SEI-based air traffic control signal sorting method and system

Abstract

The invention relates to an air traffic control signal sorting method based on SEI, and the method comprises the steps: obtaining the RF and PW of each sample through measurement, carrying out the normalization and down-sampling of a time-frequency diagram of each sample signal, and obtaining a signal time-frequency diagram with the size of 64 * 64; respectively training a designed convolution auto-encoder and a full-connection auto-encoder by using the PDW parameter characteristics of the sample and the time-frequency graph; discarding decoding parts of the two trained auto-encoders, and splicing features output by the two auto-encoders to obtain a joint feature; k-means clustering is carried out by using the joint features to obtain pseudo labels of the samples; finely adjusting network parameters; and clustering and network fine tuning operations are alternately carried out for multiple times, so that the network extracts features suitable for clustering, and a last clustering result is taken as a final radiation source sorting result. According to the method, safety monitoring and individual identification in the flight process of the aviation target can be realized, and the problem of identification and sorting of aircrafts of the same type in the same airspace at present is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of aviation control, and in particular relates to an aviation control signal sorting method and system based on SEI. Background Art

[0002] There are subtle differences between individual electromagnetic radiation sources, and they are unique, which is the manifestation of the differences between individual electromagnetic radiation sources on the signal carrier. Due to the physical differences in hardware between individual electromagnetic radiation sources, the signal emitted by the RF end is rich in individual characteristics that are different from other electromagnetic radiation sources. This is called Specific Emitter Identification (SEI).

[0003] SEI relies on the characteristics of the physical layer of the device. The hardware of the transmitter is affected by manufacturing, process, and processing. Even if the equipment is from the same batch and the same model, there will be individual differences in its crystal oscillator, oscillator, and phase-locked loop. Since these physical differences do not depend on the specific content information transmitted by the transmitter, its characteristics are stable and effective. The RF feature recognition process mainly consists of three parts: acquisition and mechanism analysis of electromagnetic radiation source signals, selection and extraction of electromagnetic radiation source features, and classification and recognition of electromagnetic radiation source features.

[0004] In the air traffic control emitter signal sorting method based on machine learning, the selection and improvement of clustering algorithm is the focus of research.

[0005] Clustering algorithm is an unsupervised learning method that groups data according to certain clustering criteria. Compared with supervised classification methods, clustering methods require less prior information and are more in line with the requirements of non-cooperative signal sorting in air traffic control reconnaissance.

[0006] Common clustering methods can be divided into three types: prototype clustering represented by k-means, density clustering represented by DBSCAN, and hierarchical clustering represented by AGNES. The prototype clustering algorithm starts from the representative points in the sample space and iteratively updates the "key points" to solve; density clustering divides different groups according to the density of sample distribution in the feature space; the principle of hierarchical clustering is to divide the data into multiple different levels to form a tree-like clustering structure.

[0007] The k-means clustering algorithm is the most commonly used clustering method. It is also the earliest clustering algorithm used in air traffic control radiation source sorting. The advantages of this algorithm are simple principle, fast convergence, and strong interpretability; the disadvantages are that it is sensitive to isolated points and can only find spherical distribution clusters. In air traffic control radiation source sorting, it is often used for signal sorting based on PDW parameters. Usually, two parameters in PDW are selected to distribute the intercepted samples in a two-dimensional plane, and then the k-means algorithm is used to group the samples.

[0008] The DBSCAN algorithm is also a commonly used clustering method in air traffic control radiation source sorting. Compared with the k-means algorithm, it has no requirements on the shape of clusters and can cluster clusters of any shape. The DBSCAN algorithm defines two "neighborhood" parameters (ε, MinPts) for the density of each sample point.

[0009] The AGNES algorithm is a classic hierarchical clustering algorithm, which is also used in the sorting of air traffic control radiation sources. The AGNES algorithm uses a "fusion" approach to cluster. In the initial stage, each sample in the data set is a separate cluster. Only by iteratively selecting the two closest clusters to merge can data clustering be achieved.

[0010] Deep clustering has been proposed with the development of deep learning in recent years. Compared with traditional clustering algorithms, it combines deep learning and traditional clustering algorithms and has better performance. Deep clustering uses neural networks to automatically extract high-quality features of data, and then uses traditional clustering algorithms to perform clustering analysis on the data. It is currently a hot research direction in the field of data clustering, and a large amount of research work has proved its effectiveness in clustering high-dimensional data.

[0011] Deep clustering consists of two parts: neural network feature extraction and data clustering. The feature extraction part uses deep neural networks to extract the features of input data and map the data to a suitable feature space. The deep networks used for feature extraction include unsupervised neural networks such as autoencoders, generative adversarial networks, and twin networks. The data clustering part is the same as the traditional clustering process. It uses data features to select a suitable clustering algorithm for clustering. Depending on whether feature extraction and data clustering are related, deep clustering can be divided into step-by-step deep clustering and joint deep clustering.

[0012] Among them, feature extraction and data clustering are performed step by step in step-by-step deep clustering. First, the features of the original data are extracted using a deep neural network, and then the data features are clustered using a clustering algorithm. The two are not related and will not affect each other. In joint deep clustering, the data clustering part uses the feature output of the neural network in the feature extraction part as the basis for clustering to obtain the clustering results. Conversely, the feature extraction part uses the clustering results to guide the neural network to optimize and improve the feature extraction capability. The two parts are performed alternately, thereby gradually improving the performance of the final clustering.

[0013] The process of step-by-step deep clustering is consistent with that of traditional clustering algorithms, except that manual feature extraction is replaced by deep neural networks to achieve automatic feature extraction. This method can make full use of various neural networks and clustering algorithms and is easy to implement. However, since there is no correlation between feature extraction and clustering, it is difficult to ensure that the extracted features are suitable for subsequent clustering tasks. Joint deep clustering uses the results of data clustering to guide the network to update parameters in order to learn better features and promote the optimization of clustering tasks. It can be seen that joint deep clustering simultaneously optimizes the feature learning network and data clustering results, thereby ensuring that the output features of the deep neural network are oriented to clustering tasks and can achieve higher clustering performance.

[0014] How to improve the existing clustering algorithm and provide an air traffic control radiation source sorting method based on deep clustering to provide guidance for the implementation of space control and anti-interference is an urgent problem to be solved. Summary of the invention

[0015] The purpose of the present invention is to provide an aviation control signal sorting method and system based on SEI, which scans the airspace in different directions through an electronic reconnaissance and surveillance system, uses a receiving antenna to receive a radiation source signal stream mixed by multiple targets, obtains radiation source signal parameters through analysis, uses a preprocessor to sort and dilute the radiation source signal, and uses a main processor to further extract the signal characteristics after sorting, and matches them with the existing radiation source signal database characteristics to realize the radiation source signal recognition operation. After the recognition link, the individual information of the space target can be obtained, providing guidance for the implementation of space control and anti-interference.

[0016] The present invention provides an aviation control signal sorting method based on SEI, comprising the following steps:

[0017] Step 1: RF and PW of each sample are obtained by measurement, and the time-frequency diagram of each sample signal is calculated by short-time Fourier transform, and then normalized and downsampled to obtain a signal time-frequency diagram with a size of 64*64;

[0018] Step 2, respectively train the designed convolutional autoencoder and fully connected autoencoder using the PDW parameter features and time-frequency graph of the sample;

[0019] Step 3, discard the decoding parts of the two trained autoencoders, and concatenate the intra-pulse features and parameter features extracted by the two autoencoders from the signal's time-frequency diagram and PDW parameters to obtain the joint features;

[0020] Step 4: Use the joint features to perform k-means clustering to obtain pseudo labels of samples;

[0021] Step 5: concatenate the output of the autoencoder network and input it into a fully connected layer and a softmax classification layer, transform the network into a classification network, and use the pseudo-label supervision network for training to achieve fine-tuning of network parameters;

[0022] Step 6: Perform clustering and network fine-tuning operations alternately for multiple times to enable the network to extract features suitable for clustering, and take the last clustering result as the final radiation source sorting result.

[0023] Furthermore, the convolutional autoencoder in step 2 is used to extract intra-pulse features from the time-frequency diagram of the signal, and the fully connected autoencoder is used to map the parameter features composed of PW and RF to a higher-dimensional space for feature fusion.

[0024] Furthermore, the convolutional autoencoder includes 6 convolutional layers, 3 pooling layers, 1 Flatten layer, 2 fully connected layers, 1 Reshape layer and 3 upsampling layers; the Flatten layer is used to convert multidimensional input into one dimension and is located between the convolutional layer and the fully connected layer; the Reshape layer is used to convert one-dimensional input into multiple dimensions; the upsampling layer is the inverse process of pooling and is used to interpolate the original data; by reasonably setting the parameters of each network layer, the input and output of each layer of the encoding and decoding process of the entire autoencoder are consistent. The distribution is called to ensure that the network can effectively reconstruct the input from the implicit features; wherein, the filling method selects same filling, the activation function selects ReLU function, the step size of the convolution layer and the pooling layer is set to 1, the convolution kernel size is 3*3, the window size of the pooling layer and the upsampling layer is set to 2*2, the number of convolution kernels of the 1st to 6th convolution layers are 16, 8, 8, 8, 16, and 1 respectively; the number of neurons in the two fully connected layers are 32 and 512 respectively, the activation function uses ReLU function, and the autoencoder feature vector output is 32 dimensions.

[0025] Further ID, the fully connected autoencoder contains 5 hidden layers, wherein the 3rd hidden layer is the feature output layer, the number of neurons in the 1st to 5th hidden layers are 8, 8, 16, 8, 8 respectively, the activation function is the ReLU function, the number of output layer nodes is 2, and sigmoid is used as the activation function.

[0026] Furthermore, in step 3, the 32-dimensional feature matrix output by the convolutional autoencoder and the 16-dimensional feature matrix output by the fully connected autoencoder are concatenated together as the joint features of the samples corresponding to the air traffic control radiation sources.

[0027] The present invention also provides an aviation control signal sorting system based on SEI, comprising a signal sorting module, wherein the signal sorting module executes the aviation control signal sorting method based on SEI.

[0028] The present invention also provides a non-transitory computer-readable storage medium, which stores computer instructions. When the computer instructions are executed by a processor, the SEI-based aviation control signal sorting method is implemented.

[0029] The present invention also provides an electronic device, comprising:

[0030] A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the SEI-based aviation control signal sorting method as described in any one of claims 1 to 5 by executing the computer instructions.

[0031] Through the above scheme, through the SEI-based aviation control signal sorting method and system, by pre-sorting, detecting and clustering aviation control signals, safety monitoring and individual identification of aviation targets during flight can be achieved, solving the current problem of identifying and sorting aircraft of the same type in the same airspace.

[0032] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 It is a flow chart of the aviation control signal sorting method based on SEI of the present invention;

[0034] Figure 2 It is a classification diagram of radiation source signal feature extraction methods;

[0035] Figure 3 This is a flow chart of deep clustering sorting data in one embodiment of the present invention;

[0036] Figure 4 This is a structural diagram of a convolutional autoencoder in one embodiment of the present invention;

[0037] Figure 5 This is a structural diagram of a fully connected autoencoder in one embodiment of the present invention;

[0038] Figure 6 The figure is a schematic diagram of an electronic device according to the present invention. DETAILED DESCRIPTION

[0039] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0040] Ginseng Figure 1As shown, this embodiment provides an aviation control signal sorting method based on SEI, comprising the following steps:

[0041] Step S1, data preprocessing:

[0042] The RF and PW of each sample are obtained by measuring, and the time-frequency diagram of each sample signal is calculated by short-time Fourier transform. The time-frequency diagram is normalized and downsampled to obtain a signal time-frequency diagram with a size of 64*64.

[0043] Step S2, network pre-training:

[0044] The designed convolutional autoencoder and fully connected autoencoder are trained using the PDW parameter features and time-frequency graphs of the samples respectively;

[0045] Step S3, feature fusion:

[0046] The decoding parts of the two trained autoencoders are discarded, and the intra-pulse features and parameter features extracted by the two autoencoders from the time-frequency diagram and PDW parameters of the signal are concatenated to obtain the joint features;

[0047] Step S4, clustering:

[0048] Use joint features to perform k-means clustering to obtain pseudo labels of samples;

[0049] Step S5, network fine-tuning:

[0050] The output of the autoencoder network is concatenated and input into a fully connected layer and a softmax classification layer, the network is transformed into a classification network, and the network is trained with a pseudo-label supervision network to achieve fine-tuning of network parameters;

[0051] Step S6, clustering and network fine-tuning operations are performed alternately for multiple times, so that the network extracts features suitable for clustering, and the last clustering result is taken as the final radiation source sorting result. The number of clustering and network fine-tuning cycle operations is a parameter of the radiation source sorting method based on deep clustering, and the number of cycles will directly affect the sorting result.

[0052] The present invention is described in further detail below.

[0053] In the research of radiation source identification, individual identification of radiation sources occupies an important position. It matches the electromagnetic characteristics of the signal with the individual device that generates it, laying the foundation for the subsequent acquisition of basic information such as the signal model and purpose, thereby achieving accurate identification of the radiation source signal.

[0054] The selection and extraction of electromagnetic radiation source features starts from the signal, analyzes its signal characteristics, explores the manifestation of the physical differences of the transmitter in the signal, and amplifies the signal-level features caused by the physical differences of the radio frequency equipment by performing a series of signal processing on the signal. The core of the classification and identification of electromagnetic radiation source features is to use the extracted feature vectors to select appropriate and effective classifiers and classification decisions to effectively classify different individuals.

[0055] 1. Classification structure of radiation source signal feature extraction method

[0056] Extraction of electromagnetic radiation source features RF signals are the embodiment of RF devices at the signal level, so RF signals carry the RF device's own identity. However, the signal is affected by the complex electromagnetic environment, and there is a lot of redundant information in the signal, so the feature information that can characterize the differences between individual electromagnetic radiation sources is very small. Therefore, the main purpose of extracting electromagnetic radiation source features is to remove irrelevant components and redundant information in the signal, and to maximize the feature differences between signals caused by physical differences in the devices.

[0057] The radio frequency signals emitted by the electromagnetic radiation source can be divided into two types according to their different states: transient signals and steady-state signals. For the two different signal states, the feature extraction methods adopted are also different, which can be divided into feature extraction based on transient signals and feature extraction based on steady-state signals. On the one hand, the transient signal feature refers to the characteristics of the radio frequency signal generated when the radio frequency transmitter is turned on and off. Its characteristics are extremely short duration, with relatively obvious signal start and end points, and the spectrum characteristics are mostly continuous spectrum, with strong time-varying and short-term basic characteristics. In addition, the instantaneous characteristics of transient signals are also called transient signal characteristics. By selecting the instantaneous information of the signal, including the instantaneous amplitude, phase, frequency, etc. of the signal. However, if only the instantaneous characteristics of the signal are used and the classifier is used as the basis for classification, there will often be problems where the feature differences between individuals are too small and individuals cannot be distinguished. The current main methods for instantaneous feature extraction include Hilbert transform to extract the instantaneous feature parameters of the signal; wavelet decomposition, wavelet coefficients as features; finding the fractal dimension of the signal; polynomial fitting coefficients, empirical mode decomposition and other methods. Time-frequency features can describe how the frequency distribution of a signal changes over time, and can map the individual characteristics of a signal from both the time and frequency dimensions. However, they also have obvious limitations. The limitations of time resolution and frequency resolution lead to uncertainty in feature representation. Empirical mode decomposition (EMD) has excellent recognition, but the time and computational costs are high.

[0058] In terms of the steady-state characteristics of the signal, compared with the transient characteristics, its mechanism analysis and feature modeling are equally difficult. Until now, there is no clear physical description. The steady-state signal is generated by multiple sub-modules of the multi-element devices inside the transmitter. Therefore, it is difficult to use a unified mathematical description to model and analyze its system, so the extraction of the steady-state characteristics of the signal is more difficult. There are some common feature extraction methods, such as selecting conventional characteristic parameters of the signal such as carrier frequency, bandwidth, modulation parameters, etc. as signal features for classification to achieve individual identification of the radiation source. Most of them use carrier frequency jitter as a steady-state feature as the research content. High-order statistics are based on statistics and have excellent characteristics. At the theoretical level, they have excellent characteristics such as suppressing Gaussian noise and characterizing frequency domain features. Therefore, there have been many studies on signal feature extraction using high-order statistics. However, due to the generally high computational cost of high-order spectra, the most commonly used high-order spectrum is the bispectrum. Since the bispectrum has a relatively large dimensional feature, the conventional feature extraction algorithm based on the bispectrum generally needs to be processed by a dimensionality reduction algorithm to reduce the dimension of the feature space. The mainstream methods include integral bispectrum, etc. Figure 2 shown.

[0059] 2. Classifier design:

[0060] The core of classification is to construct a classification model based on the existing data feature set. The model can map its data features into any of the classes it belongs to, and is often used in data classification and data prediction. Nowadays, machine learning algorithms have been applied to various fields, among which the most popular deep learning algorithms are a natural and good classifier.

[0061] The multi-level network architecture is the core framework of deep learning. It can fit data through a large number of computing units, has good self-learning ability, and undertakes part of the work of manually screening data features in traditional methods. The network can self-explore the differences in the data set. In addition, deep learning has the characteristics of many parameters and deep structure, which can adapt well to large data sets. In many fields such as image recognition, speech recognition, text judgment, and medical image analysis, we can already see that deep learning has shined and achieved many breakthrough results.

[0062] Convolutional neural network is a classifier involved in the common field of deep learning. It is also a widely used network structure. It contains convolution computing units and has the ability of feedforward and representation learning. It can classify the input information features according to its hierarchical structure.

[0063] For classifiers, we need to consider the model's fitting ability and generalization ability, the adaptability of the data size dimension to the model, and the relationship between data features (discrete, continuous, independent, mutual). The selection of neural network classifiers requires analysis of data application scenarios and data feature sets. It is not a simple and crude choice of large and complex networks. In simple classification tasks, complex networks are often more prone to overfitting than simple networks, and the training time cost and GPU consumption cost of complex networks will be higher. Therefore, we should choose appropriate and reasonable neural network classifiers for specific scenarios.

[0064] The radiation source sorting method based on deep clustering first uses two autoencoder networks to extract intra-pulse features and parameter features from the signal's time-frequency diagram and PDW parameters, then fuses the two features to form a joint feature, and then uses the k-means clustering algorithm to cluster the sample features. The clustering labels obtained by clustering guide the network to perform supervised fine-tuning. After repeated clustering and network fine-tuning, the last clustering result is intercepted as the final sorting output. The sorting data processing flow is as follows: Figure 3 As shown in the figure, the process includes five key parts: feature extraction, feature fusion, k-means clustering, supervised classification and back propagation.

[0065] In the feature extraction part, two autoencoder networks, a convolutional autoencoder and a fully connected autoencoder, are constructed. The convolutional autoencoder is used to extract intra-pulse features from the time-frequency diagram of the signal, and the fully connected autoencoder is used to map the parameter features composed of PW and RF to a higher-dimensional space for feature fusion.

[0066] The feature fusion part fuses the output features of the encoding networks of the two autoencoders to form a joint feature containing intra-pulse modulation information and PDW parameter information;

[0067] The main task of the k-means clustering part is to cluster the data according to the joint features. This part runs the k-means algorithm. After clustering, each sample can get its corresponding cluster label, which is not the final sorting result and is called a "pseudo label".

[0068] The supervised classification and back propagation parts are mainly used to fine-tune the network. Supervised classification inputs the joint features into a classification layer with a Softmax activation function to obtain the classification output of the network, and then uses the pseudo labels obtained by k-means clustering as the labels of the samples, calculates the loss of the network, and uses the back propagation algorithm to update the parameters of the encoding network of the autoencoder to improve the feature extraction ability of the network.

[0069] In the air traffic control signal radiation source sorting method based on deep clustering, the k-means clustering of the clustering module and the parameter update of the network are performed alternately, and the parameters are updated using the clustering results. At the same time, the network after the parameter update can also learn better features to improve the correctness of the clustering results. Finally, the result of the k-means clustering output after multiple iterations is the sorting result. From the above analysis, it can be seen that the air traffic control signal radiation source sorting result based on deep clustering realizes end-to-end clustering from the original signal to the final clustering result, without the need for manual feature extraction, making the entire sorting process more intelligent.

[0070] In order to realize automatic feature extraction, two autoencoders are designed. The autoencoder designed for PDW parameter features is a common fully connected layer autoencoder, while the autoencoder for extracting intra-pulse features is a convolutional autoencoder.

[0071] In order to extract the intra-pulse features of the air traffic control radiation source signal, a convolutional neural network-based autoencoder network is designed to extract signal features from the time-frequency diagram in an unsupervised manner. The specific network structure is as follows: Figure 4 . The network contains 6 convolutional layers, 3 pooling layers, 1 Flatten layer, 2 fully connected layers, 1 Reshape layer and 3 upsampling layers. The Flatten layer is used to "flatten" the input and convert the multi-dimensional input into one dimension. It is often used between the convolutional layer and the fully connected layer. The function of the Reshape layer is opposite to that of the Flatten layer, that is, to convert the one-dimensional input into a multi-dimensional one. The upsampling layer is the inverse process of pooling and is mainly used to interpolate the original data. By reasonably setting the parameters of each network layer, the input and output of each layer of the encoding and decoding process of the entire autoencoder are symmetrically distributed, ensuring that the network can effectively reconstruct the input from the implicit features. Among them, the padding method uses "same" padding, the activation function uses the ReLU function, the step size of the convolutional layer and the pooling layer is set to 1, the convolution kernel size is 3*3, the window size of the pooling layer and the upsampling layer is set to 2*2, and the number of convolution kernels of the 1st to 6th convolutional layers is 16, 8, 8, 8, 16, and 1 respectively. The number of neurons in the two fully connected layers is 32 and 512 respectively, and the activation function also uses the ReLU function. Since the number of neurons in the fully connected layer 1 is 32, the output of the autoencoder feature vector is 32-dimensional.

[0072] The pulse width and carrier frequency of the air traffic control radiation source signal are used as the PDW parameter feature of the radiation source. This feature is only a 2-dimensional vector, which is far from the 32-dimensional intra-pulse feature. In order to better fuse the intra-pulse features and avoid the large difference in the dimensions of different features, which leads to the parameter feature information being ignored in the sorting process, a fully connected autoencoder network is designed to map the 2-dimensional features to a higher dimension. Figure 5The figure is a fully connected autoencoder structure diagram. The autoencoder contains 5 hidden layers, of which the third hidden layer is the feature output layer. The number of neurons in the first to fifth hidden layers is 8, 8, 16, 8, and 8, respectively. The activation function is the ReLU function. The number of output layer nodes is 2, and sigmoid is used as the activation function.

[0073] In order to improve the accuracy of air traffic control emitter sorting, an air traffic control emitter sorting method based on deep clustering is proposed. From the two aspects of feature selection and network fine-tuning, relevant feature extraction and network fine-tuning strategies are formulated to enable the network to mine the relevant information in the signal as much as possible while converting the mined information into features suitable for clustering algorithm classification.

[0074] In order to make the features input to the k-means clustering module contain the intra-pulse and PDW parameter information of the signal, the intra-pulse and parameter features of the air traffic control signal are fused, and the intra-pulse and parameter information of the signal is used for sorting. Feature fusion refers to the fusion of the features output by the encoder part of the two autoencoders. The specific implementation method of feature fusion is to achieve feature fusion by splicing feature matrices, that is, splicing the 32-dimensional feature matrix output by the encoder of the convolutional autoencoder with the 16-dimensional feature matrix output by the encoder of the fully connected autoencoder as the joint features of the samples corresponding to the air traffic control radiation source; in order to further improve the network performance, the network is fine-tuned after each clustering is completed, that is, the parameters of the encoder part of the autoencoder network are fine-tuned to make the output features of the network more suitable for the clustering algorithm.

[0075] The decoder parts of the two trained autoencoder networks are discarded, leaving only the encoding layer. The outputs of the encoding layer are spliced ​​together to form joint features, and then the joint features are input into a fully connected layer of 32 neurons to further fuse the intra-pulse features and parameter features. Finally, a classification layer is connected after the fully connected layer as the network output, and the activation function is set to Softmax. According to the output of the network Softmax layer and the pseudo-label output of the k-means clustering, the network loss is calculated. Since it is a classification problem, the network loss function is set to cross entropy loss. Finally, the back propagation algorithm is used to fine-tune the network to improve the network's feature extraction ability.

[0076] The present invention realizes safety monitoring and individual identification of aviation targets during flight by pre-sorting, detecting and clustering aviation control signals, thus solving the current problem of identifying and sorting aircraft of the same type in the same airspace.

[0077] This embodiment also provides an aviation control signal sorting system based on SEI, including a signal sorting module, and the signal sorting module executes the aviation control signal sorting method based on SEI.

[0078] This embodiment also provides a non-transitory computer-readable storage medium, which stores computer instructions. When the computer instructions are executed by a processor, the SEI-based aviation control signal sorting method is implemented.

[0079] Ginseng Figure 6 As shown, this embodiment also provides an electronic device, including:

[0080] A memory 201 and a processor 202, wherein the memory 201 and the processor 202 are communicatively connected to each other, the memory 201 stores computer instructions, and the processor 202 executes the SEI-based aviation control signal sorting method by executing the computer instructions.

[0081] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims

1. A method for sorting aviation control signals based on SEI, characterized in that: The steps include: Step 1: RF and PW of each sample are obtained by measurement, and the time-frequency diagram of each sample signal is calculated by short-time Fourier transform, and then normalized and downsampled to obtain a signal time-frequency diagram with a size of 64*64; Step 2, respectively train the designed convolutional autoencoder and fully connected autoencoder using the PDW parameter features and time-frequency graph of the sample; Step 3, discard the decoding parts of the two trained autoencoders, and concatenate the intra-pulse features and parameter features extracted by the two autoencoders from the signal's time-frequency diagram and PDW parameters to obtain the joint features; Step 4: Use the joint features to perform k-means clustering to obtain pseudo labels of samples; Step 5: concatenate the output of the autoencoder network and input it into a fully connected layer and a softmax classification layer, transform the network into a classification network, and use the pseudo-label supervision network for training to achieve fine-tuning of network parameters; Step 6: Perform clustering and network fine-tuning operations alternately for multiple times to enable the network to extract features suitable for clustering, and take the last clustering result as the final radiation source sorting result.

2. The method for sorting aviation control signals based on SEI according to claim 1, characterized in that: The convolutional autoencoder in step 2 is used to extract intra-pulse features from the time-frequency diagram of the signal, and the fully connected autoencoder is used to map the parameter features composed of PW and RF to a higher-dimensional space for feature fusion.

3. The method for sorting aviation control signals based on SEI according to claim 2, characterized in that: The convolutional autoencoder includes 6 convolutional layers, 3 pooling layers, 1 Flatten layer, 2 fully connected layers, 1 Reshape layer and 3 upsampling layers; the Flatten layer is used to convert multi-dimensional input into one dimension and is located between the convolutional layer and the fully connected layer; the Reshape layer is used to convert one-dimensional input into multiple dimensions; the upsampling layer is the inverse process of pooling and is used to interpolate the original data; by reasonably setting the parameters of each network layer, the input and output of each layer of the entire autoencoder encoding and decoding process are symmetrically distributed. The network is configured to ensure that the network can effectively reconstruct the input from the implicit features; wherein, the filling method is same filling, the activation function is ReLU function, the step size of the convolution layer and the pooling layer is set to 1, the convolution kernel size is 3*3, the window size of the pooling layer and the upsampling layer is set to 2*2, the number of convolution kernels of the 1st to 6th convolution layers are 16, 8, 8, 8, 16, and 1 respectively; the number of neurons in the two fully connected layers are 32 and 512 respectively, the activation function is ReLU function, and the autoencoder feature vector output is 32 dimensions.

4. The method for sorting aviation control signals based on SEI according to claim 3, characterized in that: The fully connected autoencoder contains 5 hidden layers, wherein the 3rd hidden layer is the feature output layer, the number of neurons in the 1st to 5th hidden layers are 8, 8, 16, 8, 8 respectively, the activation function is the ReLU function, the number of output layer nodes is 2, and sigmoid is used as the activation function.

5. The method for sorting aviation control signals based on SEI according to claim 4, characterized in that: In step 3, the 32-dimensional feature matrix output by the convolutional autoencoder is concatenated with the 16-dimensional feature matrix output by the fully connected autoencoder as the joint features of the samples corresponding to the air traffic control radiation source.

6. An aviation control signal sorting system based on SEI, characterized in that: It comprises a signal sorting module, and the signal sorting module executes the aviation control signal sorting method based on SEI as described in any one of claims 1-5.

7. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, and when the computer instructions are executed by a processor, an SEI-based aviation control signal sorting method as described in any one of claims 1 to 5 is implemented.

8. An electronic device, characterized in that: include: A memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the SEI-based aviation control signal sorting method as described in any one of claims 1 to 5 by executing the computer instructions.

Images