Mathematical Problems in Engineering

Volume 2017, Article ID 6129120, 9 pages

https://doi.org/10.1155/2017/6129120

## A Parameter Estimation Algorithm for Multiple Frequency-Hopping Signals Based on Sparse Bayesian Method

Institute of Information and Navigation, Air Force Engineering University, Xi’an, Shanxi 710077, China

Correspondence should be addressed to Kun-feng Zhang; moc.oohay@gnahz_gnefnuk

Received 26 February 2017; Revised 1 June 2017; Accepted 28 June 2017; Published 6 September 2017

Academic Editor: Zhike Peng

Copyright © 2017 Kun-feng Zhang et al. 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

Parameter estimation and network sorting for noncooperative wideband frequency-hopping (FH) signals have been essential and challenging tasks, especially in the case with little or even no prior information at all. In this paper, we propose a nearly blind estimation approach to estimate signal parameters based on sparse Bayesian reconstruction. Taking the sparsity in the spatial frequency domain of multiple FH signals into account, we propose a sparse Bayesian algorithm to estimate the spatial frequency parameters. As a result, the frequency and direction of arrival (DOA) parameters can be obtained. In order to improve the accuracy of the estimation parameters, we employ morphological filter methods to further clean the data poisoned by the noise. Moreover, our method is applicable to the wideband signal models with little prior information. We also conduct extensive numerical simulations to verify the performance of our method. Notably, the proposed method works well even in low signal-to-noise ratio (SNR) environment.

#### 1. Introduction

Frequency-hopping (FH) communication is widely used in the military communications, due to its high confidentiality, antijamming capacity, low susceptibility, and strong networking capabilities [1–3]. However, these features also make it difficult to estimate the parameters of multiple FH signals, especially when there is no prior information about the hopping patterns. In general, the parameter estimation of FH signals often involves sorting and identifying of possible spatial FH signals belonging to different FH networks [2, 4, 5]. Parameter estimation and network sorting are two fundamental but not well-solved problems in the field of FH signals reconnaissance.

Among various methods of parameter estimation, estimation methods based on maximum likelihood (ML) are practically intractable with few or no prior information, and thus several ML based methods have been proposed [6, 7]. Recently, another two techniques have attracted a lot of attentions, that is, time-frequency analysis and sparse reconstruction, since they can estimate the parameters by recovering clear time-frequency spectrum. Methods based on time-frequency analysis are attractive due to its simplicity and fast computation, but their performance degrades in low signal-to-noise ratio (SNR) scenarios [8, 9]. By contrast, sparse reconstruction based methods achieve good time-frequency spectrum and accurate parameter estimation in low SNR conditions [10–12]. However, all these aforementioned methods suffer from degrading performance and mismatching of signals and parameters with multiple FH signals.

In order to achieve accurate estimation and network sorting results simultaneously, direction of arrival (DOA) information and blind separation (BS) is utilized with the assumption of multiple array elements. In many cases, FH signals are actually wideband, while many existing methods of DOA estimation consider the FH signals to be narrow-band signals [13, 14]. To solve this issue, an expectation maximization (EM) approach is proposed for blind hop timing DOA and frequency estimation in the situation of possible bandwidth mismatch in [15]; the success of this approach mainly depends on the selection of the initial value. Unfortunately, it is difficult to obtain a reliable initial time-frequency spectrum using short-time Fourier transform (STFT) especially in low SNR environments. Paper [16, 17] proposes a novel method for parameter estimation and network sorting of wideband FH signals, which can only be used for multiple orthogonal signals but fails in asynchronous signals.

With the assumption of wideband multiple FH signals, we propose a new method based on sparse reconstruction and morphological filtering to achieve better parameter estimation and network sorting results. We exploit the sparsity of the spatial frequencies of multiple FH signals to build redundant sparse dictionaries, based on which spatial frequencies are estimated for the reconstruction of FH signals and the DOA estimation can also be achieved. To improve the estimation accuracy in low SNR conditions, morphological filtering is used to achieve more accurate time-frequency and time-space spectrum. Therefore, parameters of FH signals can be estimated and then networks can be sorted with DOA. Moreover, the proposed method can be used for different multiple FH signals, different parameters, and different SNR conditions.

The rest of the paper is organized as follows. In signal model, the FH signal model and the sparse signal model are formulated. In proposed method, parameter estimation and network sorting algorithms are formulated mainly including sparse Bayesian reconstruction method and morphological filtering. In Algorithm Procedure, the procedure of the algorithm is summarized step by step. Several simulations and conclusions are given in Simulation and Analysis and Conclusion, respectively.

#### 2. Signal Model

##### 2.1. FH Signal Model

Assume that FH signals impinge onto a uniform linear array (ULA) with interantenna spacing ; the DOA is , and the bandwidth of receiver is , where , , and and are the maximum and minimum frequency of the signals, respectively. The th FH signal can be expressed as follows:where denotes the amplitude of the th signal, is the frequency of the th FH signal at the moment , and is the initial phase of the th FH signal.

The model after discrete sampling can be expressed as follows:where is the sampling frequency.

Assume elements of the array are isotropic, ignoring the effect of mismatch and mutual of antenna. Assume the incidence angle of the th signal is , and its steering vector at th discrete moment can be described as

Both frequency and incidence angle can be deduced from (3). Define as “spatial frequency”, where , when , and , so (3) can be rewritten as

Therefore, the signal vector at the th sampling moment can be expressed as follows:where is the array manifold array of the th sampling moment and denotes incoming spatial signals with incidence angle vector and spatial frequency vector . The steering vector changes with the signal frequency of FH signal. denotes the Gaussian White Noise vector with zero mean and variance , where denotes identity matrix.

##### 2.2. Spatial Frequency Estimation Based on Sparsity

In this paper, the interelement spacing meet , and the incidence angle of signal is , so that the range of spatial frequency is (−0.5, 0.5). If the sets of FH frequencies are finite, the possible combination of spatial frequencies are also finite, which satisfy the sparse condition in the spatial frequency domain. Spatial frequencies can be divided evenly into a complete discrete sets , whose spacing of division is . The quantization error of each element in the spatial frequency corresponding to each signal is less than or equal to , which makes the spatial frequencies at any moment constitute a small subset. In this way, (5) can be extended to the output model of redundant arrays as follows:where ’s columns are steering vectors corresponding to each element in the sets . are the extended version of vector with frequencies from , which also extended from . if and only if . can be renamed as for simplicity.

snapshots of received data are grouped as an operational frame. Data in two consecutive frames are not overlapped. The th frame can be expressed aswhere and and denotes snapshots number. The position of nonzeros in is related to .

usually have small sampling interval, whose number of samples satisfies . This means is with great sparsity. In order to estimate parameters from (5), should be bound to sparsity while obtaining best fitting of to . Therefore, the optimization target function can be written as follows: where is the penalty factor and .

By the optimization above, the sparse solution can be obtained, whose positions and amplitude of nonzero rows indicate the spatial frequencies of signals and signal amplitudes, respectively.

#### 3. Proposed Method

The proposed method based on sparse Bayesian reconstruction is presented as the following steps. Firstly, signal frequency and DOA are roughly estimated by spatial frequencies obtained from reconstruction processing. Then, the estimated frequency and DOA are processed by morphological filtering, which produces precise parameter estimation and network sorting.

##### 3.1. Spatial Frequency Estimation

According to sparse Bayesian theory and (5), the probability density function of amplitudes from output data can be expressed as follows:where denotes the variance of the th frame.

To facilitate notation, intermediate variable is introduced to denote the power spectrum of incidence signal, and where and of different moments are mutually independent, and denotes Gaussian distribution with zero-mean variance of . According to Bayesian Probabilistic Theory, the posterior probability density function (PDF) of with respect to can be expressed aswhere and denote the first and second moments of . Given and , and can be estimated as

From (12) and (13), it can be seen that the precision greatly depends on the prior distribution of and estimation of . Precise results can be obtained only when and can correspond to the energy of incidence signal.

To obtain more precisely, the observation data is used for optimization. The likelihood function of with respect to is

Taking logarithm of (14) and omitting constant terms, the target function with respect to iswhere is the estimated array output covariance matrix. By minimizing (15), , the power spectrum of incidence signal on spatial frequency sets , can be obtained.

In practice, it is difficult to optimize directly on (15), so expectation maximization (EM) algorithm is used to estimate . Every iteration of EM includes an E-step and an M-step. In E-step, the first-order moment and second-order moment are estimated by maximizing (11). In the M-step, the iteration of is estimated by calculating given :where denotes the th column of the matrix after the th iteration and denotes the element on the th row and th column of the matrix after the th iteration, which can be deduced from (12) and (13).

To improve the speed of iteration and avoid exception caused by many zeros in , this iteration can be modified aswhere is a small positive value.

The estimated power can be obtained after th iteration. The number of sources is known to be . The first peaks ordered by amplitude are reserved for further processing. Every peak of the spectrum corresponding to every incidence signal usually contains multiple consecutive nonzero amplitudes because of quantization error of . The spatial frequency can be roughly estimated via the linear interpolation of the main peak and the second largest amplitude beside the peak, as follows:where and denote the spatial frequency of the largest and the second largest value of the th peak in the spectrum , whose power is and , respectively.

Based on the rough estimation above, the spatial frequency estimations in the th iteration can be obtained. The unbiased estimation of corresponding noise iswhere , the notation denotes pseudo-inverse operation. When and meet the requirement for convergence, the iteration is terminated, and the power spectrum is estimated to be , which is taken into (12) to estimate the first-order moment .

##### 3.2. Precise Estimation of Spatial Frequency

The reconstruction in Section 3.1 would lead to estimation error in the range of . To reduce the quantization error, a precise estimator for spatial frequency is given in this section, based on the results of previous reconstruction.

By the algorithm in Section 4(1), the noise variance , the range of spatial frequencies , and the covariance matrix of signal waveform can all be obtained, where and denote spatial frequency for the largest and the second largest peaks in the th peak of the spectrum . Let and . can be obtained by crossing out two rows and two columns corresponding to and . . Equation (15) can be modified as follows in order to estimate spatial frequency precisely:where is the power of the th signal.

The target function for signal power and spatial frequency is

It can be inferred from ,

Taking (22) into , (23) can be deduced, where :

The precise estimation of spatial frequency is as follows:

In practical applications, the peak position can be obtained by searching in the smaller range , so that the spatial frequency can be precisely estimated. The computational complexity of the process above is moderate because it is conducted consecutively for signals; the search range of which is only .

##### 3.3. DOA Estimation and Network Sorting

The spatial frequencies of signals can be obtained by algorithms listed in Sections 3.1 and 3.2. amplitudes can be obtained from peak positions from the estimated first-order moment , where are signal vectors. The number of samples is small, so phase-difference method is used to estimate frequency for signals, with the following formula:where means calculating the phase of a complex number.

Therefore, the incidence angle can be estimated by with signal frequencies:

The incidence angle and frequency of signals can be obtained from every frame. Considering the spatial incidence angle of the signal in the processing time is basically invariant, those angles are divided into different clusters with known number of sources as . The frequencies corresponding to different spatial incidence angles are also categorized, so that different signals of different networks are sorted.

##### 3.4. Parameter Estimation Based on Morphological Filtering

In the frequency estimation of FH signals, one frame data may include information about two FH frequencies of the same signal; therefore, the estimation results are often in low credulity, since they suffered from great fluctuation in low SNR condition.

Many image filtering algorithms have been proposed to exclude these data points. Among these methods, median filter, wiener filter, and morphological filter are conventional methods to deal with binary image. Median filter is often used to suppress salt-and-pepper noise and it is a nonlinear filter. But the image we are processing has the feature that there is only one nonzero value in every time index; the median filter method is inefficient with the image of this type. Wiener filter is also a nonlinear filter. Under the criterion of minimum mean-square error, wiener filter is the best filter. Since the wiener filter needs iteration, it has a high computational complexity. morphological filtering method is used to improve the time-frequency spectrum since it has lower computational complexity and it is easier to be realized.

Consider frequency estimates , ; , which is converted to binary image data after quantization by the following formulas:where and denote interval and bandwidth.

The obtained binary image is going through erosion and dilation operation. The scatter points of estimates can be eliminated by erosion, while the boundary can be restricted. The void in the signal can be filled by dilation, but the boundary would also be expanded. The erosion and dilation operations using structural element on the image are expressed aswhere is the domain of .

There is only one nonzero value in each column of the image, and the probability of presence of consecutive points with greater fluctuation is small, so that the structural element is as . By erosion and dilation, outliers can be excluded while normal values are unaffected, generating a clarified time-frequency matrix . The matrix is then turned back to the original estimates by coordinate transformation, with eroded scatter points modified by neighboring points. This transformation can be expressed as follows:where

The modified time-frequency spectrum is turned into binary form and analyzed for component connectivity in order to get the hop moment, period, and frequency sets. The binarization is the same as (27). As the time-frequency points are consecutive by time, connected component detecting could be used to extract every hop of every signal. Eight connected components’ detecting can get a better selection of the hop sequence than four connected components’ detecting for its loose condition. The four connected components and the eight connected components could be seen in Figure 1.