TestGuru
发表于 2016-5-27 18:14
dsp2008请问哪里有错,俺倒是灰常地愿意讨论{:{05}:}
TestGuru
发表于 2016-5-27 18:15
你光发个点评,有无任何论据,让俺怎么回复你呢?
czl_gl
发表于 2016-5-27 20:13
为了了解加窗和栅栏效应,今天还重新看了一遍测试技术,发现这些东西还是一知半解的,大致原因可能是高等数学没学好
TestGuru
发表于 2016-5-27 21:41
TestGuru 发表于 2016-5-27 18:14
dsp2008请问哪里有错,俺倒是灰常地愿意讨论
发现没法回复评论,没办法只好摘录回答了。
“...你错误的根源在于把DFT当成了一种新的傅立叶变换。 要理解DFT的频谱泄露和栅栏效应,须搞清楚DTFT的概念。DFT是从DTFT引出来的,DTFT的频域离散化,并将时域和频域均去主周期,即得DFT...."
作为学术交流和争论,你需要明确指明别人的描述错误,而不能以为别人不懂而随便扣帽子。能不能请你明确指出我的描述具体哪句话是错误的,反映出俺错误地把DFT当成了一种新的傅立叶变换了{:{05}:}。俺只是想通俗易懂地解释这个现象,至少比百度那个准确而详细吧。你的确可以往更深一层理论上去推,从DFT,返回到DTFT,再到(DTFS,CTFT,CTFS,特注:无先后顺序{:{05}:}),期待你更精彩的解释。
hcharlie
发表于 2016-5-28 16:32
本帖最后由 hcharlie 于 2016-5-28 17:46 编辑
栅栏效应是指离散采样以后,有些频率成分被挡住,像隔了栅栏看东西一样,会漏掉一些信息。这是个比方,有相像处也有不同之处,相像处是,某个频率的信号被挡住,而不同之处是,在临近频率处漏点光,这是栅栏效应没有说到的。所以就不必太较劲了!
mxlzhenzhu
发表于 2016-5-28 18:18
楼主回答挺好。
hcharlie点评很精彩。
这是dsp2008想要大家看的书
学习无止境。
dsp2008
发表于 2016-5-28 19:22
楼上的兄弟提供的洋文书里,第3.11节DFT RESOLUTION, ZERO PADDING,AND FREQUENCY-DOMAIN SAMPLING从原理到图示,把栅栏效应描述得非常清楚。把这一节文字反复看几遍,应该就能明白了。
dsp2008
发表于 2016-5-28 19:29
3.11 DFT Resolution, Zero Padding, and Frequency-Domain Sampling
One popular method used to improve DFT spectral estimation is known as zero padding. This process involves the addition of zero-valued data samples to an original DFT input sequence to increase the total number of input data samples. Investigating this zero-padding technique illustrates the DFT’s important property of frequency-domain sampling alluded to in the discussion on leakage. When we sample a continuous time-domain function, having a continuous Fourier transform (CFT), and take the DFT of those samples, the DFT results in a frequency-domain sampled approximation of the CFT. The more points in our DFT, the better our DFT output approximates the CFT.
To illustrate this idea, suppose we want to approximate the CFT of the continuous f(t) function in Figure 3-20(a). This f(t) waveform extends to infinity in both directions but is nonzero only over the time interval of T seconds. If the nonzero portion of the time function is a sinewave of three cycles in T seconds, the magnitude of its CFT is shown in Figure 3-20(b). (Because the CFT is taken over an infinitely wide time interval, the CFT has infinitesimally small frequency resolution, resolution so fine-grained that it’s continuous.) It’s this CFT that we’ll approximate with a DFT.
Figure 3-20 Continuous Fourier transform: (a) continuous time-domain f(t) of a truncated sinusoid of frequency 3/T; (b) continuous Fourier transform of f(t).
dsp2008
发表于 2016-5-28 19:30
Suppose we want to use a 16-point DFT to approximate the CFT of f(t) in Figure 3-20(a). The 16 discrete samples of f(t), spanning the three periods of f(t)’s sinusoid, are those shown on the left side of Figure 3-21(a). Applying those time samples to a 16-point DFT results in discrete frequency-domain samples, the positive frequencies of which are represented by the dots on the right side of Figure 3-21(a). We can see that the DFT output comprises samples of Figure 3-20(b)’s CFT. If we append (or zero-pad) 16 zeros to the input sequence and take a 32-point DFT, we get the output shown on the right side of Figure 3-21(b), where we’ve increased our DFT frequency sampling by a factor of two. Our DFT is sampling the input function’s CFT more often now. Adding 32 more zeros and taking a 64-point DFT, we get the output shown on the right side of Figure 3-21(c). The 64-point DFT output now begins to show the true shape of the CFT. Adding 64 more zeros and taking a 128-point DFT, we get the output shown on the right side of Figure 3-21(d). The DFT frequency-domain sampling characteristic is obvious now, but notice that the bin index for the center of the main lobe is different for each of the DFT outputs in Figure 3-21.
Figure 3-21 DFT frequency-domain sampling: (a) 16 input data samples and N = 16; (b) 16 input data samples, 16 padded zeros, and N = 32; (c) 16 input data samples, 48 padded zeros, and N = 64; (d) 16 input data samples, 112 padded zeros, and N = 128.
TestGuru
发表于 2016-5-28 20:23
我这个MIT的小抄本更简单明了,有没有错俺就不知道,只大概瞄了一下。
TestGuru
发表于 2016-5-28 20:28
我公司也有英文版的FFT文章,图文并茂地解释涉及混叠、泄漏、栅栏效应、补零等常见问题,可参考https://en.wikipedia.org/wiki/Fast_Fourier_transform中的文档,俺在这里就不暴露目标了。{:{05}:}
TestGuru
发表于 2016-5-28 21:46
前面说了,尽管有栅栏效应存在,但大能量的谱线并不容易被遮挡/淹没而看不到(这个跟真正的从栅栏缝看东西还不太一样),除非旁边也有大能量的谱线;但是小能量的谱线则比较容易被噪音和/或旁边的有一定能量的谱线淹没。
下面以三角窗的频谱来演示一下栅栏效应,以及通过补零来让那些被掩盖的细节显露出来。时域数据点1024点。
时域点数=1024点,FFT点数=1024,未补零
时域点数=1024点,FFT点数=4096,尾部补零
时域点数=1024点,FFT点数=16384,尾部补零
时域点数=1024点,FFT点数=65536,尾部补零
时域点数=1024点,FFT点数=131072,尾部补零
TestGuru
发表于 2016-5-28 21:52
上面分析用的三角窗数据:文本文件和WAV文件。
hcharlie
发表于 2016-5-29 09:16
LZ已经表示满意了。讲通俗点就够了。
我来讲点通俗的,讲点大家能看到的现象。
栅栏效应讲像隔着栅栏看东西,不讲频域,讲讲现象本身。
讲讲拍电影,就像时域上采样,一秒钟拍24帧,但每一帧只感光很短时间比如1ms,所以它拍的画面是离散的,跟人的视觉是有区别的,人有“视觉残留”,大概几十ms,所以人看到的是几十ms画面的“平均值”,与电影拍的“瞬时值”是有区别的,我们看电影的回放,是放的离散值,相当于隔着“时间的栅栏”看世界,是看的24个1ms的影像。所以看电影有时候跟看实物是有区别的,比如我们看转动电扇的叶片,直升机的转动的旋翼,人眼是看不清的,但在看电影看电视时,有时就能看到直升机的旋翼会静止不动,或者会缓慢正转,甚至反转,这也应该是栅栏效应的结果吧。
TestGuru
发表于 2016-5-29 10:34
本帖最后由 TestGuru 于 2016-5-29 10:36 编辑
其实有些人想说的,应该就是书本上那一句话,DTFT在频域是连续谱,DFT是在对DTFT的抽样,抽密点,被隐藏的谱线就露出来了,密到无穷大,就等于DTFT了,这貌似上过大学的都知道吧,但是初学者容易造成一个错误理解,就是没抽到的就完全被挡住了,看不到了。
好吧,我只踩我自己,我刚学数字信号处理时也这么理解,但是后来发现不完全是这么回事。一个N点的时域的采样序列,通过DFT转换到频域,没有丢失任何能量,就是说没有任何能量被挡在栅栏之外,在频域被抽到的点,其能量集中于一条谱线上,没抽到的点,通过泄漏方式展现自己的能量,或者说被抽到的频点,都正二八经地站在自己的位置上拍照,没被抽到的频点,靠偷影的方式把自己的身体分成几块以叠罗汉的方式挤到别人的位置上,最终能量守恒。唉,这个书本上又没这样说过.....{:{28}:}