An ADS-B emitter identification method based on convolutional neural network and ensemble learning

Abstract

The application discloses an ADS-B radiation source identification method based on a convolutional neural network and ensemble learning, and comprises the following steps: acquiring an actual signal received by an ADS-B, and truncating actual ADS-B data according to the data characteristics of a synchronization header, a DF bit, a CA bit, an ICAO address code, a ME field message bit and a parity check bit; establishing two neural network structures; adopting a non-end-to-end and end-to-end processing mode for different data bits, inputting into the two neural network structures for training respectively, and extracting classification features from multiple dimensions; and utilizing multiple deep learning devices, and adopting a stacking method to fuse neural network feature results. Through the fusion of neural network feature results, the collective ability advantage of the individual learning device is obtained by adjusting the weight value, the obtained classifier has the advantages of high identification accuracy and strong generalization ability.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of radiation source identification, and particularly relates to an ADS-B radiation source identification method based on a convolutional neural network and ensemble learning. BACKGROUND

[0002] Research on signal receiving, analysis and identification of electromagnetic radiation sources is becoming a current hotspot. According to the received signal, the radio frequency fingerprint of the equipment is extracted, and the hardware difference of different equipment is caused by the difference in transmitter circuit design and the manufacturing tolerance of electronic devices in the production process. Due to the difference in hardware, the characteristics of the transmitted radio frequency signal can be extracted, so as to indirectly represent the transmitter characteristics and identify the identity of the radiation source. A complete fingerprint identification system includes fingerprint collection, signal preprocessing, signal feature extraction, similarity calculation and finally fingerprint identification. The main methods are divided into 1) artificial feature extraction, and 2) deep learning. Since deep learning does not require human intervention and has high accuracy, it is gradually becoming the mainstream method of electromagnetic radiation source identification.

[0003] The traditional ADS-B specific radiation source identification method has the following disadvantages:

[0004] 1. Low data utilization, only the synchronous head is used for identification, and most of the available data in the ADS-B signal is discarded, resulting in low data utilization and low identification probability;

[0005] 2. Weak generalization ability, only a single feature extraction method and identifier are used, the signal feature extraction ability is weak, and a large amount of ADS-B transmitters and massive received data cannot be accurately processed.

[0006] 3. Inaccurate input data, the ADS-B data with labels is introduced into the neural network training and device identification, causing the classification identifier to identify not the hardware characteristics of the transmitter circuit, and bringing errors to the system identification.

[0007] Therefore, an ADS-B radiation source identification method based on a convolutional neural network and ensemble learning is proposed to solve the above problems. SUMMARY

[0008] Therefore, the purpose of the present application is to provide an ADS-B radiation source identification method based on a convolutional neural network and ensemble learning, which uses a combination of non-end-to-end and end-to-end processing methods to build a convolutional neural network individual learner, and fuses and learns the primary classification results, which is beneficial to improve the identification accuracy and enhance the generalization ability of the classifier.

[0009] To achieve the above purpose, the present application provides the following technical solutions:

[0010] The application provides an ADS-B radiation source identification method based on a convolutional neural network and ensemble learning, comprising the following steps:

[0011] Step one: obtaining the actual received signal of ADS-B, and truncating the actual ADS-B data according to the data characteristics of the synchronization header, DF bit, CA bit, ICAO address code, ME field message bit and parity check bit;

[0012] Step two: establishing two neural network structures;

[0013] Step three: for different data bits in step one, using a non-end-to-end and end-to-end processing method, inputting into two neural network structures for training, and extracting classification features from multiple dimensions;

[0014] Step four: using multiple deep learningers, and using a Stacking method to fuse neural network feature results.

[0015] Further, in step one, the baseband signal y(t) actually received by the ADS-B is represented as:

[0016] y(t) = (t) + · (t) = [H(t)*(x(t))] + n(t)

[0017] In the formula, H(t) represents the additional influence of the channel, R(.) represents the radio frequency additional influence of the receiver when receiving, I(t) is the signal in-phase component, i is the imaginary unit, Q(t) is the signal quadrature component, n(t) represents the noise influence, and S(.) represents the radio frequency additional influence of the transmitter.

[0018] Further, in step two, the neural network structure comprises a first neural network and a second neural network, and the training steps of the first neural network comprise: inputting data into two one-dimensional convolution layers with a channel number of 32, two one-dimensional convolution layers with a channel number of 64, three one-dimensional convolution layers with a channel number of 128, and finally transmitting to a Softmax layer through a full connection layer Fc, wherein a ReLU function output layer is required after each convolution layer, and the size of all convolution kernels is 6;

[0019] The training steps of the second neural network comprise: A1: inputting data into a two-dimensional convolution layer with a channel number of 128, wherein the size of the convolution kernel is 8; A2: inputting the data output in step A1 into four one-dimensional convolution layers with a channel number of 128, wherein the size of the convolution kernel is 3; A3: stacking the data output in steps A1 and A2, and then transmitting to a Softmax layer through a full connection layer Fc.

[0020] Further, in step three, the specific implementation steps comprise:

[0021] B1: the second neural network is trained for the synchronization header and DF bit data, input is three-dimensional data of I, Q data and transformed frequency domain features F(w), and a Softmax layer output result S1 is obtained;

[0022] B2: for 24-bit data of the data bit being the ICAO address code, since the data bit has a label property, an ideal baseband signal is reconstructed by averaging all pulses according to the received y(t) signal The influence of the channel H(t) and the receiving R(.) in the transmission process is ignored, that is, the difference between the baseband signal transmission and the receiving of the complete radio frequency link S(.) is caused, so that there is a residual error The residual error x c (t) is transformed into the frequency domain, and the residual error frequency domain amplitude mag(FFT(x c (t))) and the phase ang(FFT(x c (t))) are used as two-dimensional input, the second neural network structure is trained, and a Softmax layer output result S2 is obtained.

[0023] B3: for 40-bit data of the data bit being the ME field and the PI field, the data is divided into two segments of 20 bits, three-dimensional data of I, Q and transformed frequency domain features F(w) are used as input to the first neural network for training, and Softmax layer output results S3 and S4 are obtained.

[0024] Further, the Softmax layer output results S1, S2, S3 and S4 are fused by using a stacking method, that is

[0025]

[0026] In the formula, w i is the weight allocated to each base learner.

[0027] The present application has the following beneficial effects:

[0028] 1. The present application is a kind of based on convolutional neural network and integrated learning ADS-B specific radiation source identification method, according to the data characteristics of ADS-B signal itself, the signal is truncated, different data in different positions is processed in different ways, and all known information is fully and reasonably utilized.

[0029] 2. The present application is a kind of based on convolutional neural network and integrated learning ADS-B specific radiation source identification method, which proposes two kinds of convolutional neural network structures for different data in different positions, uses non-end-to-end and end-to-end processing methods for different data bits, inputs the neural network for training, and extracts classification features from multiple dimensions.

[0030] 3. The ADS-B specific radiation source identification method based on the convolutional neural network and the ensemble learning utilizes multiple deep learning devices, fuses multiple feature classification results, and enhances the classification accuracy and the generalization ability of the classifier.

[0031] Additional advantages, objects, and features of the application will be set forth in part by the description that follows, and in part will become apparent to those skilled in the art upon examination of the following or can be learned from practice of the application. The objects and other advantages of the application can be realized and attained by the structure particularly pointed out in the written description and claims hereof as well as the appended drawings. BRIEF DESCRIPTION OF DRAWINGS

[0032] To make the objectives, technical solutions and beneficial effects of the present application clearer, the present application provides the following drawings for illustration:

[0033] Figure 1 The figure is an ADS-B signal transmission block diagram of the embodiment of the present application;

[0034] Figure 2 The figure is an ADS-B signal transmission waveform diagram of the embodiment of the present application;

[0035] Figure 3 The figure is a neural network architecture diagram of the embodiment of the present application;

[0036] Figure 4 The figure is an ensemble learning processing flowchart of the embodiment of the present application. DETAILED DESCRIPTION

[0037] As shown in the figure, the present application provides an ADS-B radiation source identification method based on the convolutional neural network and the ensemble learning, which comprises the following steps: Figures 1-4

[0038] Step one: acquiring the actual received signal of ADS-B, and truncating the actual ADS-B data according to the data characteristics of the synchronization header, the DF bit, the CA bit, the ICAO address code, the ME field message bit and the parity check bit;

[0039] Specifically, the process block diagram of ADS-B transmission and reception is as shown in the figure Figure 1 ​As shown, the ideal signal ADS-B transmitter ideal baseband signal is x(t), after the transmitter modulation is transformed into M(t)=S[x(t)], wherein S(x) represents the baseband signal via the transmitter, the antenna emits the system transformation, at the receiving end, the receiving signal is R(t)=R[H(t)*S(x(t))], wherein H(t) represents the additional influence of the channel, R(t) represents the demodulated baseband signal after receiving; considering the influence of noise, for the actual received signal y(t), y(t)=I(t)+i·Q(t)=R[H(t)*S(x(t))]+n(t), wherein the actual ADS-B data bit is shown in table 1:

[0040] Table 1 ADS-B message format table

[0041]

[0042] Among them, taking the most widely used ADS-B data civil flight downlink data chain format 17 message form as an example, the first five bits of the message data are fixed as 10001, the CA bits of 6-8 are the transponder device capability bits, 9-32 are 24-bit ICAO address code, 33-88 are ME field message bits, including the position, heading, speed and other flight and weather information of the aircraft, which is a variable field, and the PI bit is the parity check bit, which is different according to different transmission information, and is a variable field;

[0043] Step two: establishing a neural network structure, the neural network structure includes a first neural network and a second neural network, wherein the first neural network training step is as shown in Figure 3 a, including: input data to two channel number 32 one-dimensional convolution layer, two channel number 64 one-dimensional convolution layer, three channel number 128 one-dimensional convolution layer, and finally transmitted to the Softmax layer through the full connection layer Fc, wherein each convolution layer needs to pass through a ReLU function output layer, and the size of all convolution kernels is 6;

[0044] The second neural network training step is as shown in Figure 3 b, including: A1: input data to a two-dimensional convolution layer with 128 channels, wherein the size of the convolution kernel is 8; A2: input the data output in step A1 to four one-dimensional convolution layers with 128 channels, wherein the size of the convolution kernel is 3; A3: stack the data output in steps A1 and A2 and transmit to the Softmax layer through the full connection layer Fc;

[0045] Step three: for different data bits in step one, the non-end-to-end and end-to-end processing methods are adopted, respectively input into two kinds of neural network structures for training, and the classification features are extracted from multiple dimensions;

[0046] Specifically, the second neural network is used for training of the synchronization header and the DF bit data, input is three-dimensional data of I, Q data and transformed frequency domain features F(w), and a Softmax layer output result S1 is obtained.

[0047] For 24-bit data of the data bit being an ICAO address code, since the data bit has a label property, an ideal baseband signal is reconstructed by averaging all pulses according to the received y(t) signal The influence of the channel H(t) and the receiving R(.) in the transmission process is ignored, that is, the difference between the baseband signal transmission and the receiving of the complete radio frequency link S(.) is caused. Thus, there is a residual error The residual error x c (t) is transformed into the frequency domain, and the residual error frequency domain amplitude mag(FFT(x c (t))) and phase ang(FFT(x c (t))) are used as two-bit inputs, a second neural network structure is trained, and a Softmax layer output result S2 is obtained.

[0048] For 40-bit data of the data bit being the ME field and the PI field, the data is divided into two segments of 20 bits each, three-dimensional data of I, Q and transformed frequency domain features F(w) are used as inputs, a first neural network is trained, and Softmax layer output results S3 and S4 are obtained.

[0049] Step four: a plurality of deep learning devices are used to fuse neural network feature results by using a stacking method.

[0050] Specifically, the Softmax layer output results S1, S2, S3 and S4 are fused by using a stacking method, that is

[0051]

[0052] In the formula, w i is a weight assigned to each base learning device.

[0053] The beneficial effects of the above technical solutions are as follows: by fusing neural network feature results, combining the capabilities of a plurality of individual learning devices, adjusting weight values, the collective capability advantage of the individual learning devices can be obtained, and the obtained classifier has the advantages of high recognition accuracy and strong generalization capability.

[0054] Finally, it should be explained that the above preferred embodiments are only used to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present application.

Claims

1. An ADS-B radiation source identification method based on convolutional neural networks and ensemble learning, characterized in that, Includes the following steps: Step 1: Obtain the actual signal received by ADS-B, and truncate the actual ADS-B data according to the characteristics of each data segment: synchronization header, DF bit, CA bit, ICAO address code, ME field message bit, and parity bit. Step 2: Establish two neural network structures; Step 3: For different data bits in Step 1, non-end-to-end and end-to-end processing methods are used to input them into two different neural network structures for training, and classification features are extracted from multiple dimensions. The specific implementation steps in step three include: B1: The synchronization header and DF bit data are trained using a second neural network, with I and Q data and transformed frequency domain features as inputs. The 3D data is used to obtain the output result S1 of the Softmax layer. B2: For 24-bit data whose data bits are ICAO address codes, due to their inherent tagging properties, based on the received... The signal is averaged over all pulses to reconstruct the ideal baseband signal. Ignore the channel during transmission With Receiving The influence of this, resulting in residuals residual After transforming to the frequency domain, the amplitude of the residual in the frequency domain is... With phase As two inputs, the second neural network structure is used for training to obtain the output result S2 of the Softmax layer; B3: For the 40-bit data consisting of the ME and PI fields, divide it into two 20-bit segments and use I, Q, and transformed frequency domain characteristics. Three-dimensional data is input into the first neural network for training, and the output results S3 and S4 of the Softmax layer are obtained. Step 4: Utilize multiple deep learners and employ the Stacking method to fuse the neural network feature results.

2. The ADS-B radiation source identification method based on convolutional neural networks and ensemble learning according to claim 1, characterized in that, In step one, the baseband signal actually received by ADS-B Represented as: In the formula, Indicates the additional effects of the channel. This indicates the additional radio frequency effects of the receiver during reception. For the in-phase component of the signal, For imaginary units, For the quadrature components of the signal, Indicates the impact of noise. This indicates the additional radio frequency effects of the transmitter.

3. The ADS-B radiation source identification method based on convolutional neural networks and ensemble learning according to claim 2, characterized in that, In step two, the neural network structure includes a first neural network and a second neural network. The training steps of the first neural network include: input data is sequentially fed to two one-dimensional convolutional layers with 32 channels, two one-dimensional convolutional layers with 64 channels, and three one-dimensional convolutional layers with 128 channels, and finally transmitted to the Softmax layer after passing through the fully connected layer Fc. Each convolutional layer passes through a ReLU function output layer, and the size of all convolutional kernels is 6. The training steps of the second neural network include: A1: Input data into a two-dimensional convolutional layer with 128 channels and a kernel size of 8; A2: Input the data output from step A1 into four one-dimensional convolutional layers with 128 channels and a kernel size of 3; A3: Stack the data output from steps A1 and A2, pass them through a fully connected layer Fc, and then transmit them to the Softmax layer.

4. The ADS-B radiation source identification method based on convolutional neural networks and ensemble learning according to claim 3, characterized in that, The outputs S1, S2, S3, and S4 of the Softmax layer are fused using the Stacking method. In the formula, The weights assigned to each base learner.

Images