real time dsp professors: eng. diego barral eng. mariano llamedo soria julian bruno
TRANSCRIPT
Real time DSP
Professors: Eng. Diego Barral Eng. Mariano Llamedo Soria Julian Bruno
Filters
conventional filters time-invariant fixed coefficients
adaptive filters time varying variable coefficients
adaptive algorithm function of incoming signal
exact filtering operation is unknown or is non-stationary!
Random Processes
random != deterministic concepts
realization ensemble ergodic
tools mean variance correlation/autocorrelation
stationary processes & WSS
Adaptive Filters
parts digital filter adaptive algorithm
filter FIR IIR (stability problems are difficult to handle)
Adaptive Filters
d(n) desired signal y(n) output of the filter x(n) input signal e(n) error signal
FIR Filter
wl(n) adaptive filter coefficients
Performance Function
coefficients are updated to optimize some predetermined performance criterion mean-square error (MSE)
for FIR R: input autocorrelation matrix p: crosscorrelation between d(n)
and x(n)
Performance Function
MSE surface One global minimum
point!
Gradient Based Algorithms
properties convergence speed steady-state performance computation complexity
method of steepest descent greatest rate of decrease (negative gradient) iterative (recursive)
LMS Algorithm
statistics of d(n) and x(n) are unknown estimation of MSE
avoids explicit computation of matrix inversion, squaring, averaging or differentiating
Performance Analysis
stability constraint μ controls the size of the incremental correction
λmax is the largest eigenvalue of the autocorrelation matrix R
Px input signal power large filters => small μ strong signals => small μ
Performance Analysis
convergence speed large μ => fast convergence
λ => relation between stability and speed of convergence estimation
Performance Analysis
excess mean-square error the gradient estimation prevents w from staying at wo
in steady state w varies randomly about wo
trade-off between the excess MSE and the speed of convergence
trade-off between real-time tracking and steady-state performance
Modified LMS Algorithms
normalized LMS algorithm μ varies with input signal power optimize the speed of convergence and maintain
steady-state performance independent of reference signal power
c is a small constant μ(n) is bounded
0 < α < 2
Modified LMS Algorithms
leaky LMS algorithm insufficient spectral excitation may result in divergence
of the weights and long term instability
where v is the leakage factor 0 < v ≤ 1 equivalent of adding low-level white noise degradetion in performance
(1 - v) < μ
Applications
operate in an unknown enviroment track time variations
identification inverse modeling prediction interference canceling
Applications
adaptive system identification experimental modeling of a process or a plant
Applications
adaptive linear prediction provides an estimate of the value of an input
process at a future time in y(n) appear the highly correlated components of
x(n)
i. e. speech coding and separating signals from noise
output is e(n) for spread spectrum corrupted by an additive narrowband interference
Applications
adaptive linear prediction
Applications
adaptive noise cancellation (ANC) most signal processing techniques are developed
under noise-free assumptions the reference sensor is placed close to the noise
source to sense only the noise, because noise from primary sensor and reference sensor must be correlated
the reference sensor can be placed far from the primary sensor to reduce crosstalk, but it requires a large-order filter
P(z) represents the transfer function between the noise source and the primary sensor
uses x(n) to estimate x’(n)
Applications
adaptive noise cancellation (ANC)
Applications
adaptive channel equalization transmission of data is limited by distortion in the
transmission channel channel transfer function C(z)
design of an equalizer in the receiver that counteracts the channel distortion
training of an equalizer agreed sequence by the transmitter and the receiver Decision device
Applications
adaptive channel equalization
Implementation considerations
finite-precision effects prevent overflow
scaling of coefficients (or signal)
quantization & roundoff => excess MSE => stalling of convergence
depends on μ threshold of e(n) -> LSB