ARTIFICIAL NEURAL NET BASED NOISE CANCELLOR P.K. Dash, S. a h a , P.K. Nanda, Electrical Engineering Department, Regional Engineering College, Rourkela 769 008 H.P. Khincha, Electrical Engineering Department, Indian Institute of Science, Bangalore 560 012 ABSTRACT T h i s paper p r e s e n t s a new method for n o i s e c a n c e l l a t i o n with an Artificial Neural Network. The network used is a feedforward one with t h r e e l a y e r s . The backpropagati on and st a s t i cal Cauchy ’s 1ear n i ng a1gor i t hms are empl oyed for a d a p t a t i o n of t h e i n t e r n a l parameters of t h e network. The constrained tangent h y p e r b o l i c f u n c t i o n is used t o a c t i v a t e t h e neurons and to provide the desired non-1 i n e a r i t y . Promising s i m u l a t i o n r e s u l t s for noise cancellation intensify the val i d i t y of s u p e r s e d i n g t h e proposed scheme To for many existing techniques. demonstrate t h e e f f e c t i v e n e s s . t h e proposed different input method is a p p l i e d t o . SNRs. With conditions with varying incomplete s i g n a l samples t h e n e t is found t o produce o u t p u t having a striking resemblance with t h a t of t h e d e s i r e d ones. A performance comparision of the two a l g o r i t h m s is p r e s e n t e d i n t h e paper f o r better appraisal. I NTRODUCTI ON The usual method of estimating a s i g n a l c o r r u p t e d by a d d i t i v e n o i s e is t o p a s s i t through a f i l t e r t h a t t e n d s to suppress t h e noise while leaving t h e signal r e l a t i v e l y unchanged. The f i l t e r s for !.he above purpose may be f i x e d or a d a p t i v e , where t h e l a t t e r one h a s t h e a b i l i t y t o a d j u s t t h e i r own p a r a m e t e r s a u t o m a t i c a l l y and t h e i r d e s i g n r e q u i r e s l i t t l e or no a p r i o r i knowledge of signal or noise c h a r a c t e r sti cs. In c i rcumstances where a d a p t i v e n o i s e c a n c e l l a t i o n is a p p l i c a b l e C11 levels of noise rejection are a t t a i n a b l e t h a t would be d i f f i c u l t or i m p o s s i b l e t o a c h i e v e by d i r e c t f i l t e r i n g . A good number of a d a p t i v e a l g o r i t h m s have been r e p o r t e d i n t h e l i t e r a t u r e over t h e p a s t f e w decades t o u p d a t e t h e i n t e r n a l p a r a m e t e r s w i t h a view t o o b t a i n h i g h t h e o u t p u t s i g n a l i n many f i d e l i t y of complex envi r onments. However t h e major f o c u s of r e s e a r c h over t h e p a s t few y e a r s is t o d e v e l o p robust algorithms for a d a p t i v e systems. H o w e v e r t h e past f e w years witnessed t h e f l i n c h i n g of t h e r e s e a r c h e r s t o t h e Neural NetCANN3. The f i e l d of A r t i f i c i a l p o t e n t i a l f e a t u r e s of t h e ANN motivated t o p u r s u e r e s e a r c h i n t h e f i e l d with much a c c e l e r a t e d v i geur w i t h a -hope of a c h i e v i n g human l i k e performance i n t h e f i e l d s of speech and image r e c o g n i t i o n 12.31. In above f e a t u r e s n e u r a l addition to t h e networks can provide solution to o p t i m i z a t f o n , h a n d w r i t t e n d i g i t - r e c o g n i ti o n , radar target detection. text- to- speech s y n t h e s i s and computational models of b r a i n f u n c t i o n s . Adaptive non- parametric n e u r a l networks designed for pattern c l a s s i f i c a t i o n w o r k well for many real world problems C41. Neural n e t models are spcified by the net topology.. node c h a r a c t e r i s t i c s . and t r a i n i n g or l e a r n i n g r u l e s . Both t h e d e s i g n p r o c e d u r e s and t h e training rules for supervised and unsupervised t r a i n i n g a r e t h e t o p i c of much c u r r e n t r e s e a r c h C5.6.71. The a p p l i c a t i o n of f eedf or w a r d mu1ti 1a y e r ed neur a1 net e s t i m a t i o n have heen models t o s p e c t r a l r e p o r t e d i n few r e s e a r c h p a p e r s 18.91The a p p l i c a t i o n of n e u r a l n e t t o t h e f i e l d of n o i s e c a n c e l l a t i o n is h a r d l y found i n t h e 1i t e r a t u r e . N e t based N o i s e I n t h i s paper a Neural C a n c e l l e r i s proposed. A t h r e e layered f eedf o r w a r d ne t w o r k i s employed f o r noi se c a n c e l l a t i o n . The backpropagation a l g o r i t h m and S t a t i s t i c a l Cauchy’s a l g o r i t h m w i t h Bol tzrmnn’ s p r o b a b i l i t y d i s t r i b u t i o n is employed for updating the internal Since parameters i . e . weight of t h e n e t . both p o s i t i v e and n e g a t i v e t i m e domain output s a m p l e r are of equal importance. t h e tangent hyperbolic function. constrained to o p e r a t e i n n o n - l i n e a r zone. is used t o a c t i v a t e t h e neurons and t o p r o v i d e t h e d a r i rod non-1 i near i t y . The tangent hyperbolic function s a t u r a t e s a t u n i t y for small magnitudes of b o t h p o s i t i v e and n e g a t i v e v a l u e s . hence t h e maximum v a l u e of tho activation function is suitably a d ‘ u o t o d and tho o u t p u t of a neuron is roishhod b e e r n a C r & r r i C &he nm-LLnear zone. pr-edictei‘ o p u t bs o t for inout samplesThe contaminated wui tt h h harmonic . - i n t o r f e r e n c e and additive uncorrelated w h i t e n o i s e b e a r s a s t r i k i n g resemblance w i t h t h a t of t h e d e s i r e d o n e s . The fundamental s i g n a l components p r i r m r i l y of power f r q u e n c y and i t s c o r r e s p o n d i n g t h i r d and f i f t h harmonic compononts are s u c c o r sf u l 1 y e x t r a c t e d f r o m t h e c o r r u p t e d signal with l o w SNRs. The n e t is also exposed t o many never s e e n d a t a o t h e r t h a n t h e t r a i n e d ones and t h e n o i s e c a n c e l l a t i o n f e a t u r e is commendable. The n e t p o s s e s s e s potentiality to extract the desired components even w i t h 50% s u p p r e s s e d i n p u t d . c and w i t h d a t a . b o t h with predominant appreciable harmonic s t r e n g t h , without computer signal degradation. Sufficient s i m u l a t i o n r e s u l t s aro p r e s e n t e d f o r a wide yar i e t y of cases. 37 ARTIFICIAL. NEURAL NETWORK I n t e r e s t i n a r t i f i c i a l n e u r a l networks h a s grown r a p i d l y over t h e p a s t f e w years. This i n s u r g e n c e of i n t e r e s t fired by b o t h theoritical and application successes m o t i v a t e d t o m a k e machines t h a t l e a r n and remmber i n w a y s t h a t boar a s t r i k i n g p r o c e s s e s and resemblance t o human mental a s s i g n e d a new s i g n i f i c a n t meaning to A r t i f i c i a l I n t e l l i g e n c e . I n 1QSOs and 1Q600 a g r o u p of researchers combined the b i o l o g i c a l and p s y c h o l o g i c a l insights to produce t h e f i r s t a r t i f i c i a l n e u r a l network model. Due to limitation in r e p r e s e n t a t i o n a l c a p a b i l i t y of t h e s i n g l e neuron and single layer neurons, mu1 t i 1a y e r ed per c e p t r o n s wor e d e v e l oped of which invoked the development a1gor i t hm back pr o p a g a t i on t r ai n i ng by P a r k e r C103. Development of d e t a i l e d mathematical models began more t h a n f o u r decades a g o w i t h t h e w o r k of McCulloch and P i t t s . Hebb. Rasenbl a t t W i d r o w and o t h e r s . More r e c e n t w o r k by H o f f i e l d . Rumelhart McCelland C111. Sejnowski , Feldman. Grossberg. Widrow t l 2 . 1 3 1 and o t h e r s l e d t o a new i n s u r g e n c e i n t h e f i e l d . Over t h e p a s t few y e a r s t h e r e s e a r c h about n e u r a l n e t h a s been p e r s u e d w 1 t h accel er a t e d v i gour s , si n c r t h e advent of new n e t t o p o l o g i e s , robust algorithms. a n a l o g VLSI implementation and t h e c o n c e p t of p a r a l l e l p r o c e s s i n g . These models are s p e c i f i e d by t h e net topology, node c h a r a c t e r i s t i c s and t r a i n i n g & l e a r n i n g set r u l e s . These r u l e s s p e c i f y a n i n i t i a l of w e i g h t s and i n d i c a t e how w e i g h t s s h o u l d the b e a d a p t e d d u r i n g u s e t o improve performance. A d a p t a t i o n or l e a r n i n g is a major f o c u s of n e u r a l net research. The a b i l i t y t o a d a p t and c o n t i n u e l e a r n i n g is essential in areas such as speech P 0, W 3 a C rt Bidden output layer layer j k i Fig. 1, A typical 3 layered feedforward Artificial Neural Network. TRAI N I NG ALGORITHMS The n e t t r a i n e d to j u s t i f y the noise cancellation capability i s shown i n F i g . C l > . The s u b s c r i p t s i,.j and k r e f e r to t h e input. h i d d e n and o u t p u t layer r e s p e c t i v e l y and t h e s u b s c r i p t s n. m and 1 c o r r e s p o n d t o a n y u n i t irr I n p u t . h i d d m and output layer. The n e t is t r a i n e d with s u p e r v i s i o n by t h e r e q u i r e d i n p u t and t a r g e t vectors. The t w o a l g o r i t h m s employed for t r a i n i n g are b a c k p r o p a g a t i o n and s t a t i s t i c a l Cauchy’s a l g o r i t h m which are p r e s e n t e d b e l o w . . r e c o g n i t i o n where t r a i n i n g d a t a i s Backpropagation Algorithm: For e a c h t i m e s t e p t h e f o l l o w i n g s t e p s are to bo exacuted. I n i t t l i s a t i o n of weights. t h e w e i g h t s of b o t h t h e l a y e r s random V a l uos. SteD-1: Stev-2: Tho i n p u t vector t h e t a r g e t v e c t o r do, Xo,. . . . , dN-l Set t h e to small . . . .XN-l are and formed and a p p l i e d t o t h e n e t . Sterr-8: C a l c u l a t e t h e a c t u a l o u t p u t s of b o t h limited t h o layers and new t a l k e r s . new words. new p h r a s e s and new envi r onments are c o n t i nousl y encountered. with However t h e m u l t l l a y e r e d ANN back pr opagat i on a1 gor i t hm is mst common1y used for many p r a c t i c a l problems. A t y p i c a l is t h r e e l a y e r e d network shown i n F i g . C l 3 t r a i n e d w i t h s u p e r v i s i o n . The t h r e e l a y e r e d network h a s i n p u t l a y e r . h i d d e n l a y e r and output layer consisting of multiple n o n - l i n e a r neurons associated w i t h s u i t a b l e a c t i v a t i o n f u n c t i o n . Each i n p u t p a t t e r n t o t h e n e t f e e d s up t h r o u g h t h e hidden l a y e r s upto t h e output l a y e r . Each neuron forms t h e weighted sum of t h e i n p u t s and p a s s i t t h r o u g h t h e hidden l a y e r s u p t o t h e o u t p u t l a y e r . Each neuron forms t h e weighted sum of t h e i n p u t s and p a s s i t t h r o u g h t h e n o n - l i n e a r a c t i v a t i o n f u n c t i o n t o produce a o u t p u t which s e r v e s as t h e i n p u t t o t h e next l a y e r i.e. output l a y e r . The n e t is t r a i n e d t o minimize t h e o b j e c t i v e f u n c t i o n which is d e f i n e d a s t h e h a l f of t h e sum of t h e s q u a r e s of t h e d i f f e r e n c e s between t h e pr edi cted o n e s and the cor r espondi ng d e s i r e d component. Once t h e network is t r a i n e d it h a s t h e c a p a b i l i t y of producing o u t p u t v e r y close t o t h e d e s i r e d ona w i t h i n p u t p a t t e r n s other than t h e t r a i n e d ones t h u s e s t a b l i s h i n g t h o v a l i d i t y for real t i m o implemontation. A y = f< k . x . cw1> c13 A y = E s t i m a t e d Output = Weighting factor to constrain t h e a c t i v a t i o n function i n t h e non-1 i n e a r zone. CWI= Weight matrix between any t w o layers. ?. fC.3= A.tanhC.> S t e ~ - 4 : C a l c u l a t e t h e error function. e * = d - y E =Cl&> P 1 Cdk - & c 23 objective r: 31 yk> A 2 c 4> Ep = o b j e c t i v e f u n c t i o n Adapt t h e weight. tW1 C t + l 3 = CWlijCt> + v . 6 X where i J CWI C t > i s t h o w i g h jtj SteD-5: C5) from iJ hidden node i or f r o m a n i n p u t to node j a t tima t. X i r e i t h w r t h o output of node i or -j is a n i n p u t , r) is a convergence c o - e f f i c i e n t varied from 0.1 t O 1.0. bJ is an error term f o r node j. If node j is an output node thela 38 when d J i s t h e d e s i r e d o u t p u t of node j 12 are the target waveforms for f undamental , 3r d harmonic and 5 t h harmonic respectively. Fig. 5 & 1 3 exhibit tha i n a d e q u a t e i n p u t samples and t h e p r e d i c t e d o u t p u t s which ara u p t o t h o m a r k s and hence i t is c l a i m e d t h a t Neural Net based n o i s e c a n c e l l e r is supcriar to the exiistiny mot hods. CONCLUSION Exhaustive computer simulation results e s t a b l i s h t h o v a l i d i t y t o o p t for the proposed schenw over t h e e x i s t i n g d i g i t a l f i l t e r s . The performance i n e x t r a c t i n g t h e signal components from i n a d e q u a t e c o r r u p t e d d t a excel o v e r t h e e x i s t i n g t e c h i n i q u e r . .;?a local minima t r a p p i n g e f f e c t of the objective function is circumvented by s w i t h i n g over t o Cauchy's algorithm. The combined backpropagation and Cauchy's a l g o r i t h m i s b e i n g s t u d i e d t o overcome local r i n i m a t r a p p i n g and t o achieve faster convergence. The n e t is found t o ' be f a u l t t o l e r a n t and real t i m e implementation of t h e proposed scheme seems t o be more promising. I k v e l opmant and i mpl e m e n t a t i on of the g e n e r a l i r e d filter is in Progress by employing b o t h a d a p t i v e P a t t e r n Racognit.ion concept and r o b u s t t r a i n i n g a l g o r i t h m . & and yJ is t h e a c t u a l o u t p u t . If node j is a n i n t e r n a l hidden t h e n c ?> '- 6 CA- Xy/A3 6kCW1 Jk - J where k is ovor a l l nodes to t h e r i g h t of node j. The s t e p s f r o m 1 t o , 5 are r e p e a t d t i l l t h e o b j u t i v e f u n c t i o n is minimized. Cauchy's Algorithm: The f o l l o w i n g s t o p s are adopted for t r a i n i ng t h e network Step-1: A v a r i a b l e T t h a t represents an i n i t i a l a r t i f i c i a l t e m p e r a t u r e i s chooson. U s u a l l y T is i n i t i a l i z e d t o a large v a l u e . Also a n o t h e r variable. t = a r t i f i c i a l t i m e analogous t o a s t o p . is initialised. U s u a l l y t i s set t o a small v a l u e . Here T=0808.0, t = 2 . 0 . Step 2: A set of i n p u t is a p p l i e d t o t h e network, t h o c o r r e s p o n d i n g o u t p u t and t h e o b j e c t i v e f u n c t i o n is e v a l u a t e d . Step-3: The w e i g h t s are changed by a random v a l u e which is g i v e n as x = pCTCt>.tanCFYX>l C83 . where p co-ef f i c i e n t x = the learning rate ACKNOWLEDGEMENT = t h e weight change The a u t h o r s acknowledge w i t h t h a n k s t h e M i n i s t r y of Human Resource Development , Go-. of I n d i a for p r o v i d i n g necessary PCX> = a radom numbor s e l e c t e d f r o m uniform d i s t r i b u t i o n over the open i n t e r v a l -n& t o n&. Step- 4: If the objective function is improvedCreduced>, r e t a i n t h e weight change. a funds . REFERENCES B.Widrow et a1,"Adaptive noise c a n c e l l i n g p r i n c i p l e s and applications" . IFEE. vo1.63. pp. 1892-1716. k c . Proc. 1875. c 8 1 R. P. Lippmann. " An Introduction to computing w i t h Neural Nets".IEEE ASSP Mag. 4C23. pp 4-22, A p r i l , 1987. t31 A. Waibel. H.Sawai and K. Shi kind. "Modul ar i t y and Scal i ng i n Large phonemic Neural N e t w o r k " . I E E E Trans. on ASSP V o l . 37. N o . 1 2 . Dec.1989. C41 R. P. Lippmann."Pattern classification network" , IEEE Communication using neural Magazine PP 27-62. Nov. 1989. C53 M.R.Azimi-Sadjadi.S.Citrrn and S a s s a n Shaedr ash. "Super v i sed L e a r n i ng P r o c e s s of Multi- Layer P e r c e p t r o n Neural N e t w o r k using Fast L e a s t Squares'., ICASSP, April 3-6, 1900.N e w Mexico. U. S. A. [GI P. Eurrascano and P. Lucci ,"A Learnlng r u l o d o m i n a t i n g Local minima i n multiloyap p e r c e p t o n s " ICASSP A p r i l 3-6 1N e w Mexico C13 C Q> Step-5: If t h e weight change r e s u l t s i n a n increase in the objective function. calculate t h e p r o b a b i l i t y of a c c e p t i n g t h a t change from t h e Boltzmann's d i s t r i b u t i o n as f 01lows. PCc> = exp C-cAl3 c10> where PCc>=the p r o b a b i l i t y of a change of c i n t h e obiective function. k = a constanL analogous to Boltzmann's c o n s t a n t t h a t must b e choosen. Here k = is s e l e c t e d . T = t h e artificial temperature. The above s t e p s are repeated u n t i l t h e o b j e c t i v e f u n c t i o n i s minimized. RESULTS & DISCUSSION I n t h i s research t h e network is t r a i n e d i n d e p e n d e n t l y for e a c h case namely. fundamental, 3rd Harmonic & 5 t h Harmonic etc. The s i g n a l g e n e r a t e d is goverened by yCt> = 2 sinCkot> + VCt> 0.S. A. C 71 S. Y . Kung and K. I . Diamantras, "A Neural N e t w o r k Lear n i ng a1gor i thm for a d a p t a t i VIZ Principal Componenet Extraction CAPEX>" ICASSP April 3-6. 1QQO.New Mexico U . S . A . C81 Chia-Jio Wang & M a r k A.Wickert and C . H . W i e . "Three layer Neural Network for. s p e c t r a l e s t i m a t i o n " ICASSP. April 3-6. 1990. N e w Mexico, USA. t01 P. K. Dash, P. K. Nanda and S. S h e . '* Spectral e s t i m a t i o n E i n g & L i f i C i r l Neural Network" . International Conference on SIE%i!-S 5 SYSEYS. Dec. 10-12. T i r u p a t i . India. C 101 Phi 11i p D. Warsor man, "Neura L Computing. T h e o r y and Practice" Van Nortrand & Reinhold 1989. C 11I D. E. R u m e l h r r ' t and J. L. McCe11and. "Para.1 l e l D ~ ~ t r i ~ t e d . P r o c c s s tV~a"l .. 1 & 2 Cambridge. MA. MIT Press 1-. c11> . where k = 1 . 3 , 5 . 7 . . n 8, V C t > is zero mean w h i t e n o i s e . The o b j e c t i v e f u n c t i o n s for t h e fundamental, 3 r d harmonic and 5th harmonic are p r e s e n t e d i n F i g s . 2.6.7 & 11. Fig. 6 shows t h e local minima t r a p p i n g w i t h backpropagation algorithm for third harmonic which i s a11 evi a t e d u s i ng Cauchy' s algorithm although it is slower in convergence as d e p i c t e d i n ' Fig. 7. The input with varying a.Rs and the corresponding o u t p u t s bearing a close resemblence to t h a t of t h e desired o n e s . a r e r e p r e s e n t e d i n F i g s . 4.9.10 & 1 4 Fig. 3, 8 39 /? B. Widrow. R. G. Winter and R. A. B a x t e r , "Layered Neural Nets for pattern No. 7. r e c o g n i t i o n " . IEEE Trans. on ASSF'. J u l y 1988. 1 1 3 1 M. Stevenson, R. Winter and B. Widrow. "Sensi vi t y of f eedf or w a r d Neur a1 Networks t o weight errors " . IEEE Trans. on Neural N e t w o r k . V o l . 1. No. 1. March I Q Q O . C121 I I v) aJ a Input output I 1.6 1.0 0.6 1 2tn -1.0 25.001 0.00 * ~. .lJ 0.0 4 C - 06 A ! 2.01 I -1.6 -2.0 -2.6 0 ie 12 8 4 Samples --> Fig. 5 I I - - 10.00 m ' 'El 125.00 0 1500 C 4500 3000 Adaptations --> Fig. 2 -20.00 500 250 0 Adaptations --> Fig. 6 A -2.6 - ' . O i + y - y - - 12 Samples IO.OO[ 18 --> Fig. 3 -.. A * 2.01 Input - Output' -50.00' 0 2000 4000 8000 BOO0 Adaptations --> Fig. 7 0 12 8 4 Samples 16 --> Fig. 4 Fig. 2: L e a r n i n g Curve for Fundamental Samples u s i n g b a c k p r o p a g a t i o n a l g o r i t h m c 3 sets of t r a i n i n g s i g n a l s ) , r)=O. 45. X=O. 0068 A = 3 . 0 . f = 800 Hz. S F i g . 3: Tar g e t Wavrf or m for fundamental C SoHz3 F i g . 4: I n p u t s i g n a l w i t h SNR=OdB the p r e d i c t e d o u t p u t for fundamental. 4. 40 F i g . 5: I n p u t s i g n a l . SNR=OdB w i t h a w n samples suppr errcdC i . e 2 . 4 , . .16> and t h e horresponding predicted output for fundamental . , Fig. 6 : L e a r n i n g Curve for 3 r d HarmonicCBack p r o p a g a t i o n Al. > .r)=O. 8 . A=O. 009. A=2.4. F i g . 7 : L e a r n i n g Curve for 3 r d Harmonic u s i n a Cauchy ' s A l . ,p=O. 01 , A =O. 0 7 , A = 2 . 0 . 4 -16 -2.0 I - 4 : ' ' ' : ' ' ' : ' ' ' 0 ; 8 4 ' 12 ' ' ! 18 Samples --> Fig. 12 -+-Input 0 -2.6' : ' ' ' 0 : I I 0 4 , I 8 Output 8 4 12 16 Samples -->Fig. 13 , , , - 18 12 Samples --> Fig.. 9 A -+- Input 2.0 A - 2.6 I I Output 0 P) a :I -+- - Input .Output 3 CI -4 c 0, Id L -2 - -3 t 0 -2.9 -2.6: ' 0 ' ' : ' 4 ' ' : ' ' 8 ' ' ' 12 ' 4 8 12 18 Samples -->. Fig. 14 ! 16 Samples --> Fig. 10 Fig. 8:Target for 3rd Harmonic. F i g . Q: Input s i g n a l , SNR=-2OdB & t h e output Cbackpropagation Al> Fig. 10: Input S i g n a l , S)rlR=-20dB & Output CCauchy's Algorithm3 Fig. 11: Learing Curve for 5 t h harmonic Cbackpropagation Al3 F i g . 1 2 : Target for 5 t h harmonic. Fig. 13: Input s i g n a l %R=odB with even samplr'suppreesd t h e output. F i g . 14: Input s i g n a l SNR=OdB and t h e output -ao.oo 0 is00 MOO Adaptations--> Fig. 11 41