南京理工大学数字信号处理英文课件 Lecture14.ppt

197 * Lecture 14 Digital Filter Structures ( IV ) Structures for discrete-time filters (IV) ? Effects of round-off noise and overflow – equivalent linear model – scaling ? Zero-input limit cycles Nth-order linear time-invariant system: Consider implemented with finite-length ---- non-linear equation A. Round-off Noise: model Suppose each quantizer can be replaced by noise sources that are equal to the quantization error at the output of each quantizer, example: e[n] = Q( b x[n] ) － b x[n] i.e: Q x[n] b x[n] b e[n] Round-off Noise: model ( II ) Round-off Noise: model ( III ) Assumption: independent white noises (independent from signals) ? May be regrouped in one noise source Variance for truncation to B + 1 bits: σe2 = 2－2B/12 ( Assume the noise probability is uniform distribution ) Round-off Noise: model ( IV ) ? Effect of an LTI on a white noise Total noise power due to round-off computed by isolating the transfer function of the noise The equivalent model of direct form I of IIR filter: B. Round-off Noise: DF I N.B.: FIR . N = 0 and h e f f [n] =δ[n] Similar, the quantization noise model of direct form II is below: And the variance of output noise is: Round-off Noise: DF II ★ Round-off Noise: Remarks ? with a (2B + 1)-bit adder internally, one only has to quantize once before the memory elements and/or output ? DF I : ? DF II : Overflow and scaling: The smallest output noise variance (subject to dynamic range constraints ) is the most important-----overflow E.g. If | x | 1,we may multiply it by a suitable small number s less than 1, i.e. 0 s 1, Then | s x | 1 C. Overflow and scaling ? Overflow may happen in “addition” nodes ? Scaling must be used to avoid it ? Scaling reduces SNR ! ? scaling: B. Scaling reduces SNR ( signal to noise ratio) Example: x[n] System function: H ( z ) → s H ( z ), and the output of unquantized sys- tem is y