Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier

Measurement xxx (xxxx) xxx

Contents lists available at ScienceDirect

Measurement journal homepage: www.elsevier.com/locate/measurement

Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier Ahmed. K. Ali ⇑, Ergun Erçelebi Department of Electric and Electronic Engineering, Gaziantep University, Gaziantep 27310, Turkey

a r t i c l e

i n f o

Article history: Received 24 August 2019 Received in revised form 1 November 2019 Accepted 7 November 2019 Available online xxxx Keywords: Radius-constellation Pattern recognition APSK DVB-S2X standard Satellite communications standard

a b s t r a c t Modulation recognition of transmitted signals remains a significant concern in smart, modern communication systems such software-defined radios. Applying machine learning algorithms to interesting features extracted from the input signals are widely used for classification of such signals. Here, we propose leveraging novel higher order spectra features (HOSF) to a classification algorithm based on neural network properties to deal with modulation recognition problems specified for the M-APSK DVB-S2X modulation signals standard. The HOSF characteristics of signals are extracted under Additive white Gaussian noise (AWGN) channel, and four individual parameters are defined for distinguishing modulation signals from the set, {16, 32, 64} -APSK. This approach makes the recogniser more intelligent and improves its success rate of classification. The results illustrate excellent classification accuracy obtained at a low SNR of 0 dB, which demonstrates the potential of combining these proposed features with a neural network classifier for the application of M-APSK modulation classification. Ó 2019 Elsevier Ltd. All rights reserved.

1. Introduction The last decade witnessed significant development in satellite communication systems that offer data-oriented services, including TV-broadcasting, internet services, voice call services, and high-quality video streaming. As a result, the quantity of orbits and the volume of broadband data traffic has increased to such an extent that new automatic monitoring systems are required. A core part of such monitoring systems is an automatic modulation classifier, and comprehensive details on existing classification techniques have been documented [1]. Amplitude Phase Shift Keying (APSK) is an attractive modulation type for digital transmission compared with nonlinear satellite channels due to its mapping of symbols on constellation shapes and spectral efficiency as well as its correlating robustness against nonlinear distortion. Correspondingly, it has recently been extensively reliable as seen with its superior performance with the second generation of the digital video broadcasting satellite standard, DVB-S2 [2]. On the other hand, an essential feature of the DVB-S2 standard is that it supports higher-order modulation schemes, such as 16-APSK and 32-APSK, which are designed specifically for nonlinear channels and are sensitive to the characteristics of satellite transponders [3].

⇑ Corresponding author at: Department of Electrical and Electronics Engineering, University of Gaziantep, 27310 Gaziantep, Turkey E-mail address: [email protected] (A.K. Ali).

Modulation recognition is an intermediate step before signal detection and demodulation, and it demonstrates an important option for contemporary radio systems to prepare modulation information of signals prior to signal demodulation and decoding. Numerous real-time applications exist, such as software-defined radio, signal recognition, threat assessment, and spectrum monitoring. On the other hand, many military applications also demand the same automatic detection of modulation types selected by adversarial radio systems, including signal interception and jamming. In non-cooperative communication channels, efficient recognition of modulation schemes from the transmitting to the receiving terminals is significant for subsequent signal demodulation and information extraction. This perspective that represents a major aspect to automatic modulation classification (AMC) has become an interesting topic for communication engineers while remaining challenging to researchers working in the field of signal processing. Weaver et al. [4] formalised the foundational theory of AMC in 1969 when they suggested the AMC scheme in a Stanford University internal report, and has become well-known by communication engineers. AMC plays an essential role in many applications, such as military and civilian applications involving electromagnetic countermeasures and spectrum management [5,6]. Modulation identification of digital and analogue signals are distinguished into two categories of algorithms. The first is the maximum likelihood (ML) guess method based on decision theory

https://doi.org/10.1016/j.measurement.2019.107257 0263-2241/Ó 2019 Elsevier Ltd. All rights reserved.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

2

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

[7], which was also studied in [8] that computes the likelihood functions of all candidate modulation schemes, and the modulation type is selected on the value of maximal likelihood. This approach is challenging to implement in real-time due to the complexity in calculating the ML functions [9]. The second method is the traditional feature-based AMC algorithm studied in [10], which is well-known in many published materials. In non-cooperative communication channels, the lack of prior information in the received signals’ modulated types along with the requirement for a practical application to have instant recognition, pattern recognition methods are typically used with the feature-based AMC algorithm [8]. The recognition method extracts a distinct feature prior to the datasets for each pattern recognition and applies a machine learning method to perform the recognition of the received pattern type [11]. Feature-based (FB) algorithms could be realised in the steps of pre-processing data, signal feature extraction, and the classifier, which decides the correct pattern. As statistical features could share many advantages with robust performance and low complexity [12], FB-AMC methods have been investigated in many recent articles [13,14] along with addressing cyclic feature[15] and wavelet transforms [16], while the statistical higher-order moments and cumulants have also been studied [17–19]. Typical feature extraction methods include instantaneous time domain-based features, such as the maximum value of the power spectral density (PSD) of normalized-centered instantaneous amplitude features, which is a well-documented approach [20–22]. Similarly, the maximum value of the magnitude of the th

discrete Fourier transform (DFT) of the k power of the received signal (Ck ) feature was presented in [23] for the classification of M-PSK signals, as well as in [24] with orders k ¼ 2; 4 for a blind guess of the frequency offset of PSK and QAM modulation schemes. Using amplitude variance in combination with only one time of 4th power transformation and fast Fourier transform (FFT) could ensure that accuracy reaches 99% for recognising signals such as QPSK, 8PSK, 16QAM, and 32QAM at the lower value of optical signal noise ratio (OSNR). The simulation signals are transmitted through two types of channel additives, Gaussian noise channel and long-distance standard single-mode fibre channel [25]. However, these feature-based methods include very high order statistics [26], and multiple features were studied in [27,28] regarding the 16-APSK and 32-APSK schemes. Some of these features are correlated with the channel noise-to-power ratio, so it becomes necessary to estimate the SNR before modulation recognition. Nevertheless, accurate estimation of the SNR remains difficult, especially when it is low. Also, other features are sensitive even at a constant SNR. Hence, the features mentioned above are unsuitable for the recognition of APSK signals. Some investigations confirm that the APSK modulation types have difficulty in nonlinear amplifiers due their constellation shape [29–31], which subsequently challenges the identification process of the pattern in the classifier block. Conventional classifiers feature random forests [32] while intelligent classifiers have been developed with machine learning algorithms. Classifiers based on artificial neural networks (ANN) combined with genetic algorithms have also been studied [33] along with ANN-based classifiers that include multi-layer perceptron [34,21] and support vector machines (SVM) [35,36]. The classification methods designed on ANN and SVM have shown excellent performance for blind modulation recognition where there exists less prior knowledge on the transmuted modulation signals. Likewise, the SVM machine learning algorithm to classify QAM modulation signals transmitted through optical transmission channel was studied with details in [37]. Nevertheless, in FB-AMCs, the machine learning algorithms perform merely as a mapping function between the extracted signal features and a pattern of

multi-hypotheses. Some investigations confirmed that to obviate the extracted signal features, machine learning methods must be adopted with deep learning methods [38,39] to accomplish sufficient learning. However, these methods all demand a great deal of memory and require lengthy processing times due to their computational complexity. Most of the previous research described have concentrated on modulation recognition algorithms that are corrupted by additive white Gaussian noise (AWGN), which is a typical noisy channel model widely used in studies related to information theory for imitating the influence of naturally occurring random processes. Also, this assumption provides a good understanding of the behaviour of systems and reflects major trends. Therefore, in this paper, we consider the transmission channel corrupted by AWGN and derive a new instantaneous relationship between the maximum amplitude and the parameters of (a) the Power Spectral Density (PSD) in terms of the maximum amplitude with a normalized-centered instantaneous amplitude, and (b) the maximum value of the magnitude of the discrete Fourier transform (DFT) of the transmitted signal embedded with k powers 2; 3; 4; 5; 7; and 9, to the signal amplitude. To overcome the problem that occurs at APSK nonlinear amplifier caused by its constellation shape by enhancing the performance of the APSK classifier, two new instantaneous features are derived. The first feature is based on the maximum value of the PSD and the second on dividing the k power order as {4th/2th, 7th/3th, and 9th/5th} to create three new distinct instantaneous parameters. These features are attractive because they have a value less with influence from channel noise and remains nearly constant even with a low SNR. Another significant advantage with these features is their low computational complexity, as it only requires the computation of the PSD and DFT to the received signal. All proposed features are less sensitive to symbol synchronisation as well as being insensitive to the carrier frequency and phase offsets. As a result, the proposed APSK classifier is comprised of four features extracted from the received signals, including {cMax ; C4 =C2 ; C7 =C3 ; C9 =C5 }. The remainder of this paper is arranged as follows. Section 2 introduces a typical representation of the APSK modulation scheme. The details of the APSK-classifier block diagram, the proposed higher-order spectral features (HOSF), and the distribution of signals with the new extracted features are presented in Section 3. In Section 4, we propose our AMC-based MLP method with its computational complexity detailed. A detailed description of the classification method follows with a comprehensive analysis in Section 5. Simulation results are introduced and discussed in Section 6 with a discussion followed by the conclusions of the research presented in Sections 7 and 8, respectively. 2. Signal model and APSK constellation The typical frame of a sampled signal after passing the matched filter is expressed as [1]

rðnÞ ¼ aejð2pf o nTþho Þ

1 X

WðlÞhðnT  lT þ eT T Þ þ gðnÞ

ð1Þ

l¼1

where a is the attenuation factor, f o is the frequency offset, and ho and gðnÞ are the phase offset and additive noise, respectively. The transmitted signal WðlÞ is assumed to have unity energy with complex envelope constellation points on the I -Q plane and h is the residual channel effect caused by timing errors eT with the symbol timing of T. The parameters T, f o , and eT have known assumptions. These hypotheses simplifies the signal model of the matched filter so that it can be written as

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

r ðnÞ ¼ aWðnÞejð2pf o nTþho Þ þ gðnÞ

WðnÞ ¼

N X

ti qðN  iÞ

ð2Þ ð3Þ

i¼1

where r(n) is the complex envelope of the received samples. The real and imaginary parts of w(n) are the I and Q components of the transmitted signal, which could be constructed in a twodimensional scatter plot to obtain the constellation diagram of the transmitted signal as each M-APSK type has a constellation diagram. Also, the ti constellation points at N include symbols inside each data block and the signal power function qð‘Þ is usually a root-raised cosine with a constant roll-off factor. The constellation points of the APSK modulation of {16, 32, and 64} order are considered in this work and given in Table 1, and the constellation diagrams for the APSK modulation types are presented in Fig. 1.

Table 1 The constellation points of the {16, 32, and 64} APSK modulation type. Constellation points of 16, 32, and 64 APSK 8 9 r 2 ½1; 2; Ms 2 ½4; 12; m 2 ½0; . . . :; 11; if M ¼ 16 < = m ti 2 rej2pMS ; where r 2 ½1; 2; 3; M s 2 ½4; 12; 16; m 2 ½0; . . . :; 15; if M ¼ 32 : ; r 2 ½1; 2; 3; 4; M s 2 ½4; 12; 16; 32; m 2 ½0; . . . :; 31; if M ¼ 64

3

3. APSK modulation recognition method We propose a new concept for the APSK classification algorithm using an MLP-classifier and instantaneous time domain-based features, as outlined in Fig. 2 and detailed in the following. All aspects of our proposed features and classification mechanism are verified through simulated signal characteristics, as listed in Table 2. These generated signals are compressed from 1080 independent signal realisations. 3.1. Feature extraction Different types of digital signals have various properties, so identifying the suitable features for their recognition, especially those at higher-order or that include a close constellation diagram, is a critical challenge. However, time-domain features are hidden in the instantaneous values of the digital signals. In this research, we compute and experiment with four novel instantaneous features based on the normalised-centred instantaneous amplitude of a digital signal, described below. 3.1.1. The maximum value of the PSD The maximum value of the PSD of a normalized-centered instantaneous amplitude is written as [40]

Fig. 1. The {16, 32, and 64} APSK constellation shapes under the effect of AWGN (where SNR = 30 dB).

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

4

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

Fig. 2. The proposed APSK -classifier system.

k are used, where k = 2, 3, 4, 5, 7, and 9. Three new features based on the magnitudes (C4/C2, C7/C3, C9/C5) of the received signal are derived and assigned with the parameters and written as

Table 2 Simulation signal parameters.

cMax ¼

Signal Parameters

Value

Carrier frequency f c Symbol rate f d Sampling rate f s Number of symbols N Total number of samples Ns

384 kHz 38.4 kHz 307.2 kHz 512*8 32,768

MaxjDFTðacn ðnÞÞj2 Ns

ð4Þ

f 1 ¼ cMax where N s is the total number of samples in the signal, Ns P aðnÞ, aðnÞ is the absolute value acn ðnÞ ¼ amðnaÞ  1, ma ¼ 1=Ns

f2 ¼

C4 C2

ð6Þ

f3 ¼

C7 C3

ð7Þ

f4 ¼

C9 C5

ð8Þ

Fig. 3(b), (c), and (d) show the mean value of the variation of f2, f3, and f4 against the SNR. These features enhance the classification accuracy and distinguish the 16-APSK, 32-APSK, and 64-APSK modulations at SNR = 0 dB. 3.2. The distribution and vector orientation of the proposed features

n¼1

of the analytic form of the transmitted signal, and ma is the sample mean. Previous research confirmed that this feature has a good separation of PSK in one group [10]. However, for our purposes, we find that it is useful for distinguishing between {16, 32, and 64} APSK signals. Fig. 3(a) presents the mean value of the feature against the SNR (in dB) for {16, 32, and 64} APSK signals showing it can separate APSK modulation signals at an SNR as low as 3dB.

3.1.2. The maximum value of the DFT magnitude with the kth power of the analytic expression of the transmuted signal The maximum DFT instantaneous feature is documented in [23,24] and is represented as

Ck ¼

2   k  MaxDFTðaðiÞ Þ

ð5Þ

Ns th

Many magnitudes of the k power of the maximum value of the DFT used in the literature exploit different aspects of this feature, such as those utilised for the classification of FSK, PSK, and th

QAM schemes [41–44]. Dividing the values of the k order provides additional benefits along with the Ck magnitude feature from the received signal, including good classification near varying amplitude modulations and recognising the order of varying amplitude modulations. In this work, six values of the parameter

A fixed threshold algorithm is not an optimal solution for making classification decisions. Instead, the distribution of the features extracted from the received signal samples over the original twodimensional space can identify the intersection between the classes. Fig. 4(a) and (b) show the features in the original dimensional space at SNR = 5 dB for 1080 singles realisations. Based on Fig. 4(a), the 16-APSK is separable from the 32-APSK and 64APSK according to the features f1 and f2. On the other hand, Fig. 4 also suggests there is no optimal threshold value for f3 or f4 that can exactly distinguish the three modulation schemes. 4. APSK classifier-based MLP In this work, an MLP neural network architecture is used as a classifier for pattern recognition. Conventional MLPs are comprised of an input layer of source nodes, i.e., the dimension of feature vectors, one or more hidden layers that function as computation nodes, and an output layer [45]. In this outer layer, the number of nodes depends on the number of variables representing classes or signals. A feed-forward artificial neural network featured in our proposed pattern recognition system mimics the MLP structure, and consists of a double hidden layer, as illustrated in Fig. 5. Pattern recognition methods typically include two stages of training and testing. The weights are calculated at each node during training, and the selected learning algorithm plays a crucial

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

5

Fig. 3. Graphs of the SNR versus the mean value of (a) f 1 , (b) f 2 , (c) f 3 , and (d) f 4 .

role in the calculation of optimum weight values. The learning algorithm and its convergence speed are critical to an ANN, and numerous training algorithms have been proposed. The backpropagation (BP) algorithm is a family of learning methods used for achieving efficient training in the feed-forward ANN. Under some circumstances, the BP-classifier algorithm can produce non-robust results and take a long time to converge of the global minimum. On the other hand, other learning algorithms, such as the cascade correlation and the Levenberg–Marquardt, require extensive computing memory to accomplish the best training. For our proposed pattern recognition system, the resilient backpropagation (RPROP) algorithm is used as a heuristic function, which includes a first-order optimisation [46]. According to the Manhattan weights update rule [47], RPROP enacts only the sign of the partial derivatives as the indication of the direction of the weight update while acting on each weight independently. For performing this computation, the size of the partial derivative does not impact the weight step. The following expression clarifies the adaptation of the update values of Dij (the weight tuning) for the RPROP algorithm.

8 þ dE dE g  Di j ðt  1Þ; if dW ðt  1Þ dW >0 > > ij ij < dE dE  Di j ðtÞ ¼ g  Di j ðt  1Þ; if dW ðt  1Þ dW < 0 ij ij > > : g0  Di j ðt  1Þ; otherwise Fig. 4. The distribution of the features (a) f 1 vs f 2 and (b) f 3 vs f 4 in the original dimensional space at a signal-to-noise ratio of 5 dB.

ð9Þ

where g0 ¼ 1; 0 < g < 1 < gþ ; g;0:þ denotes the weight multiplied by the update factors, Wij indicates the weight value between

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

6

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

Fig. 5. The MLP network architecture.

the neuron j of the current layer and neuron i of the previous layer, and E is the error function for the network output. The partial derivative corresponding to the error function for each weight changing its sign indicates that the previous update value for that weight is excessive and has diverged over the minimum. Therefore, the updated value is modified by multiplying by the factor g as expressed in Eq. (9). If the partial derivative yields the same sign as the previous iteration, then the updated value is increased by multiplying by the factor gþ . This process accelerates the convergence in shallow areas. To avoid over acceleration, each epoch following a multiplication with the factor gþ , the new update value is neither increased nor decreased (g0 ) from the previous values, and the results of Dij remain non-negative at each epoch. The neurons’ weight update values are then followed by an adaptation process to update the actual weight determined by

DW i

8 dE ðtÞ > 0 Di j ðtÞ; if dW > > ij < dE ðtÞ ¼ þDi j ðtÞ; if dW ðtÞ < 0 j ij > > : 0; otherwise

W ij ðt þ 1Þ ¼ W ij ðt Þ þ DW i j ðt Þ

an MLP. The architecture of the {16, 32, 64} APSK classifier is shown in Fig. 6, which depicts the ANN block combined classifier. An advantage of this structure is that it achieves higher classification accuracy compared to conventional pattern recognition systems. All neurons are fully connected, as shown in Fig. 5, where those in the input layer do not function as part of any calculation. Instead, these four neurons correspond to the number of input features and distribute these to the calculating neurons included in the hidden layer. The neurons in the hidden layer, comprised of 12 in the first layer and 16 in the second, perform computations on their inputs and pass the results to the three neurons in the output layer. These output neurons correspond to the number of targets of the three-signal scheme, 16-APSK, 32-APSK, and 64-APSK. Each signal is labelled as a 3-bit binary pattern vector.

ð10Þ

ð11Þ

4.1. The proposed classifier The goal of this study is to develop an intelligent AMC system to enable a receiver to recognise the length of the possible modulation scheme used by the transmitter for modulating the datasets. We consider the modulation classifier as a pattern recognition problem, as described above. The distinct group of interest features are extracted from the transmitted signal creating a feature vector used by the classifier to decide the modulation types used by the transmitter. Any pattern recognition scheme applied to the selected features also applies to the classifier and its characteristics. We evaluate classifier performances according to accurate recognition rate probabilities and calculation speed, and we present results for the {16, 32, 64} APSK classifier combinations of

Fig. 6. The architecture of the proposed {16, 32, 64} APSK classifier algorithm.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

7

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

5. The classifier entity With the complexity of emerging developments in digital systems and the trend towards digital telecommunications beyond analogue telecommunication, most modern communication systems use digital signals. Considering the changes in message parameters, three digital signal schemes exist for the DVB-S2 standard, including 16-APSK, 32-APSK, and 64-APSK signals [2]. To evaluate the performance of our proposed classifier under noisy channel or variable data length scenarios, we evaluate our AMC method under AWGN using the invoked signal tools available in MATLAB simulations. We created 1080 samples of each signal type, which included AWGN added to the symbols at various SNRs (0–20 dB). Each signal type has 512*8 symbols with which the digital information (the message) was created randomly for each 1080 simulated sample to confirm independent results. Classification accuracy assessments of the classes were provided by the classification matrix result with two parameters defined. First, the classification accuracy, AV , shows the analysis of different classes in percentages, and, second, the probability of the correct classification, Pcc , is formulated as

PCC ¼

j X

pðxk jxk ÞPðxk Þ:

ð12Þ

k¼1

For example, a ‘‘j” refers to the modulation signal candidates to be classified ðx1 ; x2 ; x3 ; xj Þ, where Pðxk Þ is the probability that the modulation scheme xk occurs, and Pðxk jxk Þ is the probability of a correct classification when a xk signal is sent [18].

5.1. Network training As seen in the architecture of the proposed classifier in Fig. 6, a four-key feature provides the input, and the neural network-based classifier is a feed-forward network referred to as an MLP. Table 3 lists the neural network architectural specifications adopted for (training and classification) scenarios. In a computer simulation to train the classifier, a training database comprised of 1080x4 data samples with four input feature sets and three target outputs are used with the following training phase steps. 1) The created data consists of an input matrix and a target matrix. 2) The generated data is divided into training, validation, and testing datasets with 70% of the dataset used for training to update the weights of the network. This training process continues until the performance metric of the mean squire error (MSE) converges to 6 1e - 6 of the total training data. Another 15% of the dataset is used to validate the network’s efficiency to generalise and halt the training before overfitting occurs. The remaining 15% offers independent testing data to assess the network’s generalisation ability.

5.2. Classifier training and test criteria For the APSK classifier test stage, the only data required from the trained network are the synaptic weights and biases that must be available to on the trained network. Our training stage is performed using 1080 samples from the realisations for each APSK signal type with varying SNRs from 0 and 20. The performance is evaluated on 720 samples of the new realisation set of interest from an APSK signal scheme with SNRs ranged from 0 and 20. The classifier test and performance evaluations are processed through the following steps.

a. The vector of the extracted features set is shown in Eq. (13), where the nr is a total of the realisation elements. This vector is used in the test stage of the trained classifier.

2

f1

3

6f 7 6 7 f0 ¼ 6 2 7 4f3 5 f4

ð13Þ

nr

b. The target matrix, Tr, is defined for each class of the modulation signal, as explained in Fig. 6. For each realisation of the test class, the corresponding output vector, based on the number of hidden layers used, is calculated as follows: i. The location equivalent to each signal pattern in the output vector is set as

2

16APSK : f½001T gLSB

3

7 6 T 7 Tr ¼ 6 4 32APSK : f½010 gLSB 5 64APSK : f½100T gLSB

ð14Þ

nr

ii. The adapted output vector should agree with one of the columns of the target matrix Tr, and this agreement referred to as the {16, 32, 64} APSK modulation scheme. iii. The complete criteria are applied for each realisation in the test category. c. The probability of correct recognition is calculated as the percentage of realisations from the testing samples that have the correct modified output vector. 6. Experimental results The results of the degradation under the simulation scenarios applied to the APSK classifier are presented in this section, where we assume that the carrier frequency is constant, as mentioned in Table 2. Therefore, we only consider complex base-band signals that randomly generate data over each scheme and expose them to a noisy channel of pure AWGN, according to testing ratios of the signal-to-power noise ranged from 0 to 20 dB. All signal samples in the training and testing are labelled with the corresponding modulation type and SNR. Samples in the training dataset are uniformly distributed through SNR range and padded by 1 dB. 6.1. Accuracy evaluation The most striking results from the data are summarised in Table 4, which depicts that the best performance is achieved with correct classifications of the {16, 32, 64}-APSK signal classes at SNR 0; 5; 15; and P 20 dB. The confusion matrix in Table 4 is for N = 512*8 and an SNR of 0 dB, and the number of overlapping signals, in this case, is higher compared to the SNR of 5 dB, which is expected because channel SNR is lower. Furthermore, most of the modulation misclassification occurs between 16-APSK and 32APSK because they have similar constellation diagrams, particu-

Table 3 Specification of the ANN-based classifier structure. Network structure

Numbers of neurons in each layer Input

MLP

Performance function

Hidden 1

4 12 Activation function ~ Tanh MSE

Hidden 2

Output

16

3

Logsig

~

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

8

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

Table 4 The confusion matrix of the {16, 32, 64} APSK classifier at SNR = 0; 5; 15; and P 20 dB. Es =N s ¼ 0 dB Simulated signal type

16-APSK 32-APSK 64-APSK

Es =N s ¼ 5 dB Predicted signal type 16-APSK

32-APSK

64-APSK

16-APSK

32-APSK

64-APSK

86.0 14.0 0.0

0.1 93.9 6.0

0.0 1.3 98.8

99.9 0.1 0.0

0.4 99.6 0.0

0.0 0.0 100.0

Es =N s ¼ 15 dB Simulated signal type

16-APKS 32-APSK 64-APSK

Es =N s ¼P 20 dB Predicted signal type 16-APSK

32-APSK

64-APSK

16-APSK

32-APSK

64-APSK

100 0 0

0 100 0

0 0 100

100 0 0

0 100 0

0 0 100

Predicted signal type: PCC %.

larly at low SNR. The efficiency of the classifier increases at an SNR > 5 dB. However, this superior performance could be attributed to the robustness of the internal structure of the APSK classifier and the merits of the proposed features. The influence of using different lengths of signal symbols per data block on the classification accuracy is shown in Fig. 7. Using a long row of data symbols to extract the classification features enhances the correct classification rate. For example, when using N = 5000 symbols leads to an improvement in classification accuracy nearly exceeds 98% at an SNR of 5 dB compared to N = 1000 symbols. The results in Fig. 7 also suggest that the classification accuracy rate is a function of the SNR of the channel and the number of symbols in the signal. This relationship is due to the fluctuation in the extracted feature values based on these parameters, which may suggest that misclassification increases with lower SNR and smaller block lengths. In our evaluation of the APSK classifier over different channels, we set the sample length to 512*4 and the SNR between 0 and 30 dB. For the simulation scenarios, the phase and frequency offset are assumed to be constant, and the classifier performance is investigated in AWGN-only, multi-path, and unknown channel conditions with a symbol rate of 20 MHz, fading variables, average path gains of 0; - 15; and - 20 dB, path delays of 0; 1107 ; and 2107 seconds, a factor k = 3 dB, and maximum Doppler shift of 15.5 Hz.

Fig. 8 depicts the classification accuracy rate at N = 512*4 with three different channel scenarios. For this test, the maximum Doppler shift and the SNR of the channel are known at the receiver site, and these limitations are used to train the classifier. The accuracy of classification is good for AWGN-only channels compared to multi-path fading channels due to the existence of a strong signal attributable to the line of sight component. Also, an increasing SNR is seen to results in better classification accuracy. The classifier performance under unknown channel noise provides a significantly high correct classification rate reaching 95%. 6.1.1. Flat-fading channel with phase and frequency offset In the flat-fading channel, we set the signal length at 512x4 symbols and SNR 0 to 20 dB. The signal realizations of the 720 samples were tested for each class. The phase noise and frequency offset were considered separately. For phase offset, we used five 









angles (0 ; 5 ; 10 ; 20 ; 30 Þ. In the simulation, a relative frequency f o (computed as the ratio between the carrier frequency and the symbol sampling frequency) was used to indicate different levels of frequency offsets. It was limited by reduction ratios ð0% ; 1% ; 2% ; 3% ; 6% Þ of the actual carrier frequency offset f c .The results are presented in Figs. 9 and 10. In the case of the phase offset, Fig. 9 shows that the phase offset has a small effect on the classifier performance at SNR levels below 5 dB. However, it had almost no significant influence at SNR > 6 dB levels. The reason

Fig. 7. Classification accuracy for different quantities of symbols.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

9

Fig. 8. The accuracy of classification for different channels environments.

Fig. 9. Accuracy of correct classification versus SNR for different values of phase offset.

Fig. 10. Accuracy of correct classification versus SNR for different values of frequency offset.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

10

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

for the superior performance of the APSK-classifier under the phase th

offset could be attributed to the powerful capability of the k power of the magnitude of signal to derive the features that have a high discrimination between modulation signals. In the case of the frequency offset, Fig. 10 shows that the frequency offset has an influence on classifier performance, particularly when SNR is below 5 dB. At SNR > 5 dB, the effect ratios of 1% f c , 2% f c ; and 3% f c showed a slightly better performance than the effect ratio 6%fc. At SNR > 7 dB, the performance gaps become nearly equal.

exceeds the template matching classifier in terms of the correct classification rate. However, the correct classification accuracy is not the only factor for favouring one classifier over another. Other important aspects include the intricacy of the system and the order of computations involved in making the classificatory decision. Because modulation classification is a real-time practical problem, systems with simple structures and low complexity are more attractive. The APSK classifier includes simple signal processing operations, such as the operation for computation of the magnitude of the Fourier transform, squared magnitude of signals, and other conventional mathematical operations like multiplication and addition.

6.2. Comparison with an existing method Fig. 11 illustrates a compares the classification accuracy of a template matching the MAPSK classification in [14] with the proposed APSK classifier. To ensure a flexible comparison, the simulation scenario is performed by applying the same parameters used in [14]. Specifically, we use the same modulation schemes in our comparison of the three M-APSK types of 16-APSK, 32-APSK, and 64-APSK with a data length of 1000 symbols per signal. The proposed classifier provides a correct classification accuracy of nearly 100% at an SNR > 5 dB. At the same time, the classifier with the template matching delivers a classification accuracy near 100% at an SNR > 8 dB. So, the proposed APSK classifier

100

Av (%)

90

Template Matching Method Proposed classifier

80

70

60 0

5

10

15

20

Es/No (dB) Fig. 11. Comparison of the APSK classifier with the template matching recognition method.

6.3. The complexity of the training We conduct two stages of experiments to concentrate on the effects of the training samples on the classification accuracy of the proposed APSK classifier. In the first set of experiments, we focus on the classification accuracy of the proposed APSK classifier by varying the SNR, which has been the prime criteria for measuring the performance of an AMC system. We set the signal length at 512*4 symbols, and the SNR varies between 0 and 20 dB. Following this, the modulated signals are segmented into 240, 360, 480, 720, and 1080 samples to form the training dataset while the testing dataset contains 720 samples. The results presented in Fig. 12 show that the best performance is achieved when using L = 720 for each signal generated. However, this performance advantage is highest between 6 and 20 dB. The performance curves for L = 240, 480, 360, and 1080 signal samples show nearly equal accuracy in the range of SNR from 5 to 20 dB. As expected, for a low SNR below 5 dB, the difference between the trained and tested data becomes more prominent, so the classifier accuracy deteriorates. However, by increasing the SNR, the performance curve improves considerably, especially with larger signal samples. Furthermore, the classification accuracy curve is better for L = 240, 480, and 720 signal samples. The second set of experiments focuses on the influence of testing signal sample length to the classification accuracy of the proposed APSK classifier. In this scenario, similar settings are applied as before, except the testing signal includes 1080*2 samples for each APSK class. The performance curves are seen in Fig. 13, and, again, the APSK classifier outperforms for a training length of L = 720 samples. However, the proposed APSK classifier shows

Fig. 12. Accuracy of correct classification vs SNR for different numbers of training samples while testing with 720 signal realisations from 16-APSK, 32-APSK, and 64-APSK. Each consists of a signal length N = 512*4 under the AWGN channel.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

11

Fig. 13. Correct classification accuracy vs the SNR for different numbers of training samples while testing 1080 *2 signal realisations from 16-APSK, 32-APSK, and 64-APSK. Each consists of a signal length of N = 512*4 under the AWGN channel.

Fig. 14. The classification accuracy rate vs unknown SNR conditions with 720 testing signal samples for each modulation scheme containing 512*4 symbols.

nearly equally superior and robust accuracies for the training samples with length L = 240, 360, 480, and 1080.These results prove the enhanced robustness of the proposed APSK classifier with few training samples, which is important for good AMC classification.

from 0 to 20 dB padded by 2 dB are in close proximity. The gap between the two curves decreases to 5 dB SNR, after which the performance becomes equivalent. Also, the classifier trained at a range of SNR from 0 to 15 dB maintains high accuracy even when the test data is generated at a high SNR of 15 dB.

6.4. Performance on an unknown channel SNR

7. Discussion

Next, we evaluate the robustness of our proposed APSK classifier based on the MLP network by training our signal samples at a different SNR range than the test signal. In this test, we only focus on the AWGN channel for a better understanding of the performance with an unknown SNR at the receiver. First, the training dataset is uniformly distributed in two ranges of SNR from 0 to 15 dB padded by 2 and from 0 to 20 dB with a step of 1. Next, the test dataset is created with an SNR range from 0 to 20 dB. A signal realisation of 1080 samples is used for training while testing on 720 samples, and each modulation scheme in both datasets is comprised of the same 512*4 symbols. The performance curves are seen in Fig. 14 showing that the data trained at a range of SNR from 0 to 20 dB with a step of 2 dB has an advantage of around 5% over the data trained at a range of SNR from 0 to 15 dB. This result can be attributed to the data generated to train and test the APSK classifier at range of SNR

While this paper does not provide a final solution, we highlight a promising line of research for automatic modulation recognition algorithms. Due to the complexity of this issue, an increasing demand exists for the use of automatic digital radio signals for a variety of applications. Nevertheless, this work provides interesting results with the development of a novel method to recognise DVB-S2X standard-specific to the MAPSK modulation type. An APSK classifier is introduced by using an MLP-based neural network architecture and kth order spectrum features. However, better results could be accomplished by performing the training and testing on larger input datasets, so future work should include additional datasets to provide new insight. Further options for future work could use other machine learning methods, such as SVM or decision tree, in connection with the proposed feature sets to create a hybrid classifier.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

12

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx

8. Conclusion A novel algorithm for an AMC is presented using an MLP neural network architecture. In addition to sharing the advantages of conventional neural network architecture, the combination with the MLP network structure and instantaneous signal features provides further merits: 1. Four spectrum ratios are extracted, which reduce the complexity of the learning algorithm. 2. Five training dataset inputs are used for the classification of the APSK schemes using the MLP neural network structure. 3. Results show that the classifier provides good results even when the SNR values of the training samples are different than those of the test samples. Also, the simulation results show that the performance of the proposed recognition method significantly outperforms existing methods. The proposed APSK classification method is not influenced by the frequency offset and can be used with an AWGN channel and multi-path fading channel while also offering good accuracy with unknown channel parameters. Finally, to the best of our knowledge, there exists little work in the literature on APSK modulation classification using conventional MLP neural network architecture. Thus, we demonstrate the significant potential for MLP networks in APSK signal identification. Declaration of Competing Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. References [1] O.A. Dobre, A. Abdi, Y. Bar-Ness, W. Su, Survey of automatic modulation classification techniques: classical approaches and new trends, IET Commun. 1 (2007) 137–156, https://doi.org/10.1049/iet-com:20050176. [2] ETSI, Digital Video Broadcasting (DVB); modulation systems for Broadcasting , other broadband satellite applications (DVB-S2), 2009. [3] T. Specification, V. Tr, Digital Video Broadcasting (DVB); Second generation framing structure, channel coding and modulation systems for Broadcasting, Interactive Services, News Gathering and other broadband satellite applications, Part 1 (2011) DVB-S2. [4] E. Agirman-Tosun, H., Liu, Y., Haimovich, A. M., Simeone, O., Su, W., Dabin, J., & Kanterakis, Method to reduce the signal-to-noise ratio required for modulation recognition based on logarithmic properties, IET Commun. 12 (2018) 1360– 1366. doi:10.1049/iet-com.2018.0030. [5] J.A. Sills, Maximum-likelihood modulation classification for PSK/QAM, in: Mil. Commun. Conf. Proceedings, 1999. MILCOM 1999. IEEE, 1999: pp. 217–220. [6] F. Hameed, O.A. Dobre, D.C. Popescu, On the likelihood-based approach to modulation classification, IEEE Trans. Wirel. Commun. 8 (2009) 5884–5892, https://doi.org/10.1109/TWC.2009.12.080883. [7] W. Wei, J.M. Mendel, A new maximum-likelihood method for modulation classification, in: Signals, Syst. Comput. 1995. 1995 Conf. Rec. Twenty-Ninth Asilomar Conf., 1995: pp. 1132–1136. [8] J.L. Xu, W. Su, M. Zhou, Likelihood-Ratio Approaches to Automatic Modulation Classification, 41 (2011) 455–469. doi:10.1109/TSMCC.2010.2076347. [9] M.Z. Xu, L. Jefferson, Su Wei, Software-defined radio equipped with rapid modulation, Recognition 59 (2010) 1659–1667, https://doi.org/10.1109/ TVT.2010.2041805. [10] E.E. Azzouz, A.K. Nandi, Automatic identification of digital modulation types, Signal Process. 47 (1995) 55–69, https://doi.org/10.1016/0165-1684(95) 00099-2. [11] H. Hosseinzadeh, F. Razzazi, A. Haghbin, A self training approach to automatic modulation classification based on semi-supervised online passive aggressive algorithm, Wirel. Pers. Commun. 82 (2015) 1303–1319, https://doi.org/ 10.1007/s11277-015-2284-7. [12] A. Hazza, M. Shoaib, S.A. Alshebeili, A. Fahad, An overview of feature-based methods for digital modulation classification, 2013 1st Int. Conf. Commun. Signal Process. Their Appl. ICCSPA 2013. 1 (2013). doi:10.1109/ ICCSPA.2013.6487244. [13] A. Ali, E. Ergun, An M-QAM Signal Modulation Recognition Algorithm in AWGN-Channel, Sci. Program. 17 (2019), https://doi.org/10.1155/2019/ 6752694.

[14] Y. Deng, Z. Wang, Modulation recognition of MAPSK signals using template matching Electron, Lett. 50 (2014) 1986–1988, https://doi.org/10.1049/ el.2014.2700. [15] O.A. Dobre, M. Öner, S. Rajan, R. Inkol, Cyclostationarity-based robust algorithms for QAM signal identification, IEEE Commun. Lett. 16 (2012) 12– 15, https://doi.org/10.1109/LCOMM.2011.112311.112006. [16] B.S. Boutte, David, A Hybrid ICA-SVM Approach to Continuous, IEEE Signal Process. Lett. 16 (2009) 402–405, https://doi.org/10.1109/LSP.2009.2016444. [17] O.a. Dobre, Y. Bar-ness, Robust QAM Modulation Classification Algorithm Using Cyclic Cumulants, Commun. Soc. (2004) 745–748, https://doi.org/ 10.1109/WCNC.2004.1311279. [18] A. Swami, B.M. Sadler, Hierarchical digital modulation classification using cumulants, IEEE Trans. Commun. 48 (2000) 416–429, https://doi.org/10.1109/ 26.837045. [19] A. Smith, M. Evans, J. Downey, Modulation classification of satellite communication signals using cumulants and neural, Networks (2019), https://doi.org/10.1109/CCAAW.2017.8001878. [20] A.E. Sherme, A novel method for automatic modulation recognition, Appl. Soft Comput. J. 12 (2012) 453–461, https://doi.org/10.1016/j. asoc.2011.08.025. [21] M.L.D. Wong, A.K. Nandi, Automatic digital modulation recognition using artificial neural network and genetic algorithm, Signal Process. 84 (2004) 351– 365, https://doi.org/10.1016/j.sigpro.2003.10.019. [22] E. Azzouz, A. Nandi, Automatic Modulation Recognition of Communication Signals, IEEE Trans. Commun. 46 (2013) 218, https://doi.org/10.1109/ 26.664294. [23] X. Fangjuan, L. Chisheng, W. Guojin, An efficient and simple method of MPSK modulation classification, in: 2008 Int. Conf. Wirel. Commun. Netw. Mob. Comput. WiCOM 2008, 2008: pp. 1–3. doi:10.1109/WiCom.2008.339. [24] P. Ciblat, M. Ghogho, Blind NLLS Carrier Frequency-Offset Estimation for QAM, PSK, and PAM Modulations: Performance at Low SNR, 54 (2006) 1725–1730. [25] A. Yi, H. Liu, L. Yan, L. Jiang, Y. Pan, B. Luo, Amplitude variance and 4th power transformation based modulation format identification for digital coherent receiver, Opt. Commun. 452 (2019) 109–115, https://doi.org/10.1016/j. optcom.2019.07.016. [26] W. Su, Feature space analysis of modulation classification using very highorder statistics, IEEE Commun. Lett. 17 (2013) 1688–1691, https://doi.org/ 10.1109/LCOMM.2013.080613.130070. [27] E.E. Azzouz, A.K. Nandi, Modulation, Automatic of, Recognition Signals, Communication, Kluwer Academic Publishers, 1995. [28] S. Zhou, Z. Wu, Z. Yin, Z. Yang, Noise-Robust Feature Combination Method for Modulation Classification Under Fading Channels, 2018 IEEE 88th Veh. Technol. Conf. (2018) 1–5. doi:10.1109/VTCFall.2018.8690662. [29] R. De Gaudenzi, A. Guillén, i. Fàbregas, A. Martinez, Performance analysis of turbo-coded APSK modulations over nonlinear satellite channels, IEEE Trans. Wirel. Commun. 5 (2006) 2396–2407, https://doi.org/10.1109/TWC.2006. 1687763. [30] B. De Gaudenzi, R., i Fabregas, A. G., Vicente, A. M., Ponticelli, APSK coded modulation schemes for nonlinear satellite channels with high power and spectral efficiency, 20th AIAA Int. Commun. Satell. Syst. ICSSC 2002. (2002) 1–11. [31] A. Puengnim, N. Thomas, J. Tourneret, J. Vidal, Classification of linear and nonlinear modulations using the Baum – Welch algorithm and MCMC methods, Signal Process. 90 (2010) 3242–3255, https://doi.org/10.1016/j. sigpro.2010.05.030. [32] Z. Zhang, Y. Li, X. Zhu, Y. Lin, A method for modulation recognition based on entropy features and random forest, (2017) 243–246. doi:10.1109/QRSC.2017.47. [33] S. Norouzi, A. Jamshidi, A.R. Zolghadrasli, Adaptive modulation recognition based on the evolutionary algorithms, Appl. Soft Comput. J. 43 (2016) 312– 319, https://doi.org/10.1016/j.asoc.2016.02.028. [34] E.E. Azzouz, a.K. Nandi, Automatic modulation recognition—II, J. Franklin Inst. 334 (1997) 275–305, https://doi.org/10.1016/S0016-0032(96)00070-1. [35] J.D.H. Agirman-Tosun, Yu Liu, A.M. Haimovich, Osvaldo Simeone, Wei Su, Modulation classification of MIMO-OFDM signals by Independent Component Analysis and Support Vector Machine, in: 2011: pp. 1903–1907. doi:10.1109/ ACSSC.2011.6190354. [36] D. Wang, M. Zhang, Z. Cai, Y. Cui, Z. Li, H. Han, M. Fu, B. Luo, Combatting nonlinear phase noise in coherent optical systems with an optimized decision processor based on machine learning, 369 (2016) 199–208. doi:10.1016/j. optcom.2016.02.029. [37] C. Wang, J. Du, G. Chen, H. Wang, L. Sun, K. Xu, B. Liu, Z. He, QAM classification methods by SVM machine learning for improved optical interconnection, Opt. Commun. 444 (2019) 1–8, https://doi.org/10.1016/j.optcom.2019.03.058. [38] Y. Lecun, Y. Bengio, G. Hinton, Deep learning, (2015). doi:10.1038/ nature14539. [39] A. Ali, F. Yangyu, Unsupervised feature learning and automatic modulation classification using deep learning model, Phys. Commun. 25 (2017) 75–84, https://doi.org/10.1016/j.phycom.2017.09.004. [40] A.K. Nandi, E.E. Azzouz, Algorithms for automatic modulation recognition of communication signals, IEEE Trans. Commun. 46 (1998) 431–436. [41] B.G. Mobasseri, Constellation shape as a robust signature for digital modulation recognition, Mil. Commun. Conf. Proceedings, 1999. MILCOM 1999. IEEE. 0 (1999) 442–446. doi:10.1109/MILCOM.1999.822722. [42] C. Yin, B. Li, Y. Li, B. Lan, Modulation classification of MQAM signals based on density spectrum of the constellations, Proc. 2010 2nd Int. Conf. Futur.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257

A.K. Ali, E. Erçelebi / Measurement xxx (xxxx) xxx Comput. Commun. ICFCC 2010. 3 (2010) 57–61. doi:10.1109/ ICFCC.2010.5497682. [43] Y.H. Tanc, A. Akan, Modulation identification of digital M-ary QAM signals by Hilbert-Huang Transform, Proc. IEEE Int. Conf. Electron. Circuits, Syst. (2013) 678–681, https://doi.org/10.1109/ICECS.2013.6815505. [44] N. Ahmadi, R. Berangi, Modulation classification of QAM and PSK from their constellation using Genetic Algorithm and hierarchical clustering, 2008 3rd Int. Conf. Inf. Commun. Technol. From Theory to Appl. ICTTA. (2008). doi:10.1109/ICTTA.2008.4530242.

13

[45] W.W. Hsieh, O. Sci-, Neural Networks and Learning Machines, 3rd ed., Pearson Prentice Hall, Ontario, Canada, 1999. [46] G.P. Zhang, Neural Networks for Classification: A Survey, IEEE Trans. Syst. 30 (2000) 451–462, https://doi.org/10.1109/5326.897072. [47] M. Riedmiller, Advanced supervised learning in multi-layer perceptrons — From backpropagation to adaptive learning algorithms, 16 (1994). doi: https://doi.org/10.1016/0920-5489(94)90017-5.

Please cite this article as: A. K. Ali and E. Erçelebi, Automatic modulation recognition of DVB-S2X standard-specific with an APSK-based neural network classifier, Measurement, https://doi.org/10.1016/j.measurement.2019.107257