Documente Academic
Documente Profesional
Documente Cultură
Jos e C. M. Bermudez
Department of Electrical Engineering Federal University of Santa Catarina Florian opolis SC Brazil
Linear Digital Filters Statistical Linear Filtering Optimal Linear Filtering Adaptive Filtering Adaptive Filter Applications The LMS Adaptive Algorithm
FIR or IIR. Most useful for frequency selective applications. Desired signal and interference are in dierent frequency bands. Example: To extract high frequency noise (> 20 kHz) from speech.
FIR IIR
FIR Filters
Simple to design Can implement linear phase Are always stable Require coecients than IIR lters
IIR Filters
Harder to design than FIR lters Introduce phase distortion May become unstable Require coecients than FIR lters
Many applications more than just frequency band selection Signal and interference are frequently within the same freq. band
We assume some statistical properties of the useful and unwanted signals are available Mean Autocorrelation Cross-correlations etc.
We dene some statistical criterion for the optimization of the lter performance
u(n) and d(n) stationary and zero-mean d(n): information plus some interference y (n): estimate of d(n) given some observations of u(n) (n|Un ) wk u(n k) = d
y (n) =
Performance index: Mean-Square Error w = [w1 , . . . , wN ]T y (n) = uT (n)w J (w) = E {e2 (n)} Mean-Square Error u(n) = [u(n), . . . , u(n N + 1)]T
Rwo = p
(Normal Equations)
10
0 w2
10
20
20
10 w1
10
20
wo : Minimum of this surface Solution of the Normal Equations Computationally intensive even with ecient algorithms
Normal equations must be solved for each n!?! The algorithms used for the stationary case become inecient What about Kalman Filters?
Requires a dynamics model (state-space) for d(n) Computationally heavy for real-time
Signal statistics may be unknown, and there may be no time to estimate them
Computational complexity between 2 input samples limited by processor speed and by cost
Adaptive Filters
Change their weights as new input samples arrive Weight updating is controlled by an adaptive algorithm Optimal solution is approached by improving performance a little bit at each iteration
Optimal solution is approximated after several iterations (iteration complexity convergence time)
Filter weights become random variables that converge to a region about the optimum weights
speaker A H
speech from A
speaker B
echo from A
H speech from B
speech from A
x(n)
Alg.
w(n)
_
h echo y (n)
+ +
y (n)
+
r (n)
e(n)
d(n)
speech from B
Transmitted symbols
Channel + + _ + Adaptive Equalizer
noise
decision error
wo y (n) + x(n)
input signal
+ +
w(n)
Steepest descent needs p and R !? Instantaneous approximation of the gradient E [e2 (n)] e2 (n) = 2e(n)u(n) w(n) w(n)
E [v (n)] = E [w (n)] w o
0.15 0.13 0.11 0.09 0.07 0.05 0.03 0.01 0.01 0.03
E[V(n)]
2e3 4e3 6e3 8e3 10e3 12e3 14e3 16e3 18e3 20e3
iterations
30 10 10
= 0, 001 = 0, 002
30
= 0, 004
50 70 90 110 0 2e3 4e3 6e3 8e3 10e3 12e3 14e3 16e3 18e3 20e3 iterations