Introduction
(View more DSP information or download the preliminary implementation.)
It is common to have a number of analog signals, sampled at
moderate rates, where time relationships between channels are
important. For example, in beamforming applications with a
microphone array, time shifts are related to the location of the
audio source, which is the critical information. You don't want
those time shifts confused with artificial time delays caused by
discrete time sampling. A direct way to address this problem is
to use simultaneous sampling hardware. This works well, avoiding
the time and phase anomalies. However, simultaneous sampling
hardware comes at a price, because electronic components must be
dedicated to each signal line. Sometimes economics dictate a
search for a less expensive alternative.
Data acquisition devices that can multiplex their conversion
electronics provide a much more economical approach to data
sampling, but these devices must operate serially. There is an
incremental time-shift as each data channel is sampled, but this
shift is very precise and predictable. Within certain limitations, a
straightforward application of established digital processing
techniques can provide an accurate and economical way to correct for
the time and phase shifts introduced into the data by the sampling
process.
Time-Shift Corrections Using the MTSFILT Processing Command
A processing command to apply time-shift filtering to the
practical problem of aligning data samples skewed in time by
multiplexed sampling is available. The MTSFILT processing command is a
Multiple channel Time-Shift Filtering command that acts
upon multiple data channels to produce sample streams in which the
time delays inherent in multiplexed sampling are eliminated. The
command has the following form.
MTSFILT( inpipe, nchan, xover, outpipe )
where the parameters are
- inpipe
- a pipe containing multiplexed word data samples
- nchan
- the number of multiplexed channels in the data stream
- xover
- the oversampling present in the data set to reduce
the time-shift between channels
- outpipe
- a pipe where the multiplexed, reconstructed word
data samples are placed
Each input signal from inpipe must be at least 2x
oversampled; that is, the effective sampling rate for each channel
(sampling rate / the number of multiplexed channels) should be at
least 4 times the highest frequency present in the signal at the
input sample rate. This restriction results from the properties
of the lowpass filtering used to build the interpolation filters.
The data sampling rate can be increased beyond what is strictly
necessary to operate the time-shift filtering, to reduce the time
separation between physical samples, thus reducing the magnitude of
the corrections needed, and reducing sensitivity to certain
practical concerns such as disturbances and noise. The
xover parameter is a positive "oversampling factor"
specifying the amount of additional oversampling present in the
signal. The output pipe <outpipe> receives the multiplexed results
of the time-shift filtering at the desired rate after disposing of
the extra samples.
Example: Suppose that a value is
needed from each of 4 channels in 25 milliseconds. That is 40
microseconds to cover the four channels, or a sampling interval of
10 microseconds between data channels. A Data Acquisition Processor
has plenty of capacity to sample faster than that. So the value of
xover can be set to 4, and the sampling interval set to
2.5 microseconds.
Test
A configuration with five data channels requires simultaneous
samples at 400 microsecond intervals, with the full theoretical
Nyquist band preserved at this sample rate. Apply 4x oversampling;
at this rate, the sampled data stream well below the theoretical
band limits, and sampling is 100 microseconds on each channel.
Because there are five channels, the multiplexing sample interval
must be set to 20 microseconds. After capturing the multiplexed
data, the data stream is sent to the MTSFILT command for time-shift
correction. The reconstructed simultaneous-time data stream, with
samples at 400 microsecond intervals on each channel, is transferred
to the PC for logging.
The MTSFILT command was tested with a randomized signal
appropriately filtered to enforce the bandlimit restrictions.
The same signal is sampled on all channels, so if the sampling
were simultaneous, the values captured in each channel should
be the same. When multiplexed sampling is used, it is easy
to see the progression of signal changes from one channel to
the next.
After processing by the MTSFILT processing command, the
sample values should be all the same. Due to the limitations
of the filter design, there can be small variations, but these
are at the level of bit chatter in the low order bit of
the analog to digital converter hardware.
You can verify this test in the downloadable configuration for
DAP Measurement Studio. After
time-shift correction, the values of all samples are
indistinguishable from independent simultaneous measurements.
How Does Time-Shift Filtering Work?
Transversal or "shift register" filters are completely natural
for time shifting. A finite impulse response (FIR) filter is a
practical implementation of this idea with a finite length array for
the filter coefficients. Consider for example a degenerate FIR
filter characterized by the following set of coefficients:
( 0.0 0.0 1.0 0.0 0.0 )
When this set of coefficients is convolved with a data stream
to evaluate the filter, the center coefficient is aligned at the
point of interest in the data stream, terms of the data stream
and filter characteristic are multiplied pairwise, and the
resulting products are summed to produce the filtered output.
Clearly, the only term that has any effect is the one aligned
to the 1.0 coefficient, so the new data stream is unchanged by
the filtering operation.
Consider now what happens if the 1.0 value is swapped with the
0.0 value on its left side. Now the only term that aligns with a
nonzero coefficient is one position older. The filter result is
almost the same, except that the filter output sequence matches
the original stream one sample displaced into the past. In a
similar fashion, swapping the 1.0 value with the 0.0 value on its
right side shifts the data stream so that output terms appear one
position sooner. Shifting the filter coefficients has the effect of
time-shifting the output data stream.
This works perfectly, provided that the desired time shift
matches an exact integer number of sample positions. What if it
does not?
Non-Integer Shifts: Signal Reconstruction and Interpolation Filters
Let's quickly review some basic theory. Shannon, Nyquist, and
other pioneers of digital signal processing considered the problem
of determining when a set of discrete samples of a signal is
sufficient to completely characterize the original analog signal.
The first fundamental and in some ways astonishing result is:
An analog signal can be perfectly and
completely reconstructed, anywhere, from its discrete-time
samples provided that the original signal is suitably
band-limited.
That restriction is subtle. Usually the bandlimit condition
cannot be guaranteed. But even so, where the violations are
small, deviations from theory are small and the results
are very good, as good as the original samples.
The problem of signal reconstruction is closely related to the
"interpolation problem" of classical numerical analysis — to
determine the approximate value of a function at any point given the
function values at a discrete set of locations. For this
reason, the filters for reconstructing an analog signal are also
called interpolation filters.
A second extremely important theoretical result is the explicit
specification of a digital filter form to perform the reconstruction
perfectly. Omitting the details [1], we can
simply point out:
the theoretically ideal
reconstruction filter to rebuild the original signal value at any
point in time is a transversal shift register filter.
It would in fact be a FIR filter except for the annoying
difficulty that it is infinite. A filter with an infinite number of
terms is not the sort of thing you want to evaluate in real time.
Most of the terms are very small, and that suggests truncating the
infinite theoretical filter to a finite length as an approximation.
This must be done with care, as dropping an infinite number of terms
can have unexpected side effects.
The reconstruction problem becomes much easier if we insist that
the original signal is bandlimited well below theoretical Nyquist
limit. Suppose that the frequency band for the original system is
restricted as shown in figure 1, and then a lowpass filter as
shown in figure 2 is applied. What is the result?
The filter has no effect, because the frequency bands that it
attenuates are already completely empty. As long as this property is
satisfied, other design details are arbitrary, so pick a symmetric
FIR filter with a small number of filter coefficients and a flat,
accurate passband.
Now pick a desired time location for reconstructing the signal.
Construct the ideal interpolation filter according to theory, and
apply this to the discrete signal — after the change-nothing
lowpass filtering is applied. Now, the two filter operations,
lowpass and interpolation, can be combined mathematically. The
combination yields a filter form that has the interpolation
performance of the theoretical reconstruction filter when operating
upon the restricted-band signals, and can be truncated to a finite
length with very little loss of accuracy.
Knowing the conditions under which efficient FIR filters can
perform the signal reconstruction at any desired point in time, we
are not forced to start from the general theory for designing
interpolation filters. Any method that produces a good approximation
to the theoretical ideal will also work well. The
accompanying worksheet [2] shows a simple design method based on
least squares best approximation, and this can generate practical
interpolation filter designs for any desired sample location.
From Interpolation to Time Shift
Time shifting is simply a matter of selecting the position
in a time sequence where the interpolation filtering is applied.
By selecting an appropriately staggered time-shift for each data
channel, the values of the reconstructed samples can be made to
align in time, correcting the artificial time delays caused by
sampling.
To summarize briefly, the values produced by the interpolation
filter can be indistinguishable from values obtained by physical
simultaneous sampling at those instants. However, suitable
applications must :
- make sure that the signal is clean and properly band limited,
- oversample so that the rate is not too close to the theoretical
Nyquist limit,
- allow for processing delays as filtering is performed,
- not expect accurate reconstructions near signal discontinuities.
The World's Most Terrifying DSP Operation
It is a well known fact that common digital filtering operations
such as Butterworth IIR filtering [3] can
replace the original signal samples with new sample values at the same locations [4] such that the signal is improved [5] but in all other respects
there is minimal disturbance. [6] Anybody who
just finished the previous sentence without flinching should check
the footnotes right now to find out how it is
wrong [3], wrong[4], wrong[5], and (not
surprisingly) wrong[6].
In contrast to Butterworth-like filters, which distort data
in a manner that is seldom well understood, the interpolation
filters have a sound theoretical basis, are extremely
predictable, and perform efficiently.
So please explain. Why is Butterworth-like digital filtering,
despite all of its known deficiencies, applied almost without a
second thought; while time-shift filtering (without those
deficiencies) remains too terrifying to consider?
We certainly are baffled. But time shifting remains perhaps
the most under-utilized DSP technique.
Conclusions
The MTSFILT command makes time-shift corrections easy to
apply. It can't replace true simultaneous sampling, but
sometimes there is good economy possible from using the digital
time-shift filters with lower-cost multiplexed sampling devices.
If the limitations are respected and the command is used properly,
the results are indistinguishable from measurements captured by
actual simultaneous sampling. In some cases, with data corrupted
by low levels of random noise, despite the fact that this
violates the prerequisite conditions, measurements are actually
improved due to the noise reduction inherent in the filtering.
There should be no terror in applying well-suited and
well-established DSP technology properly. Be very afraid about
too casually applying other limited technologies presumed to
be "safe" for no particularly good reason.
MTSFILT uses established DSP technology, not anything new
and experimental. The only thing new is an improved package that
makes it very easy to use. This demonstrates that "applying DSP"
does require staying mired in mathematical theory.
If you were looking for the benefits of simultaneous sampling
without the price tag, consider the MTSFILT command for your
project.
Footnotes and References:
- For the full details, you can try
Digital Signal Processing, Alan V. Oppenheim and
Ronald W. Schafer, Prentice Hall Inc., 1975, section 1.7.
Any other good textbook on digital signal processing will
cover this fundamental material.
-
A copy of a Mathcad 12 worksheet for time-shift filter designs
using least squared methods is provided with the command download.
(Mathcad is a registered trade name of Mathsoft
Engineering and Education, Inc.) An example application of
the worksheet is shown in the HTML snapshot, and you can grab all of the formulas there.
-
While Butterworth filters are certainly IIR filters, they are
continuous-time filters and will always remain so. Some discrete
time filters, notably the "max-flat designs" share some of the same
properties. It is possible to map analog designs into discrete time
IIR designs, but there are limits to what can be achieved this way.
-
The same locations? Well, not really. The shifts produced by
Butterworth filters are far more drastic than you might think,
and because these are a nonlinear function of frequency, you must
pay the price of waveform distortions.
-
Filtering can never add information to a signal. Sometimes
Butterworth filters "improve" the signal in the sense that the
results "look smoother" and have attenuation where you need
it. But it's not all improvement, if you take the distortions into
account.
-
Don't be fooled. The only thing that a Butterworth-like
filter minimizes is passband flatness in the neighborhood of zero
frequency. The rest is a compromise.
Return to the DSP page. Or download the preliminary implementation.
|