THE PERFORMANCE ANALYSIS OF STFT-ANFIS CLASSIFICATION METHOD ON PULSED RADAR TARGET CATEGORIZATION

Transcription

1 ISTABUL UIVERSITY JOURAL OF ELECTRICAL & ELECTROICS EGIEERIG YEAR VOLUME UMBER : 006 : 6 : 1 (97-105) THE PERFORMACE AALYSIS OF STFT-AFIS CLASSIFICATIO METHOD O PULSED RADAR TARGET CATEGORIZATIO Engn AVCI 1 İbrahm TURKOGLU Mustafa POYRAZ 3 1, Frat Unversty, Department of Electronc and Computer Scence, 3119, Elazg, TURKEY 3 Frat Unversty, Engneerng Faculty, Department of Electrc and Electronc, 3119, Elazg, TURKEY 1 E-mal: engnavc@frat.edu.tr E-mal: turkoglu@frat.edu.tr E-mal: mpoyraz@frat.edu.tr ABSTRACT In ths paper, a pattern recognton system s developed for automatc classfcaton of the radar target sgnals. Feature extracton s an mportant subset of the pattern recognton system. For feature extracton s used Short Term Fourer Transform (STFT) tme-frequency dstrbuton of the pulse radar target sgnals. Adaptve etwork Based Fuzzy Inference System (AFIS) classfer s used at classfer part of the pattern recognton system. Radar sgnals are obtaned from pulse radar system for varous targets. The classfer performance s evaluated accordng to the proposal method. Keywords: Feature extracton, AFIS, radar target classfcaton, STFT. 1. ITRODUCTIO Radar s a mcrowave system for detectng objects and determnng ther dstance, or range. The word radar s an acronym of the words rado detecton and rangng. An elementary radar conssts of a transmtter, a recever, transmt and receve antennas, and an ndcator. A rado sgnal s generated by the transmtter and radated by the transmttng antenna. Part of the transmtted sgnal strkes a reflectng object, or target, and s scattered n all drectons. Some of the sgnal from the target s reflected back to the radar. Ths sgnal, called the echo, s captured by the recevng antenna. The recever then detects the echo sgnal whch s demodulated to produce a vdeo sgnal. The vdeo sgnal s sent to the ndcator whch ndcates the presence of a target. The ndcator may also ndcate target dstance, drecton, velocty etc. In practce, the transmtter and recever usually share the same antenna, as shown n Fgure 1. the antenna usually drects the sgnal nto a narrow beam whch s systematcally swept through the regon where targets are expected Pulsed Radar Systems Radars are most often classfed by the types of waveforms they use, or by ther operatng frequency. Consderng the waveforms frst, radars can be Contnuous Wave (CW) or Pulsed Receved Date : Accepted Date:

2 98 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton Radars (PR). CW radars are those that contnuously emt electromagnetc energy, and use separate transmt and receve antennas. Unmodulated CW radars can accurately measure target radal velocty (Doppler shft) and angular poston. Target range nformaton cannot be extracted wthout utlzng some form of modulaton. The prmary use of unmodulated CW radars s n target velocty search and track, and n mssle gudance. Pulsed radars use a tran of pulsed waveforms (manly wth modulaton). In ths category, radar systems can be classfed on the bass of the Pulse Repetton Frequency (PRF), as low PRF, medum PRF, and hgh PRF radars. Low PRF radars are prmarly used for rangng where target velocty (Doppler shft) s not of nterest. Hgh PRF radars are manly used to measure target velocty. Contnuous wave as well as pulsed radars can measure both target range and radal velocty by utlzng dfferent modulaton schemes. The pulsed radar s most used n radar types [1], []. In a pulsed radar system, short bursts of rado frequency (RF) energy are generated for transmsson. Ths s usually accomplshed by frst generatng a tran of narrow, rectangularshape pulses and usng these to ampltudemodulate a snewave RF carrer. The pulsed RF sgnal s transmtted by antenna. If the sgnal strkes a target, a porton of the sgnal wll be reflected back to the radar as an echo. The antenna captures the echo pulses whch are sent to the recever. The receved pulses are then demodulated and converted to a vdeo sgnal for dsplay. Antenna Target Azmuth Angle Duplexer Transmtter Recever Indcator Fgure 1. Smplfed radar system. 1.. Short Term Fourer Transform (STFT) Tme-Frequency Dstrbuton Fgure shows a block dagram of the target dentfcaton used n ths work. For accuracy of the early-tme response, we use the nverse Fourer transform. Feature Extracton from STFT: The STFT s the basc method for analyzng no statonary sgnals. In STFT, the sgnal s dvded nto small segments, where these segments of the sgnal can be assumed to be statonary [3]. For ths purpose, a wndow functon s chosen. The STFT s defned n the tme doman as follows: STFT where ( t) [ ] exp ( jω t ) ( τ, Ω) = f ( t) w ( t τ) dt (1) f s a tme-doman sgnal and w t s a wndow functon. From the gven frequency- () Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

3 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton 99 doman backscatterng data, we can obtan the transent response f ( t) usng IFFT. Then, an M STFT matrx of M frequency ponts and tme ponts s computed usng Equaton 1 [4]. But, an M matrx s too large to act as the neural network nput. Consequently, some form of data reducton s requred. Frst, the STFT matrx s dvded nto J tme bands and K frequency bands. Generally, J and K are determned consderng the tme and frequency resolutons. Then, the STFT s ntegrated wthn each of these tme and frequency bands. The component of the feature matrx n the kth frequency band and j th tme band s denoted as F k, j, and s defned as Δτ = j k,j 1 Δ kδ Ω ( ) F STFTτ, Ω dωdτ. ( j ) τ ( k 1) ΔΩ for k = 1,...,K and j = 1,...,J () where Δ τ = T fnal / J and Δ Ω = BW / K (3) are the wdth of each tme and frequency band, wth T fnal and BW beng the fnal tme and bandwdth, respectvely. Usng the above process, an M STFT matrx can be represented by a K J (where K, J << M, ) feature matrx. The fnal feature vector of K J dmenson s represented by X = T [ F, F,..., F, F, F,..., F, F ]. 1,1 1, 1,J,1, K,J 1 K,J (4) We can use these feature vectors as the neural network nput. However, a further data transform s needed because a K J -dmensonal nput space s stll large, and each of these vectors has many redundances, especally n the hghfrequency regon of the late-tme perod Pattern Recognton Pattern recognton s the research area that studes the operaton and desgn of systems that recognze patterns n data. It encloses subdscplnes lke dscrmnant analyss, feature extracton, error estmaton, cluster analyss (together sometmes called statstcal pattern recognton), grammatcal nference and parsng (sometmes called syntactcal pattern recognton). These mportant applcaton areas are radar target recognton, mage analyss, character recognton, speech analyss, man and machne dagnostcs, person dentfcaton and ndustral nspecton [5]. The ablty to recognze patterns s fundamental to computer vson. Here, term pattern refers to a quanttatve or structural descrpton of an object. In general, one or more descrptors form patterns. The pattern space corresponds to a measurement or an observaton space. A pattern vector s referred to as an observaton vector. A pattern vector often contans redundant nformaton; hence, the pattern vector s mapped to a feature vector [6-9]. Pattern recognton systems usually consder a feature space onto whch feature vectors are mapped frst. The feature vector s used to decde the class to whch the nput sample belongs. The purpose of feature extracton s to reduce data by retanng certan features or propertes that dstngush nput patterns. Features correspondng to dfferent shapes, textures, and spectral sgnatures are used, as are features such as varous tme-frequency dstrbutons. In ths developed pattern recognton study, Short Term Fourer Transform tme-frequency dstrbutons of real radar targets Doppler sgnals were used for feature extracton process. Then, Adaptve etwork Based Fuzzy Inference System (AFIS) was used n ths study as classfer. Ths used pattern recognton block dagram s shown n Fgure [10].. PROPOSED METHOD An effcency feature extracton method was developed for sx target objects whch are shown n Fgure 3 to separate one from the others. There, pulse radar Doppler sgnals were used as real nput space. Expermental applcaton was realzed on havng educatonal purpose and mult functon 960/1 Model Lab-Volt radar experment set. Pulse echo sgnals were receved to computer meda by usng data accusaton card has 1 Khz sample frequency. The pulsed radar system parameters were adjusted as bellow: Pulse wdth: 1 ns, RF oscllator: 9.4 Ghz, Pulse Repeat Frequency (PRF):16 Hz Constant Target Range: 85 cm Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

4 100 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton Used pattern recognton mechansm and calculate scheme whch were gven n Fgure 3. We can see that the feature extracton s the most mportant part of pattern recognton system, and drectly mpresses accomplshed of classfer. Inputs : atural states Perceves Feature Extracton by STFT AFIS classfer Output: Recognton Fgure. Block Dagram of the Pattern Recognton System Small metal plaque Large metal plaque Plexglas plaque Corner reflector Sphere Cylnder Fgure 3. Used radar targets. The feature extracton s the most mportant part of pattern recognton and correct pattern classfcaton key. The purposes of feature extracton from sgnals are to rse the accomplshment of classfer whle the classfcaton tme s reduced, to reduce the data quantty wll be processed to mnmum level, and to prove safe of recognton system. For Extracted features sn t mpressed from not be controlled parameters n system, the extracted features should be determned. Thus, the features may be generalzed and safe of systems may be rased [11]. For features extracton of unstable sgnals commonly s nterested n composton of the tme-frequency regon [1], [13]. Thus, defntely data whch ncludes both transent alteraton and frequency alteraton can be extracted from radar Doppler sgnals. In ths study for feature extracton, frstly each of the dfferent targets whch were shown n Fgure 3. Real Doppler sgnals were receved from Lab- Volt Radar educaton set. Secondary, Short Term Fourer Transform tme-frequency dstrbuton (STFT) whch was gven at Equaton was appled to have been obtaned target Doppler sgnals. Thrdly, Gauss whte nose whch was gven at Equaton 5 was appled to ths have been obtaned STFT. SR ratos of Ths Gauss whte noses were changed 1, 3, 5, 7, 9 respectvely [14], [15]. Maxmum value, mnmum value, arthmetc average and geometrc average of ths nosy STFT tme-frequency dstrbutons whch were obtaned for each of target were calculated to form feature vector. umercal values of them were regarded as for feature vector at classfcaton stage [14], [15]. Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

5 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton 101 σs c = (5) SR /10 σ 10 w σ s σ w There, s sgnal varance, s nose varance, SR s sgnal / nose rato, c s nose scale constant STFT tme-frequency dstrbutons of radar targets Doppler sgnals whch were obtaned Lab-Volt radar experment set are gven n Fgure 4. small metal plaque large metal plaque plexglas plaque corner reflector ampltude sphere cylnder frequency tme Fgure 4. To be obtaned STFT tme-frequency dstrbutons of constant targets pulsed radar Doppler 3. THE CLASSIFICATIO STAGE In ths study, the pattern recognton nput space was obtaned by usng feature extracton method whch was gven n Secton. Then, Adaptve etwork Based Fuzzy Inference System (AFIS) classfer algorthm was used for radar targets classfcaton Adaptve-etwork-Based Fuzzy Inference System Algorthm Both artfcal neural network and fuzzy logc are used n AFIS's archtecture. AFIS s conssted of f-then rules and couples of nput-output, for AFIS tranng s used learnng algorthms of neural network [16], [17], [18]. For smplcty, we assume the fuzzy nference system under consderaton has two nputs (x, y, t, and k) and one output (z). For a frst order Sugeno fuzzy model, a typcal rule set wth base fuzzy f-then rules can be expressed as If x A 1 y B 1 t C 1 k D 1 then f 1 = p x + q y + r t + s k + u (6) Where, p, r, q, s, u are lnear output parameters. The AFIS s archtecture whch has four nputs and one output s showed n Fgure [8]. Ths archtecture s formed by usng fve layer and sxteen f-then rules: Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

6 10 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton Layer-1: Every node n ths layer s a square node wth a node functon. th rules frng strength to the sum of all rule s frng strengths: O 1, =μ A (x), for =1,, O 1, =μ C-4 (t), for =5,6 O 1, =μ B- (y), for =3,4 O 1, =μ D-6 (k), for =7,8 (7) Where x, y, t, k are nputs to node, and A, B, C, D are lngustc label assocated wth ths node functon. In order words, O 1, s the membershp functon of A, B, C, D. Usually we choose μ A (x), μ B (y), μ C (t), μ D (k) to be bell-shaped wth maxmum equal to 1 and mnmum equal to 0, such as nmum equal to 0, such as μ c A (x), μ B (y), μ C 4 (t), μ D 6 (t) = exp(((x )/(a )) ) (8) Where a, c s the parameter set. These parameters n ths layer are referred to as premse parameters. Layer-: Every note n ths layer s a crcle node labelled Π whch multples the ncomng sgnals and sends the product out. For nstance, O, = w = μ A (x).μ B- (y).μ C-4 (t).μ D-6 (k), =1,, 3,.,16 (9) Each node output represents the frng strength of a rule. (In fact, other T-norm operators that performs generalzed AD can be used as the node functon n ths layer). Layer-3: Every node n ths layer s a crcle node labelled. The th node calculates the rato of the O3,= w =w /(w 1 +w + +w 16 ), =1,,3,.,16 (10) Layer-4: Every node n ths layer s a square node wth a node functon O 4, = w.f =w.( p x + q y + r t + s k + u ), =1,,3,.,16 (11) Where, w s the output of layer 3, and {p, q, r, s, u } s the parameter set. Parameters n ths layer wll be referred to as consequent parameters. Layer-5: The sngle node n ths layer s a crcle node labelled that computes the overall output as the summaton of all ncomng sgnals: w f O 5, = overall output = w f = (1) w In ths study, to be obtaned 4 x 30 feature vector whch was stated n Secton was gven nputs of AFIS classfer as nput sets. Outputs of ths AFIS classfer formed from a decson space = {small metal plaque, large metal plaque, Plexglas plaque, corner reflector, sphere, cylnder} that represents sx number dfferent real radar targets. Ths real radar targets are shown n Fgure 3. Ths AFIS classfer was tested by usng hundred numbers nosy test data for each of sx targets. Be obtaned classfcaton performance results of the AFIS classfer are gven on Table1. Table 1. Achevement of AFIS Classfer (%) Target Object Small Metal Plaque Large Metal Plaque Plexglas Plaque Corner Reflector Sphere Cylnder Small Metal Plaque Large Metal Plaque Plexglas Plaque Corner Reflector Sphere Cylnder Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

7 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton 103 Layer-4 Layer- Layer-3 w1 w1 x y t k w w 1 * f 1 Layer-1 w 3 w 4 Layer-5 x A 1 A w 5 w 6 w 7 f y B 1 B w 8 w 9 t C 1 C w 10 w 11 k D 1 D w 1 w 13 w 14 w 15 w 16 x y t k Fgure 5. Used AFIS archtecture of 4-nputs and 16-rules n ths study. 4. COCLUSIOS In ths paper, proposng feature extracton method n Secton was appled to real pulsed radar sgnals. A hundred percent determnaton functons were composed by usng AFIS classfer. In addton to, there are clear dfferences among the determnaton functons are understand as to be seen n Table 1. These ndcators show to have been extracted feature whch strongly and effectvely from natural nputs. The determnaton functons of system at decson space are very clear. The features whch were selected for feature vector very good summarze. In addton to, AFIS s selected as classfer. Because ths classfer add learnng and decson extracton feature from learned to system. Thanks to proposed method n ths study, to be realzed basc ntellgent recognton systems may be appled at wde areas. In radar pattern recognton studes at the future, systems whch are less affected by nose and envronment may be realzed. REFERECES [1]. Mahafza, B., R., Radar Systems Analyss and Desgn Usng, Chapman& Hall/CRC, Unted States of Amerca, p.p.59, 000. []. J. Ahern, G. Y. Delsle, etc., Radar, Lab- Volt Ltd., vol. 1, p.p. 4-7,Canada, Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

8 104 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton [3]. Turkoglu, I., Hanbay D., Artfcal eural etwork and An Intellgent DTMF Sgnal Code Decoder on Based FFT, 10. Sgnal Processng and Comuncaton Applcatons (SİU00), p.p , Pamukkale/Denzl, Turkey, 00. [9]. Schreer, P., J., Scharf, L., L., Stochastc Tme-Frequency Analyss Usng the Analytc Sgnal: Why the Complementary Dstrbuton Matterts, IEEE Transactons on Sgn Processng, (003), Vol.51, o.1.p.p , 003. [4]. Cho, I., S., Km, H., T., Effcent feature Extracton from tme-frequency analyss of transent response for target dentfcaton, Mcrowave and Optcal Technology Letters Vol. 6, o: 6, p.p , 000. [5]. Madrd J.J. M., Corredera J. R. C., Vela G. M., A neural network approach to Doppler based target classfcaton, Radar 9. Internatonal Conference, pp , Brghton, England, 199. [6]. Swatnck Z., Semklo R., The artfcalntellgence tools utlzaton n radar sgnal processng, 1th Internatonal Conference on Mcrowaves and Radar (MIKO '98), vol. 3, pp , Krakow, Poland, [10]. Stankovc, L., On the Realzaton of the Polynomal Wgner-Vlle Dstrbuton for Multcomponent Sgnals, IEEE Sgnal Processng Letters, Vol. 5, o. 7, [11]. Beastall W. D., Recognton of radar sgnals by neural network, Frst IEE Internatonal Conference on Artfcal eural etworks, (Conf. Publ. o. 313), pp , London, UK, [1]. Applcaton of pattern recognton technques for early warnng radar, asa Techncal Reports, AD-A99735, [13]. Guangy C., Applcatons of wavelet transforms n pattern recognton and de-nosng, Concorda Unversty (Canada), [7]. Jakubak A., Arabas J., Grabczak K., etc., Radar clutter classfcaton usng Kohonen neural network, Radar 97 (Conf. Publ. o. 449), pp , Ednburgh, UK, [14]. Sowelam S.M., Tewfk A.H., Waveform selecton n radar target classfcaton, IEEE Transactons on Informaton Theory, vol. 46, pp , 000. [8]. Turkoglu, I., Arslan, A. I., Erdoğan, An expert system for dagnose of the heart valve dseases, Expert systems wth applcatons, (00), 3(3), 9-36, 00. [15]. Turkoğlu I., Arslan A., Optmzaton of the Performance of eural etwork Based Pattern Recognton Classfers wth Dstrbuted Systems, IEEE Computer Socety, Internatonal Conference on Parallel and Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ

9 The Performance Analyss Of Stft-Anfıs Classfcaton Method On Pulsed Radar Target Categorzaton 105 Dstrbuted Systems (ICPADS 001), pp , Kyong Ju, Korea, Jun. 6-9, 001. [16]. Passno, K., M., and Yurkovch, S., Fuzzy Control, Adsson Wesley Longman, Inc., Calforna, pp.59-67, [17]. Avc, E., Turkoglu, I., Modelng of Tunnel Dode by Adaptve-etwork-Based Fuzzy Inference System, Internatonal Journal of Computatonal Intellgence, ISS , (003), Volume:1, umber:1, p.p [18]. Kosko, B., eural etworks and Fuzzy Systems, A Dynamcal Systems Approach, Englewood Cffs., J: Prentce Hall, [19]. Jang, J. S. R., Sun, C. T., euro-fuzzy Modelng and Control, proceedngs of the IEEE, vol. 83, o. 3, [0]. Jang, J. S. R., AFIS: Adaptve networkbased fuzzy nference systems,ieee, Trans. Syst., Man. and Cybern., vol. 3, pp , Engn AVCI, Ibrahm TURKOGLU, Mustafa POYRAZ