment, the improved MCE loss function is expressed as:

[0051] (4);

[0052] (5);

[0053] The improved PL loss function is expressed as:

[0054] (6);

[0055] Wherein, represents the training sample, represents the training sample corresponding feature vector, represents the class prototype of the feature vector of, represents the feature vector the nearest heterogeneous prototype of, represents the first parameter of the MCE loss function; represents the second parameter of the MCE loss function; represents the parameter of the PL loss function; represents the norm operation.

[0056] It should be noted that, only serves as an intermediate parameter and has no actual meaning.

[0057] Optionally, takes a value of 0.25, takes a value of 1, takes a value of 0.5, takes a value of 0.1.

[0058] Step f: Return to step b for iterative training until the joint loss function converges, and obtain the trained working mode recognition network.

[0059] Among them, in each round of training process, for each class prototype, among the Euclidean distances corresponding to the prediction results of the MFR working modes that are consistent with the working mode class labels, select the maximum Euclidean distance as the rejection threshold of this class prototype, and update the rejection threshold of each class prototype during the iterative training process. That is, in each round of training process, for each class prototype, a rejection threshold will be obtained, and it is used as the new rejection threshold to replace the rejection threshold obtained in the previous round of training process. Until the joint loss function converges, the rejection threshold at this time is used as the final rejection threshold of the class prototype.

[0060] When using the trained working mode recognition network to recognize the working mode of the MFR to be measured, after obtaining the feature vector of the PDW parameters of the MFR to be measured, first, calculate the Euclidean distance from this feature vector to each class prototype, and determine the category of the class prototype corresponding to the shortest Euclidean distance; then, compare the shortest Euclidean distance with the rejection threshold of the determined class prototype, and obtain the MFR working mode recognition result according to the comparison result.

[0061] Among them, if the shortest Euclidean distance does not exceed the rejection threshold of the determined class prototype, then use the category of this class prototype as the MFR working mode recognition result; if the shortest Euclidean distance is greater than the rejection threshold of the determined class prototype, then determine that the MFR working mode is an unknown working mode.

[0062] In this embodiment, the MFR working mode recognition result can be mathematically expressed as:

[0063] (7);

[0064] (8);

[0065] Among them, represents the feature vector corresponding to the PDW parameters of the MFR to be measured, represents to the class prototype the shortest distance, represents the class prototype of the nearest distance represented by represents the unknown working mode.

[0066] The MFR working mode recognition method for an open environment according to the embodiments of the present invention utilizes the designed parallel attention-temporal feature perception network with a larger receptive field, which can better process PDW parameters with temporal characteristics, enabling the working mode recognition network to have a higher recognition accuracy and a stronger rejection ability. Moreover, the joint loss function makes the working mode recognition network pay more attention to difficult-to-train samples during the training process, thereby improving the recognition accuracy of the network. Applying it to a reconnaissance aircraft can improve the recognition accuracy of the reconnaissance aircraft.

[0067] Furthermore, the effect of the MFR working mode recognition method for an open environment of this embodiment is illustrated through simulation experiments.

[0068] Taking the signal intercepted from a non-cooperative multi-functional radar as an example, the parameter selection for its different working modes is shown in Table 1.

[0069] Table 1 PDW parameter table for eight working modes of a multi-functional radar

[0070]

[0071] To verify the recognition effect of the method of the present invention, the network is trained with known working modes and tested with unknown and known working modes. It should be noted that although the network model structure remains unchanged, in order to verify the recognition accuracy and reliability of the working modes of the multi-functional radar in an open environment, different working modes are used as known modes and unknown modes for verification respectively. The verification results are shown in Table 2. It can be seen from Table 2 that in the three groups of experiments, the recognition accuracy of the network for known working modes is above 94%, and the rejection rate for unknown working modes is above 98%, meeting the requirements of engineering practice and having high engineering application value.

[0072] Table 2 Verification results

[0073]

[0074] It should be noted that in this text, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant is intended to cover non-exclusive inclusion, so that an article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed. Without further limitation, an element defined by the statement "comprising an..." does not exclude the presence of additional identical elements in the article or device comprising said element. Similar words such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect.

[0075] The above content is a further detailed description of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is only limited to these descriptions. For those of ordinary skill in the technical field to which the present invention pertains, without departing from the concept of the present invention, several simple deductions or substitutions can be made, and all should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for identifying the working mode of MFR for an open environment, characterized in that, it includes: Step 1: Obtain the PDW parameters of the MFR to be measured, and perform maximum-minimum normalization processing on each dimension of the PDW parameters to obtain normalized PDW parameters; Step 2: Input the normalized PDW parameters into the trained working mode recognition network to obtain the MFR working mode recognition result; where, the working mode recognition network includes a parallel attention-temporal feature perception neural network and a classifier. The parallel attention-temporal feature perception neural network is used to extract features from the normalized PDW parameters to obtain a feature vector corresponding to the PDW parameters, and the classifier classifies the feature vector to obtain the MFR working mode recognition result; the parallel attention-temporal feature perception neural network includes: a first linear layer, an activation function layer, a position encoding layer, multiple encoder modules, a DANet module, and a second linear layer. Among them, the normalized PDW parameters pass through the first linear layer and the activation function layer in sequence to obtain non-linearly transformed features. The position encoding layer adds position information to the non-linearly transformed features. The multiple encoder modules sequentially extract features from the non-linearly transformed features with added position information to obtain a feature matrix corresponding to the PDW parameters. The DANet module extracts the channel-position features of the feature matrix, and the channel-position features pass through the second linear layer to obtain a feature vector corresponding to the PDW parameters; the DANet module includes a position attention module and a channel attention module. Among them, the position attention module extracts position features from the input feature matrix to obtain corresponding position features; the channel attention module extracts channel features from the input feature matrix to obtain corresponding channel features; the channel features and the position features are added to obtain the channel-position features.

2. The method for identifying the working mode of MFR for an open environment according to claim 1, characterized in that, the structures of the multiple encoder modules are the same, and each includes: a first normalization layer, a multi-head attention layer, a first Dropout layer, a second normalization layer, a first dilated convolution layer, a first ReLU activation layer, a second Dropout layer, a second dilated convolution layer, a second ReLU activation layer, a third Dropout layer, a third dilated convolution layer, a third ReLU activation layer, and a fourth Dropout layer, where, the first normalization layer, the multi-head attention layer, and the first Dropout layer are connected in sequence; the output of the first Dropout layer is added to the input of the first normalization layer and used as the input of the second normalization layer; the second normalization layer, the first dilated convolution layer, the first ReLU activation layer, the second Dropout layer, the second dilated convolution layer, the second ReLU activation layer, the third Dropout layer, the third dilated convolution layer, the third ReLU activation layer, and the fourth Dropout layer are connected in sequence; The output of the encoder module is obtained by adding the input of the second normalization layer to the output of the fourth Dropout layer.

3. The method for identifying the MFR working mode for an open environment according to claim 1, wherein, the training process of the working mode recognition network includes: Step a: Obtain a dataset of PDW parameters, perform min-max normalization processing on each dimension of the PDW parameters in the dataset to obtain a training sample set. The dataset includes PDW parameters corresponding to multiple categories of MFR working modes, and each PDW parameter is attached with a corresponding working mode category label; Step b: Input the training sample set into the constructed working mode recognition network, and extract the feature vectors corresponding to the training samples; Step c: Calculate the class prototypes according to the feature vectors; Step d: Calculate the Euclidean distance from the feature vector of each training sample in the training sample set to each class prototype, and use the category of the class prototype corresponding to the shortest Euclidean distance as the prediction result of the MFR working mode; Step e: According to the prediction result of the MFR working mode corresponding to the training sample and the working mode category label, use the joint loss function to backpropagate and update the network parameters of the working mode recognition network; Step f: Return to step b and iterate the training until the joint loss function converges to obtain the trained working mode recognition network; wherein, in each round of training process, for each class prototype, among the Euclidean distances corresponding to the prediction results of the MFR working modes that are consistent with the working mode category labels, select the maximum Euclidean distance as the rejection threshold of the class prototype, and update the rejection threshold of each class prototype during the iterative training process.

4. The method for identifying the MFR working mode for an open environment according to claim 3, wherein, the calculation formula of the i-th class prototype is as follows: Among them, a i represents the i-th class prototype of the base class, and T i represents the set composed of the training samples of the i-th class of the base class. |T i | is the total number of samples of the i-th class of training samples. x represents the training sample, and f(x) represents the feature vector corresponding to the training sample x.

5. The method for identifying the MFR working mode for an open environment according to claim 3, wherein, the joint loss function is composed of an improved MCE loss function and an improved PL loss function, and the joint loss function is expressed as: loss = loss1 + ωloss2; where, loss represents the joint loss function, loss1 represents the improved MCE loss function, loss2 represents the improved PL loss function, and ω represents the weight value of the improved PL loss function.

6. The method for identifying the MFR working mode for an open environment according to claim 5, wherein, the improved MCE loss function is expressed as: the improved PL loss function is expressed as: Among them, x represents the training sample, f(x) represents the feature vector corresponding to the training sample x, and a yi represents the class prototype of the feature vector f(x), and a rj represents the nearest outlier prototype of the feature vector f(x), and margin 1 represents the first parameter of the MCE loss function; λ represents the second parameter of the MCE loss function; margin 2 represents the parameter of the PL loss function; || || 2 represents the norm operation.

7. The method for identifying the MFR working mode for an open environment according to claim 1, wherein, the classifier classifies the feature vectors to obtain the MFR working mode recognition result, including: Calculate the Euclidean distance from the feature vector to each class prototype, and determine the category of the class prototype corresponding to the shortest Euclidean distance; Compare the shortest Euclidean distance with the rejection threshold of the determined class prototype, and obtain the MFR working mode recognition result according to the comparison result.

8. The MFR working mode recognition method for an open environment according to claim 7, characterized in that, the shortest Euclidean distance is compared with the rejection threshold for determining the class prototype, and according to the comparison result, the MFR working mode recognition result is obtained, including: if the shortest Euclidean distance does not exceed the rejection threshold for determining the class prototype, the class of this class prototype is used as the MFR working mode recognition result; if the shortest Euclidean distance is greater than the rejection threshold for determining the class prototype, it is determined that the MFR working mode is an unknown working mode.

Images