Specific Emitter Identification via Bispectrum-Radon Transform and Hybrid Deep Model

Speciﬁc emitter identiﬁcation is a technique that distinguishes diﬀerent emitters using radio ﬁngerprints. Feature extraction and classiﬁer selection are critical factors aﬀecting SEI performance. In this paper, we propose an SEI method using the Bispectrum-Radon transform (BRT) and a hybrid deep model. We propose BRTto characterize the unintentional modulation of pulses due to the superiority of bispectrum distributions in characterizing nonlinear features of signals. We then apply a hybrid deep model based on denoising autoencoders and a deep belief network to perform further deep feature extraction and discriminative identiﬁcation. We design an automatic dependent surveillance-broadcast signal acquisition system to capture signals and to build dataset for validating our proposed SEI method. Theoretical analysis and experimental results show that the BRT feature outperformed traditional features in characterizing UMOP, and our proposed SEI method outperformed other feature and classiﬁer combination methods.


Introduction
Specific emitter identification (SEI) is a technique that distinguishes different emitters using radio frequency fingerprints.
e technique has attracted much attention in recent decades, meeting many application requirements in electronic intelligence, cognitive radio, wireless network security, and Internet of things.To realize SEI, there are two key problems to be solved: effective feature extraction and classifier selection.
ese transforms are appropriate for analyzing nonlinear and nonstationary signals.However, they are timeconsuming when used as input features for SEI due to their large dimensions.Furthermore, the choice of kernel and basis functions influences their performance considerably.Other researchers have explored modeling the imperfections and distortions of some transmitter components to realize SEI, including the imperfections and distortions of oscillators, receivers, and transmitters [13][14][15][16].However, this kind of method offers only incomplete results because UMOP is a systematic result caused by all components of the transmitter, not just those chosen for modeling.
Since the theory of higher-order spectral analysis (HOSA) was developed in the late 1980s, research has shown its superiority in analyzing non-Gaussian and nonstationary processes, such as modeling of nonlinear systems [17,18] and signal classification [19,20].e higher-order spectrum refers to the multidimensional Fourier transform of higherorder cumulants.Compared with feature extraction methods in the time and frequency domains, the higherorder spectrum has three characteristics: (a) it contains amplitude and phase information simultaneously, (b) it includes the nonlinear component of signals, and (c) the higher-order cumulant of Gaussian noise equals to zero.Owing to these characteristics, it is possible to extract UMOP features using the higher-order spectrum.
Deep learning has garnered much attention because of its superiority for feature learning and feature representation.Deep learning stems from the simulation of the information processing process of human brain.Humans often describe the input information in hierarchical ways, with multiple levels of abstraction.e brain also appears to process information through multiple stages of transformation and representation [21].To simulate the information processing process in human brain, deep learning methods extract abstract feature in hierarchical ways.Deep learning models are built using the basic modules including convolutional neural network (CNN), restricted Boltzmann machine (RBM), and autoencoder (AE) and its variants.By combining these simple but nonlinear modules, deep learning methods are able to transfer the feature representation at one level into a higher and more abstract level.By combining enough such transformations, very complex and nonlinear features can be learned [22].According to the architectures and techniques intended for use, deep learning models can be categorized into three main classes: generative deep architectures, discriminative deep architectures, and hybrid deep architectures.Among the three architectures, the hybrid deep architectures stack several generative and discriminative modules, making them superior in some fields [23].Hybrid deep models have been used extensively to perform image classification [24,25] and video classification [26,27].
Motivated by the preceding discussion, we propose a feature extraction method called the Bispectrum-Radon transform (BRT).Furthermore, we propose an SEI method based on the BRT and a hybrid deep model.In the proposed SEI method, the BRT extracts UMOP features while the hybrid deep learning model performs further feature learning and identification.We summarize the main contributions of this paper as follows: (iv) e thorough experiments and analysis show the effectiveness and superiority of the proposed method.
e rest of this paper is organized as follows: in Section 2, we analyze the UMOP generation mechanism.We introduce the BRT in Section 3. We present the hybrid deep learning model and the proposed SEI method in Section 4. In Section 5, we present our experimental results and discussions.Section 6 concludes the paper.

Generation Mechanism of UMOP
e UMOP is caused by various components of the transmitter circuit, including oscillators, power amplifiers, and modulators.Figure 1 [28] shows the general structure of a radio transmitter.
As shown in Figure 1, before being transmitted by the antenna, the signal is processed by a series of electronic devices, including many nonlinear electronic devices, such as power amplifiers, intermediate frequency (IF) filters, and radio frequency (RF) filters. is process causes the unintentional generation of various nonlinear components that remain in the transmitted signals as UMOP.
UMOP has the characteristics of universality, uniqueness, and stability.First, UMOP is ubiquitous because differences in the performance of electronic devices and circuits are inevitable, even when for identical devices are operating under the same conditions.Meanwhile, the uniqueness of UMOP is attributable to differences in device manufacturing processes, component tolerances, material properties, and environmental factors [29].Finally, UMOP is stable because these differences between devices and circuits are subtle but consistent over a considerable period.Owing to the characteristics mentioned above, UMOP is extremely useful in solving SEI problems, making the choice of features for characterizing UMOP a significant research focus for SEI.

Bispectrum-Radon Transform
In this section, we introduce our proposed BRT to characterize UMOP.We first introduce the bispectrum distribution and the Radon transform.We then present the BRT and the algorithm for extracting the bispectrum projection vectors via the BRT.Finally, we analyze the discrimination performance of the BRT feature.

Bispectrum Feature of UMOP.
Based on HOSA theory, a bispectrum distribution is the two-dimensional Fourier transform of the third-order cumulants.It has the lowest order and the fewest calculations among higher-order spectrums.We express the bispectrum distribution of a pulse x(t) as where w 1 and w 2 are frequencies in the two-dimensional Fourier transform.C 3x (τ 1 , τ 2 ) is a third-order autocorrelation function calculated as Figure 2 shows bispectrum distributions of a theoretical pulse and a received ADS-B pulse.e subtle UMOP features are visible, demonstrating the effectiveness of the bispectrum distribution in describing UMOP.

Bispectrum-Radon Transform.
As presented in Figure 2, the bispectrum distribution shows strong potential for characterizing UMOP.However, it is time-consuming to use two-dimensional bispectrum matrices to perform SEI directly.Hence, we introduce the Radon transform to project the two-dimensional bispectrum matrices into one-dimensional vectors.
In a two-dimensional Euclidean space, a matrix can be projected in a certain direction by applying the Radon transform.Figure 3 shows the geometrical interpretation.
For a line f(w 1 , w 2 ) in a plane (w 1 , w 2 ), the Radon transform is given by where θ is the given angle and p is the distance from the origin to the line f(w 1 , w 2 ).We define the delta function δ(x) as us, we can express the delta function in equation (3) as erefore, we can understand the Radon transform as a projection process, where each point in R(θ, p) corresponds to a line in the plane.
Based on the bispectrum theory and the Radon transform, we propose a feature extraction method called Bispectrum-Radon transform (BRT).BRT projects the bispectrum matrix into a projection vector, BR(θ), using the Radon transform.BR(θ) is the linear integral of B(w 1 , w 2 ) with θ, written as us, the process for extracting bispectrum projection vectors using the BRT decomposes into two processes: (a) the estimation of the bispectrum distribution and (b) the projection using the Radon transform.We use an indirect method for bispectrum estimation.Algorithm 1 presents the whole algorithm.Figure 4 shows the process of extracting bispectrum projection vectors, and Figure 5 shows the bispectrum projection vectors from actual ADS-B signals of three aircraft.Subtle and stable differences are clearly visible in Figure 5, proving the usefulness of the projection vectors for characterizing UMOP, distinguishing different emitters and accomplishing SEI.

Hybrid Deep Model
In this section, we present an SEI method with a hybrid deep model performing identification using the bispectrum projection vectors as input.We first introduce the basic modules of the hybrid model and their hierarchical feature learning mechanism.
en we present the proposed SEI method.

RBM and DBN.
First proposed by Hinton et al. [30], DBN is a class of the generative model constructed from stacked restricted Boltzmann machines (RBMs).It learns hierarchical feature representations from complex input data.RBMs are probabilistic generative models able to learn a joint probability distribution of the training data [31].ey are good at extracting hidden and abstract features from training data.Le Roux and Bengio [32] have given the theoretical proof that RBMs are universal approximators of discrete distributions.e proof means that RBMs are excellent for modeling complex distributions.
As Figure 6 shows, an RBM consists of a visible layer and a hidden layer.e visible layer processes the input data, and the hidden layer detects the feature.
ere are full and undirected connections between visible and hidden units but none between pairs of visible or pairs of hidden units [33].In Figure 6, the visible units v � [v 1 , v 2 , . . ., v n ] T and the hidden units h � [h 1 , h 2 , . . ., h m ] T are in binary states (i.e., for Mathematical Problems in Engineering arbitraries i and j, v i ∈ 0, 1 { } and h j ∈ 0, 1 { }).w ij is the weight between unit v i and h j .
To describe the state of an RBM, the energy function of E(v, h) is defined as where a i and b j are the biases of units v i and h j , respectively.E(v, h) is determined by the weights and biases under a given v and h.According to the given energy function, a joint probability distribution p(v, h) can be obtained for all of the possible pairs of visible and hidden vectors: where Z is a normalization term and Z �  v,h exp(− E(v, h)).e probability distribution p(v) of visible layer v is the marginal distribution of p(v, h).us, we can express p(v) as As mentioned above, there are no visible-visible or hidden-hidden connections in RBM.Hence, when given a visible layer v, the state of each hidden unit is conditionally independent, and vice versa.
erefore, the conditional distributions of a visible unit and a hidden unit take the forms where σ � (1 + e − x ) − 1 .Hinton [34] proposed a learning algorithm called the contrastive divergence (CD) algorithm to learn the optimal distribution for RBM.In the CD algorithm, the original state of the visible layer is set to an arbitrary training sample with the binary state of each hidden layer deduced by equation (10).After obtaining all of the states of hidden units, the probability of each visible unit can be deduced by equation (11).erefore, a reconstructed visible layer is obtained, and a reconstruction error exists between the reconstructed visible layer and the initial input visible layer.e application of a stochastic gradient ascent algorithm optimizes the reconstruction error.With sufficient training, the reconstruction error gradually decreases and the RBM parameters stabilize.As noted, a DBN is composed of several stacked RBMs.To effectively train a DBN, Hinton et al. [30] proposed a (2) Let x k (0), x k (1), . . ., x k (M − 1) be the M points of k-th segment.Calculate the third-order cumulants of K segments , where i and j are the indexes of the third-order cumulants.
(3) e average of the third-order cumulants of all segments,  c(i, j), is calculated and is taken as the estimation of the third-order cumulant of the signal S  c(i, j) where L < M − 1. w(i, l) is the two-dimensional lag window function; (5) Calculate the bispectrum projection vector BR(θ) by conducting (5).e training process has two stages: pretraining and fine-tuning.In the first stage, the RBMs are trained from the bottom-up layer by layer.e CD algorithm is used to train and optimize the RBM parameters.In the second stage, the accumulated error in the layers after pretraining leaves the parameters less than optimal, so the backpropagation is then used to fine-tune the DBN parameters.e fine-tuning stage optimizes the entire DBN.

AE and DAE.
Motivated by the training strategy of DBN, Bengio et al. [35] proposed replacing the RBM modules with autoencoders to build the stacked autoencoder (SAE) model.SAE is also a generative model, and it could be used for unsupervised feature learning.An AE consists of an encoder and a decoder that perform the encoding and decoding process separately.Figure 7 shows the basic architecture of an AE.
In the encoding process, the hidden layer effectively extracts the intrinsic expression of the input data.e encoding function can be described as where x is the input data and h is the code obtained.W represents the weight matrix between the input layer and the hidden layer.p is an offset vector.s f ( ) is the activation function of the encoder.In this paper, we choose the sigmoid function as the activation function: In the decoding process, the output layer reconstructs the input data based on the hidden layer.
e decoding function can be described as where W T represents the weight matrix between the hidden layer and the output layer.q is an offset vector.s g ( ) is the activation function of the decoder.Hence, the learning task of AE can be abstracted into a convex optimization problem that minimizes the reconstruction error between the reconstruction and the input data.We describe the optimization problem as where θ is the parameter set and θ � W, W T , p, q  .Equation (15) shows that the AE attempts to learn an identity function such that the reconstruction is as close as possible to the input data.
Based on the basic AE model, some variants are proposed with different constraints, such as sparse autoencoders (SAE), denoising autoencoders (DAE), contractive autoencoders, and convolutional autoencoders.With different constraints, the AE variants can be applied for data denoising, dimensionality reduction and feature extraction.Among the variants of AE, DAE reconstructs the original input data from the noise-contaminated input data by minimizing the reconstruction errors.In DAE, part of the input data is randomly replaced with noise.In practice, they are usually set to 0. Hence, the added noise allows the autoencoder to extract more expressive features.erefore, the optimization problem of DAE can be described as where  x is the noise-contaminated input data.e training of the hybrid deep model has two phases: the pretraining phase and the parameter fine-tuning phase.In the pretraining phase, we choose different training methods according to the modules of each layer.e RBM layers are trained using the CD algorithm, and the DAE layer is trained using the method proposed by Vincent et al. [36].In the parameter fine-tuning phase, we use backpropagation to fine-tune the parameters of the entire model.Hence, the proposed SEI method includes the following four steps: Step 1 (feature extraction): calculate the bispectrum projection vectors of the received signal via the proposed BRT method.en, normalize the bispectrum projection vectors and divide them into training and testing sets.
Step 2 (pretraining): use the unsupervised training method mentioned above to train the hybrid deep model layer by layer and obtain the initial parameters of the model.
Step 3 (fine-tuning): use the backpropagation to finetune the parameters of the hybrid deep model based on the expected results and the output.
Step 4 (individual identification): use the testing set as the input of the hybrid deep model, and obtain the identification result.Compare the identification results with the original labels, and then calculate the identification accuracy.

Experimental Settings.
In this section, we first introduce the ADS-B signals and our signal acquisition system.en, the preprocessing method of the received signals and the parameter settings are introduced.

Signal Acquisition.
To evaluate the performance of the proposed SEI model, we conducted a signal acquisition experiment and the received signals were used to test the performance of our SEI model.
We chose ADS-B transmitters for civil aircraft as the objects to be identified.ADS-B is a type of surveillance technology using satellite and data link technology.Aircraft equipped with ADS-B systems periodically and automatically broadcast their real-time flight information, such as speed, altitude, heading, and climb rate, into the surrounding airspace.Ground stations monitor air targets in the airspace by receiving and decoding ADS-B signals.
e most widely used data link in ADS-B systems is the 1090 MHz extended squitter Model S (1090ES) data link, using pulse position modulation (PPM) recommended by the International Civil Aviation Organization (ICAO).ICAO has drafted general technical standards for ADS-B transmitters and stipulates that each aircraft has a unique aircraft registered number (ARN).erefore, it is possible to obtain the ARN of an ADS-B pulse train by decoding the received pulse train according to the general standard.After obtaining the ADS-B signals and its ARNs, we can build a dataset to test the SEI method.Figure 9 shows a received ADS-B pulse train.
We designed a signal acquisition system, shown in Figure 10, to receive and store ADS-B signals.e SEI of the ADS-B transmitters is subsequently performed offline on a computer.e sampling rate and the bandwidth of the signal acquisition system were 5 GHz and 500 MHz separately.
We carried out the experiment in Xi'an, China, over two weeks.We collected pulse trains from six aircraft to create a dataset for our SEI experiments.ere are three types of aircraft within the six aircraft: Boeing 737-89P (A and B), Airbus A321-214 (C and D), and Airbus A320-232 (E and F). Figure 12 shows the number of pulses from the six aircraft.

Parameter Settings.
According to the general technical standards, the pulse width of ADS-B signals is 0.5 ± 0.05 μs.Considering the differences in device manufacturing processes, component tolerance, and environmental factors, the duration length of a single pulse is set to be 0.8 μs, which is en the position of the point with an amplitude threshold αA max could be searched.According to the pulse width L which is specified in the general standard of the ADS-B system, the start and the end points of the pulse could be determined based on β 1 and β 2 .Finally, an ADS-B pulse with the length of L s � (β 1 + β 2 )L is obtained.In our experiment, α � 0.4, L � 2500, β 1 � 0.2, and β 2 � 1.4.We set the duration of the rising edge to be 0.3 μs, so the first 1500 points were intercepted as the rising edge of the pulse.
Using the received signals, we built a dataset to evaluate the SEI method.
e dataset consisted of the bispectrum projection vectors and their labels.e label of each ADS-B signal, i.e., the ARN of the aircraft, was deduced by decoding We used the indirect method with a 128-point fast Fourier transform to compute the bispectrum distribution of the rising edges.e projection angle in the Radon transform is a key parameter affecting the performance of BRT features, so we sampled at intervals of 5 °between 0 °and 180 °.
e performance of BRT features with different projection angles are shown in Figure 14.Finally, we set θ to 45 °, which achieved the highest identification performance.In addition, before the bispectrum projection vectors input into the hybrid deep model, we applied a normalization process.
e performance of the hybrid deep model is highly dependent on the amount of available training samples, the architecture, and the hyperparameters.e dimension of the bispectrum projection vectors is 185 in the experiment, so we used 185 units in the input layer.We applied a DAE layer to map the data to a dimension space three times as large as the original dimension, so there are 555 units in the output of the DAE.After that, the DBN consists of two RBM layers that were applied to compress and project the data to 300 and 100 dimensions.Finally, we used a logistic layer to determine the category of the input data.
Among the hyperparameters, we found the number of the epochs, the minibatch size, the learning rate, the zeromasked fraction, and the momentum have a significant influence on the performance of the model.We therefore applied a gird search method to tune and select the hyperparameters of the hybrid deep model.Finally, in the pretraining stage, we used a learning rate of 0.03, a minibatch size of 25, an epoch number of 25, and a momentum of 0.9.e zero-masked fraction is 10%, which means that 10% of the input units were randomly set to zero.Besides, in the fine-tuning stage, we used a learning rate of 0.3, a minibatch size of 25, and an epoch number of 200.

Experiments.
In this part, we performed comparative experiments to evaluate the feature extraction and identification performance of the proposed SEI method.First, we presented the output of each layer to observe the processing of the hybrid deep model.Second, we compared the feature extraction and identification performance of the existing method and the proposed model.Furthermore, we analyzed  the SEI ability of the proposed model based on the confusion matrix.Finally, we compared the performance of the features under different SNRs.

Comparative Experiment.
e comparative experiments evaluate the performance of the proposed model from two aspects: feature extraction and identification.
AIB, SIB, RIB, and CIB are integral bispectrum features with different integration paths.Figure 16 shows the integral paths of AIB, SIB, RIB, and CIB.As mentioned earlier, bispectrum has shown good performance in extracting UMOP features.However, the existing bispectrum-based features have some limitations.First, the integration paths of AIB and SIB could contain all the bispectrum points, but subtle energies caused by UMOP are easily drowned in integral calculation.Second, the comprehensiveness of RIB and CIB is related to the density of integration paths.e integration paths of RIB and CIB make it difficult to achieve uniform sampling in the process of discrete integral summation.As shown in Figure 16, if the integration paths are sparse, some points will be missed, or, if they are dense, some points will be reused in the adjacent path.Finally, the BDS features contain only the energy on the diagonal line, so a considerable amount of energies scattered in other regions is lost.
RF-DNA features are statistical RF fingerprint features that extracted from instantaneous amplitude, phase, and frequency responses [29,37,38].In [29,37,38], the RF-DNA sequences are extracted by (a) segmenting the instantaneous amplitude, phase, and frequency of the received signals, then centering and normalizing the segments; (b) calculating the metrics (i.e., variance, skewness, and kurtosis) of each segment and the entire region; and (c) arranging the metrics in a vector form.In our experiments, the received signals were segmented into 10 segments, so finally we obtained a full-dimensional N � 99 RF-DNA feature set.
HHT is the well-known technique to deal with nonlinear and nonstationary signals [39].In [39], an SEI feature set named VMD-SF is proposed based on variational mode decomposition (VMD) and spectral features.e VMD-SF feature set includes temporal and spectral features.e temporal features are extracted from the Hilbert spectrum of the decomposed VMD modes, and the spectral features are extracted from the spectral modes obtained using VMD.In our experiments, signals were decomposed into three modes using VMD, so we obtained a full-dimensional N � 18 VMD-SF feature set.
Fractal features have potential in characterizing complexity and irregularity of signals.In [40], fractal features based on box-counting dimension and variance dimension are used to distinguish phase noises between emitters.In our experiments, we applied the method proposed in [40] to obtain the fractal features of the received signals.e received signals were segmented into 10 segments, so we obtained a full-dimensional N � 20 fractal feature set.
Figure 17 shows the features extracted from actual ADS-B signals of six aircraft.
Second, to evaluate the identification ability of the hybrid deep model, six classifiers were applied for a comparison: DAE, DBN, conventional neural network (CNN), k-nearest neighbor (KNN), support vector machine (SVM), and LDA.
e DAE model used in our experiment has two stacked autoencoders and a backpropagation layer.e DAE model has an architecture of (185, 100, 60).
e DBN model consisted of two stacked RBMs with an architecture of (185, 120, 60).e LeNet-5 used in this paper consists of two sets of conventional and average pooling layers, followed by two fully connected layers and finally a softmax classifier.e size of the input image in LeNet-5 is 28 × 28, so we extended the one-dimensional feature to 1 × 784 using cubic spline interpolation and further reshaped to 28 × 28.In addition, we reduced the dimension of input data via the PCA method before inputting the data into the SVM and LDA. e ratio of energy loss in the PCA method is 0.01.A gird search method was used to tune and select the hyperparameters of classifiers.e parameter settings of the classifiers are listed in Table 1, and the results of the comparative experiments are shown in Table 2.
Table 2 shows the performance of different features with different classifiers.Among all the combinations of features and classifiers, the proposed method based on BRT features and the hybrid deep model performs best.
As mentioned above, a bispectrum is a type of higher-order spectrum that is superior in expressing intrinsic nonlinear features of signals and reducing environmental noise.e bispectrum distribution contains amplitude and phase information simultaneously and accurately characterizes nonlinear components within signals.Gaussian noise can be suppressed in the bispectrum calculation because the higher-order cumulant of Gaussian noise is equal to zero.Among the traditional bispectrum-based features, the RIB and CIB features have a problem of nonuniform sampling that introduces additional errors.e AIB and SIB have fixed integration paths, which means it is difficult to maximize the UMOP difference of different emitters' signals.Furthermore, the UMOP features are 10 Mathematical Problems in Engineering Mathematical Problems in Engineering easily drowned in integral calculation.Our previous studies [41,42] have proven that the BDS is useful for identifying signal modulations.However, although a considerable amount of energy is accumulated on the diagonal line, there are still significant scattered energies gathered in other regions.ese scattered energies contain abundant individual information, which are lost in the BDS feature.It should be noted that the six bispectrum-based features are continuous features, while the RF-DNA, VMD-SF, and fractal features are discrete features consisting of different metrics.Discrete features such as RF-DNA, VMD-SF, and fractal features characterize UMOP through different metrics.But its characterization ability depends on the validity of the selected metrics.In the experiments, we found that the variance metrics, within RF-DNA, of different aircraft were not distinguishable so does the spectral roll-off within VMD-SF and the box-counting dimension within the fractal features.For SEI problems like the identification of ADS-B emitters, the ultimate goal is to identify individual emitters discriminatively, even if these emitters operate on the same frequency and have the same modulation.Mathematical Problems in Engineering Hence, to achieve stable and effective individual recognition performance, it is necessary to extract subtle UMOP features adequately and maximize the distinction between UMOP features of different emitters.e proposed BRT extracts features using the projection of the bispectrum distribution and characterizes the UMOP within the signals.It not only contains the amplitude, the phase, and the spatial information of the distribution but also contains all of the distribution details.Hence, it offers significantly more complete extraction of UMOP features.In addition, the difference in different emitters' signals is maximized by selecting the optimal projection angle.erefore, the BRT extracted features contain more individual signal characteristics and enable the best individual recognition performance.
Among all the classifiers, the hybrid deep model performed best.Actually, KNN, LDA, and SVM are shallow learning methods, and it is difficult to use them to extract the deep structural features of high-dimensional input data.KNN and LDA are traditional linear classifiers.e SVM classifier has some nonlinear learning ability by applying kernel functions, but it remains a shallow learning algorithm when compared with the DBN and DAE.Compared with traditional classifiers, DBN, DAE, and CNN better perform deep learning of high-dimensional input data.CNN is more suitable for processing images and may not take full advantage of the BRT features.e hybrid model combines the advantages of DAE and DBN.
e experimental results demonstrate that the hybrid deep architecture performed better than the single deep architecture.In the hybrid model, the application of DAE mapped the data into a high-dimensional space and increased the data separability.In this process, DAE's recoverability ensured the stability of the projection result.en, based on the output of the DAE layer, the two-tier DBN further extracted intrinsic and nonlinear features hierarchically.Furthermore, the pretraining process provided reasonable initial parameters for the model, and the backpropagation mechanism improved the discrimination performance.
Consequently, theoretical analysis and experimental results show that the proposed model achieved satisfactory SEI performance.

Analysis of Confusion Matrix.
e experimental results in Table 2 show that the proposed model performed satisfactorily in identifying ADS-B signals of aircraft.To further evaluate and analyze the individual identification ability of the method, especially the identification ability of the same type of emitters, we analyzed the identification result in detail.As noted previously, there were three types of aircraft within the six aircraft in our experiment.Figure 18 shows the normalized confusion matrix of the identification result of the proposed model.It shows that the proposed model satisfactorily performed SEI.
e identification accuracy ranged between 85.63% and 94.29%.It is worth noting that the identification error rate of aircraft of the same type is higher than that between different types of aircraft, which agrees with our expectations.Based on the confusion matrix, we use four evaluation metrics to analyze the performance of the model further: precision, recall, specificity, and F1 score.
ese metrics are defined based on the concepts of true positive (TP), false negative (FN), false positive (FP), and true negative (TN), as illustrated in Figure 19.
e definitions of the metrics are as given below: where precision refers to the proportion of the correctly predicted samples in the positive-predicted samples.Recall refers to the proportion of the correctly predicted samples in the positive-labeled samples.Specificity refers to the proportion of the correctly predicted samples in the negativelabeled samples.e F1 score is a comprehensive indicator that combines the results of precision and recall.ese metrics value ranges from 0 to 1, with 1 representing the best output of the model.ese are commonly presented as percentages between 0 and 100.e metrics of the six aircraft are shown in Table 3.All the metrics indicate good success, demonstrating that the proposed method exhibited satisfactory performance for individual identification of the six aircraft.

Further Analysis.
We considered evaluating the performance of different features by superimposing Gaussian noise on the received signals.e function awgn.m in Matlab was applied to add white Gaussian noise to the received signals.Its calling format is y � awgn(x, snr, "measured"), where x is the received signals, snr is the signal-to-noise ratio, and the parameter "measured" means that the signal level of x is computed to determine the appropriate noise level based on the value of snr.However, it should be noted that the SNR of the processed signals is lower than the true SNR of the ADS-B signals due to the noise carried in the receiving process.erefore, we define the snr in awgn.m function as RSNR, which means the received-signal-to-noise ratio (RSNR).Figure 20 shows the identification accuracies of features under different RSNRs.e classifier is the hybrid deep model.It can be seen from Figure 20 that the performance of the features decreased as the RSNR decreases.Compared with other features, the BRT feature maintained higher identification accuracy than other features under different RSNRs.Excellent UMOP features combined with different classifiers achieved satisfactory SEI performance, while the performance using poor feature expression always remained low.

Further
In future research, we will consider further improvements to the BRT feature extraction method.Furthermore, the experimental results from Figure 18 and Table 2 show that the individual identification accuracy of the proposed method averaged 90.25% using BRT features of a single ADS-B pulse.erefore, a significant advantage of the proposed method is that the individual feature of the aircraft can be obtained based on a single pulse without decoding the entire ADS-B pulse sequence.If some ADS-B pulses are missing or overlapping, the proposed method can still identify the airplane while the traditional method based on PPM decoding may not be able to do so.Furthermore, the receiving process can accept multiple ADS-B pulses received continuously.erefore, it is possible to improve the confidence of the identification results by fusing the identification results of several continuous pulses.is is part of our future work.
Our proposed method is not only applicable to identifying ADS-B emitters in civil aviation but is also usable for other SEI application, such as the identification of airborne fire-control radar and shipborne warning radar.Many other emitters behave differently from ADS-B transmitters.eir individual information is not contained in their intentional modulation, so the identifying information of these transmitters cannot be determined solely from their modulation modes.However, with sufficient training samples, our proposed method can successfully identify these emitters.

Conclusions
Feature extraction and classifier selection are critical factors for realizing SEI.In this paper, we propose an SEI method using bispectrum theory and deep learning models.e proposed method extracts the UMOP features via BRTand uses a hybrid deep model to perform deep feature extraction and discriminative identification.Experimental results show that BRT was superior in characterizing UMOP and that our proposed method achieved satisfactory SEI performance with ADS-B emitters.Owing to the satisfactory performance of the proposed method, we have considered further investigations improve the confidence of the identification results by fusing the identification results of several continuous pulses.We also plan to investigate the performance of our proposed method in other SEI applications.

Mathematical Problems in Engineering
(i) We propose a feature extraction method based on bispectrum theory and used it to characterize UMOP.(ii) We propose an SEI method.e method extracts the UMOP features via BRT and uses a hybrid deep model to perform deep feature extraction and discriminative identification.(iii) We design a signal acquisition system to receive the automatic dependent surveillance-broadcast (ADS-B) signals of civil aircraft and use the received signals to verify the identification performance of the proposed method.

Figure 1 :
Figure 1: General structure of a radio transmitter.

Figure 2 :Figure 3 :
Figure 2: e bispectrum distributions of (a) a theoretical and (b) a received ADS-B pulse.
Figure 11   shows the structure of the system.e system includes a 1090 MHz antenna, a Tektronix DPO7054 oscilloscope, and an ADS-B RF front end.e antenna receives ADS-B signals transmitted by aircraft in surrounding airspace in real time, then the RF front end's power amplifiers and bandpass filters process the signals, amplifying the ADS-B signals and filtering out environmental noise and other signals.After that, an oscilloscope shows and stores the ADS-B signals synchronously.

Figure 8 :
Figure 8: Architecture of the hybrid deep model.

Figure 9 :Figure 10
Figure 9: A received ADS-B pulse train.

.
To observe the classification process of the hybrid deep model, the output of each layer was projected onto the 2-D plane via the linear discriminant analysis (LDA) method.Figures 15(a)-15(d) show the 2-D reduction of the input and output data of each layer.Figure 15(a) shows that although the input data overlapped, they were still distinguishable and discriminative.It means that the BRT features are effective for extracting the UMOP features.Figures 15(b)-15(d) show that the data are gradually separated and increasingly differentiated layer by layer.It shows that the hybrid deep model successfully performed further extraction of the input data hierarchically.

Figure 15 :Figure 16 :
Figure 15: 2-D reduction of the input data and output of each layer.(a) 2-D reduction for the input data, (b) 2-D reduction for output of DAE, (c) 2-D reduction for output of RBM1, and (d) 2-D reduction for output of RBM2.

Figure 18 Figure 19 :
Figure 18: e normalized confusion matrix of the experimental result.

Figure 20 :
Figure 20: e performance of features under different RSNRs.
In this part, we propose a hybrid deep model to realize deeper feature extraction and expression in SEI process.e hybrid deep model is built based on RBM and DAE.Figure8shows its architecture.A DAE is used as the first layer, two RBMs as the middle layers, and a logistic layer acts as a classification layer.e hybrid model combines the advantages of DAE and DBN both.First, the structure of the AE model and its variants maintain the recoverability of the input data, which is very important for the extraction of primary feature.Furthermore, the DAE maps the data into a high-dimensional space, which effectively increases the separability of the data.In this process, the recoverability of DAE ensures the stability of the pro- 6.3.Hybrid DeepModel.jectionresult.Moreover, DAE also extracts abstract features better than the basic AE model.Two stacked RBMs make up a two-tier DBN that used to extract deep feature from the output of the DAE layer.erefore, in the hybrid deep model, the DAE maps the input data into a high-dimensional space, and then two RBM layers project the data into the low-dimension space and extract deep features layer by6Mathematical Problems in Engineering layer.Hence, the hybrid model is able to further extract the intrinsic feature of the input data hierarchically and is resistant to noise within the input data.4.4.SEI Method via BRT and the Hybrid Deep Model.We propose an SEI method using the BRT and the hybrid deep model.e BRT is applied to extract bispectrum projection vectors of emitters' signals, and then the hybrid model is applied to realize hierarchical feature detection and classification.

Table 1 :
e parameter settings of classifiers.

Table 2 :
e results of the comparative experiments (identification accuracy in %).
Analysis.Feature extraction and classifier selection are two critical factors affecting SEI performance.It is easy to see in Table 2 that the combination of excellent individual features and classifiers resulted in satisfactory SEI performance.But it is worth noting that the difference in

Table 3 :
e confusion matrix metrics (%).Engineering performance between features was greater than the difference between classifiers.is phenomenon indicates that the selection of UMOP features is the more influential factor for SEI performance.e discriminative UMOP features could provide a good start for the identification of classifiers.