Wireless Communications and Mobile Computing

Volume 2018, Article ID 1372439, 13 pages

https://doi.org/10.1155/2018/1372439

## Multilayer Learning Network for Modulation Classification Assisted with Frequency Offset Cancellation in Satellite to Ground Link

College of Electronic and Information Engineering, Shenyang Aerospace University, Shenyang 110136, China

Correspondence should be addressed to Guan Qing Yang; moc.361@yq_naug

Received 18 September 2017; Revised 27 February 2018; Accepted 31 March 2018; Published 10 May 2018

Academic Editor: Ioannis Krikidis

Copyright © 2018 Guan Qing Yang. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

#### Abstract

A multilayer learning network assisted with frequency offset cancellation is proposed for modulation classification in satellite to ground link. Carrier frequency offset greatly reduces modulation classification performance. It is necessary to cancel frequency offset before modulation classification. Frequency offset cancellation weights are established through multilayer learning network based on MSE criterion. Then the weight and hidden layer of multilayer learning network are also established for modulation classification. The hidden layers and weight are trained and tuned to combat the interference introduced by frequency offset. Compared with current modulation classification algorithm, the proposed multilayer learning network greatly improves the Probability of Correct Classification (PCC). It has been proven that the proposed multilayer learning network assisted with frequency offset has higher performance for modulation classification within the same training sequence.

#### 1. Introduction

Satellite to ground link adopts different modulation technology to satisfy different requirement. With rapid development of communication technology, modulation classification for satellite communication becomes an important research topic in signal recognition, especially in lager frequency offset environment [1–3].

Doppler frequency offset exists between users and satellite, and it greatly damages the link performance and especially affects modulation classification. As satellite is moving at high speed, frequency offsets are also changing. This process requires that modulation classification should overcome frequency offsets within larger SNR dynamic range. So it is very important to adopt frequency offset cancellation method for modulation classification.

In early research work, the literatures were researching on cancellation algorithms for frequency offsets, which were estimated at receivers and then sent back to respective transmitter. But this is not suitable for long-distance transmission in satellite to ground link. Reference [4] analyzed single user in earliest frequency offset cancellation algorithm. Least squares (LS) and minimum mean-square error (MMSE) were applied in [5]. LS method required frequency offset, while MMSE algorithm required noise powers as well. Reference [6] proposed an iterative cancellation for canceling frequency offset interference, which should need large complex iterations. Huang gave frequency offsets cancellation method using circular convolution in [7], which should need relatively large complexity.

Reference [8] proposed an alternative method in time-domain method for frequency offset compensation, but the precision was not very satisfactory. Reference [9] proposed a joint iterative detection algorithm, which should require ICI matrix inversion in each iteration. In [10], successive interference cancellation (SIC) algorithm was proposed to cancel frequency offset interference. The interference was cancelled by reconstructing signals. SIC algorithm depending on power order was proposed in [11]. Reference [12] gave SIC algorithm based on the signal-to-interference-plus-noise ratio (SINR) order. Reference [13] proposed an iterative parallel interference canceller (PIC) solution. Based on this, selective parallel interference cancellation (SPIC) algorithm was presented in [14] but had a significant implementation complexity. Based on signal-to-interference ratio (SIR) analyzing, a parallel interference canceller for mitigation of interference due to CFOs was presented in [15].

Modulation classification for communication signal in satellite link is the multivariate pattern classification problem with multiple unknown parameters. Modulation classification algorithms should extract the feature from sampling signal. These algorithms can be classified into the following 5 types.

*Type 1 Histogram Features*. In 1984, Liedtke et al. proposed a pattern recognition method to classify digital modulated signals. Liedtke et al. used amplitude histogram, frequency histogram, and phase histogram as characteristic parameters. Too many feature dimensions lead to lager computational complexity. However, reducing feature dimensions would affect the ability of different modulation classification with similar distribution functions.

*Type 2 Statistical Cumulant Features*. In view of the larger number of histogram features, literature [16] proposed a sixth-order cumulants feature in order to improve the performance. However, the sixth-order cumulants algorithm was only based on the energy model, which was not effective for frequency offset. In [17–19], cyclic high-order cumulants were obtained for modulation classification, and the performance in the Gaussian white noise channel was given. But the performance in large frequency offset status was not considered. References [20, 21] used high-order cumulant feature for modulation classification. These feature extraction operations were relatively easy to obtain. However, computational complexity for higher-order cumulants was improved. As modulation classification algorithms obtained features based on cumulants, classification performance within larger frequency offset was not in consideration.

Asoke K. Nandi and EEAzzouz et al. used the statistical features of instantaneous envelope, phase, and frequency of signal data and applied decision theory for modulation classification [22, 23]. A similar algorithm was proposed by Chan and Gadbois. According to the features of the signal envelope, the ratio of the signal envelope variance to the signal mean square was used as the decision criterion [24]. Louis used digital signals such as 2FSK, 4FSK, 8FSK, OQPSK, MSK, BPSK, QPSK, 8PSK, 16QAM, 64QAM, and other digital signals based on instantaneous frequency, phase, and amplitude [17]. Swami adopted normalized cumulative power of the symbol-synchronous sampling sequence to classify QAM, PSK, ASK, and other signals [20]. In 2001, Wong and Nandi [25] used statistical cumulant and spectral features in classification.

*Type 3 Transformation Feature Criterion*. Cumulant features for satellite signal are only to identify several common modulations. In addition to directly using the statistical parameters and the histogram as a classification feature, the signal can be transformed into other feature spaces and the satellite data analysis can be accomplished by using the feature parameters in the new feature space. With the development of signal processing theory, the signal could be transformed into various forms. Wavelet transformation feature for modulation classification was proposed in [26]. Haar wavelet feature transform had better antinoise features than the high-order cumulants. But it had weak anti-frequency offset features. In [27], modulation classification algorithm based on cycle spectrum features was proposed. Based on extending this algorithm, modulation classification algorithm based on bispectrum feature was used to classify MPSK signal in [28]. In [29], wavelet packet was proposed, which was decomposed by the wavelet packet. The partial decomposition vector of the average energy was obtained. Then average energy was arranged in a certain order to construct classification feature. In [30], MFSK and MPSK are classified through wavelet transform, and better PCC was obtained when the SNR was greater than 6 dB. In [31], an automatic classification algorithm based on spectral analysis was proposed, which used the statistical characteristics assisted with maximum likelihood estimator.

*Type 4 Bayesian Classification Criterion*. A classical method for modulation classification is the maximum likelihood method. Modulation classification is obtained through maximum likelihood function. In [32], MPSK signal based on phase maximum likelihood function was proposed. In [33], a joint likelihood function using amplitude and phase was proposed within larger SNR, and it was on the premise that the amplitude and phase were assumed to be independent. In [34], classification for MFSK signal was discussed. After averaging the unknown parameters in the average likelihood function, the integral expression of zero-order modified Bessel function was obtained. The high-order correlation analysis based on Bessel function was derived for modulation classification. In literature [35], a generalized likelihood ratio function for classification framework was proposed, which firstly expanded the power for the likelihood ratio function and then made the expected average processing for the unknown parameters and then formed the classification statistics based on higher correlation order.

The advantage of transformation domain theory for modulation classification is theoretically guaranteeing that the result is optimal under Bayesian cost criterion. The classification performance curve can be obtained by theoretical analysis. The likelihood ratio performance can be used as a theoretical upper limitation. Of course, the maximum likelihood algorithm also has limitations and shortcomings. Firstly, compared to the modulation classification method, it requires more a priori knowledge, such as the form of the distribution function including the mean, variance, and signal-to-noise ratio. Secondly, the existence of unknown parameters results in the complex calculation, which is also difficult for real-time processing.

In order to cancel frequency offset, the eigenvalue is very limited to suppress the noise based on Bayesian classification criterion. Current methods such as feature extraction, classification, and regression can be used as shallow structure algorithms. These methods are limited to complex functions in finite samples and computational units, and their generalization ability is restricted to complex classification problems.

*Type 5 Multilayer Neural Network*. With improvement of artificial intelligence, researchers introduced multilayer neural network algorithms for modulation classification. Multilayer learning could represent the input data distribution by learning a multilayer nonlinear network structure and implementing complex function approximation. The multilayer learning motivation lies in the establishment of a neural network, which could simulate the human brain for signal analysis and mimic human brain mechanisms to interpret the data.

Multilayer learning algorithms obtain data features by building models with multiple hidden layers to improve classification or prediction accuracy. Multilayer learning algorithm is different from the traditional shallow learning structure; the difference in multilayer learning lies in the following: (1) emphasizing the depth of the model structure usually with many hidden nodes and (2) the importance of feature learning being clearly highlighted. By layer-by-layer feature transformation, the feature of data in the original space is transformed into a new feature space to make classification or prediction easier. Compared with the method of constructing data features manually, the use of data to learn features can better express intrinsic information. Identification part needs to determine the appropriate decision rules and classifier structure.

Common classifiers include tree structure classifiers [36, 37], neural network classifiers [38, 39], and support vector machine classifiers [40, 41]. The tree structure classifier used a multilevel classifier structure, each structure according to one or more of the characteristic parameters. Neural network classifier uses different structures of artificial neural network for a variety of ways to achieve training and testing. SVM classifier inputs feature vector into high-dimensional space and constructs the optimal classification in the high-dimensional space to achieve data classification. In [42], a classifier based on tree structure was adopted. The multilevel tree classifier structure is based on one or more feature parameters. In [43], a multilayer clustering algorithm is used for modulation classification. In [44], -nearest neighbor algorithm classifier was used to identify modulation mode. Literature [45] used the distance distribution function to optimize and obtain better classification performance. However, this algorithm did not give analysis in larger frequency offset. In [46], an improved KNN multilayer learning architecture was proposed to classify modulation. However, this algorithm was applicable for Gaussian white noise channel environment, and the performance of the algorithm was limited for larger frequency offset.

#### 2. System Model and Problem Formulation

##### 2.1. Satellite to Ground Link Model

Due to high-speed movement of the satellite, the satellite to ground link is established between station and satellite. In this process, it is necessary to accurately calculate the constellation. The process has been shown in Figure 1. In this process, frequency offset of the satellite to ground link is changing, so it is necessary to accurately cancel the interference of frequency offset before modulation classification.