A multi-algorithm ADS-B emitter individual identification method and system

Abstract

The application discloses a multi-algorithm ADS-B radiation source individual identification method and system, and relates to the technical field of radiation source individual identification. The method comprises the following steps: inputting a pretreated target ADS-B radiation source signal into an ADS-B identification model to obtain a corresponding ADS-B radiation source individual; and the training process of the ADS-B identification model is as follows: training an integrated model by using a pretreated sample data set, to obtain the ADS-B identification model; wherein the integrated model is a network obtained by integrating a neural network algorithm and a random forest algorithm by using an integrated algorithm. The application can still maintain a high recognition rate in the case that the number of radiation source individuals increases.

Description

Technical Field

[0001] The present invention relates to the technical field of individual radiation source identification, and in particular to a multi-algorithm ADS-B radiation source individual identification method and system, specifically an ADS-B radiation source individual identification method and system based on a neural network algorithm and a random forest algorithm. Background Art

[0002] The Automatic Dependent Surveillance-Broadcast (ADS-B) system is an automated surveillance system that integrates wireless data communication technology and satellite navigation and positioning technology and is widely used in various aircraft. With the rapid development of air traffic in recent years, the number of aircraft has increased significantly. Accurately identifying individual ADS-B signal emitters and enabling dynamic flight monitoring can effectively prevent airspace safety accidents and provide valuable judgment basis for relevant personnel to determine the attributes of unidentified targets, analyze and locate targets, and restore air situation status. However, with the surge in airspace traffic and the increasingly complex channel environment, the stability of individual ADS-B emitter identification has significantly decreased. This problem has become one of the difficult problems that urgently need to be overcome in this field.

[0003] At present, the technology of individual radiation source identification is developing rapidly, and a large number of new and better methods for individual radiation source identification have emerged. For example, the method of using the improved extreme learning machine (ELM) to train the complexity characteristics of radiation source signals can more accurately identify signal categories. However, as the number of individual radiation sources increases, the recognition effect is poor.

[0004] With the rise and maturity of convolutional neural networks (CNNs), CNNs are often combined with constellation map features. Convolutional neural networks are used to extract features from constellation maps, which has greatly improved the recognition effect. For example, a method using a residual network (ResNet) to perform comprehensive feature extraction on I-channel data and Q-channel data improves data utilization and achieves higher accuracy. The method of introducing an attention mechanism and a soft threshold function into a deep residual network model enables the improved residual network model (Deep Residual Shrinkage Networks, DRSN) to have higher recognition accuracy and noise resistance. The SEI method based on vector map uses a neural network to extract features from the vector map for recognition, achieving good recognition results. The method of directly inputting IQ data for end-to-end individual recognition has good recognition results, but due to the uncertainty of the extracted features, the generalization performance of the model is poor.

[0005] While the above methods improve the model's recognition performance to a certain extent, in practice, their recognition performance decreases significantly as the number of individual emitters increases. Therefore, it is crucial to improve the model's stability and robustness so that it can maintain a high accuracy rate despite varying numbers of emitters. Summary of the Invention

[0006] The purpose of the present invention is to provide a multi-algorithm ADS-B emitter individual identification method and system, which can maintain a high recognition rate when the number of individual emitters increases.

[0007] To achieve the above object, the present invention provides the following solutions:

[0008] The present invention provides a multi-algorithm ADS-B emitter individual identification method, comprising:

[0009] Acquire a target ADS-B emitter signal and pre-process the target ADS-B emitter signal;

[0010] Inputting the pre-processed target ADS-B emitter signal into an ADS-B recognition model to obtain an ADS-B emitter individual corresponding to the target ADS-B emitter signal;

[0011] The training process of the ADS-B recognition model is as follows:

[0012] Acquire an original sample data set; the original sample data set includes a plurality of sample data, the sample data including input data and corresponding label data; the input data is a sample ADS-B emitter signal, and the label data is an ADS-B emitter individual corresponding to the sample ADS-B emitter signal;

[0013] Preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set; the preprocessed sample data set includes a plurality of preprocessed sample data, and the preprocessed sample data includes a sample ADS-B signal characteristic graph and corresponding label data;

[0014] The preprocessed sample data set is used to train the integration model to obtain an ADS-B recognition model; wherein the integration model is a network obtained by integrating the neural network algorithm and the random forest algorithm using an integration algorithm.

[0015] The present invention also provides a multi-algorithm ADS-B emitter individual identification system, comprising:

[0016] A data acquisition module is used to acquire a target ADS-B emitter signal and pre-process the target ADS-B emitter signal;

[0017] A prediction module is configured to input the pre-processed target ADS-B emitter signal into an ADS-B recognition model to obtain an ADS-B emitter individual corresponding to the target ADS-B emitter signal;

[0018] The training process of the ADS-B recognition model is as follows:

[0019] Acquire an original sample data set; the original sample data set includes a plurality of sample data, the sample data including input data and corresponding label data; the input data is a sample ADS-B emitter signal, and the label data is an ADS-B emitter individual corresponding to the sample ADS-B emitter signal;

[0020] Preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set; the preprocessed sample data set includes a plurality of preprocessed sample data, and the preprocessed sample data includes a sample ADS-B signal characteristic graph and corresponding label data;

[0021] The preprocessed sample data set is used to train the integration model to obtain an ADS-B recognition model; wherein the integration model is a network obtained by integrating the neural network algorithm and the random forest algorithm using an integration algorithm.

[0022] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0023] The present invention provides a multi-algorithm ADS-B emitter individual identification method and system. The integrated model involved is a network obtained by integrating a neural network algorithm and a random forest algorithm using an integrated algorithm. Compared with a single neural network, the recognition accuracy has been significantly improved, and a high recognition rate can still be maintained when the number of individual emitters increases. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0025] Figure 1 A schematic diagram of a flow chart of a method for identifying individual ADS-B emitters based on a neural network algorithm and a random forest algorithm provided in an embodiment of the present invention;

[0026] Figure 2 A schematic diagram of the ADS-B recognition model training process provided by an embodiment of the present invention;

[0027] Figure 3 ADS-B emitter signal waveform format diagram provided by an embodiment of the present invention;

[0028] Figure 4 I and Q path data diagrams of ADS-B emitter signals from two aircraft provided in an embodiment of the present invention;

[0029] Figure 5 A diagram showing the splicing of I and Q data of two different ADS-B emitter signals provided in an embodiment of the present invention;

[0030] Figure 6 A data result diagram after amplification of sample data provided by an embodiment of the present invention;

[0031] Figure 7 A schematic diagram of the structure of the integrated model provided by an embodiment of the present invention;

[0032] Figure 8 A schematic diagram of the structure of a DRSN network provided in an embodiment of the present invention;

[0033] Figure 9 A schematic diagram of the hardware voting integration provided by an embodiment of the present invention;

[0034] Figure 10 A graph showing the training loss and validation loss of the integrated model provided by an embodiment of the present invention as a function of the number of iterations;

[0035] Figure 11 A comparison chart of the accuracy of five different network models and an integrated model provided in an embodiment of the present invention;

[0036] Figure 12 A schematic diagram showing the accuracy of the integrated model as the type of radiation source increases according to an embodiment of the present invention;

[0037] Figure 13 A schematic diagram of the structure of an ADS-B emitter individual identification system based on a neural network algorithm and a random forest algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0039] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0040] Example 1

[0041] Ensemble learning can combine the advantages of each model to make the model more robust. It has been used in the fields of text and image recognition and has achieved good recognition results. For example, a lightweight incremental ensemble learning algorithm achieves linear computational complexity with respect to the number of categories, while reducing training costs and maintaining a relatively high accuracy within a certain signal-to-noise ratio range. Another example is a twin emitter individual recognition method based on ant colony feature selection and parallel classification ensemble learning, which is expanded from binary classification to multi-classification, improves the applicability of the model, and achieves better recognition results. Combining the idea of ​​ensemble learning, an embodiment of the present invention provides an ADS-B emitter individual recognition method based on a neural network algorithm and a random forest algorithm. The constructed ensemble model VRDGD-RF (ensemble model of VGG19, ResNet34, DenseNet, GoogleNet, DRSN18and Random Forest) has a significantly improved recognition accuracy compared to a single neural network, and can still maintain a high recognition rate when the number of individual emitters increases.

[0042] like Figure 1 As shown, an embodiment of the present invention provides an ADS-B emitter individual identification method based on a neural network algorithm and a random forest algorithm, which specifically includes the following steps.

[0043] Step 100: Acquire a target ADS-B emitter signal and pre-process the target ADS-B emitter signal.

[0044] Step 200: Input the pre-processed target ADS-B emitter signal into an ADS-B recognition model to obtain an ADS-B emitter individual corresponding to the target ADS-B emitter signal.

[0045] Among them, such as Figure 2 As shown, the training process of the ADS-B recognition model is:

[0046] Step A: Obtain an original sample data set; the original sample data set includes multiple sample data, and the sample data includes input data and corresponding label data; the input data is a sample ADS-B emitter signal, and the label data is the ADS-B emitter individual corresponding to the sample ADS-B emitter signal.

[0047] In the embodiment of the present invention, the input data of the original sample data set is based on the ADS-B radiation source signal in the 1090ES data link. The ADS-B radiation source signal operating at 1090MHz consists of two parts: the header and the data bit. The 8us header has a total of 4 pulses, the pulse width is 0.5us, and the positions of the 4 pulses are fixed, with intervals of 1.0us, 3.0us, and 4.5us respectively. The data bits are composed of a 5-bit downlink format field (DF), a 3-bit additional identifier segment (CA), a 24-bit unique identifier segment (ICAO), a 56-bit extended message segment (ME), and a 24-bit parity check field (PI), as shown in the following figure. Figure 3 shown.

[0048] The data bits use pulse position modulation (PPM), and since the data rate is 1Mbps, each message lasts 120μs. To eliminate the impact of the address pattern on the identification results and protect data privacy, the signal fragment corresponding to the unique identifier (ICAO) has been removed from the data according to the signal frame structure.

[0049] Step B: preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set.

[0050] First, in order to reduce the impact of data amplitude on the characteristics, the sample ADS-B emitter signal needs to be normalized; the data dimension of each sample ADS-B emitter signal is 2*4800 I-channel data and Q-channel data, and a two-dimensional picture of the I-channel data and Q-channel data is drawn; among them, Figure 4 Figure 1 shows the I-channel and Q-channel data diagrams of the ADS-B radiation source signals of two aircraft. Figure 4 The horizontal axis is the sampling point, Figure 4 The unit of the vertical axis is volts; Figure 4 (a) is the first type of I-way data diagram, Figure 4 (b) is the first type of Q-path data diagram, Figure 4 (c) is the second type of I-way data diagram, Figure 4 (d) is the second type of Q-path data diagram.

[0051] Secondly, after normalization, through the observation of the normalized I-channel data and Q-channel data, it is found that there are certain similarities and differences between the normalized I-channel data and Q-channel data, and the normalized I-channel data and Q-channel data are both characteristic data of the sample ADS-B emitter signal. In order to improve the utilization of data, the normalized I-channel data and Q-channel data are integrated together by left-right splicing to form an I-channel and Q-channel data splicing diagram, as shown below. Figure 5The I and Q data splicing diagrams of two different types of ADS-B emitter signals are shown as follows: the time of the I and Q data of the sample ADS-B emitter signal is used as the horizontal axis, and the normalized I and Q data are used as the vertical axis. The normalized I and Q data are spliced ​​together in a left-right splicing form to obtain the I and Q data splicing diagram, also known as the sample ADS-B signal characteristic diagram. Figure 5 (a) is the first type of splicing data graph, Figure 5 (b) is the second type of splicing data graph.

[0052] Finally, in order to prevent overfitting, the existing sample data should be amplified without distortion. The sample data amplification method used in the embodiment of the present invention is a method that combines brightness enhancement, rotation (90° and 180°), flipping (90° and 180°), dimming and contrast enhancement. After amplification using the above method, the sample ADS-B emitter signal can be amplified to 12 times the original value. Taking one of the sample data as an example, the amplified data result is as follows: Figure 6 Specifically, a sample data amplification method is used to amplify the sample ADS-B signal characteristic graph to obtain a preprocessed sample data set. The preprocessed sample data set includes multiple preprocessed sample data, and the preprocessed sample data includes the sample ADS-B signal characteristic graph and the corresponding label data. Figure 6 (a) is the original image, Figure 6 (b) is the image obtained by Gaussian blurring the original image. Figure 6 (c) is the image obtained after brightness enhancement of the original image. Figure 6 (d) is the image obtained by rotating the original image 90 degrees counterclockwise. Figure 6 (e) is the image obtained by dimming the original image. Figure 6 (f) is the image obtained by horizontally flipping the original image. Figure 6 (g) is the image obtained by flipping the original image horizontally and vertically. Figure 6 (h) is the image obtained by vertically flipping the original image. Figure 6 (i) is the image obtained by horizontally flipping, vertically flipping, and dimming the original image. Figure 6 (j) is the image obtained by vertically flipping and Gaussian blurring the original image. Figure 6 (k) is the image obtained by rotating the original image 180 degrees clockwise and enhancing the brightness. Figure 6 (m) The original image is rotated 180 degrees clockwise and Gaussian blurred.

[0053] Step C: Use the pre-processed sample data set to train the integration model to obtain the ADS-B recognition model; wherein, the integration model is a network obtained by integrating the neural network algorithm and the random forest algorithm using an integration algorithm. Specifically:

[0054] The preprocessed sample dataset is divided into training set, validation set and test set according to the ratio of 8:1:1.

[0055] The neural network algorithm and the random forest algorithm are integrated using the integration algorithm to build an integrated model.

[0056] The integrated model is trained using the training set and the validation set to obtain a trained integrated model.

[0057] The performance of the trained ensemble model is evaluated using the parameters obtained by training the trained ensemble model, and the trained ensemble model whose performance meets the set conditions is determined as the ADS-B recognition model.

[0058] The integrated model includes a network input terminal, a neural network integration module, a random forest integration module, a voting module and a network output terminal;

[0059] The network input end is connected to the input end of the neural network integration module and the input end of the random forest integration module respectively; the input end of the voting module is connected to the output end of the neural network integration module and the output end of the random forest integration module respectively; the output end of the voting module is connected to the network output end.

[0060] Furthermore, the neural network integration module includes a bootstrap algorithm module and a VGG network, a DRSN network, a ResNet network, a GoogleNet network and a DenseNet network respectively connected to the output end of the bootstrap algorithm module; the network input end is connected to the input end of the bootstrap algorithm module; the input end of the voting module is respectively connected to the output end of the VGG network, the output end of the DRSN network, the output end of the ResNet network, the output end of the GoogleNet network and the output end of the DenseNet network.

[0061] The DRSN network includes an input layer, a convolutional layer, a residual shrinkage layer, a normalization layer, an activation function layer, an average pooling layer, a fully connected layer and an output layer connected in sequence; the residual shrinkage layer is a layer constructed by adding an attention mechanism algorithm and a soft threshold function on the basis of the ResNet residual module.

[0062] The random forest integration module includes a bootstrap algorithm module and multiple identical random forests; the network input end is connected to the input end of the bootstrap algorithm module; the output end of the bootstrap algorithm module is respectively connected to the input end of each of the random forests; the input end of the voting module is respectively connected to the output end of each of the random forests.

[0063] The voting module has a built-in relative majority voting method.

[0064] The framework of the integrated model VRDGD-RF provided by the embodiment of the present invention is as follows Figure 7 As shown in the figure, the overall framework consists of five neural networks and a random forest. When designing the integrated model framework, the embodiment of the present invention attempted to use a variety of different deep neural networks to build the integrated model. Taking into account the convergence and recognition rate performance of the integrated model, it was finally determined that VGG19, DRSN18, ResNet34, GoogleNet, and DenseNet121 were used to build the integrated model VRDGD-RF. The number after the neural network name represents the number of layers corresponding to the network. Figure 7 Bootsraping means repeatedly sampling samples to create new samples, which is similar to resampling and can improve the robustness of the model; block1, block2, block3, block4, and block5 represent the five neural networks mentioned above respectively; block6 consists of five identical random forest RFs.

[0065] From the above integrated model framework, it can be seen that the data set is first repeatedly sampled using the bootstrap algorithm to form new data samples, which can effectively improve the robustness of the model. Then, the data are sent to the neural network module and the random forest module for recognition and classification respectively. Each block contains 5 sub-neural networks, and each sub-neural network will output a prediction result. The integrated model VRDGD-RF will obtain 30 (6*5) prediction results at this stage, and then these prediction results will be voted using the majority voting method to obtain the final recognition result of the integrated model.

[0066] In the embodiment of the present invention, the integrated model is mainly divided into two modules: neural network module and random forest. Among them, VGG, ResNet, DenseNet, and GoogleNet in the neural network module are relatively well known and will not be described in detail here. The main introduction is the DRSN18 module, whose network structure is as follows Figure 8As shown. Among them, the input picture specification is 120*120, the number of channels is 3. Conv is a convolution layer, BN is a normalization layer, Relu is an activation function, K is a convolution kernel, S is a step, Avgpool is an average pooling layer, FC is a full connection layer, RS-Block is a residual shrinkage module, and the RS-block mainly adds an attention mechanism algorithm and a soft threshold function on the basis of the ResNet residual module, which can effectively improve the anti-noise performance and robustness of the model.

[0067] Another module is a random forest module, which can randomly sample and randomly extract features, and can find the most stable and reliable results through a large number of basic tree models, so as to improve the robustness and generalization performance of the model.

[0068] The optimization of the integrated model is to optimize the combination of the sub-learners in the entire integrated model, and the embodiment of the application uses an integrated evolutionary algorithm to select the optimal sub-learner combination in the model fusion process. The embodiment of the application takes the neural network and the random forest as the sub-learner, and obtains the final classification result through majority voting of the output results of the six sub-learners, and the system structure is as shown in Figure 9 .

[0069] The embodiment of the application adopts a plurality voting method, and the mathematical model is:

[0070]

[0071] In formula (1), the represents the output of the input data x in the category C j , the recognition result with the most votes is selected as the final recognition result, and for the case of the same number of votes, one is randomly selected as the final recognition result.

[0072] Among them, the step 100 comprises:

[0073] The target ADS-B radiation source signal is normalized; the data dimension of the target ADS-B radiation source signal is 2*4800 I path data and Q path data.

[0074] The I path data and the Q path data of the target ADS-B radiation source signal are taken as the horizontal axis, and the normalized I path data and Q path data are taken as the vertical axis, and the normalized I path data and Q path data are spliced together in the form of left and right splicing to obtain a target ADS-B signal feature map.

[0075] The target ADS-B signal feature map is the target ADS-B radiation source signal after preprocessing.

[0076] The following is an example to verify the technical solution provided by the embodiment of the present invention.

[0077] This example uses an ADS-B dataset derived from the ADS-B emitter signals of 100 civil airliners. The number of samples per category in the training and test sets is roughly equal, with approximately 120 ADS-B data samples per aircraft, each consisting of I / Q data with a dimension of 2 x 4800. The training, validation, and test sets have a sample ratio of 8:1:1, with approximately equal data for each sample across the training, validation, and test sets. The ensemble model is trained and tested using the PyTorch framework, running Windows 11 on an NVIDIA RTX A6000 GPU. The initial learning rate is set to 0.001, with updates occurring every 20 epochs. The cross-entropy loss function is used, and the Adam optimizer is used.

[0078] Loss convergence analysis

[0079] Figure 10 The curves of the training loss and validation loss of the integrated model changing with the number of iterations, from Figure 10 As can be seen from the figure, the training loss of the ensemble model converges at 0.03, while the validation set loss of the ensemble model converges at 0.19. This shows that the ensemble model performs well on the test set and has good convergence performance.

[0080] Comparative analysis of recognition rates

[0081] In order to verify the effectiveness of the proposed model, the classification and recognition performance of 30 types of ADS-B signals by the integrated model VRDGD-RF and five different neural network models were simulated. The results are shown in the figure below. Figure 11 As shown in the figure, the ensemble model achieves the best recognition performance, with an accuracy of 94.52%. Next in line are DR SN18 and ResNet34, which have similar accuracy rates, reaching 91.21% and 90.56% respectively. VGG19 performs slightly worse, reaching 80%. DenseNet121 performs nearly 18% worse than the ensemble model, with an accuracy of only around 77%. The worst performing network is GoogleNet, with an accuracy of less than 70%. This demonstrates that ensemble models significantly improve performance compared to single deep neural networks.

[0082] Analysis of the impact of the number of radiation sources on recognition rate

[0083] In order to verify the change of the recognition performance of the proposed model with the number of individual radiation sources, the recognition performance of the integrated model VRDGD-RF was simulated under different numbers of individual radiation sources. The results are as follows: Figure 12 As shown. Figure 12 As can be seen, although the accuracy of the ensemble model decreases with the increase in the number of individual radiation sources, it still reaches over 80% when the number of radiation sources increases to 50. This shows that the ensemble model VRDGD-RF has good generalization ability and can maintain a high recognition rate even when the number of individual radiation sources increases, and the decline in the recognition rate is relatively gentle.

[0084] Comparative analysis of comprehensive indicators

[0085] In order to evaluate the comprehensive performance of the integrated model VRDGD-RF, the performance of the integrated model is measured by multiple evaluation indicators such as accuracy Acc, precision P, recall R and F1 value, and compared with a single neural system. The calculation formulas of each indicator are as follows:

[0086]

[0087]

[0088]

[0089]

[0090] Where tp, fp, tn, and fn are the number of samples in the true positive, false positive, true negative, and false negative classes, respectively. A true positive (tp) is a sample whose label and prediction result are both positive; a true negative (tn) is a sample whose label and prediction result are both negative; a false negative (fn) is a sample whose label is positive but whose prediction result is positive; and a false positive (fp) is a sample whose label is negative but whose prediction result is positive.

[0091] Table 1 shows the performance parameters of the ensemble model and the single model for identifying individual emitters of 30 aircraft.

[0092] Table 1 Performance comparison of different networks

[0093]

[0094] As shown in the table above, the performance indicators of the ensemble model are all improved compared to those of any individual deep neural network model. The accuracy of the ensemble model VRDGD-RF is 3.31%, 18.09%, 5.96%, 15.86%, and 4.75% higher than that of DRSN18, VGG, ResNet34, GoogleNet, and DenseNet121, respectively. Its F1 value is 3.79%, 15.95%, 5.39%, 12.19%, and 5.96% higher than that of the other five models, respectively. Both precision and recall are superior to those of a single neural network, demonstrating that the ensemble model significantly improves recognition rate, precision, recall, and F1 value compared to a single neural network.

[0095] Example 2

[0096] In order to execute the method corresponding to the above-mentioned embodiment 1 and achieve the corresponding functions and technical effects, an ADS-B emitter individual identification system based on a neural network algorithm and a random forest algorithm is provided below.

[0097] like Figure 13 As shown, the system includes:

[0098] The data acquisition module 1 is used to acquire the target ADS-B emitter signal and pre-process the target ADS-B emitter signal.

[0099] Prediction module 2 is used to input the pre-processed target ADS-B emitter signal into an ADS-B recognition model to obtain an ADS-B emitter individual corresponding to the target ADS-B emitter signal.

[0100] The training process of the ADS-B recognition model is as follows:

[0101] Acquire an original sample data set; the original sample data set includes a plurality of sample data, the sample data including input data and corresponding label data; the input data is a sample ADS-B emitter signal, and the label data is an ADS-B emitter individual corresponding to the sample ADS-B emitter signal;

[0102] Preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set; the preprocessed sample data set includes a plurality of preprocessed sample data, and the preprocessed sample data includes a sample ADS-B signal characteristic graph and corresponding label data;

[0103] The preprocessed sample data set is used to train the integrated model to obtain an ADS-B recognition model; wherein the integrated model utilizes a set.

[0104] To address the problem that the recognition accuracy of a single neural network in individual emitter identification decreases significantly as the number of individual emitters increases, the present invention provides a multi-algorithm ADS-B emitter individual identification method and system, specifically a method and system for ADS-B emitter individual identification based on a deep neural network and random forest ensemble model. The present invention uses multiple deep neural network models and random forests to train an enhanced data set, and then uses the hard voting method in the ensemble learning method to integrate the results obtained by each network model and random forest identification, making the recognition results more convincing and maintaining a high recognition rate even when the number of individual emitters increases. Experimental results show that after fusing five types of neural networks (DRSN, VGG, ResNet, GoogleNet, DenseNet) and random forests, the recognition accuracy can be improved by 3%-20% compared to a single neural network, and the recognition accuracy can still be maintained at a high level even when the number of individual emitters increases.

[0105] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0106] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A multi-algorithm ADS-B emitter individual identification method, characterized in that: include: Acquire a target ADS-B emitter signal and pre-process the target ADS-B emitter signal; Inputting the pre-processed target ADS-B emitter signal into an ADS-B recognition model to obtain an ADS-B emitter individual corresponding to the target ADS-B emitter signal; The training process of the ADS-B recognition model is as follows: Acquire an original sample data set; the original sample data set includes a plurality of sample data, the sample data including input data and corresponding label data; the input data is a sample ADS-B emitter signal, and the label data is an ADS-B emitter individual corresponding to the sample ADS-B emitter signal; Preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set; the preprocessed sample data set includes a plurality of preprocessed sample data, and the preprocessed sample data includes a sample ADS-B signal characteristic graph and corresponding label data; The preprocessed sample data set is used to train the integration model to obtain an ADS-B recognition model; wherein the integration model is a network obtained by integrating the neural network algorithm and the random forest algorithm using an integration algorithm; The integrated model includes a network input terminal, a neural network integration module, a random forest integration module, a voting module and a network output terminal; The network input end is connected to the input end of the neural network integration module and the input end of the random forest integration module respectively; The input end of the voting module is connected to the output end of the neural network integration module and the output end of the random forest integration module respectively; The output end of the voting module is connected to the network output end; The voting module has a built-in relative majority voting method.

2. The multi-algorithm ADS-B emitter individual identification method according to claim 1, characterized in that: Preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set specifically includes: Normalizing the sample ADS-B emitter signal; the data dimension of the sample ADS-B emitter signal is 2*4800 I-channel data and Q-channel data; The time of the I-channel data and Q-channel data of the sample ADS-B emitter signal is used as the horizontal axis, and the normalized I-channel data and Q-channel data are used as the vertical axis. The normalized I-channel data and Q-channel data are spliced ​​together in a left-right splicing form to obtain the sample ADS-B signal characteristic diagram; The sample data amplification method is used to amplify the sample ADS-B signal characteristic graph to obtain the preprocessed sample data set.

3. The multi-algorithm ADS-B emitter individual identification method according to claim 2, characterized in that: The sample data augmentation method is a method that combines brightness enhancement, rotation, flipping, dimming and contrast enhancement.

4. The multi-algorithm ADS-B emitter individual identification method according to claim 1, characterized in that: The neural network integration module includes a bootstrap algorithm module and a VGG network, a DRSN network, a ResNet network, a GoogleNet network and a DenseNet network respectively connected to the output end of the bootstrap algorithm module; The network input end is connected to the input end of the bootstrap algorithm module; The input end of the voting module is respectively connected to the output end of the VGG network, the output end of the DRSN network, the output end of the ResNet network, the output end of the GoogleNet network and the output end of the DenseNet network.

5. The multi-algorithm ADS-B emitter individual identification method according to claim 4, characterized in that: The DRSN network includes an input layer, a convolutional layer, a residual shrinkage layer, a normalization layer, an activation function layer, an average pooling layer, a fully connected layer and an output layer connected in sequence; The residual shrinkage layer is constructed by adding the attention mechanism algorithm and the soft threshold function to the ResNet residual module.

6. The multi-algorithm ADS-B emitter individual identification method according to claim 1, characterized in that: The random forest integration module includes a bootstrap algorithm module and multiple identical random forests; The network input end is connected to the input end of the bootstrap algorithm module; The output end of the bootstrap algorithm module is respectively connected to the input end of each random forest; The input end of the voting module is connected to the output end of each random forest respectively.

7. The multi-algorithm ADS-B emitter individual identification method according to claim 1, characterized in that: The pre-processing of the target ADS-B emitter signal specifically includes: Normalize the target ADS-B emitter signal; the data dimension of the target ADS-B emitter signal is 2*4800 I-channel data and Q-channel data; The target ADS-B emitter signal's I-channel data and Q-channel data's time are plotted on the horizontal axis, and the normalized I-channel data and Q-channel data are plotted on the vertical axis. The normalized I-channel data and Q-channel data are spliced ​​together in a left-right splicing manner to obtain the target ADS-B signal characteristic diagram. The target ADS-B signal characteristic diagram is the target ADS-B emitter signal after preprocessing.

8. A multi-algorithm ADS-B emitter individual identification system, characterized by: include: A data acquisition module is used to acquire a target ADS-B emitter signal and pre-process the target ADS-B emitter signal; A prediction module is configured to input the pre-processed target ADS-B emitter signal into an ADS-B recognition model to obtain an ADS-B emitter individual corresponding to the target ADS-B emitter signal; The training process of the ADS-B recognition model is as follows: Acquire an original sample data set; the original sample data set includes a plurality of sample data, the sample data including input data and corresponding label data; the input data is a sample ADS-B emitter signal, and the label data is an ADS-B emitter individual corresponding to the sample ADS-B emitter signal; Preprocessing the sample data in the original sample data set to obtain a preprocessed sample data set; the preprocessed sample data set includes a plurality of preprocessed sample data, and the preprocessed sample data includes a sample ADS-B signal characteristic graph and corresponding label data; The preprocessed sample data set is used to train the integration model to obtain an ADS-B recognition model; wherein the integration model is a network obtained by integrating the neural network algorithm and the random forest algorithm using an integration algorithm; The integrated model includes a network input terminal, a neural network integration module, a random forest integration module, a voting module and a network output terminal; The network input end is connected to the input end of the neural network integration module and the input end of the random forest integration module respectively; The input end of the voting module is connected to the output end of the neural network integration module and the output end of the random forest integration module respectively; The output end of the voting module is connected to the network output end; The voting module has a built-in relative majority voting method.

Images