Automatic recognition of sonar targets using feature selection in micro-Doppler signature

As Sffri , Seyed-Hmid Zhiri ,*, Mohmmd Khishe

a Department of Electrical Engineering, University of Birjand, Birjand, Iran

b Department of Electrical Engineering, Imam Khomeini Marine Science University, Nowshahr, Iran

Keywords:Micro-Doppler signature Automatic recognition Feature selection k-NN PSO

ABSTRACT Currently, the use of intelligent systems for the automatic recognition of targets in the fields of defence and military has increased significantly.The primary advantage of these systems is that they do not need human participation in target recognition processes.This paper uses the particle swarm optimization(PSO)algorithm to select the optimal features in the micro-Doppler signature of sonar targets.The micro-Doppler effect is referred to amplitude/phase modulation on the received signal by rotating parts of a target such as propellers.Since different targets' geometric and physical properties are not the same,their micro-Doppler signature is different.This Inconsistency can be considered a practical issue(especially in the frequency domain) for sonar target recognition.Despite using 128-point fast Fourier transform(FFT)for the feature extraction step,not all extracted features contain helpful information.As a result, PSO selects the most optimum and valuable features.To evaluate the micro-Doppler signature of sonar targets and the effect of feature selection on sonar target recognition, the simplest and most popular machine learning algorithm, k-nearest neighbor (k-NN), is used, which is called k-PSO in this paper because of the use of PSO for feature selection.The parameters measured are the correct recognition rate,reliability rate,and processing time.The simulation results show that k-PSO achieved a 100%correct recognition rate and reliability rate at 19.35 s when using simulated data at a 15 dB signal-tonoise ratio (SNR) angle of 40°.Also, for the experimental dataset obtained from the cavitation tunnel,the correct recognition rate is 98.26%,and the reliability rate is 99.69%at 18.46s.Therefore,the k-PSO has an encouraging performance in automatically recognizing sonar targets when using experimental datasets and for real-world use.

1.Introduction

The growing usage of automatic target recognition systems in various military and civilian industries has elevated the subject of automatic target identification and detection to a challenge for craftspeople in this field[1].The primary advantage of using these systems is that they eliminate the need for human intervention in target identification and detection processes [2,3].The primary reasons for replacing these systems with human-operated systems are the slow human reaction time, low dependability, and heavy reliance on mental states [4,5].

Automatic recognition of various sonar targets is critical for defensive and military missions at sea[6].Due to the complexity of real-world sonar targets' physical qualities, intelligent systems for recognizing this sort of data offer unmatched capabilities [7,8].

There are two general approaches to classifying sonar datasets.The first method is to use definitive computational methods,which have high reliability and always produce the best answer;however,the disadvantage of this method becomes apparent as the size of the data increases (as with sonar data), leading to an increase in spatial and temporal complexity [9,10].The second approach is to use stochastic methods.Despite giving a near-optimal solution,these strategies have a lower temporal and spatial complexity than certain others [11,12].The most well-known stochastic approach has been established to be the use of artificial intelligence techniques [13].Machine learning is the most fundamental subset of artificial intelligence [14].Supervised learning is the most suited technique for classification problems[15].In this way,the learning algorithm estimates the output of the training data in different iterations.Then, an observer corrects these outputs, and when the algorithm achieves an acceptable performance, the learning process ceases [16].

Analyzing targets'emitted audio signals is critical for classifying sonar targets using different artificial intelligence systems[17].The majority of the sound propagated is produced by the propeller area and its spinning components [18,19].With the extensive and detailed study that has been done so far on the methods of automatic recognition of sonar targets, it can be said that the only method that has a suitable mathematical model and theory to simulate the reference target database is the method of using modulation of rotating parts of sonar targets.Modulating rotating parts in the science of automatic recognition of targets is known as the micro-Doppler phenomenon [20,21].Extensive investigations on the use of radar micro-Doppler for detection and recognition of ground and air targets have been done in recent years [22,23].However,this phenomenon is seldom employed for sonar targets in the sea and ocean.

The sound reflected from the rotating part (propeller) of sonar targets,on the other hand,has various frequency signatures,which are used to recognize different targets.Despite frequency signatures, the extraction of optimum features minimizes complexity and, as a result, considerably decreases processing time.Such a system in the defence industry requires real-time processing with high reliability.Therefore,the best technique for selecting optimal features is to use meta-heuristic algorithms.In recent years, the field of optimization has witnessed a real tsunami of metaheuristic algorithms [24,25].For example, reference [26]clearly shows that the Gray Wolf, Firefly, and Bat Algorithms are all metaphors from the PSO.This reference provides evidence that these algorithms use new metaphors, change terms, and confusingly present the algorithm, obscuring the fact that they are all PSOs.As a result, the optimization algorithm used to feature selection is PSO.

Therefore,the main contribution in this paper is as follows:

(1) Preparation of sonar simulated datasets for 25 real reference targets with different SNR ratios and different viewing angles;

(2) Obtaining practical sonar datasets using cavitation tunnel;

(3) Use micro-Doppler signature of sonar targets for automatic target recognition usingk-NN;

(4) Use PSO to feature selection in sonar targets micro-Doppler signature for automatic target recognition usingk-PSO;

(5) Investigate correct recognition rate, reliability, and processing time in both datasets.

The paper's second section is dedicated to related works involving sonar target classification.The third section discusses how to simulate a micro-Doppler-based dataset as well as how to obtain and collect practical datasets using the cavitation tunnel.Feature extraction and feature selection are described in section fourth.Section five examines the experimental results.The sixth section is concerned with the conclusion.

2.Related works

Automatic sonar target recognition's vast commercial and military applications have recently attracted many researchers on the subject.Table 1 summarizes some recent studies.

Table 1 shows the use of the mel-frequency cepstral coefficient(MFCC) for feature extraction.Obviously, the extraction of 40 different feature vectors in the MFCC method is associated with an increase in the correct recognition accuracy.Furthermore, the complexity increases significantly, increasing by processing speed.Therefore, the MFCC approach is not suitable for real-time applications.

Processing speed, recognition accuracy, and reliability are essential for defence and military applications, especially during the mission.As a result, this article selects the optimal features using PSO and constructs the recognition system most simply and efficiently possible.

3.Obtain sonar datasets

In this section, the simulated dataset based on the micro-Doppler signature and the data obtained from the cavitation tunnel in the practical experiment are used as follows.

3.1.Simulated sonar dataset based on Micro-Doppler signature

This section starts with a brief overview of the micro-Doppler effect.Then, the sonar micro-Doppler simulation steps are described by adjusting the model parameters according to the dimensions of the actual floating propellers.

Table 1Some recent related work research.

3.1.1.Micro-Doppler phenomenon

Micro-motions, such as vibrations or rotations of an object or structure on an object, cause changes in the extra frequencies on the signal,leading to side-bands on the object's Doppler frequency[21].This phenomenon is called micro-Doppler sonar.Recent research has shown that micro-Doppler techniques can identify or classify a target with its micro-Doppler properties.To explore the micro-Doppler properties of an object, time-frequency analysis is used to provide information about these local properties over time and frequency [18].In most cases, the micro motion has a unique signature object.Micro-motion is created directly by the dynamic motion properties of an object, and the micro-Doppler properties are a direct reflection of micro-motion.So we can classify an object with unique dynamic motion properties using its micro-Doppler signatures.The theory of this phenomenon is discussed below:

The analytic signal of a pure tones(t) is defined as the signals^(t), such that s(t) =Real{s^(t)}, and is generally expressed in the polar format as [42].

A target moving at a constant radial velocityvhas the following Doppler shift relative to the sonar (or radar)system

wheref0is the carrier frequency of the active sensor andCsis the speed of propagation of sound in water (or air).If the target has a numberMparts and each part moves at a velocity component vi（t）,the Doppler shift is the sum of each single Doppler shift.

For such a target, the analytical signal of the echo return is as follows:

The above relation makes it possible to extract the Doppler signature from the data.This signal component contains the micro-Doppler information on the target, which can be used for target recognition and classification.

3.1.2.Sample rate and reference targets

In some ways,the sampling rate is a critical component of signal processing.Without an appropriate sample rate, the sampling quality may be compromised, resulting in the loss of critical information from the signal.As a result, to preserve the signal's quality and resolution,a sampling rate of 1000 kHz in the range of 0-0.4 s was employed.

One of the most significant obstacles to sonar research is a dearth of reliable data.On the other hand,factors such as the sea's complex and varied environment and the existence of undesired signals (noise, clutter, and reverberation) motivate the development of a simulated dataset utilizing a mathematical model of the propeller's return signal.The targets examined are listed in Table 2.

Table 2Information about reference targets.

3.1.3.Mathematical model for extracting Micro-Doppler signature

To generate a dataset of return signals from the rotating part(propeller) of sonar targets, discussed in subsection 4.2, a suitable mathematical model(Eq.(6))has been used to simulate the return signal from the propeller.How to obtain a return signal using Eq.(6)is shown in Fig.1.

The parameters used in Eq.(6) are described in Table 3.

This model assumes that the target is in the range of the transmitted signal, and the signal hits the target propeller.

The parameterArplay an essential role in Eq.(6).For example,if the target is along the sonar receiver (Ar=1), the signal is almost wholly modulated, whereas if the target propeller is located right next to the receiverAr= 0), no signal is generated (sr（t） = 0).

There are many random processes in seawater and oceans.Each of these random processes has its own distribution.According to the central limit theorem, when random processes increase with any type of distribution, their distribution is Gaussian.Therefore,noise mixed with the return signal from the target is assumed to be white Gaussian noise,and its random samples are simulated using uniform random variables.We change the noise power by changing its variance.Different SNR ratios are performed by separately changing the noise power for each target.The signal strength for each target is its corresponding power at the same viewing angle.

The reference classes correspond to the twenty-five objectives of Table 2 samples of each class include feature vectors in nine SNRs(20 dB,15 dB,10 dB,5 dB,0 dB,-5 dB,-10 dB,-15 dB and-20 dB)and eight viewing angles(10°,20°,30°,40°,50°,60°,70°and 80°).Each class contains 30 samples in SNR and specified viewing angles.Thus, there are 2160 samples in each class (corresponding to each target)for all SNRs and viewing angles.In total,the dataset contains 54,000 samples.

Fig.1. How to obtain a return signal using Eq.(6).

Table 3Parameters Eq.(6).

3.2.Sonar dataset obtained from cavitation tunnel

The micro-Doppler signature is obtained from the rotating parts of the propellers of ships and submarines.As a consequence, four different propeller models were installed in the cavitation tunnel test location, and sonar dataset were obtained using two hydrophones to validate the system.According to this,it can be inferred that since the collected sonar was obtained directly from the target propeller, then there is a micro-Doppler effect in the collected signal.The NA-10 England cavitation tunnel was used to create and gather realistic sonar data; Fig.2 and Table 4 details the tunnel's size and characteristics.Four propellers with different diameters and specs were utilized in this experiment.Table 5 lists the parameters of each propeller.The propeller rotates at 1800 RPM,and the flow velocity is 4.5 m/s in this test.

Two model hydrophones, B&K 8103, have been used to receive sonar in the cavitation tunnel.The hydrophones are coupled with the data accusation board model UDAQ_Lite.In Table 6, Specifications of the data accusation board are reported.Fig.3 shows samples of signals obtained with different propellers in the cavitation tunnel.

4.Feature extraction and feature selection

This section has two subsections.The first sub-section discusses how feature vectors are extracted.The second sub-section explains about using PSO to select the optimal features.

Fig.2. Cavitation tunnel model NA-10 England.

Table 4Technical specifications of cavitation tunnel model NA-10 England.

Table 5The parameters of each propeller.

Table 6The data accusation board model UDAQ_Lite specification.

4.1.Feature extraction

Due to the need for real-time processing in defence and military applications, the feature extracted from these signals is the 128-point components from Fast Fourier Transform (FFT).Eq.(7) expresses the structure of the target property vector for the simulated data at the angle of view(θ)and the specified SNR ratio as follows:

Each of its components corresponds to the point of 128-point FFT in the angle of view and the specified SNR.Fig.4 shows samples of simulated acoustic signals and frequency signatures from sonar micro-Doppler at different SNRs No.8, and Fig.5 shows the effect of the viewing angles on the return signal at SNR = -20 for target No.8 (see Fig.5).

4.2.Feature selection

Feature selection can be recognized as the process of identifying useful features and removing useless and duplicate features.The feature selection aims to obtain a subset of the feature that solves the problem well with minimal performance reduction.

One of the powerful feature selection methods is the use of meta-heuristic algorithms [43,44].Swarm intelligence (SI) is an intriguing subfield of multi-solution meta-heuristic approaches.Natural colonies,herds,and groupings serve as the primary source of inspiration.Some of the advantages of using SI algorithms are as follows:

SI algorithms store search space information over a period of repetition.Evolutionary algorithms (EAs), on the other hand,discard information about previous generations.SI algorithms often use memory to store the best solution available.SI algorithms usually have fewer parameters to adjust [45,46].Therefore, this essay makes use of the most widely used and well-known SI subset,the PSO.The following is a summary of the principles of PSO.

4.2.1.PSO

PSO is one of the most widely used SI algorithms and a population-based self-adaptive optimization technique developed by Kennedy and Eberhard[47,48].This method starts by creating a group consisting of completely random components, and the search is done in the main loop with continuous repetitions.Two perspectives have been considered to model the existing order in the collective movement of these creatures.One dimension is the existing social interactions between group members,and the other dimension is the individual privileges that each group member may have.In the first dimension, all group members are obliged to always change their position by following the best person in the group.In the second dimension,it is necessary for each member to memorize the best situation they have personally experienced so far and to be inclined towards such a situation.Therefore, each member can become the group leader so that the others have to follow them.Eq.(8)determines the new velocity for each group of particles,and Eq.(9)shows how the spatial position of the particles changes using the calculated velocity.

In these equations i=1,2,…,pop represents each particle,and pop represents the size of the population.Vi=（v1.v2.….vi）andXi=（x1.x2.….xi）represent the velocity and position of theith particle.whilepbest.iandgbestrepresent the historical best solution of the ith particle and the global best solution, respectively.Besides that,c1andc2denote the position constants,whilerand1andrand2are two random values generated from [0,1].Moreover, ω andtrepresent the inertia weight and current iteration number, respectively.

Since the search space is binary when it comes to feature selection, PSO must be converted to a binary version of PSO before use.Fig.6 shows the pseudo-code converting the PSO to its binary version.Therefore, feature selection is done using the PSO binary version in this article.The PSO flowchart used is shown in Fig.7.

4.2.2.Feature selection using PSO

Subsection 3.1,which describes how to simulate the signal,uses 25 reference classes.Each class is simulated at a specific SNR and angle of view.Due to the random nature of the noise,each signal in a class is simulated 30 times with SNR and a specified viewing angle.Of the 20 samples, it is used to select features, and the remaining 10 samples are used as unknown targets.On the other hand, as stated in, feature extraction is based on 128-point FFT.Therefore,the feature vector contains 128 members.Therefore,the extracted feature matrix for each class with SNR and specified viewing angle and different noises are

Fig.3. Samples of signals were obtained with different propellers in the cavitation tunnel.

Fig.4. samples of simulated acoustic signals and frequency signatures of sonar micro-Doppler in different SNRs for target No.8.

Fig.5. Effect of the viewing angles on the return signal at SNR = -20 for target No.8.

Fig.6. Pseudo-code converting PSO to the binary version.

There are 25 matrixes 20×128 for the simulated dataset,which adds complexity.Therefore, to avoid complexity, data fusion is performed by averaging each column in the matrix 20×128 in Eq.(10).The feature matrix for each class with SNR and specified viewing angle after data fusion is

Then, the new feature space (Eq.(11)) is converted to binary feature space using the pseudo-code of Fig.6.As shown in Table 7,the number of possible modes for selecting the feature is 2128- 1.The reason for deleting one of the modes is not to use any of the features.

Each particle in the PSO is a 25×128 feature matrix for every 25 classes, with a specified SNR and viewing angle, and in the binary feature space.Then,according to the number of particles,and in the dimensions (dimension refers to the number of selected features)of one to 128 dimensions, the features are selected completely randomly.To clarify the initialization of particles,assuming a threedimensional state, the initialization of particles in first run is as described in Table 8:

The fitness function is exactly equal to the correct recognition rate.k-NN calculates the correct recognition rate for each particle as a fitness function.The calculation of the correct recognition rate is done in dimensions 1 to 128.Finally, according to the correct recognition rate in different dimensions, the output (gbest) is obtained with the best selected features.For example, it is assumed that in the four-dimensional state the best features are achieved and in this case is as follows:

As shown in Eq.(12),gbestperforms best on the first, second,fifth,and seventh selected features to identify targets labeled Class 1.As a result, in dealing with unknown targets, the frequency components associated with these features are most effective.

In this paper, a simulated dataset based on micro-Doppler signature and data obtained from the cavitation tunnel are used in a practical experiment.For the simulated dataset,out of every 30 signal samples with SNR and specified viewing angle,10 samples are used as unknown targets.In addition, the practical dataset obtained from the cavitation tunnel is also used for unknown targets.Thek-PSO approach when facing an unknown target is that after the feature selection step(obtaininggbest)from the simulated micro-Doppler signature database, the nearest neighbor value for each class is calculated according to the selected features.The unknown target is assigned to the class with the shortest distance.

Thek-PSO classifier then performs the classification using the selected features.

5.Experimental results

In this section, the measured indicators are described in subsection 5.1.k-PSO performance is then tested in two steps.Therefore,in the first step and subsection 5.2,the simulation results are discussed for the simulated unknown targets.In the second stage,the performance evaluation ofk-PSO was measured using an experimental dataset obtained from the cavitation tunnel.Then,the classifiers were trained using sonar micro-Doppler, and to test it, practical sonar data obtained from cavitation tunnel were used.

5.1.Measured parameters

Measured indicators are correct recognition rate, processing time, and reliability.Reliability is another essential indicator in pattern recognition.It somehow determines the validity of the final decision of the classifier in the face of a pattern.Occasionally, a classifier may accurately recognize all instructional patterns inside a given class.However,due to the inclusion of patterns from other classes into the scope of that class, the reliability of the decision is reduced.For example,in Fig.8,two different classes whose samples are distinguished from each other by white and black bullets in two-dimensional feature space are separated byd(x).The decision functiond(x)assigns one of the black patterns to the white class.Therefore, the detection rate of black bullets byd(x)equals 80%.This is while the reliability of the black class is 100%.Thus,although the rate of correct detection of white bullets byd(x)is 100%,the rate of confidence in the classifier's decision to assign a pattern to the white class is 5/6 = 83%.

Fig.7. PSO flowchart.

Table 7Different feature modes.

The reliability of each class is defined as Eq.(10).

whereRiindicates the reliability of classith,piis the number of classith training points that are correctly classified andpTotalis the total number of training points placed by a hyperplane in the area belonging to this class.

Table 8The initialization of particles PSO in first run for three-dimensional state.

Fig.8. White class detection rate is 100%,its reliability is 83%,and black class detection rate is 80%, and its reliability is 100%.

5.2.Simulation results for unknown sonar targets using k-PSO

This section investigates the simulation results obtained from thek-PSO.The classification results are obtained from an average of 20 runs for each of the experiments.Each experiment assumes that the target angle of view and the SNR ratio is known.Due to the random nature of the noise, in order to make the reference information in each signal as comprehensive as possible,the production of random noise samples is repeated 30 times.20 samples are used to form a reference class for a specific target, and another 10 samples are used as test data (unknown targets).

Table 9 shows the correct recognition rate, reliability, and processing time usingk-PSO andk-NN classifiers for the simulated dataset.The value ofkink-NN andk-PSO is assumed to be equal to 3.

Fig.9 shows the correct recognition rate in terms of viewing angle at SNR=15 dB,and Fig.10 shows the correct recognition rate in terms of SNR at a viewing angle of 40°for both classifiers.

Fig.11 shows the reliability rate in terms of viewing angle at SNR=15 dB,and Fig.12 shows the reliability rate in terms of SNR at a viewing angle of 40°for both classifiers.

Fig.13 shows the correct recognition ratek-PSO of the classifier in all SNRs and the angles of view of the simulated dataset.Fig.14 shows the reliability ratek-PSO of the classifier in all SNRs and the angles of view of the simulated dataset.

In general,the simulation results show that thek-PSO classifier has an encouraging performance in terms of correct recognition rate, reliability, and processing time.However, to prove the efficiency ofk-PSO as well as to confirm the mathematical model used in the next step, the simulation results were performed using experimental sonar data obtained from the cavitation tunnel.Theclassifiers were trained using simulated data, but the sonar experimental data discussed in subsection 3.2 was used to test the classifier.Table 10 shows the average correct recognition rate,reliability rate,and processing time for the classifiers used.

Table 10 shows that the performance ofk-PSO is promising for the experimental data obtained from the cavitation tunnel.Due to the addition of a subsystem for extracting optimal features, the amount of processing time has increased compared to thek-NN classifier.However, 18.46 s is appropriate for recognizing sonar targets.On the other hand, achieving a correct recognition rate of 98.26% with a reliability of 99.69% has led to this classifier being recommended for use in the real world.

6.Conclusions

Fig.9. Comparison of correct recognition rate in terms of viewing angle at SNR=15 dB for both classifiers.

Fig.10. Comparison of correct recognition rate, in terms of SNR at 40° viewing angle for both classifiers.

This paper presents a new method for automatically identifying sonar targets by feature selection from the sonar micro-Doppler signature.In thek-PSO method, in addition to the classification of targets in the feature space based on the signal returned from the sonar targets propeller, PSO is used to select the most optimal features.In this way, the types of sonar targets can be recognized accurately and reliability and in a considerable time.The simulation results in working with experimental data indicate thatk-PSO have significant performance for automatic detection of sonar targets.Although due to military and commercial applications as well as the novelty of the micro-Doppler signature for sonar targets located in a complex ocean and sea environment,it is necessary to research and study various machine learning solutions.However, based on the findings ofk-PSO on experimental data presented in this paper,the employment of this classifier in real-world applications appears to be quite promising.A few research directions can be proposed for future work by automatically recognizing sonar targets: using artificial neural networks, using hybrid classifiers to achieve more accurate accuracy,and using other machine learning algorithms to improve accuracy, reliability, and processing time.This paper was conducted in three general phases.The first phase was the preparation of two sonar datasets.The first dataset is obtained from a suitable mathematical model, which has the ability to simulate sonar in SNRs and different angles of view.The second dataset is obtained from practical experiments using a cavitation tunnel.In the second phase, feature extraction was performed using 128-point FFT.The PSO then selects the most optimal features.In the third phase,k-PSO andk-NN performed automatic sonar target recognition.

Fig.11. Comparison of reliability rate in terms of viewing angle at SNR=15 dB for both classifiers.

Fig.12. Comparison of reliability rate, in terms of SNR at 40° viewing angle for both classifiers.

Fig.13. Correct recognition rate, k-PSO classifier in all SNRs, and viewing angles of the simulated dataset.

Fig.14. Reliability rate, k-PSO classifier in all SNRs, and viewing angles of the simulated dataset.

Table 10Simulation results with classifiers used for the practical test dataset.

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.