1. Introduction
Direction of arrival (DOA) estimation has been an important research field for scientific researchers because of its wide application scenarios. Its application in 5G and internet of things communication [
1,
2], medical imaging [
3], earthquake monitoring [
4] and underwater acoustic signal processing [
5,
6] greatly facilitates people’s production and life. With the rapid development of information technology, more and more application scenarios and complex application environment bring more challenges to DOA estimation technology. In real life, the signals are often mixed with noise and the effective signals are limited, so it is necessary to study the DOA estimation algorithm under the condition of small snapshot and low signal-to-noise ratio (SNR).
The early direction-finding algorithm is mainly developed around beamforming technology, which can enhance the desired signal and suppress interference by weighting each element in spatial filtering. However, the technology cannot distinguish the space target within a beam width, which limits the estimation accuracy of the technology. In order to improve the resolution of this technology and make the algorithm adapt to the changes of signal environment, researchers use various criteria to determine the weights [
7]. The representative article is Capon’s minimum variance method [
8]. However, this kind of method assumes that the signal is a spatially stationary random process, and this assumption is not valid in real signals, the algorithm has a large error or even failure in low SNR environment, which leads to some limitations in application.
The emergence of subspace algorithms greatly improves the resolution of DOA estimation. The most classic method is multiple signal classification algorithm (MUSIC) [
9], which uses the orthogonal characteristics of signal subspace and noise subspace to construct spatial spectral peaks, and then achieves super-resolution DOA estimation. Another classical algorithm uses the rotation invariance of signal subspace to achieve super-resolution estimation [
10]. The above method and the improved method [
11,
12,
13,
14] depend on the statistical characteristics of the data, so it needs a lot of snapshots data to achieve accurate estimation. In addition, the above method cannot effectively distinguish the incident direction of coherent signal. The maximum likelihood based DOA estimation algorithm is a typical representative of subspace fitting algorithms, but it needs multidimensional search to solve the optimal solution, and the amount of computation is huge.
Sparse representation of signals is a signal processing technology that has attracted much attention in recent years. It shows the information contained in the signal as comprehensively as possible with a simple mathematical model, which makes the signal processing more convenient [
15]. DOA estimation algorithm based on compressed sensing (CS-DOA) is proposed in recent years to achieve high-resolution estimation using signal sparsity [
16]. Because it can realize DOA estimation for single snapshot data and coherent signals, this algorithm has been widely concerned. At present, CS-DOA is mainly divided into two strategies according to different signal reconstruction algorithms, one of which is greedy algorithm and its improvement algorithms [
17,
18,
19]. This kind of method search for the support set of unknown signals iteratively, that is, the position of non-zero elements. However, when the wrong target is selected in the process, the target will be retained until the end, making the result biased. Another kind of strategy is based on convex optimization method, including basis pursuit de-noising (BPDN), Least absolute shrinkage and selection operator (LASSO) and
algorithm [
20,
21,
22]. This kind of algorithms needs to choose appropriate regularization parameters to ensure the performance of the algorithm. In addition, because the DOA estimation algorithm based on CS principle needs to divide the spatial domain into grids in advance. The angle of the real signal may not be within the dividing grid, leading to the bias of the result. A grid evolution DOA (GEDOA) estimation method is proposed in literature [
23]. Literature [
24] solves the mesh mismatch problem by introducing bias parameters into the signal model. In the process of grid fission, sparse Bayesian learning method is used to correct the position of spatial spectrum, so as to achieve fast and accurate DOA estimation. These above algorithms based on CS cannot guarantee the estimation performance at low SNR. There are also many literatures that focus on the research on the principle of CS in the mutual-prime array and the location of far and near field signals, and many fruitful works have been done in the above research area [
25,
26,
27,
28,
29].
For the DOA estimation algorithms under the condition of small snapshot number, researchers have also explored other methods [
30,
31]. In reference [
30], a Toeplitz matrix, using which to obtain the signal subspace and noise subspace, is constructed using any row of the received signal’s covariance matrix. Then, the DOA estimation of single snapshot signal is realized by using the subspace method. However, this method needs to know the number of signal sources, and it is still unable to achieve the correct estimation under the condition of low SNR.
Based on the above analysis, the DOA estimation is difficult under the condition of small snapshot number and low SNR. Therefore, this paper proposes a DOA estimation algorithm based on CS and density-based spatial clustering (DBSCAN). Firstly, calculate the power spectrum entropy of the signal, and select the appropriate regularization parameters; Secondly, the CS-DOA method based on the basis pursuit noise reduction strategy is used to estimate the DOA of the small snapshot data of the signal, and the peak data of the estimation result is saved; Finally, the saved peak data is clustered by DBSCAN, and calculate the cluster center value of each cluster, the number of clusters is the number of the signal sources and the cluster center value is the estimated value of DOA. Under the same conditions of low SNR and small snapshots, the proposed method can get effective DOA estimation, which has more advantages than the traditional super-resolution algorithms. In the real underwater acoustic experiment environment, when there is a short pulse signal, the DOA estimation still can be obtained by this method. The method proposed in this paper does not need to predict the number of signal sources. In the case of low SNR, small snapshots, coherent signals and a small number of abnormal points, it can still obtain effective DOA estimation, which has more advantages than traditional DOA methods.
The rest of this paper is arranged as follows: in the second section, the principle of DOA estimation based on CS and its characteristic analysis are introduced, as well as the power spectrum entropy, regularization parameter selection method and the clustering principle based on density. In the third section, the specific steps of the proposed method are introduced. In the fourth and fifth sections, the effectiveness and advantages of the proposed method in simulation experiments and real underwater acoustic experiments of micro electro mechanical system (MEMS) vector hydrophone are verified, respectively. Finally, a conclusion of this paper is given in the sixth section.
2. Theoretical Basis and Analysis
2.1. Principles and Characteristics of DOA Estimation Based on CS Theory
2.1.1. DOA Estimation Principle Based on CS Theory
CS theory is a set of new theory about sparse signal acquisition and compression which was put forward in 2006. It uses the characteristics of sparse signal to sample the signal and compress the data properly, which can effectively reduce the burden of data transmission and processing. It has attracted the attention of the majority of scientific researchers and has been widely used in various fields of signal processing. Now it has formed a set of perfect theoretical basis [
32]. The DOA estimation based on CS theory exactly uses the sparsity of the signal to process the signal.
Suppose that there are
far-field narrow-band signals incident on a uniform linear array composed of
antennas, and the incident angle of the
signal is
. The single snapshot data vector received by the array at time t can be expressed as
where,
represents
source vectors;
is the direction matrix;
is receiving noise for array.
The potential incidence angles of space are divided, and the divided space angles are expressed as
, As shown in
Figure 1, a solid circle represents an actual incidence angle of space, and the hollow circle represents the incidence angle of potential space that does not exist. When the number of potential incidence angles
is much larger than the number of actual incidence angles, namely
, the incident signal can be regarded as a sparse signal in the spatial domain.
In order to take advantage of the sparsity of the signal, the array manifold matrix
is extended to form a super complete redundant dictionary
, which contains all possible angles, namely
Combining Formulas (1) and (2), the Formula (1) can be transformed into a sparse representation problem, that is
where,
is the
dimension coefficient vector. According to the sparse theory, the number of non-zero elements in
in the dictionary is
, which means that the dictionary
is
sparse. The angle of the vector corresponding to the position of the non-zero element represents the value of the incident angle
, and the amplitude of the non-zero element is the amplitude of the signal at the sampling time.
Obviously, Formula (3) is an underdetermined equation, which cannot be solved directly. However, when the sensing matrix satisfies certain sparse reconstruction conditions, the signal can be reconstructed with a high probability. The sparser the signal is, the higher the reconstruction accuracy will be. Candes proved that the signal reconstruction can be solved by solving the minimum
norm problem [
33], which can be solved by the following equation
But
norm minimum optimization algorithm is a non-deterministic polynomial hard problem. In order to ensure the sparsity of signal and the accuracy of signal reconstruction, Formula (4) is usually transformed into an optimization problem with inequality constraints
where
represents a possible noise standard deviation. In general, the solution of Formula (5) is transformed into the solution of the following unconstrained optimization problem
In the Formula (6),
is a regularization parameter related to noise. In the process of signal reconstruction, it affects the sensitivity of DOA estimation to signal noise and largely affects the DOA estimation result is whether correct. The existing literature do not provide good suggestions for the selection of
[
34,
35].
2.1.2. Features of CS-DOA
In order to visually show the characteristics of the DOA estimation method based on the CS theory, and provide suggestions for the selection of the regularization parameter
and the improvement of the method, this article designs some experiments for analysis, considering that different signal reconstruction methods have a great impact on the accuracy of DOA, the BPDN method has higher stability and accuracy than other reconstruction algorithms, therefore, this paper chooses BPDN method to reconstruct the signal, and uses the software for disciplined convex programming (CVX toolbox) to realize the reconstruction process [
36].
Experiment 1: CS-DOA experiment under the same regularization parameter and different SNR. The number of array elements is 10, and the three incident angles are −10°, 15°, 50°. In this experiment, the regularization parameter is 10, random Gaussian white noises are added, the SNRs are 30 dB, 10 dB, −5 dB, −10 dB, respectively. The results of the experiments are shown in
Figure 2.
It can be seen from
Figure 2 that the lower the SNR, the falser peaks, and the spectral value of the false peaks becomes stronger. This indicates that the noises make CS-DOA unable to obtain the correct DOA estimation, and the lower the SNR, the greater the error of DOA estimation. Due to the existence of a large number of false peaks, the number of sources is also difficult to accurately estimate. Even though the algorithm allows some noises in the model, it is still difficult to obtain good DOA estimation results in a low SNR environment.
Experiment 2, CS-DOA experiment with the same SNR and different regularization parameters. The number of array elements is 10, and the three incident angles are −10°, 15°, 50°. In this experiment, the SNR is 0dB, and the regularization parameters are 0.5, 5, 15, and 30 respectively. The results of the experiments are shown in
Figure 3.
It can be seen from
Figure 3 that under the same SNR environment, the size of the regularization parameter is related to the adaptability of the CS-DOA algorithm to noise. The larger the regularization parameter, the stronger the adaptability of the CS-DOA algorithm to noise. When the regularization parameter is large enough, the correct DOA estimate will even be filtered out, as shown in
Figure 3d. It can be seen that the complexity of signal noise can provide a reference for the selection of regularization parameters. How to determine the noise complexity of the signal and provide suggestions for the selection of regularization parameters will be discussed in the next section.
In addition, combining
Figure 2 and
Figure 3, it can be seen that in experiments with different SNRs and regularization parameters, the CS-DOA algorithm can obtain the estimation result near the correct DOA value with a high probability. Even if there are some false peaks in the signal reconstruction process, and even the correct solution is lost, as the number of snapshots increases, if we can use the CS-DOA estimation results obtained from a small number of snapshots for comprehensive analysis, it seems that the correct number of sources and the correct DOA estimation value can be obtained. How to use these results for analysis will be further discussed in the following content.
2.2. Selection of Regularization Parameters
Combining the content of the previous section, we propose that the appropriate regularization parameters can be selected by determining the complexity of the signal noise. In fact, no specific signal component must be regarded as noise. Noise is just a part of the signal that we do not need in the process of signal processing. For example, the sound made by a ship engine when it is running is regarded as a kind of noise for tourists or underwater creatures. When we use a sound to judge the type, size and direction of the ship, the sound emitted by its engine is the useful signal we need, and other sounds become useless noise [
37]. In reality, we only need part of the information in the signal for DOA estimation, and the signal is usually composed of various frequency components and distributions.
Power spectral entropy (PSE) is an extension of information entropy in the frequency domain. Because it can be used to describe the characteristics of frequency changes, it has been widely used in the analysis of the frequency complexity of various signals [
38]. Generally, the smaller the power spectrum entropy of the signal, the stronger the signal sparsity. This paper selects the power spectrum entropy to represent the frequency complexity of the signal, which can indirectly reflect the magnitude of signal noise.
The calculation of power spectrum entropy includes the following three steps:
Step 1, Calculate the power spectrum of signal
:
where
is the signal length and
is the Fourier transform of
;
Step 2, Calculate the probability density function of the spectrum of all frequency components by normalization:
where
is the spectral energy of frequency component
,
is the corresponding probability density, and
is the number of frequency components in the fast Fourier transform of the total probability density;
Step 3, Calculate the value of PSE:
Different from the traditional spectrum analysis, power spectrum entropy quantifies the frequency complexity of the signal in its calculation process, which makes the analysis of the signal characteristics more convenient.
Figure 4 shows the trend of power spectrum entropy of signals in different SNR environments obtained through 100 independent Monte Carlo tests. It can be seen from
Figure 4 that the power spectral entropy of the signal decreases as the SNR increases, and eventually tends to a stable state. When the SNR is greater than 5 dB, the power spectrum entropy of the signal is between 0.3 and 0.4. When the SNR changes from 5 dB to −10 dB, the stronger the noise, the greater the power spectrum entropy of the signal.
Through the analysis in
Section 2.1.2. we know that when the regularization parameters are randomly selected, the signal reconstruction results may be affected by pseudo peaks. In order to analyze the influence of different regularization parameters on the orientation accuracy of the CS-DOA model, we repeated a large number of experiments for different regularization parameters under different noise intensity environments. Save the average value of the effective DOA estimates in the experiment, and find the root mean square error from the true angle.
Figure 5 shows the root mean square error of the CS-DOA model with different regularization parameters in different noise environments.
It can be seen from
Figure 5 that when the SNR is high, the DOA estimation error is small, and the selection of the regularization parameter has little effect on the DOA estimation result. In a low SNR environment, the smaller the regularization parameter, the greater the root mean square error of the DOA estimate. When the regularization parameter increases, the error gradually decreases. When the regularization parameter is greater than 30, the error has a rising trend. Combined with the analysis in
Section 2.1.2, if the real signal energy is weak and the regularization parameter is too large, the correct DOA estimation may be lost.
Based on the above analysis, in the DOA estimation model introduced in this article, the regularization parameters are selected according to the noise complexity represented by the power spectrum entropy. When the power spectrum entropy value is about 0.3, the regularization parameter is selected between 5–10, and when the power spectrum entropy is large, the regularization parameter is selected between 10–20.
2.3. Density-Based Spatial Clustering
It can be seen from
Section 2.1.2, that in experiments with different SNRs and regularization parameters, the CS-DOA algorithm can obtain the estimation results near the true angle with a high probability, and the false peaks caused by noises and other unstable factors are Scattered in different locations. When there are a small number of snapshots, the peaks data can be analyzed to determine where the CS-DOA estimation are denser, and the number of sources and the DOA estimation can be obtained. It is unrealistic to analyze large amounts of data manually. The clustering method seems to be a good choice. It can analyze data with similar attributes in the collection. General clustering methods usually need to input the number of clusters in advance. In reality, we cannot determine the number of classifications in advance, which may bring wrong results. In this article, we use DBSCAN to process the peaks data of CS-DOA.
DBSCAN characterizes the data according to the tightness of the data distribution [
39]. It uses a set of parameters
called “neighborhood” to examine the connectivity between samples, and continuously expands clusters based on connectable samples to obtain the final clustering results.
In order to describe the DBSCAN algorithm in detail, the following definitions need to be understood:
neighborhood: For , its neighborhood contains samples in the sample set whose distance from is less than or equal to . That is , where is the sample data set, and represents the Euclidean distance;
Core object: If the neighborhood of contains at least samples, that is, , then is a core object;
Directly density-reachable: If is located in the neighborhood of , and is the core object, then is directly reachable by density;
Density reachable: For and , if there is a sample sequence , where , and is directly reached by the density of , then is reachable by the density of ;
Density-connected: For and , if there is so that both and are reachable by the density of , then and are density connected.
DBSCAN is a collection of the largest density-connected samples derived from the above-mentioned density reachability relationship. The clustering process of DBSCAN can be roughly divided into two steps: Step 1. Find all the core objects according to the given neighborhood parameter
; Step 2. Take any core object as the starting point to find the clusters generated by the samples whose density is reachable. Stop rule for all core objects are accessed. The pseudo-code flow of DBSCAN is shown in
Figure 6.
Based on the above analysis, the sample set is the DOA estimation result of the small snapshot data, and the neighborhood parameters can also be selected according to the characteristics of the signal data. The selection idea will be discussed in the following part.
3. The Method Proposed in This Paper (CS-DOA-DBSCAN)
Combined with the above basic theories and analysis, an estimation algorithm of the DOA based on CS theory and DBSCAN is proposed (CS-DOA-DBSCAN). The specific steps of the algorithm are as follows:
Step1: The power spectrum entropy of the signal is calculated, and then the regularization parameter range is determined;
Step2: Based on the principle of CS, the DOA of small snapshots data was estimated, and the first 15 maximum peaks data coordinates of each snapshot were saved for the further processing;
Step3: Cluster the peaks data saved in step 2 through DBSCAN to obtain the number of clusters , which are the number of signal sources;
Step4: Calculate the cluster center value of each cluster, which is the final DOA estimation.
Here are the more details for this algorithm: Although step 1 cannot lock the regularization parameter into an accurate value, the regularization parameter can be locked within a relatively small value range. Selecting the regularization parameter within this range can ensure that there are relatively few false peaks in the subsequent estimation of the DOA, and the estimation results are relatively stable, so as not to produce large fluctuations.
Combined with the analysis in
Section 2.1.2. In step 2, the peaks data obtained in the CS-DOA estimation of each snapshot is generally not too much, so the 15 maximum peaks data before each snapshot are retained. In general, DOA estimates near the true angles will be retained. If too much peaks are saved, the workload of the algorithm will increase and the efficiency of the algorithm will be reduced.
In step 3, the number of clusters represents the actual number of signal sources, and the cluster center value represents the final DOA estimation. In general, we pay more attention to the information of incident angles. In order to facilitate clustering, we normalized the amplitude of the data saved in step 2. In the process of clustering, the choice of neighborhood parameter will affect the resolution of DOA estimation. If is too large, the DOA resolution will be reduced a lot, and noise will have more influence on the clustering results, leading to the deviation of DOA estimation. If is too small, in the case of low SNR, noise points may also be clustered into a DOA estimation, resulting in wrong estimation of the algorithm. In the experiment, the value of can be determined according to the complexity of the noise.
represent the minimum number of samples contained in the
neighborhood. The selection of
are related to the number of snapshots of the data. When the noise level is moderate or mild, CS-DOA is not easy to lose the effective solution. The value of
can be equal to or slightly greater than the number of snapshots. It is recommended that no more 1.5 times than the number of snapshots. When the noise is large, CS-DOA may lose the effective solution. The value of
can be equal to or slightly smaller than the number of snapshots. The CS-DOA-DBSCAN algorithm flow is shown in
Figure 7.