"Event Compression Using Recursive Least Squares Signal Processing" Webster Pope Dove TECHNICAL REPORT 492 July 1980 Massachusetts Institute of Technology Research Laboratory of Electronics Cambridge, Massachusetts 02139 This work was supported in part by the Advanced Research Projects Agency monitored by ONR under Contract N00014-75-C-0951-NR 049-328 and in part by the National Science Foundation under Grants ENG76-24117 and ECS79-15226. tINCLA_,S IC0:.CI'Y Ctt_AC.' FiCATttl I IIC D q OF T nRnf EPORT aaREPORT S s C.- (I?,'.. C ,. l,-nOoiltd DOCU MENTATION PIwOCU.mETATIO 1 TITLE (d P AGE READL rA INSTRUCTIONS BEFORE COMPLETING FORM __ 12. GOVT ACCESSION NO. 3. RECIPIENT'S CATALOG NUMBER 1. REPORT NUMBER 4. PAGE . TYPE OF REPORT & PERIOD COVERED Subtitle) Technical Report EVENT COMPRESSION USING RECURSIVE LEAST SQUARES SIGNAL PROCESSING 7. 6. PERFORMING ORG. REPORT NUMBER Webster P'. Dove N00014-75-C-0951 10. PERFORMING ORGANIZATION NAME AND ADDRESS S. Research Laboratoiy of Electronics Massachusetts Institute of Technology 02139 Cambridge, MA 12. REPqRT DATE July 1980 Advanced Research Projects Agency 1400 Wilson Boulevard Arlington, Virginia 22217 13. NUMBER OF PAGES 151 MONITORING AGENCY NAME & ADDRESSif different from Controllin Ofice) IS. Office of Naval Research SECURITY CLASS. (of thi 15s. DECLASSIFICATION/DOWNGRADING Arlington, Virginia 22217 SCHEDULE DISTRIBuTION STATEMENT (of this Report) Approved for public release; distribution unlimited t7. DISTRIBUTION STATEMENT (of Tho 1. SUPPLEMENTARY NOTES t9. KEY WORDS (Contnu on rersmo bstrct ont-d in Block 20, if id differant rot Rport) if neceswy end dentify by block nmmbe) recursive least squares event compression linear prediction 20. ABSTRACT (Continu- on r".,se aid. if nceearmy and identt b block nu.r) see other side DD ,,R. 1473 report) Unclassified Information Systems Program Code 437 16. PROGRAM ELEMENT. PROJECT. TASK AREA &WORK UNIT NUMBERS NR 049-328 t'. CONTROLLING OFFICE NAME AND ADDRESS 14. CONTRACT OR GRANT NUMBER(s) s. AUTHOR(s) UNCLASSIFIED -----_II I - _ ____ ATlCj _ O or Tw' Tk . 1 0i. ___ ___ -C-·--·------·_C-· --- - 20. Abstract This work presents a technique for time compressing the events in a multiple event signal using a recursive least squares adaptive linear prediction algorithm. Two event compressed signals are extracted from the update equations for the predictor; one based on the prediction error and the other on the changes in the prediction coefficients as the data is processed. Using synthetic data containing three all-pole events, experiments are performed to illustrate the performance of the two signals derived from the prediction algorithm. These experiments examine the effects of initialization, white gaussian noise, interevent interference, filtering and decimation on the compressed events contained in the two signals. ___ UNCLASSIFIED SECURITY CLASSIFICATION OF THIS PAGE(Whof Dote Entered) - 2 EVENT COMPRESSION USING RECURSIVE LEAST SQUARES SIGNAL PROCESSING by Webster Pope Dove Submitted to the Department of Electrical Engineering and Computer Science, on 15 July 1980, in partial fulfillment of the requirements for the Degree of Master of Science in Electrical Engineering. ABSTRACT This work presents a technique for time compressing the events in a multiple event signal using a recursive least squares adaptive linear prediction algorithm. Two event compressed signals are extracted from the update equations for the predictor; one based on the prediction error and the other on the changes in the prediction coefficients as the data is processed. Using synthetic data containing three all-pole events, experiments are performed to illustrate the performance of the two signals derived from the prediction algorithm. These experiments examine the effects of initialization, white gaussian noise, interevent interference, filtering and decimation cn the compressed events contained in the two signals. THESIS SUPERVISOR: TITLE: Professor Alan V. Oppenheim f Electrical Engineering - 3 - I/ To Monica f - 4 - Acknowledgements I give my thanks to all the people in the Digital Signal Processing Group at MIT. Many fruitful discussions with them at odd hours contributed to this work. Of particular help were Dr. Lloyd Griffiths with his insight into least squares algorithms and Steve Lang for infinite patience in listening to my ramblings. Last and most I thank Prof. Alar. Oppenheim for the original problem suggestion, much needed guidance during the investigation and the pressure necessary to produce this document. If I can acquire a tenth of his clarity of thought while I am here, I will consider myself lucky. - 5- CONTENTS Abstract ............ Dedication 2 ........................... 3 .......... Acknowledgements 1. ........................... INTRODUCTION .................... ........... ................................... 1.1 Previous Work ..... 2. 3. LINEAR PREDICTION 6 ,...... 9 ............................. 13 2.1 The Covariance Method .... ................. 2.2 Recursive Least Squares .. 2.3 Discussion ............... 14 17 38 EVENT COMPRESSION WITH RLS 39 3.1 3.2 3.3 3.4 3.5 3.6 4. .................. 4 .................... The Data Model .......... .................. Event Compression Signals ......... Experimental Results .... .................. Linear Filtering ........ ................. Decimation .............. ................. Summary ................. ................. CONCLUSIONS I....... .................................. 40 48 54 112 130 144 146 - 6 1. INTRODUCTION various In which contain detected. signal pulses or Problems analysis. All to improved of Such processing a signal of events in an separations, the and event make an or and SONAR seismic data need for a signal in the This type successive events by locating increasing the events the easier. procedure, which reduces the duration input sequence without is event changing compression deconvolution, their relative algorithm. matched Examples linear filtering, deconvol ution. Let us compression. targets time. may cause events illustrate Consider requirements the each homomorphic predictive in share detectability impulsiveness multiple and arise located RADAR can reduce the overlap between of processing include include detection fields these be shorter events in the processed data. into lead ing pitch must that type signals problems, technique which can compress events occuring processing raw data of events this of speech rangefinding, processing with in application the the RADAR problem or of SONAR. transmission, the source finding To of the reduce pulses are event range peak of power dispersed In addition, the reflection functions of the targets the in returns the to be spread received signal out must targets are to be located accurately. even more. be Therefore compressed if the - Surface disturbance at mechanical 7 - seismograms are the surface of or an vibrator with created the by earth, explosion, generating either and a with recording a the subsequent seismic vibrations with an array of geophones also located at they are layers used the As waves surface. partially reflected to locate the As reflectivities. the source pulse and individual reflection at boundaries into the between earth, differing recordings made at the surface can be ideally the and propagate depths in the the these of boundaries RADAR situation, extension functions of leads that to the and duration of duration the their need by for the event compression of the data. One approach the time these before between to vocal pitch successive estimation pulses. glottal to measure However, pulses are filtered by the vocal tract impulse emanating from the lips, the period compress to the the next cycle. pulses of the It speech since response speech waveform does not offer well defined points at each cycle the is is from which to measure therefore waveform necessary to without changing their positions for this scheme to be effective. Another example is the which motivated this thesis. sonic well figure 1.1, log is to lower problem of sonic The procedure for a tool, shown well logging generating schematically a in down an oil well and then raise it at a fixed rate Well Log Diagram I I i HOLE 'I I ROC K I I I I % I R 61 1 CO 0 SH E AR WAT E R--1-->/ l%1 \ T I I I 5 I I I Figure 1. II__ - 9 - back to the top of the well. During its ascent the ultrasonic transmitter located bottom (roughly once the every foot for waveform at the at in receiver of the tool is pulsed of well depth) and the pressure the top of the tool is recorded for later analysis. In a problem, there take in getting model simplified very are three paths for from the transmitter which has a characteristic velocity. of the the physics ultrasonic of this to energy to tne receiver, each of The slowest path is that of a pressure wave traveling up the mud with which the well is packed (to prevent collapse); the medium velocity path is that of a shear wave coupled to the rock wall, and the fastest path is for the compression wave traveling up the rock wall. The quantities of interest to geologists are the velocities of the salear wave and compression wave, which could be determined by knowing times their of times of waves these arrival. it is To ascertain necessary to the arrival compress the in the source data since the overlap between individual events them is considerable. 1.1 Previous Work Approaching the problem of event compression typically entails modeling the physical fashion and then designing an in an appropriate algorithm which is expected situation solve the problem for data which fits that model. ____II __ to Subsequent - investigation realistic of the 10 - performance environment may then of lead the to algorithm alterations in in a the model, the algorithm or both. For detecting example, events that events they are known separated some (the set In addition, he the (this Unfortunately, in either the is event situations of of source the data's he derives event at a each point compressed are not that set the (in particular statistics a knowledge of beginning of an noise data model by assumptions the of distance and determined these being some is problem combination requires From ratio for data logging) linear the His by some minimum unknown statistics. his analyzes the context of RADAR. exponentials likelihood in are each waveform). noise in [Young,1965] signal). sonic well known, detailed information about the events is not available or the number of available exponentials from which they could be composed of compressing is large making this formulation inappropriate. A seismic common data convolving reflector is made a fixed series. source wavelet the approach source By the assuming wavelet acquiring series and (or problem the data with an was estimate to deconvolve and something [Tribolet,1977] generated by impulsive an accurate it would be possible reflector [Ulrych,1971] by to have close used to seismic of the recover it). Homomorphic - techniques to estimate [Peacock,19693 the - 11 and wavelet used linear prediction to series reflector and estimate the wavelet. Similar work has also been done on the speech pitch estimation problem the of convolution impulse excitation a vocal to recover an of The nature series. impulsive or signal speech response tract and a glottal methods these is is is applied to the data in that a single deconvolution operator order the that by assuming (Markel,19721 nearly impulsive As such these techniques will only be effective sequence. if the events in the data have similar spectra. In this thesis we apply a recursive least square (LS) adaptive linear prediction to algorithm compression, event using synthetic data from a data model based on an abstraction of the sonic well logging problem. independent data have time invariant event spectra, filtering Because the events in this would not compression effective be by (the inverse By using filter for each arrival would have to be different). an algorithm adaptive we perform time linear event varying compression on the input data sequence. To from the implement event compression we derive RLS algorithm, one is the two signals post-adaption prediction error at each point and the other is a measure of the changes in the prediction coefficients as the data two signals are compared experimentally is processed. using synthetic The data - 12 - corrupted with white gaussian noise. of other on include processing prefiltering event the To illustrate the impact compression, data, some decimation of experiments the data and postfiltering the error signal. The next chapter equations and with a issues discussion covariance method is an analytical involved of of in linear prediction. linear linear presentation of the prediction prediction and It starts describes the [Makhoul,1975]. The following section presents a derivation of the Recursive Least Squares algorithm (from the covariance method equations) and the last section of the chapter examines the problem of proper initialization chapter experimental of the comparison mentioned above, and the recursion. important of The the two third event offers compression an signals the thesis concludes with a discussion of results of these exper iments and some suggestions for future work on this problem. - - 2. 1 - LINEAR PREDICTION Linear question: prediction attempts Given a sequence coefficients for formulation answer of data, what predicting sequence with a linear to the value is of the each combination of previous is presented in equation the following best set of p sample of samples? the This (2.1) p sk= where the for {Sk} cn are L CnSk-n n=l is the the dat3 sequence, coefficients of determining squared (2.1) which prediction {gk } the over the predictions and predictor. coefficients error are are the The best chosen is error criterion the total region. Specifically E= (s k -k ) k Q (2.2) k and the coefficients are chosen to minimize the total squared error E. 1. In the general case one could use an arbitrarily chosen set of previous samples, but this discussion assumes that they are the p contiguously previous samples to the one to be predicted. __ - The choice of differentiates between 14 - the the error two most prediction. If Q extends fom - padded zeroes with or region Q is common methods of to + extrapolated what linear (the data sequence is in other some way) the resulting set of equations describe the Autocorrelation Method of prediction. linear involves In this inverting a Toeplitz case solving for the matrix of autocorrelations and is usually done by some variation of Levinson's Inversion for example of (see [Makhoul,1975]). The method ck other linear common method prediction. in As use is described the in covariance the next section, it requires that only the available data sequence be used to generate the predictor. set by requiring Therefore the limits of Q are that all the sk mentioned in eqs. (2.1) and (2.2) lie in that finite set of available signal points. 2.1 The Covariance Method The equations defining the Covariance method are: p Sk= cnsk-n (2.3) n=l ____ - 15 - k e E= (s k=k -k) 2 = minimum (2.4) S Here, p is the number of predictions coefficients and k ke are the endpoints cf the prediction region Q. and By defining the following structures S S kS 1 ks- P S -< (2.5) Agehe:e 1 Is ' -Sk -1 e ' "e S -n r' S s A = (2.6) sk e __I_ I_ __ - - 16 c = (2.7) C the covariance matrix method can be formulated as the following projection problem. S c - s (2.8) Where the desired minimizes the distance vector c which solution is that coefficient vector between Sc and s. = ST unique (SS) - 1 is no unique then some restriction (e.g. The solution is any s. (2.9) If the matrix sTs is invertible then If there which solves the normal equations STS c c = c the --------- the solution is unique. STs (2.10) solution must order of to eq. be placed the (2.9) on c predictor (STS is singular) to make p could - - the be answer lowered - 17 - reducing the number of unknowns). The covariance method generates a single predictor with the minimum possible total squared error over the finite prediction region calculating a successively sequence longer computational covariance Q. of RLS method covariance prediction savings method The compared predictor method regions to for is means for predictors over {Qi}. directly each a Qi' It offers calculating The a the derivation of this algorithm is presented in the next section. 2.2 Recursive Least Squares The RLS algorithm predictors {c[k 0] finds a sequence of for an input signal by minimizing the total up to s k squared error from a fixed starting sample sk S a signal {s k} and optimal a p coefficient predictor . For 0 c, the error sequence produced by the k 0 th predictor is given by ek[k]=sk p I s k -ncn[kl0 n=l and the quantity to be minimized is k Q (2.11) - 18 - k 2 E k0 3 = (2.12) Z1 ek[k0. k=k This matrix RLS for does predictor data point new state calculation each desired an sk calculating be predictor iterative clk 0 -l] could (see eq. calculation together with to calculate the a The successive predictors, calculation only each state advantage over O(p 2 ) inverting a 2.10). uses using the new c[k 0 ] and the method the region, method would need O(p 3 ) previous and this operations Instead, the matrix of new prediction requires Whereas the covariance which by new predictor matrix. method directly on performed is per for covariance that this predictor. operations per predictor for each matrix inversion. The derivation of the iteration can be shown in matrix form as follows. Let the signal be formed into a vector and a matrix as before, Sk 5 -1 S[k 0 k s5 -p ' ] sk O Sk s[k. k o -p (2.13) . - 19 - Then the error vector is ek Ik0 ] S e[ko] = s [k0 ]-S [ko] c[k 0] = ek (2.14) [k0] 0 with Cl [k0] (2. 15) Cp[k] ck cp [ko]J In this notation, the e total error is E [ko] e TT [k 0 and the optimal choice for c[k 0] c[ko] (2. 16) ]e[k0 ] is = (ST[koS [ko] )-ST [koS[k o] (2. 17) - 20 - The derivation of the RLS iteration from eq. performed by substituting available at efficient the manner. k 0 th The k0 + 1 point for k0 to new terms and using evaluate in the (2.17) is information c[k 0 +l] equation in after an the substitution are [ k 0k s[k0+l (2. 18) =__ 0Sk+l and S [k S k 0 I 0] +l] = Sk * (2.19) k-P+l and c[k0+l = c~k ~k-10 +1] ST[k0+ ] By introducing = i T ]S[k0+l] STk0+l]1s[k0++1] (2.20) - 21 - s k0 (2.21) kk-p+l 0 one can substitute and (2.19) to get for terms in eqs. (2.20) using eq. (2.18) -1 +rr ) 0[kS[k0 0 c[k00 +l]=(T (sTk 0 ]s[k0 +rsk +1) k0+10 (2.22) At this point introduce the following matrix identity: For any symmetric invertible matrix V and any vector r with the same size as a column of V, -1 (V 1 + r r T) = V - VrrTv l+r Vr (2.23) Now let V- and L sT[k 0 ]s[k 0 ] (2. 24) c' (2.25a) = c[k0o+l = V[k 0 +1]=(S 0 V' 22 - (2.25b) [k]S([k 0 ]+rrT) 0 (2.25c) = SSk0+. S' 0 By leaving out the indices from eq. (2.22) one can write C' Rearranging = - V terms VrrT ) (STs+rs)' l+r Vr and combining with eq. (2.26) (2.17) leads to an expression which describes how to update the predictor. c' c + Vr s-rTc) (2.27) l+r Vr The other RLS (2.24) and equation is derived by combining eqs. (2.23), (2.25b) and it shows how to update the state matrix V. V' = V - VrrTv l+rTVr. (2.28) - 23 - The full RLS iteration procedes as follows: Starting with V and c from the previous iteration and the vector of previous signal points s ko (2.29) SkO-p+l use the new signal point to calculate the new predictor c' = c + Vr (s'-rTc) (2.30) l+r Vr Finally, calculate the new state matrix T V' = V - VrrT I+rTVr (2.31) - 2.2.1 Initialization of 24 - LS The RLS algorithm offers a means to efficiently extend the error point. C0 of a Thus, given an and an calculate error region initial calculating the to predictor matrix V O, of predictors up to the the end one of the which perform both can since data those calculations. for would two an initial involve a tasks It initializing initial one can recursively of all possible data. be computationally matrix small data interval and then execute the RLS remaining by However, initial covariance predictor normally involves inversion addition, state extents method initial set of prediction coefficients the coefficients region a matrix covariance was the RLS covariance for a iteration for the to fundamentally different find an alternative means algorithm which predictor inversion In large amount of program code, involve desirable costly. calculation. did not involve Examining an a first order case suggests a solution. For a first order (single coefficient) predictor, the main variables have the following form: c[k 0] = C 1 [k0] (2.32a) - s S S[k 0o] 25 - 1 lI S (2. 32b) Sk -1 0 Sk s s[ko] = (2.32c) Sk_ - I /i U and the predictor which solves the covariance method equation (eq. 2.22) is given by k -1 c [ k] 0 ISkSk+ 1 s = k=k -1 kIk+ k -1 (2.33) k=k -1 s Because dimensional, and the (2.31) are the variables equations for C, v the RLS and r are iteration now from one (2.30) - c' = c + vr 26 - {s'-rc} (2.34a) l+r v or c + rs' c t = V (2. 34b) r2 1+ V and 2 2 VI =V- V (2. 35a) 2 l+r v or V' 1 = 1+ V (2. 35b) r 2 where r Sk (2. 36) 0 and k -1 v-1 = S T [ k ] 0 S[k ]0 2 = k=k s -1 (2. 37) Suppose one and (2.35) k 0 +l=ks). at started the first By assigning 27 the point iteration of equations in the error (2.34) region (i.e. initial values v=vinit c=Cinit we the have v = (2.38) k0-1 2~ 2 1 1+ init k=k s -1 Clearly, requirement of for eq. this iteration (2.37), V init=. to Given conform that to the result, the successive values of c are k -1 vinit init + init ctk 0 ] SkSk+ k=k -1 k -1 0 1 + Vinit making method predictors This behavior data comes then is from impulse at n=10. S22 E k=k -1 Therefore, given vinit= =, (2.33) (2.39) the subsequent values of c solve eq. identical (irrespective to of the equivalent the initial shown experimentally exciting a 1-pole in figure linear covariance value of (2.1). system c). The with The system function and time response were an - H(z) = 28 - 1 1 98 z sin] = ul[n-lO0] (.9 8 )n10 in the absence of noise should The covariance method predict this signal exactly when given more than one point of its impulse response in the chosen error As region. can be seen from the figure, RLS also predicts the response for every point after the effective means for is an This method of initialization first. matching the predictors RLS to tnose generated by the covariance method. 2.2.2 Higher Order Initialization When a multicoefficient is predictor used, the mathematical approach used in the previous section to find Vinit and products do (2.34) and Cinit is not not commute (2.35). fruitful and since in cancel the the vector fashion matrix of eqs. One reason for the difficulty of choosing the initial values of c and V in this circumstance is that the covariance specified predictor until the for error which region we is aim is at least not uniquely p points length; and even then only if the data sequence requires at . in . _. - . . . . _ VER 0.312 I . t RLS ERROR /DIv I T 1 X T T ! I t V T 1X t- T7 I i 7 1 T I V ! iI Ln r n i LHU 1.00 VER rIb/Ul /DIV 0.300 - COEFFICIENT 1 It ! I I 1 r I I T v . T I ± · ~ I ~ T · . T · . I · I T ., I I ,--t- I ;I - F 1.00 PTS/OIV Figure 2.1 I 4 I I i I I T1 1 I _. i .J I I i I 11 d~~~~~~~~~~~~~, - 30 least a p pole model to describe it. Consider the covariance matrix definition R ST S (2.40) The iteration for R is R[k+l] = R[k] + r[k] R [ks-1 S rT[k] (2.41) k>k -1 S (2.42) = 0 therefore, k -1 R[k0 ] rr T = k >k Os k=k -1 (2.43) s Now consider the iteration for V along with Vinit=Ia as an approximation approximation for R 1 . If R=V -1 then given equation (2.25b) A[ks] ] = 1 I + r~kg-l] R[k = 1I + r[k -1] and in general rTls-i] r [k -1] (2.44) - 31 - kk0-1 -l R[k)] I - r[k])r + ~a kI (2.45) k=k -1 Therefore setting a=s will make R=R and as in the first order case the RLS predictor will correspond to the covariance method predictor. In the first order case the initial c was irrelevent. Ultimately, the same is well since c true in the multidimensional case as solves R c (2.46) = ST s and once R is invertible, the value for c is fully determined. However, for the first few points of the iteration, R is singular and the trajectory followed by c depends on the data values and the value ofcinit. 2.3 show which was a 4-pole processed initialization impulse by As an example, figures 2.2 and response beg inning a 4-coefficient at sample predictor with 20 the - Vin.t init 32 - = 10100 I Cinit 0 The predictor in figure different predictors used 2.3 is in igure started initial reach non-zero data, but at data the 2.2 is started k=30 (thus values). same their value these As can after 4 trajectories the dependence of trajectory on Cinit, at ks=10; the one differ. examples use seen, both be steps into To illustrate figure 2.4 presents a 4-coefficient predictor started with V = Cinit = 10 1 the I 0 - - k Again the S 33 - = 10 same predictor value is reached after 4 steps into the data, but the trajectory differs from the previous cases. - - -~~~~_ ____ 0.281 /DOlV INPUT DRTR - Ill I fR 5.00.f. PT$/nIIv Of 5.00 PT5S/n V ER 0.315E-02 /DIV OR 5.00 PTS/OIV ER 1.17 /OIV _ 5.00 PTS/OIV VER 1.69 /DOI ER RLS ERROR COEFFICIENT lTl Tll !l I l li lll il i _:i;_:::;:_.: _: HOR al I I I I I 1 1 L5.00 COEFFICIENT 2 PTS/DIV 1.12 /DIV OR 5.00 PTS/DIV ER 0.288 COEFFICIENT 3 /DIV COEFFICIENT I J, 5,.00 4 iiiiiiii II PTS/0IV Figure 2.2 ............~ ~ ~ ~ ~ ~ ~ ~ ~ ~ ~ - I iiii I I -- ---- - L RLS ERROR 0.807E-05 /DIV (ER - -- LR 5.00 PTS/DIV .'R 1.17 /DIV MOR 5.00 PTS/OIV YE£R 1.69 /DIV 5.00 PTStDIV .0 ER 112 q -- - - -- -- - - COEFFICIENT I - - - - -- )- - I COEFFICIENT 2 COEFFICIENT 3 DIV . iTTiLl i. i. . T. 48R 5.00 PTS/DI V YER .A288 /DIV COEFFICIENT 4 9 ill I 10R 5.00 PTS/DIV Figure 2.3 ____ - . 1 . . i I I VER .RA1 /nty U; INPUT DORTA -., iMR ER .00 PTl/jlIv 5.00 PTS/orv 0.807E-05 /OIV HOR 5.00 PTS/OIV ER 1.17 /DIV OR 5.00 PTS/OIV .ER 1.69 /DIV OR 5.00 PTS/OIV ER 1.12 /DIV HOF 5.00 fER 0.288 1 T IT tI TT It 6R RLS ERROR COEFFICIENT I COEFFICIENT 2 COEFFICIENT 3 PTS/DIV /DIV COEFFICIENT 4 .1 0R 5.00 'li6il PTS/DIV Figure 2.4 il 2.2.3 Numerical Considerations The matrix in eq. change eigenvalue to of is R that the succesfully Subsequent which the choose 1/cz about 1010 for iteration would value use so that our initial the value of not cause for a be smallest the in comparison, large Fortunately, by using to of error since for numerical In of any coefficients as the data was processed. Since it the necessary iterations few clearly singular. threshold (2.44) above in require a is section offers the last first The problems. work we discovered is in equations have great potential the associated R ct derived initialization numerical potential these 37 - much it non-zero is undesirable to smaller than 1010. double precision arithmetic we were able values comparisons exceeding of the 10 20 direct for covariance calculation with the RLS algorithm initialized Vi init initialization. method at 1010 I set showed coefficient differences well under one part in 10 data lengths up to 1024. That for accuracy was deemed sufficient for this work, but for a more critical application it would be - wise to 38 - determine specifically how the roundoff error of the computer affects the initialization of the algorithm. 2.3 Discussion This chapter behind the RLS behavior. For has more references particular the basic mathematics algorithm including some simple examples of its information prediction the reader the presented can on the topic of linear is invited to examine rMakhoul,1975] he cites. be found More in information [Eykoff,1974] about and more and RLS recent ladder forms of recursive linear prediction are illustrated (Satorious ,1979]. ______ - in in 3. EVENT COMPRESSION WITH RLS The linear a of described it can how RLS in compressed the and the finally an method and input data used used of to sequence. to compress events. sonic We can which RLS for multiple event signals which introduction. signals the initialized can be of abstraction be from predictors proposing a model a very simple iteration how examines by described and series This chapter We begin chapter previous prediction generate is 39 - be well then situation describe extracted experiments are log two from performed to the event RLS illustrate the performance of these signals. The algorithm concept like RLS behind for using event an adaptive compression is prediction the following. Consider the series of predictors created by the RLS as a single time-varying prediction error processing an energy input predictor event the in the to predict of the event the to This predictor the event. predictor over expanding an containing make errors at minimizes the have error __ will coefficients as Hopefully, after passed, will stop burst the error changing. If the total region. distinct beginning of the event will by the algorithm. change which sequence expects the since predictor algorithm When events, one beginning of each not be predictable be accompanied by a the algorithm adapts the first pulse that will is few points die the away and case, then - 40 - the prediction error and the coefficient changes in a way that could be respond to the events would used to both perform event compression of the input data. 3.1 The Data Model This thesis was motivated by the problem of sonic well logging which was presented in present log Rather data fabricate log The signals introduction. in a real well log are ve y complex do to the geometry of the well. well the and than attempting (a complicated data synthetic what exhibited task which had we to accurately model in itself) the felt was we appearence an important the chose of a to well feature of well logs; namely that the signal contain multiple bursts with spectra. independent this study is each of shown three The in data model figure 3.1. delays D 1 -D for the signals A single impulse is fed to Their 3. discrete-time all-pole systems H 1 -H 3 outputs generating are added to produce the signal pulses which three pulse model used in drive three sk]. the delayed We chose a because the problem of seismic well logging can be considered a three pulse problem. To presented chosen make it later, a from experiments. the possible small data set model Most use either __ to compare f the various representative given above to figures signals was perform the a signal labeled Burstl or some ______1_1_ rKt a) i, I- - combination Burstlc. of its 42 - component pulses Burstla, Burstlb or The parameters used to generate Burstl were: Burstla: D 1 = 50 points H1 I 1-3.73z-1+5.4z-2-3.58z-3+.92 - 4 Burst b: D 2 = 150 points H 2 1-3.89z-1 +5.74z 2 -3.81z- 3 +96z -4 Burst c: G D 3 = 250 points H3 -- 3 - 3 1-1.92z -1+ 98z-2 where peak the gains powers were of 1, chosen .5, through 3.5 show the and time to .2 give the component respectively. response and log arrivals Figures 3.2 magnitude spectra for these signals.1 The complex first poles amplitude. as two bursts indicated have by 2 their superimposed gradual pairs build up of in Whereas the third burst is due to a single pair of 1. The graphs in these figures and in following have been linearly interpolated. most of those - complex poles giving 43 - it a much sharper onset. These choices were made to give the data the appearance of a sonic well log. In particular, the slow growth of the second burst is a common feature of well logs and makes finding compression or by eye a difficult task. it either by event 0.312 l00. 20D/OIV /DIV BURST I R PTS/DIV LOGIl MGN SPECTRUM Figure 3.2 ER 0.221 /DI BURSTIS rev. PTS/DIV -I -- rU,I Figure 3.3 BURST1C /IlV 0.140 -I 100. I Eli 0:: A X PTS/OIV ER 20DB/OIV LOGlO MRGN SPECTRUM 7 I · I .1R l 1 .FS/DIV Figure 3.4 I I 1 I 0.312 )R 100. 20DB/DIV /DIV EURST1 PTS/DiV LOG10 MGN SPECT9UM . IFS/DIV Figure 3.5 - 48 - 3.2 Event Compression Signals There iteration are we which compression. signals three These the prediction error examined are for available from potential use the in RLS event predictor coefficient vector c, the at the new point before the update e b, and the prediction error at the new point after the update ea. In terms of the equations for the update chapter given in the last (eqs. 2.30 and 2.31) these error sequences are defined as eb[kO+l] = s[k 0+l] - rTc (3.1) ea [k o+l] - (3.2) and = s[k 0 +l] rTc Of the two error signals we chose only to work with the post update prediction error ea because it contained compressed events as illustrated in figures 3.6 and 3.7. more The data sequence for these figures was a single pole burst and as expected both error signals have pulses at the the burst (the first point of first point of the burst can not be predicted with RLS since the previous points are all zero). However, though the second point of the burst is predictable from the first, only ea is zero for that point since eb is calculated ER 0. 312 /DIV INPUT DATA i L ~~I T I T T T HOR 1.00 VER 0.312 i A A A IA I A ~P A L LII~~~~~~~ L TTT "I T 11 i i I I i Il 1 [ I I I PTS/DIV /DIV L ~I T L HOR 'R J1 I T 1.00 0.300 1.00 A. I A T I I I A T Il I I . T A I T I. T I 7 T PTS/01V /DIV COEFICIENT I PTS/OIV Figure 3.6 I T . T T i 7I - L IL I I A T 7 i r -7 0.312 _tER E , r I [ER - ! - * - X - 1.00 0.300 1.00 INPUT DATA , I T I I T - _ I ._ I t I I I _ 1_ - I 4 - t I 1 1 ; i i I f T T T _ PTS/DIV 0.312 IL no" .A . _ ikk1 ~ At. , 1.00 HR - /DIV /DIV X , - - - * I - L1,1. . 1 PTS/DIV /DIV COEFFICIENT PTS/DIV Figure 3.7 - -- ~-.- - . - 1- 1 . I- 1 -~ _ - before updating the 51 - This predictor. effect leads to longer error bursts in eb at events in the data and led to our choice to use ea for our work. All error or the prediction error further references to in this thesis are the RLS to ea unless otherwise noted. final One signal, to the applying the signal, different left entire sequence energy compared to the this error sequence error on data of this type. The the RLS whereas to predictor each error point comes in the Because the RLS predictor need only minimize the error to RLS entire a energy the about noted error sequence comes from applying a single covariance method signal. be the covariance method by predictor should when particularly generated from point error than of the (as is the sequence the error predicted point and not over the case with the covariance method), will almost sequence of always the have lower total covariance method for the same predictor order. As a matter of observation, we have found that this reduction in energy takes the form of shorter error bursts at the events in the data; though as yet we have not proven that this must be the case. In addition to the prediction error signal, we examined the changes in the predictor coefficients as a means of generating an event compressed signal. This idea stemmed from observing how rapidly the prediction coefficients settled - 52 - after a new event occurred. Figure 3.8 demonstrates the behavior a 12 coefficient RLS predictor on the signal Burstl defined in the last section to the (some of the coefficients are omitted due lack of space). single The non-zero first event error point predictor. This behavior the event first causes very because of bursts). 250). is little the But Both and all-pole. error the error no longer coefficients led us second coefficient between activity all-pole. rapid since The or and a expected similarity note the is (burst) at point 50 causes a the The the at the event and event 150) second (point short in that settling of (presumably first third the as (point change fact in signal ettle the rapid change the coefficients time compared to the burst, despite is of the time signal of the to formulate a signal based on them which could be used for event compression. A coefficient filtering each ci [k] produce the vector Ilctk]-e[k]|t. change signal was generated by low pass with a single [tk] low pass measuring the Small random variations value are reflected in c[k] filter to distance about a fixed in the coefficient change signal as noise, with each sample having a height equal from the average predictor e[k] c[k]. and pole However, when ck] to the radial distance to the instantaneous predictor jumps to a new value due to an event ER 0.312 BURST /DIV n ["\ A A A AAA A r\, V 'wvl vivm?'V Cluv HOR VER PTS/DIV 50.0 0.355E-01 OR 50.0 PTS/DIV ER 1.23 /DIV H§R 50.0 PTS/DI¥ VER 1.88 /DIV NOR 50.3 PTS/DIV IER 2.40 /DIV I RLS ERROR DI COEFFICIENT 1 COEFFICIENT 2 COEFFICIENT 3 I ,} I (12 COEF'S) - ' I NOR 50.0 PTS/DOI COEFFICIENT 50.0 PTS/DIV Figure 3.8 I v I i i / 54 - - in the data, the coefficient exponentially decaying the distance between change signal will contain an pulse whose initial height is equal to the old and new values of c[k] arnd whose decay time is set by the low pass filters used to generate C. In effect filtered the to motion create of the the predictor coefficient is being change drawback of this scheme is that slow changes will be reduced in amplitude due to high-passed signal. One in the predictor the high-passed nature of the coefficient change signal. The RLS error and the coefficient change signal were use to perform the event compression in all remaining Figure 3.9 shows how they behave on this case the 50% decay time of the 'he change (where p is choice for signal was the order this multiple peaks 4 of points. the parameter. in the the signal to Much shorter at to empir i ally predictor) change signal burstl. filter; used We be decay figures. generate found an p/3 effective times each event In led and to longer times reduced the resolvability of closely spaced events. 3.3 EXPERIMENTAL RESULTS The error and containing last section coefficient all-pole rarely available and presented change events. it is In the behavior signals practice on of the noiseless noiseless RLS data data is quite possible that preprocessing C9f 'EM - -- BURST I /DIV .312 I . /\fVn\fnfIf\ AI - I L-_ MOM VER 1l j J " TVV - L ,I I- J "-j v f\ I t J 'm l Vv % r- 1 VJ PTS/DIV 50.0 ~~~~~~~- 1 ·- - - ------- - 50.0 PTS/OIV 5.' 6 /0 1i RLS CEFFICIENT ChRNGE _ I I I L -- nDM bU.U i PTS/DIV Figure 3.9 ! i I - 56 - (e.g. filtering) data if they could have added zeroes to the events in the were not experiments are intended what to expect from already present. The following to give the reader some insight as to these event compression signals when the input data is not ideal. 3.3.1 Additive Noise The example of event location given the in last section used noiseless data. Figure 3.10 shows what happens to that example sequence a=.001 when white Burstl giving the features appear change signal (the in this the is added deviation standard first where gaussian noise burst of this a S/N of 60db). figure: the algorithm pulse is to the input is noise Two important in the started coefficient (point 0), and the substantial difference between the coefficient response the second the starting coefficient burst (point transient response to 150) (in the and the the first (point 50). coefficients) first event are and the implied to Both large by the structure of the RLS update equations. Recall that the algorithm is trying to adjust c to get the least square error E in the equation E = IlelI 2 = Ils-Sell 2 (3.3) BURST1 S1GMR=.001 /DIV 0. .312 fER I A A (N, Abu I F HOR VFR I (i2 COEF'S) 1' \I1 PTS/DIV 50.0 0.354E-01 7 I AI i /IV ALS ERROR :-· _ R ___ A __ __ . _ 50.0 PTS/DIY 2.16 /OIV 50.0 . _ __ ___ _ Lao _ __ _ __ _ _TI _ _ ____ __J RLS COEFFICIENT CHANGE PTS/DVY Figure 3.10 A _ _ ____ __M__ __ I_ _____ · - 58 - to the matrix S and new points to Each new point adds a ro the vectors s and e giving e e' S - S = - the new c' predictor the separating i c' . (3.4) rT e'St and --- of components the 2 E'=I el minimizes new E' error 2+e'22 By into the contribution from previous points Ep and the contribution from the current represent point the Ec and in change by the introducing predictor the vector coefficients d to (i.e. d=c'-c) one has the following relations: Ep = Hell 2 + IISd11 2 (3.5) = E + d TRd E The term = Ils'-rTc', d Rd (3.6) 2 represents the cost" error of changing the predictor in terms of poorer prediction of previous points. At each new point benefits of the algorithm trades improved prediction off of that cost with the the current point - (reduction in the of cost eigenvalues Ec) R the are RLS starting error) transient in the low small. the This is the be very (and consequently the cause change coefficient (e.g. will predictor the values of Ec and be will is predictor small), responsive to the data the permit. Therefore, if that a change might chang ing of 59 - of signal both the and the sensitivity of the coefficients to the first event. To has on the illustrate event (3.1lla through offset to the impact the compressed 3.13c) right was 300 that a signals prepared points. the series using Each additive noise at a different level (i.e. and within each group the RLS different points transient starting the group of figures Burstl of a=.001 , algorithm was 9 signal three has .01 and .1) started at three ( point 0, point 200 and point 300). The starting position of the iteration has a small but noticeable effect on the compressed events. For the RLS error, the sooner the arrival occurs after the starting point of the iteration, opposite is the smaller true for the the event will be. Exactly coefficient change signal. the Equation (3.5) indicates that the longer the interval between the start of the iteration and a given arrival the more linear equations the predictor has to fit and the less itself too Since the data values generating these the new points. additional equations are it can afford to adjust noise, the equations are - 60 - independe;:t and predictor c (That would not be true or each one puts came from a noiseless all added points predictor change Fortunately, change desensitize and, this of is constraint arrival). predictor consequently, effect a starting transient leading more gradual and iteration itself the In effect, does not is not does not the to prediction less error. appear to Therefore, the the character of the compressed events. the on if the data were all zero pole the exact starting point of the as more crucial obscure as the long first arrival. In certain situations existence of the transient may be a problem due to the lack of at the start of initializing the some signals. RLS iteration In to starting transient would have to be found. starting eventless" data that case, reduce this or some means eliminate of the STRT YFR I7 :-* 0.312 T BURST 1 SIGMA=.001 /DIv AAtliAn4\ Ahf>1/ I I V VVV Ji 40R !00. PTS/DIV 0.300E-01 /DIV 100. VER 7 L 0.500 100. RLS ERROR 1 12 COEF'S PTS/DIV /[ )IV __ HOR \V _ ___ RLS COEFFICIENT CHANGE ___ ___Y_ u __ _ _ PTS/DIV Figure 3. la _Y _ I K · STRBT 0.312 -- BURST1 SIGMA. /OIV - - s T 200 - e~~~~~~~~~~~~~~~~~~~ 001 ____I_ · t _ ! _ · ·_ 1 1 1 1 ~ _ . _ ._~ f . iR , i . 11 · VVVV V AvV OR -ryER - 100. 0.300E-01 PTS/DIV /DIV ALS ERROR ( 12 COEF'S - · ;OR 100. .ER 0.500 _ ___ __ I 1_______Y___I .- - - _I· __ -- n- 'v PTS/DIV /DIv RLS COEFFICIENT CHANGE H I 100. II__ ----- PTS/D I Figure 3.1lb N________ -- - Ai I START 0.312 BURSTI /DIV T 300 SIGMR=.D01 AIf \ I 100. - VER - PTS/D1V _ 100. ER RLS ERROP 0.300E-01 /IV 0. SOP 100. L 12 COEF'S ) · ___ A ___ .-. -W- PTS/DIV /DIV RLS CEFFICIENT CHANCE PTS/DIV Figure 3.11c -~------ ~ ~ _~ A i----C----r-l-* A.- .ft STRRT I YER BURST 1 SICR=.01 /DIV 0.308 T 0 Ah~AhAA~AR . - --- -- HOR I ,- RLS ERROR /DIV -iL VEP ------ 2FJ9,~^;'-i "Af11 YT, AI --I*v -___ 100. HOR -11 11 PTS/DiV 0.300E-01 hvi^ I IT VI --- - 100. VEFR t / 0.500 rt-o-NT7U-"Vf 5f' ( 12 COEF'S S .lll, iA b .. P -e tA r ------T. '-16 0 --I .ll-1'N A N P TS/ODIV RLS COEFFICIENT CHRNGE / 'DIV 5 I J _ _ _ _ 1_ I - _- ---------------- · ---- _ m _ -l b OR 100. Pr/DoI Figure 3.12a _ --- ----------- I-------------. STARRT ER 0.308 T 200 BURSTi S!GHR=.01 /DIV i - HOR YEA IoO. . . I 'A U I - PTS/OIV 0.300E-01 /DIV RLS FRRR 12 COFF 'S -I i oR 100. YER 0.500 PTS/DI V /ODlY LS CEFFICIENT CHRNGE I L nOR 100. PTS/DIV Figure 3.12b __ ____ . START 0.308 T 300 BURSTI SIGMR=.01 /DIV p5l Ij 100. YER PTS/DIV 0.300E-01 /DIV RLS ERROR 12 COEF'S I - - i Al. A ha .~U1 A, Ijr-rVI14111TW lOR 100. 0.500 _ _ _ PTS/DIV /0V _ RLS COEFFICIENT CHANGE _ Figure 3.12c .--. IkJki~I 1 i f ip-To tr, STRRT PT 0 0. 382 ER I,,il ,.,,, ~hU.l t F .,,~Lb ~ ,,, bllrhtA I"r 'tlStf T-. OR 100. n __ -,l ~,.l~i, A. ~ fIA~ i*V4. __ _~_ __ _ 100. R 0.500 100. __ _ _I " T _~~_~__ __ _ I _· I 1 I A' t hKA V- y iV V Vf JA.,A hAi ~ yIv N"7T __ _ _ __ PT$/DIV ---- -- -- SIGAR. I BURSTI /DIV PTS/DIV RLS COEFFIC!ENT CHANGE /DIV PTS/DIV Figure 3.13a - - _ _ _ _ l _ 1 I, 1% _ · ¾/V STRRT oTR20G BUfS I SIGR= l /01V fEs 0.382 100o RLS ERGBR 0.100 2 COEF'S /DIv p /OYv 100o _ 100- . . CRNGE RLS COEFFIC ENI p?5/O1V Figure _.--------------__----_----------- 3 .13b STRRT , . T 300 BURST1 SIGMR-. d PTS/01V 100. - . . r- Ic PTS/DIV 100. r \ rEFFFICIENT CHANGE b 100. PTS/D!V Figure 3.13c _ ------- L--- - 70 - To following eliminate event 500 points the starting other to than the location the starting examples, left of the transients do not slightly changing transient the all the predictor was data consequently visible appear. the from and This sizes of has the started no effect events in the compressed signals. The change noise signal itself has at a level of -60db (figs. 3.12) the size of the is severely event little effect on the reduced is visible and (figs. third event at -20db (figs. (note the change in 3.11), coefficient but at -40db in the change signal 3.13) only the scale factor over first those three examples). As was noted previously, increasing the noise level in the region before an event reduces the sensitivity of the coefficients to that event leading to a smaller coefficient change. The RLS error degrades in a different fashion; the most prominent feature being the apparently magnified noise in the RLS error signal. If one views the problem from a filtering standpoint the equation s 1 5icI ___ (3.7) =e _____ - 71 - must be solved to lell 2 minimize or, equivalently, make e orthogonal to the rows of sis In the frequency flattening (3.8) domain the spectrum. signal Burstl this corresponds to whitening or Examination of the spectrum of the (in section 3.1) indicates that the flattening will consist largely of raising the high frequency portion of the spectrum. This high boost leads to the noise in the RLS error signal. The other apparent effect on the RLS error signal of increasing the noise is the increased events. size of the compressed Here again, because increasing the noise reduces the sensitivity of the predictor (thereby lessening the extent to which it adapts to new points); the error that the predictor makes at each new point is larger. Unfortunately, in practice, this magnification of the events in the RLS error fails to keep pace with the noise as the noise is increased. Therefore, this effect does not appear to be a useful means for enhancing the events in the RLS error. The preceding figures in this section __II______IIIIL___1111_______- were cr ated - using a pseudo random noise 72 - generator which produced exactly the same noise pattern on every graph. Since there may be some question about substantial compressed whether impact signals, on the the the last different positions. noise appearance show a fixed noise pattern with three exact three the These of pattern the figures Burstl events (figs. signal has in a the 3.14a-c) shifted to figures illustrate that the precise noise pattern has little impact on the characteristics of the events in the location signals. t c< - -- ' 0.312 /DIV BURSTI SIGMA".01 ADJ IN AA~ 0f I0' nAli 50s0 ER 0.200 _ 50.0 .y ROf 50.0 PTS/DIV /ODV RLS COEFFICIENT CHANGE -Sv P PTS/OIV Figure 3.14a .. 7L1 /DIV . 309 VER ........ m - - - - iS BURST SIGHMA-.01 \ VFR -- - - · ----- ld IY Lr ·----------AIA 1 - 50.0 .ER t V "-Q PTS/DJV OOnnF-n1 ----- ·-- ---- 1 I V V 50.0 A' l - 0.200 -I/TV l ALS ERROR · l ' -·· --- w s - vvy- r 'V A - ·- - - % T .,A . ---- · Y -7r rz n A&jAA --· Y----·-·w -) I:r A J\ u Y w --V kA -· L'Ysii-tVP I·ir-·r··- -VV1I . A -··1··I-1 fall\iA, , -I···r-- Yr~-11-n Vy" PTS/DIV RLS COEFFICIENT CHANGE /DIV I - , ~-- 50.0 - -u ---- - I-:--vs- r PTS/DIV Figure 3.14b ---- · /OIV AR 0.318 BURSTI SIGMA.01 l 50.0 PTS/DIV /DIV 0.500E-01 -I -A I - - -X.,I A . - M- RLS ERROR · .I T- P. I I - 50.0 0.200 v 50.0 I- - f _ , , - \AI 4Ax AA.di. A 'I1_ r_Y-'AfA .~·- ~I, '_~ _____ _~ _P- ·· --·I~1 1 1_·I--J - I- · · - -- -- -N, %4p - I ¥ PTS/DIV /DIv RLS COEFFICIENT CHRNGE 1 . R - n ! 1 PTS/DIV Figure 3.14c I ,· i - - NW if -· iNrA ,Y trV - Effects of Predictor Length 3.3.1.1 Figures 3.15a predictors levels. through series of first, figures: increasing predictor length the second, for the predictor to settle; time lengthens the the noise various at signal Burstl two characteristics of event location with There are increasing the eight and twelve pole show four, 3.17c on acting RLS visible in this noise 76 - the (up to a point) decreases predictor settling time. The proportion to the (see eq. 3.5) scales in covariance matrix R signal noise level. Thus the energy cost of error modifying the predictor at a new point dTRd increases with the noise level. independent adapts less to and of the level. noise the events in consequently the coefficient change level from better prediction of the event (Ec) at the current point is in error energy On the other hand the reduction Therefore, the predictor the data compressed events have longer signals as the in noise the duration RLS as increases error the and noise increases. The reason that the predictor settles faster when 8 or 12 coefficients are used rather than 4 is probably because the events occurs, little in the data contain a 4 pole model chdnge is four poles. inadequate. Once the second event Note that there is very in the predictor settling time between the 8 and - 77 - 12 coefficient predictors when the noise level is the same. 1 1. In all the experiments which were run, increasing the predictor length beyond 12 coefficients did not improve the location events. However, these signals contained at most three events and the events themselves contained only four poles (at most) each. Situations involving larger numbers of events or very complex events may benefit from larger predictor lengths. YER 0.312 BURSTI SIGHR=.0I /DIV A _ Af\P n ~~~ ~~~ ~~~~ _ A _ _ I / A 1 .1 :--I-I·-- __ - : -___ xjI *i 'II V (- PTS/ODI 50.0 RLS ERROR (4 COEF'S) 0.362E-01 /DIV YER _- · 11-1--: * _s_ X VvX V'VU OR f\ A j I A - _[OFl -"' 50.0 0.453 50.0 v- - I- -T" . IL 7 I _ v <6w>_/ -- _·L_- !- _ .-. IdY--·FCTI----·-T)h PTS/DIV /DIY RLS COEFFICIENT CHANGE PTS/DIV Figure 3.15a _ _ ./ PI. ,I i·l-dYill . n.qq9E-oi fFH . .- .. CRAS tlr51Yl. RLS ERRR /!Wv 1 COeF '$ L a.~R S;. 0 f F 0.1S9 I4 PTS/OI' tOEFF!CIENT ChANE /!iv-!l$ it i I T -I · c. PTe/D V Figure 3.15b _ ~ ,~~~~~~~~~~A __.~__. _~__. ... ~\ ( ! C. - VFR .·· n.3III Y /nlY I .iAt .al · · I_ · L-I11Y- r , vI .I BURST1 SIGMR=.1 IAl\ n I · · · I 1 I · 1 ·r . A ·1 I· v I HOR L_ lER 0.203 50. Ir 1 1· · 50. 1- I /DIV RLS ERRCR i4 COEF'SI j~iA PTS/DIV _- -_ - - -- PTS/DIV Figure 3.15c I · · ·- ,wn PTS/DIV _ 50.0 · .· ~I I1 L · t\ i u C~~~~~~~~~~~'~,J. HOR (i 1 · d -- I 'ER BURSTI SIGMf=.00 /DIV 0.312 , A I AAlI _ __ _ __ 1 _· _ 1J 1 _· ·· _ __ T _· _· § X I A * V\ JAJV0iW'\JvA I / iOR _ER E RLS ERROR 0.355E-01 /DIV A _ Iv 40R 50.0 .I /, \V ' 7w x R X f- r w 1 __ ~ ~ ~~~ I-A-A 18 COEF'S) ~ , ~ ~ ~IL ~ _ _ _ I_ _1_~_- _ .- - v_1_1___ ., __ - - -- . RLS COEFFICIENT CHRNGE /OIV K_ -- - ij ;I I 50.0 X PTS/DIV 0.514 ioR l PTS/0 V 50.0 L X PTS/OIV Figure 3.16a ^ - - i I I f I *I_'. [ER - 0.312 S /DIV BURSTI SIGMRA.01 nn~~~~~~~~~~~L~~~~~~~~ f; --- fnf \ A\ i~~~~~~~~~~~~~~~~~~~~ I E 50.0 tER . * *JS 0.211 A / /- V -- Viv /0IV RLS COEFFICIENT CHARNGE i I so0.0 ) --- RLS ERROR 18 COEF'S) h OR x PTS/DIV O.446E-01 /DIV _ | g \,J' vPTSvII lCR ,, I - - -- PTS/DIV figure 3.16b - -----I------ I- I 0.361 f/I, 50. 0 ER BURSTI SIGMR=. 1 ?TS/DIV /D v 0. 152 RLS ERROR (8 COEF'S) - I idaN b)@X.1,1i lS .nj~ T lst~~~~~~~~~IXAl ni , I Uwl~~~~~~~~lF TT Y____V·II_ Ro _ 5s0.0 O.S2E-0 IR 1OR 50. so0 · ___ __· ____I____·____·__ v1 II_ __· 1 .. _Y I f 11 fAL i.M4b, .1 iA 1I1· 1 Y·IIY r I ! PTS/DIV RLS COEFFIC!ENT CHANGE /DIV PIS/OIV figure 3.16c - - ----- Y_ Y·_·i·__ F 1 Lh _________II__ I, V ·· ·_· f /DIy C. 312 BURSTI 5IGMR=.0Cl PTS/DIv 50.0 I ER ;\A1 I n 0.356E-01 /Div RLS ERRCR i12 COEF'S) i , m .DR ER 50. PTS/DIV . 528 /Oiv -- 1-11v RLS COEFF!ICENT CNE 4 _ . _ _ . Figure 3.17a __ __ I I-v-Ia 0. 12 50.0 YER 0.432E-01 BURST! /DI¥ PTS/OIV RLS ERROC 12 COEF'S) /DIv ,,f4 A-V 'v rIv, y T" ________ I V. VI V vl iW'_~____ -.. AL AN Yv MORl SO. "y( L -vYV kA _T,'- _ L__jII w *v_4i 1h -l4 -%I- "Y~''p;Fpaq`? PTS/DIV RLS COEFF'CIENT CPNGE 0.231 50.0 S!GMn=.O: PTS/DIV Figure 3.17b I I. vvw.i~ika.I Iw F .L IT q Ark ', /DIV 0.361 BURS SI GMR= .1 IX\ 2llKI 1 PTS/DIV 50.0 ER RLS ERROR /Div 0.132 -7~ ,',0, -~~~~~~~~~~~~~~~~iur LER 0.591E-01 /DIV *pca_!-r _ _ ISushi 1) ~ ,I~I RLS COEFFICIENT CHARNCE \_ . _, I!,p !12 COEF'S) _ - _ Eigure 3.17c t N - 3.3.2 Interevent Interference As discussed an event it. 87 - in Given part a two in the determines event last section, the data preceding how the will situation, the more similar event is to the second, the less at the second event, so the adapt to it. data of to the first the prediction error will be less the predictor will change to Burstla Burstla or Burstlb as events) and a longer react Figures 3.18 and 3.19 demonstrate this fact with consisting differing predictor as the the first.l clearly shows second Figure a larger event 3.19 and either (the one coefficient with change prediction error disturbance at the second event reflecting the larger difference between the two arrivals. 1. To make the details of the second event these graphs have been magnified causing compressed events to go off scale. more some apparent of the BLRT1R YFR i-·· .448 - /(Iy SIGMF-.001 AiA WIh A 1Al 1 =- I I I . I v 50.0 VFR L.. --- I-I HOR JER /ODI A-. 1-·1· _ I O Ilv 'V¥- -I-" -- ==- l I ,A % f ; i .11 'h I1 ) r, ,., ,' " PTS/DI V 0. lOOE-01 - 15' 5 ----- RLS ERROR I-viv--- .. It IV Vlif 1:. o V I . u- . - Iv. (12 COEF$') PO1- P , W vT AI-I l"b 1 6 c- v' - 1 -- v ! * AAi. IvI 'r'V A.A -I. - . y ~ I --- TIP - V ', I. I .bIII LY~ - v PTS/CIV SO.O0 0. 100 /OIV RLS COEFFICIENT ChANGE l --g ,,, OR - -:- __j _- 3 S-- 50.0 ^- - I I /_ _ - i PTS/DIV Figure 3.18 _ _I_---------· 1- - __ L _ BURST 18 c 5S BURST' 0.514 /OiV i V *ou I VER S1GRA. 001 Y 0. IOE-O1 /Iv RLS ERROP 12 CEF'5; r14 !Ai 0 HOR 50.0 0.100 .... a I150 -A.. A, PTS/DIV /iIV I II RLS COEFFICIENT CHPNGE v, Figuire 3.19 - - - - 90 The next twelve figures (nos. 3.20a demonstrate the effect that spacing has interference. or 3.22) was used of the a Within fixed to predictor compress Burstla figures each group of events pulse illustrate length at how inter event four figures (3.20, 3.21 (4, 8 or different event 3.22d) on 12 coefficients) in data containing four the through two instances spacings. compressed These signals behave versus event spacing and predictor length. The predictor These appear error figure sequence points appears events different at differences Examining 150 compressed include 3.20d after beyond for all a first it. spacings changes reveals the the The in four coefficient up to 150 points. size small to determine the zone over shape. disturbance event which duration and extends of this which the in at least disturbance position of the second compressed event will influence its shape or size. 8 12 and coefficient disturbance, similar This led model in and size predictors they and generate shape us to believe that a given event, as error disturbance for have a much compressed all indicated shorter events spacings above the longer by the The error which 25 are points. the predictor takes to the time taken for the to die away, the further the next event must be to remain unaffected. This postulation was tested by the following - experiment: point 50, positions signals Burstlb between were at 100 91 generated point and - 350 200, with Burstlc and Burstla points; then The results are presented was performed. in starting at at various event compression figures 3.23a-h. As can be seen from the figures, the compressed events for the last two bursts only interfere time to this The question of what determines settle between them. time settling is if the predictor does not have a possible for topic future investigation. 1. Burstlc was used to desensiti ze the predictor and it To make produces a large event simply because it is first. and error the RLS clearer, the details of the remaining events first the causing magnified were graphs coefficient change compressed event to go off scale. A__ L- R O.35 /OIV 50.0 PTS/DIV 0_979E-02 /DIV VFR . -- r -. 7T ·- fOR 1V- 50.0 0.453 50.0 50 & 75 SIGMA=.00 BURSTIA IT. iT 4 COEF'SI RLS ERROR ~.1 J Y 1 ~ ~1 ,A yl 1-O. o 'ln' (i _..tE act r1r gd V-V-l -Tr PTS/DIV /DIV RLS COEFFICIENT CNGE PTS/DIV Figure 3.20a __ ,,.artA4.t. -Y- In-, ,;liAli..'..I fyVVI,VI I:uyy'~l- It. ... it ERl 0. 312 /O Iv U.RST!R 5So InGo SiG. I4 i HOP IEF t S _ . 1: v PT/i¥V 0.979E-02 /Oiv RLS ERO,; i4 COEF'S5 + -BO 4EAR 50.0 0.453 PTS/DI v /Civ RLS COEFFIC:ENT C-AoGCE t ii~,-. I T HOR 50.0 PTS' DI v Figure 3 2b or fER '. 4iq& /DIV BURSTIA .~~A,: \ -4 · - · -0 - L 50 & 50 Si GM =.CO n · ' '- -# r; A ' II,' ( J_ t II , A ,Ai II ': : L '\ / \. I\I 1 iI'VI/\ \1 \ -. I " '- ii \j V PTS/DIV 50.0 VER 0.979E-02 /OIv RLS ERROR Ir _- _-nt kV ~~ A I' 1CR rl H 50.0 JER 0.453 . ,I V li~~~~ i F. -1 .,i -~~~VR 7lr V-Ti~~~~ ii r7 N4 COEF'S) AtA,A Th. hA "I, · . - I . lw..?',v I I'- h . A W I , - -- AA- - Vy" 7" a, - I PTS/DI V /DIV 9LS CO IE Ch . --- Figure 3.20c d.. !,i,, . ".I I-101 ; i.- i. I IER 0.312 / IY BURSTIP I < VER . V ^ c 5 200 Sl=- <o - 0.979E-02 /O!v RLS ERRCR CF 'S,c I T I' It U -IiI ..i*% rI v 'ri - . -k " Alil !V~A 1 k i IA dl .,43 - YER me - . //n^' MI RLS CEFFCIENT -HrNC5 , i --- 4i HOR 50. P' /DiV Figure 3.20d ; '-, t 0.435 BURSTIR 0 50 & 75 SIGMR=.OG1 /OIV .n~~nnnn,,Tv I VE:R 0.979E-02 /OIV .- -- RLS ERROR AA , . LP--- 7 -rrr-R-r ~ _- * - * · .^.n- - IVf , HOB s0.0 LER 0. 493 I · · w v I · 'l eo I i· W' ^ I (P COEF'S) kt4hAMAA ILA -..------.-.-. r, b NA..& A-1ASiA &xA. ...,7bAA--.,Y..,,-..--.;..-I-I-. s vYY' V -V VV y- vt V "y- . -v ''v vrv,-1 n NA PIA A -4___ . _. 4.._.. y vv' -Vy hA i v; I PTS/DIV RL5 COEFFICIENT CHRNGE /DIV h -1 50.0 -1 I . I ;I-- i PTS/DIV Figure 3.21a I AJJh . a. __ Ii .. 7,i Wvs'qy -r 1 0- l 0.312 -- HOR -- /DIV BURSTIR A I I- - \ I A I 50 I 100 SIGMR=.001 k , PTS/DIV 50.0 VFR 0.979E-02 /DIV RLS ERROR 1 8 COEF'5S m I§R 50.0 PTS rER 0.493 /DIV h __ HOR IV RLS COEFFICIENT CHANGE f4_ I X i _ 50.0 PTS/DIV Figure 3.21b i ,I I - __ VER _- O. 4=8 /ODIV BURSTIl I - - 10R tER - 50.0 -I IER 0.493 _~ [~n , " ' z~s rN PTS/OIY RLS ERROR .. LAMAAPA .flt. !t * -. v I - -'U ' ' - ,-A l I1h (8 COEF'SI L1. IK * I _.1. .. , .l.,.. u , · I,- AlII Y.Y~r I v 'I - v" - r v.- I- y vv-v ..I A. . .A A 1lil. I . . ....-.. IA .Ld~ll. I- Al~;r~-AI KAn 1 A"A.. I I. -V PTS/OIV /DIY ~~~~ 50.0 .i \ V'\JVV J\ J '\ VJ\ -V , f I I '- . 97-1 t I Ij "'..V S0.0 150 SIGMp=.001 AI f AN 1A A 0.979E-02 /DIV .- I o 50 RLS COEFFICIENT CHRNGE ,__iti PTS/DI V Figure 3.21c iI I -V.-, - ~ " -.!V"Jr r -,V-f fER 1 I BURSTIR /DlV 0.312 JAF : 50 , I 200 SIGMR=.001 A f\.\-, I I A VVV v i V vv M_ PTS/DIV 50.0 HOR 4ER 0.979E-02 /DIV H RLS ERROR 18 COEF'S, I - 11 _^_I m___·l_·_ I F __ -OR -- ._ 1_1_Y.1 __-." .A-A.1 ·__·__- ___·I· -' _ r.vyv1 -A . .lkAI ^·· A ··_· _Y· r _ II_ _ _·.I ·_Y1_ ·_ __ .V-.'iIvVy1V I A . . j _ IILI:IIY^P Wv rV¶y8 -ff- _ 50.0 0.493 50.0 PTS/DIV /DIV RLS COEFFICIENT CHANGE PTS/DIV Figure 3.21d A1 ·L AII IAIL·ln I 1 'i, -WVVy-,y _iI-·- ly. n·i _1 kA···J -1-~- YI U IAAl ·III.I·YYII I ·iA · I --e) , T .v"r I0c 0.435 /DlV BURSTIA 0 50 & 75 SIGMR=.oo001 .vr AR 0.968E-02 /DIV ,r 0.508 RLS ERROR = /DIY · 12 COEF'S) RLS COEFFICIENT CHANGE .· . Figure 3.22a tot 0.312 /DIV JRSTIA o 50 & 100 SIGMR=.001 'VaVv v ,v v %. - -'- A . LER 0.968E-02 /OIV RLS ERROR I __h~~A litR so.o tER 0.508 A. N41kAA ,v v-V? 50.0 Ad -- r r I AVY.rrt WI , I - . PTS/DIl /OIv RLS COEFFICIENT CHANGE I ItF (12 COEF'S) i. Ii PTS/DIV Figure 3.22b TY _I . - A .k1It ,v. .ER 0.448 150 SIGMA=.OC1 50 BURSTIA /OIV l A - iOR I A A AIt ,T----__ _ I VER _I h. I ~-~I I Li 1 iu urrr --.--- 'V 50.0 VER - IJi. -~r ·- ruilI .,AmAFt A A -. -·-·· mu-u· ··· , vvV-l T f Pki-14,1 -kA .s v .. .. I-- "1 1 I a iA . .Il,. .k ~u~& I .~L'l J '. -`~-" I Alm-L, ', ,' __U -I I'I PTS/DIV 0.508 /DIV RLS COEFFICIENT HRNGE i h" 1 HOR 50.0 A t ,fN N lt \ ,, "=' RLS ERROR (12 COEF'S) 0.968E-02 /DIV .*' At f 1k i 1 PTS/D1V 50.0 .s j :0l I - F -_i. PTS/DIV Figure 3.22c i iA . A.i. ! A',z-... s ,,,I, Wl ' MY1 ",-_ Y-· k 1A h . I WIYL.. ,,| , I..., l,I . 1 0.312 BURSTIR /OIV 50 4 200 SICMR=.001 f\.\-, n.n 50.0 1 ER _ -Y · r i HUi PTS/DIV 0.968E-02 /DIV _ _ w -o- __ m 50.0 0.508 50.0 I I A _ I A^-__ls___ RLS ERROR d 'IfT-T- _1 Ji A - __m__R___7 '- 'lI'' (12 COEF'S) A AA ilMII________ . kA Ak& __lr__ -''V, YV,rv _ ·____ _____ _ II i- Ny , v -,- PTS/OIV /DIV RLS CEFFICIENT CHRNGE PTS/DIV Figure 3.22d __ - - ----------- r -i AoA __1_ ____________V___1 -A I __ _ ____ r v V I , v, ll· _1___ An _1_1___ V-w,,v,- - W, BURSTIA 0. 338 /D I V . O. OOE-01 SIGHR=. 001 ,JUIV /OIV r 100 RLS ERROR (12 C3EF'5) I -UlA /DIV RLS CEFFICIENT CHRiGE r',ujIV - Figure 3.23a -~ _ _ _I BURST IR © 150 0.37s /Dly SIGMf=.001 . . . YER O. IOOE-OI /IV RLS ERROR 0.100 RLS COEFFICIENT CHANGE /DIV -- L_ HOR 50.0 l~ - - I 12 COEF'SJ for~~~~~~~~~~~~~~ PTS/DIV Figure 3.23b -- L l - -- D - C I BURST iR © 175 'ER 0. 26 /DIV SIGHa=.001 1\~A ~r\ ~,\ _.. .· V __ f\ _ w \, \/- -ieOR 50.0 _ OR __ _ - _ E _ l __K i I vJLU lr· PA _ __ _ -- - . "," _ RLS ERROR /DIV - 12 COEF'S RLS COEFFICIENT CHRNGE I o50.0 - 'I'' - - ! _ - - I PTS/DIV Figure 3.23c --- i ------------ - ---- -I --A _ L PTS/DIV 0.100 _ l PTS/DIV 50.0 1 /DIV ER \.- AfV -- BURST 1R © 200 .490 ER SIGMA=. 001 /DIV AA\ V VJ \v/ ' _V AnA . v A XVI/ I v HOR 50.0 PTS/DIV O.1OOE-0i1 a- -, I, HUR j LER T -1J- . Li r r - mt m an R Ilrrrrn 1 ,111-P , · 1 50.0 RLS ERROR /DIV AL 1. A. -, - rrrlrl--p I,'ll, 12 COEF'S) -A iA 1 A-*.A kAA A.a 7 r vr uu r rrp --u.rs v ., r- r VV' .,i l+ ' ,.JAAAn. ;--- r r - --u 7 t " I - ?. . t ---- -·- .. JA . , -cn ;r -r·r-urrrrrrPurrrr-. ,1 TIT PTS/DIV /DIV 0. 100 RLS COEFFICIENT CHRNGE ..k.: , r ----4 OR 50.0 I --- PTS/DIV Figure 3.23d --I - -- --- ` © 225 2 BURSTIR /OIV 0.511 50.0 SIGMR=.001 PTS/DIV 0. 100E-01 /DIV ~ A -"'-` m ~ · - Or OR, 1s " "'" i_ _ Li - "I"" RLS ERfiCR --Y"'-"'i v 1w 50.0 PTS/DIV . ,'DIV OR 50.0 J. A.....-"'fik""';"" '""~ ~" ^ " 1-i 1 .Y -* ,V. . ^'-''" , . · ·t- 12 COEF'S) ' v " h E~l RLS COEFFICIENT CHANGE PTS/DIV Figure 3.23e I Jr A x AM A A Ar. J",.. I'V ' f.VW2Vf4 z -. Nl" ,] war-_ - f~v "I BURST 1R @ 250 0.478 /0IV SIGM=. 001 pi'/ I\ . __ 'E9 Ir HOR . , A, U ., 0.100E-01 /DIV RLS ERROR .L... .I.. I Pt' l'""1ro- ' '"'b . - -! 50.0 . I - - I J 1-IMPF · I ^~ ~ ~~ 1; v , PTS/D V 0.100 1 ,~~ . 112 COEF'S] I /DIV ALS COEFFICIENT CHRNGE ., Figure 3.23f f\ AA. A - ~I. ,[I\ li&A\. 11~ 1 A'i qu 'A"$fI VW-l LijAMj_,.AA.. *r$iw7wvl 300 BURST1R 0.511 /DIV SIGMR=.001 f 50.0 VFR I, \ PTS/DIV 1 /DIV IR A\ 50.0 RLS ERROR I12 COEF'S) PTS/DIV O. 100 /DIV RLS COEFFICIENT CHANGE I L IOR 50.0 C-- · I Y I -- PTS/DIV figure 3.23g -- - 1----6 -I -1 ! 1 350 BURSTIR 0.433 VER SIGMR=.Oo1 /OIY 2 r \fK "I-\ yl- ' V - v A V'V v _ 50.0 OR VER , I PTS/DIV RLS ERROR O.IOOE-01 /DIV I (12 COEF'S) I I -· L . HOR 50. 0 _VER a. 100 L. t I~ ~ ~ ...-. V I.''1 - - . 7 " -. - ;....* li . I ..,, , .a ~- Im~r ~ ~ AAk A. .- A.. ~ ~,* d% V. I , ,*. · ~·-.~. -r~ flr aft, vrJ PTS/DIV /DIV RLS COEFFICIENT CHRNGE 4; ___r__ B- yg I *--- :---;- -- -I I I- LR 50.0 PTS/DIV Figure 3.23h 7;L-= ~- -11 -i--e · -. Jmof 112 - - 3.4 Linear Filtering It should noise degrades enhancing these the event quality experiments were and be evident from the preceding sections that of performed post-process to compression these using event signals. signals linear (on To noisy filtering compression with as try data) a pre- RLS. Three questions were examined: Does prefiltering the data to reduce noise improve the quality of the observed compressed events? Is all-pole filtering preferable filtering for this application? Will postfiltering the in it more visible? We decided that reducing to RLS prefilter error the to make data FIR events because we thought the high frequency noise in the data might allow the predictor to adapt more quickly to the events and thereby produce faster sharper events settling in in the coefficient the RLS error. The signal with three different noise 3.24a-3.24c. frequency It band from noise. We thought reduce the error (see eq. is evident that reducing to events in the data. these .5f of thereby permit more figures is the noise energy cost d Rd 3.5) and spectra of the and Burstl levels are shown in figures from .f sample change signal that dominated the by in this band would changing the predictor responsiveness to the - Two 113 - filters were designed for the purpose o the noise power in the frequency band extending to 5fsample' spectrum Figure of a 50 of p [Deczky,1972]. criterion Figures prefiltering on the IIR first al.,1973]. event FIR filtering convolves compressed given in figure 3.25 . of the bursts making in the RLS examples. second not error and apparent seems in input contains spectrum coefficient the noisy program Figures to moel the sequence zeroes to each via an all-pole the 50 point long bursts change event The extended inadequate prediction. sequence with This causes unaffected.) the Figure the effect of FIR signals.l example them difficult (Surprisingly, burst design The convolution adds 49 technique such as RLS. the (purely recursive) filtering. filtered the inverse filter events in the compressed signals due to FIR and 3.27a-c demonstrate 3.28a-c show the effects of IIR The et. with (all-pole) filter generated by purely recursive minimum designed coefficients lfsample the time response and filter [McClellan the denominator a 6 point the illustrates FIR lowpass point algorithm Parks-McClellan 3.26 shows 3.25 from reducing for the corresponding low noise to the fact that this effect is examples is currently not 1. These figures have been adjusted to position the events in the same places as those in the IIR example. That is, they have been left shifted 25 points. - 114 - understood. The number IIR of Thus, filtering zeroes the t process each resulting data algorithm than in the modelling technique) burst, is more FIR and does case not instead introduce a large it adds easily modelled (since it is 5 poleS. by the,RLS an all-pole therefore the event location signals are better behaved. Unfortunately, comparison of these figures with 3.17 indicates that filters does not events; and reason after for this types of its of either filtering noise of the increase filtering, the because by improve the quality of both increase in the prefiltering predictor can be increased RLS cause an coefficient change in noise the FIR more correlation, change more in trying to predict it. activity or figure all-pole error signal undesirable signal. The may be that effectively predicted and the coefficients In any case, this method of enhancement seems to be inneffective. The location noise events in the therein. enhance the events) applied to the RLS the error RLS To FIR error try filter signal. signal removing shown The figure 3.29 for various noise levels. in it obscures (and figure results are the thereby 3.25 was given in In comparison with the unfiltered RLS error signal presented in figure 3.17, there is some reduction of the noise level, but the events themselves - are smeared. compress the 115 - Possibly a matched filter could be designed to events characteristics of in the RLS those events design such a filter. __ error, are but at this time the not known well enough to if Iko -R 0.312 /DIV BURSTI SIGMR=.00o A _ _ . 2008/OIV . . . _ . _ . i, . . LOGO1 MRGN SPECTRUM ., Figure 3.24a _ I17 0.313 /DIV .I -J/ul 2008/DIv _ . BURST1 SIGnR=.OI LOGlO MRGN SFECTRUM . Figure 3.24b _ ____ I_ l IS 0.367 /DIB BURSTI SIGMA=.1 lllb --:R 20DB/DIV n AAh~ --LOGIG MRGN SPECTRUM . FS/DIV Figure 3.24c ____ 1~~wiL \ t> VER T 0.692E-01 /DIV FIR LPF .IFS 5OPTS I l l I I HOR PTS/DIV VFR 5.00 XnnR/nTV r·· --· LOGlO MRGN SPECTRUM ~~~~~~~~~~~~~~~~ · 1I I_ I· I N f yI___ _ ' IX001 He X,0 It's m...... HOR .1FS/DiV Figure 3.25 [R 2.19 /DIV IIR LPF . IFS 5 POLES I j t----; Ii- ii I X I I Ii I F -H-R 1.00 PTS.'DIV .1FS/DIV Figure 3.26 _- - I 1l l VFR :- 0.330 /DIV BURSTI THROUGH FIR LOW PSS SIGMR=.C00 i7 A I. -i- - f| f I A l I. l-- -V1 J vV _ . _ _ H09 4ER _ _ 50.0 -- i\r\ V'7 -T ~ ~ , I i .A I Id % /, E I Z \J , 'I V _ PTS/DIV 0.743E-04 /DIV RLS ERROR (12 COEF'S) I -- -. ' I,I . . A, A,\ /-i~ur O0R 50.0 PTS/DIV VER 1.51 /DIV ,a- i- = ! fI iOR 50.0 -- w - -. - -- r. - V$v V w tllsm t W*4" A .AAAM t . vrv - - - L rle -- It RLS COEFFICIENT CHANGE _- I . .I Jfl ^\Ji AU .- I"r ! . . ~II . . . - -I -l .. - rl--- -- PTS/DIV Figure 3.27a I I i _ . VER /DIY 0.327 BURST1 THROUGH FIR LOW PRSS SMAR=.01 A nl . - . ? - I . I nA ! i , i ! I! , dI 1 II r\r\ / '2/'V . . . .. bL VV'V V \JIu V 50.0 VER T I ! · i_ 1_ & J L 1 · _ _ _ I \& PTS/DIV 0.183E-03 /DIV RLS ERROR 12 COEF 'S - ..... A,hit T w... V'W.,lV L--- HUR --- 50.0 G.978 50.0 m I A 8 .t~A{. ~ TAnth\ ., AL bhA O A I"w v I'yNvN Pv ...... ..... .1 _.I. . T * 1*.T-· q- i l. It .Y . II. ...rr- 1 Iv. l ' _w 1 i · -T4 PTS/DIV /DIV RLS COEFFICIENT CHRNGE PTS/DIV Figure 3.27b .l hi-oel-A--&n Ak 4LAl I MIT WI 1YI Nf I - I i .i '' I IiI''.. vu W A -A VER 0.33 :3 T I BURSTI THROUGH FIR LOW PSS SIGMA=.I /DIV -A ....................... I ---I __ I- I+OR 50.0 VER 7 I -- A MOH A V-\f · I 5u.U 0.162 50.0 RLS ERRtR O I wish y1. L _ I I\ i ·....Id I' I-1-1J, · 4 !C , /^-\ 1,I ,· . . >1 - : PTS/DIV n_786F-03 /DIV h. l I J~~~~~~~~~~~~rJ AkI kIl ONwI I lv'1 (12 COEF'S) ,1 I v PTS/DIV /DIV RLS COEFFICIENT CHRNGE PTS/DIV Figure 3.27c k ,A.J rf"P11 R 1Vvvlf r I, YFR 2 /0IV 0. 300 'A K _ _· __ 1 , A CA _ i { _ j __ __ _ nl\ * _ _ V VI iV'V i _HOR I _ THROUGH 5 POLE LOW PRSS SIGMR=.001 BURST1 YER 50.0 ----- -7- !iR I i I \'- 7 l , _I I· I: . .. I 1, I· L . S isll \JI PTS/DIV 0.231E-03 /DIV I II A RLS ERROR - -CYT---- ----· 'V 50.0 PTS/DIV 4. 46 ,'3IV _ 12 COEF'S) -N----wu I -uu-7lr p-- --·M-N 'V ' RLS COEFFICIENT CHRNGE _ Figure 3.28a 7* T---T I --7 1 VER _ 0. 298 _ _ _ _ _ /01V _ BURST1 THROUGH 5 POLE LOW PRSS _ 'A I ~~~·_~~~~ _ · _ _ I \ V-V-V y VV VI L m1o VER '1 I._. f i,, HOM 1 I..., · RLS ERROR 0.27SE-03 /DIv T, 1 Vl' _ · %I , I , _ TVJi \ /t "`V PTS/DIV 50.0 _A.. ..,.-1 _ IGMA=.Ol ....- A V AiA nir L Iiu yI N UJW 50.0 PfTS/DIV 1.36 /DIV 50.0 PTS/DIV AI..iII ·rrr·ir. -Inlu A nu i, TV I'l,'V' · 12 COEF'S) r·& -:ru yW ' vr . LA AMAA t AMH nr· ·· rii·iu yr, RLS COEFFICIENT CHRNGE Figure 3.28b 1 li.lrr· I I .nlr- ·o-l·- A ..,n· Ujl-*^ .1.wk A Ir-·-r·ul· r· WNlly uV U VWI- -,u VER 0.303 BURSTI THROUGH 5 POLE LOW PRSS SIGHR=. I /DIV I J' ' ' "fA \' ' ' 'I\" f' -" I. OR 50.0 VER i2 lV f I f\, " ' v K! "-,-V k-i' PTS/OIV .10E-02 /DIV 1 I ' ' ' ' '' RLS ERROR I12COEF'S} 1 Llwik- --- - -- --- IVf 1 w I" G HUR 50.0 'ER 0.182 ;J1 . liT jf Wv1 /DIV 50.0 ' I- P 10, I RLS COEFFICIENT CHANGE I,&\ArA-'1 __ /,.,,-__ W PT5/DIV t . HOR A Nq , jV-. 0 PTS/DIv FIgure 3.28c I . . i Im .iII, I'llf' l 0111 ly -11i I W VFR O. 312 /DIV K. BURSTI SIGMA=.001 I I A r . Ilo "-V vL\/ __ _ NOR I I I I I \ ! -- - r f\ ,t -/ _ 50.0 0.637E-02 PTS/OIV /DIV RLS ERROR THROUGH FIR LOW PSS /DIV RLS COEFFICIENT CHANGE __ - .ER 0.528 'I~' HR 50.0 - - I PTS/DIV FIgure 3.29a " I I _ fER 0.312 /DIV A -1 I -i I BURSTI SIGM=.01 A If\ J I , \,~J' -OR 50.0 PTS/OI V 0.28iE-G01 /DIV LS ERROR THROUGH FIR LOW PSS 100. 0.231 50.0 /DIV RLS COEFFICIENT CHANGE PTS/0IV Figure 3.29b I\ L VER - 0.361 BURST1 SIGMR=.1 /DiV dM ___I__ _··____ _I_ .A I I IA1 __ L AM K.''VV'V\JyYV HOR 50.0 YER 0. 124 7-I _ /-\ __· 1_· · __ AIA \ ____ V ,' JV 'v __ V · ' _I _r · · V' V 100. 0.691E-01 50.0 ; .kA- . . . . .. v. RLS ERROR THROUGH FIR LOW PSS /DIV _ S V Y'\f PTS/DIV A nir /DIV RLS COEFFICIENT CHRNGE PTS/DIV Figure 3.29c A _ A A __ V ·- _ . AA _ ___ · - 3.5 Decimation The another nature of the burstl signal suggests band-limited possible compressed low 130 - approach signals. energy The to regions .f sample (i.e. enhancement of the of event spectrum which have .5f sample) to the are amplified, to the high energy regions, in the process of linear compared prediction to (due the mentioned whitening Decimating the data would section). the in last the relative size reduce of the low energy region of the spectrum possibly reducing the impact that the compressed signals noise and in reg ion that allowing the had predictor on to the event expend more "effort" predicting the events and less predicting the noise. Figures 3.30a-c and 3.31a-c show 2/1 decimation of the Burstl signal with and without all-pole prefiltering aliasing. to reduce Figures 3.32a-c and 3.33a-c show the same examples but with 4/1 decimation. Comparison examples (figs. of 3.17 these and event figures 3.28) with show with that increasing the undecimated the coefficients decimation ratios, change more at each but neither the unfiltered nor the filtered decimation methods offers substantial improvement in the quality of the data. In 1. All-pole filtering was chosen over FIR filtering because it smears the events less. - - - 131 - fact since independent lowers the of settling the time decimation of the ratio, predictor and seems since to be decimation the interevent spacing, interevent interference is more likely if the data has been decimated. . I 9rv< ER 0.312 /DOI _ .OR BURST1 ,rtn nAIiA 'v j1S vV V 25.0 i i_ __ hi _ _ __ _ ll A - -- ---rr.v 25.0 i\ T\n I -L L __ __ __ _ r_ _ · ?'! RLS ERROR - -v --u - r ------r-. (12 COEF'S) 7 I--, w,' . -- AA- YVV Y --%/ PTS/DIY 0.995 /DOI n - 25.0 A r DECIM 2/1 PTS/DIV 'ER 0.321E-01 /DIV 4OR SIGMA=.001, RLS COEFFICIENT CHANGE -;LI - -- _R PTS/DIV Figure 3.30a i ' -- Vv - I YER O.309 /DIV BURSTI SIGMa=.0I, DECIM 2/1 ! _1 _ - - A A O _ ~~ ~~~~~ ~ · · · ~Vy'YV~iI HOR 25.0 25.0 ER _ __ _ 0.359 25.0 ~ __ · · v n · I\ I i · I r_ I i I PTS/DIY PTS/DIV /DYI A/N ",l\jI \j RLS COEFFICIENT CHANGE PTS/DIV Figure 3.30b I · /DIV 0.328 BURSTI SIGMR=.1, DECIM 2/1 I'\I 25.0 VER /DIV RLS ERROR - _ . ~A L, ·I··-LC.·L·Y 1 k I . 1 AIll · 1-··r .V j i__ HOR II PTS/D1V 0.168 T-j A! 25.0 ·· ·. ·· -1--- Y. - r flIA, V I ·- ~kv~AhaA I·YI ..1· 1-7-·-·11?II· I A A,LnA. · -····11 I vI Y I' 1\I PTS/DIV 0.733E-01 /DIV 25.0 M~,~ AA "' v 12 COEF'S) RLS COEFFICIENT CHANGE PTS/OIV Figure 3.30c I____ ---·------- _II I 1 - 1 ly· v v . ·L- " ( A .A J VI IvV V ~ ~~ ~ -- - . .ER 0.296E-02 --- -- -r-- R I 'ER -~ k-- t3li A · ' ·- · II -- - RLS ERROR /IV I-- --_ I - 5 POLE LPF. DECIM 2/1 riT\ ri\ l - e I f -1-, f 974 PTS/DIV 25.0 IHOH SIGnR.001. BURSTI /DIV 0. 300 .ER ,- 7V 25.0 PTS/DIV 25.7 /OIV ~~~' 25.0 -- --. -' \ P --- (12 COEF'S) AI'- RLS COEFF:CIENT - I II -1 A- -I - - - - - -/r.l - - -qvvT -VI I V', A --- - - -I· . CHANGE - t- i t- rT PTS/DIV Figure 3.31a ___ L- - -· I ---- ~~~~~~~~~~~~~~~~~~~~ · z [ER 0. 298 BURST1 SIGAR=.01, 5 POLE LPF, /DIV .A A fVA -~~?ITI 25.0 I n/N _ f~ PTS/DIV 0.318E-02 /DIV .-... PA A1 \A RLS ERROR A 25.0 PTS/DIV 3.41 /DIV 25.0 f\n I. DECIM 2/1 A A,. t12 COEF'S) J l RLS COEFFICIENT CHRNGE PTS/DIV Figure 3.31b -------- --- ----- ·_ - 4 · _ , _ -^ I· J .vER 0.303 ! BURSTI ,A f f TV . _ .OR /DIV 25 .0 0.130E-01 fi 25.0 .I~ . t 5 POLE LPF. DECIM 2/1 A\n , v . _ _t_ II I I _: · f l z I I /DIV RLS ERROR (12 COEF'S3 _ /OIV ·s _ "--v PTS/D1V _ 0.498 b~ SIGMR=.I. RLS COEFFICIENT CHRNGE PTS/DIV Figure 3.31c r\i\ % i / -- --- -- It I t f S;I \--J-i I. II tER /OIV 0.307 ,1 i E X BURSTI1 SIGMR=.001. J\ - _ __ VIV 12.5 DECIM 4/1 A1 L _ l_ VVVI I' 12.5 PTS/DIV i. 16 /DIV 12.5 PTS/DIV _ _ 7 a71\ \ PTS/DIV 0,849E-01 /DIV _ RLS ERROR 12 CEF'S) RLS CEFFICIENT CHANGE Figure 3.32a _r t l , -/l-\ _ I _ <I L _ BURSTI SIGMAH.1. /DIV fER DE' ol/I 0.309 S I - F OR - -- . J V VI vQA 12. S 12.5 0.553 12.5 l l; 1 ~AtAL .- - RLS ERROR (12 COEF'S] -, rt _\ J\ /l\, A PTS/DIV /DIV RLS COEFFCIENT CHANGE PTS/OIV Figure 3.32b · l · l I PTS/DIV 0.917E-01 /DIV X r fn ? / _ _ V I VFR n R --- In T ,. - --- ., , ... DUM~aun11 IGMR=,I, DECIM /1 I~ x - - - -v-1 HR 12.5 [ER 0. 178 I v- -- [R 12.5 £ER 0.134 12.5 I -- - ---- A. PTS/DIV /DIV RLS ERROR (12 COEF'S) I~~~~~~~~ A XV 1%/ I, vV A I\ 1' V ' 1f AAAv VI' E . .. v A 4v PTS/OIV /ODI RLS COEFFICIENT CHANGE f~~~~~~~~lof I m PTS/DIV Figure 3.32c IX -- !ER 0. 297 /DIV . - hOR 12.5 f BURSTi SIGMA-.001. 5 POLE LPF. CECIM 4/I . i'\ Iyj f0 -W *, , A/I I ~ I 1 mI tI ' i I ,vvjvv v I Ad . f i # 16 f , 4 _ I f _ x,.,/ \--V, PTS/DIV 0.413E-01 /DIV RLS ERROR 112 COEF'S) :ne y ER . 4OR 9.17 /DIY RLS COEFFICIENT CHANGE ,~ >--. . II 12.5 i - i -- -- 4 I PTS/DIV Figure 3.33a It 'ER . 298 /Div ' __ Afx fA _ I I i[ t ,l - U I - - - 12.5 PTS/OiV 1.40 /DIV - I' I --- -- --- - ._ HOlI L RLS ERROR ' V iRo VER 12.5 I I I I . ] i\ r \-,\Q' % q . -I 1 J . - PTS/DIV 0.525E-01 /DlV __ DECIM 4/1 10 1,N / n VV\I V VI 12.5 VER BURSTI SIGMAR.O1, 5 POLE LPF. LL -- A-- P, V - ',J,V (12 COEF'S) / , JV '-W/ f A Ax_ , - - - , V Wv\, V A - - V 11 ---\kl\ I# -# _ -4!f* ..= f, \ RLS COEFFICIENT CHANGE - -, · -· PTS/Di V Figure 3.33b ---- ----- -------- - I 1 i' K 0.303 VER , _~ _ _ /OIV , ~ A f . BURSTI SIGMA=.i, $ POLE LPF, DECIM 4/1 It VV' { v I_ I VI ,\ __ V'J' / ,li\ VI K' \J _ y' 12.5 PTS/DIV 0.818E-01 /DIV 12.5 t rER 0.245 RLS ERROR (12 COEF'S) PTS/DIV /DIV RLS COEFFICIENT CHANGE J i HOR 12.5 PTS/D IV Figure 3.33c _ __ - 3.6 Summary preceding The change signal) events when data. If pulses, the addition, order exceeds ratio despite the duration the predictor the length that and both the of these coefficient locating the positions of apparent not in the input events 40db, the compressed overlap severe events of error for are positions compressed the of S/N and visible (the RLS provide a means those the indicate examples signals compressed event are 144 - are well compressed input the In separated. events appears and of to is be on the fairly independent of the existance or nature of previous events. The following are some of the important results these experiments. The starting location of the RLS iteration the of sizes appears to only affect the a large there is compressed events. However, initial pulse in the coefficient change signal due to the sensitivity of the algorithm to the This means that the first few data points. lengths of predictor data must have several the if first event the preceding noise coefficient change signal is to be used for event compression. Additive noise in the data does not hve a substantial effect on the pulse shape of the However, it does cause compressed events. noise in the RLS error and coefficient change signals, which seems to be in proportion to Unfortunately, the noise level in the data. there is a large variation in the size of compressed events with this technique; making it difficult to establish a S/N criterion for event compression. -- ___ of - 145 - Compressed events do not influence each other substantially if the predictor has time to settle between them, even though the bursts which cause them may be severely overlapped. Figures 3.23a-h showed this by changing the relative position of the second and third bursts in an input sequence. The resulting compressed events had constant shape so long as they did not overlap. Therefore, the faster the predictor settles, the closer events can be to one another and still be resolved. Th i s al so means that events containing large numbers of zeroes (as in the FIR filtering examples) will not compress well ih this .thech-iQue unless they are widely separated. Note that a burst containing a large number of zeroes does generate a compressed event, but the event has longer duration than it would if the zeroes were not present. While filtering the data may be necessary to reduce aliasing, we found that it did not improve the quality of the event compressed signals. If filtering must be performed, all-pole filtering should be used since the zeroes introduced by FIR filtering tend to lengthen the compressed events. We found that decimation does increase the responsiveness of the predictor to each event, but the relative event to noise ratios in the location signals do not improve. In addition, the predictor settling time becomes a larger fraction of the event spacing, thereby increasing the likelihood of interevent interference. Consequently, decimation should be avoided; and in fact over-sampled data is likely to provide higher quality compressed events. _ ______I__IIIILCCC______ - 4. 146 - CONC LUSIO!S this thesis we have develope4 an event compression In technique compress model based on the events of differing by the assumed examined the RLS RLS algorithm, which provided spectra; RLS method prediction (i.e. error can effectively they fit the We all-pole events). as an event compressed signal and in addition developed the coefficient change signal for compression. event , positions minor that but or effects on should technique variations the in that S/N ratio is the ordering, of event the those high this > (e.g. event have only Therefore, this algorithm events. compressed useful be in position starting indicate experiments is effective only if the technique 40db) Our situations where the events are close to all-pole and the noise level is not high. This thesis was initial an investigation into the feasability of using the RLS algoritm for locating events with differing discover When spectra. we to synthesized be work to use the in the data. if we wanted to it did, where Given that need, the best course algorithm examples, where we of the events RLS this if the technique had any value and should further work be done. seemed began on a large knew the number These examples number of and positions indicate that is indeed a viable means for performing event compression. However, there are many questions _ touched upon in the course _______I_____IIL___C__L__ - 147 - of this thesis which need more detailed investigation. One of most the that the events in the data cannot be predictor settling time, if other. remains with issue The of unsolved. noise level, and point), dependent settling what than the spaced closer determines with decreases 3.23a-h settling time indicate that it increases predictor that indicate of this previous length a (up to it is not strongly events. the Also, on the position time for a given event is strongly dependent on the "character" of that dependence of the characteristics means thesis is they are to be resolved from each Our experiments figures this important results of for event. More predictor would be settling insofar useful; the decreasing investigation settling time on as it of the of the event the a provides predictor, thereby reducing the necessary separation between events. Another possible avenue of investigation is that of alternatives to the RLS algorithm as presented in this thesis. Some variations of this moving region Q over adaptive algorithm exist which which the total minimized, rather than an expanding Still other data. methods These involve modifications squared region energy allow E is as was done here. exponentially weighting would us.> the the past algorithm to be used over large amounts of data without the predictor becoming totally insensitive to the new points (as _______________IC______ the error region - 148 - grows the predictor tends to become less sensitive). applications that capability importantly, reducing the predicted might allow the new events, thereby might amount of algorithm reducing be essential; previous to more the minimization expanding or moving would permit the might the be settling region more data time. to be Our to own the region of effective at each more readily adapt feeling is that a means for dynamically changing error In some than iteration, simply since that region of error minimization to be reduced at a new event to shorten the settling time, and then lengthened between events to reduce noise. We found the coefficient for event compression, but our all the coefficients and a change signal signal used to equal be useful weights for fixed decay time for the filters. Is there an optimum choice for the weights of the coefficients in the change signal? Could Kalman filtering be used to track the coefficients more effectively? alternative than example adaptive signal an simply high might make decay the Perhaps passing time compressed for there is a better the coefficients. the coefficient events For change for the second burst in the Burstl signal more visible. Finally, real data. a means there is no substitute for experiments on The original motivation for this work was to find for compressing --111--- - sonic well log -- ----- data. __ Unfortunately, - subsequent 149 - tudy of the well logging problem revealed that the structure of the signals recorded in that situation does not fit the model that was assumed for this work. Experimentally we found that this those but given surprising. event technique their complicated Therefore, compression the assumed model did not compress to the data application which is needed. structure, We of that was this method more closely corresponds hope to perform using acoustic cardiac data in the near future. II signals, __ not of to experiments - 150 - References Deczky, A.G. (1972). Synthesis of Recursive Digital Filters Using the Minimum-p Error Criterion", IEEE Trans. on Audio and Electroacoustics, vol. AU-20, no. 4, pp. 257-263 Eykhoff,P. (1974). SYSTEM IDENTIFICATION: Parameter and State Estimation, John Wiley and Sons, New York, p. 235 Makhoul, J. (1975). Linear Prediction: A Tutorial Review', Proc. IEEE, vol. 63, no. 4, pp. 561-580 Markel, J. (1972). The SIFT Algorithm for Fundamental Frequency Estimation",IEEE Trans. on Audio and Electroacoustics, vol. AU-20, no. 5, pp. 367-377 McClellan, J., Parks, T., Rabiner, L. (1973). A Computer Program for Designing Optimim FIR Linear Phase Digital Filters", IEEE Trans. on Audio and Electroacoustics, vol. AU-21, no. 6, pp. 506-526 Satorious, E.H. (1979). On the Application of Recursive Least Squares Methods to Adaptive Processing", Intl Workshop on Appl of Adaptive Control, Yale University Tribolet, J. (1977). Seismic Applications of Homomorphic Signal Processing", PhD Thesis M.I.T. Ulrych, T. (1971). Applications of Homomorphic Deconvolution to Seismology", Geophysics, vol. 36, no. 4, pp. 650-660 Young, T.Y. (1965). "Epoch Detection-A Method for Resolving Overlapping Signals",Bell Sys. Tech. Journal, vol. 44, pp. 401-426