Why use an anti aliasing filter with an energy meter.?

Started by hrh1818 July 27, 2009
The following document says on page 5 an anti aliasing filter is used
on the A/D input channels of an energy meter.  Warning the document is
31 pages long, only click on the link if you have a broadband internet
connection.  http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf.  My
understanding is you only need an anti aliasing filter if you want to
reconstruct the signal or find the frequency content of a signal.
Neither criteria apply to an energy meter.   Furthermore on page 6 the
document says the sampling frequency is 4096 samples per second which
gives a Nyquist frequency of 2048 cps.   But the anti aliasing filter
is a single pole filter with the *+3db point at approximately 5000
hz.

Hence I ask did the designers make a minor mistake or am I missing
something?

Howard
On 27 Jul, 22:44, hrh1818 <hr...@att.net> wrote:
> The following document says on page 5 an anti aliasing filter is used > on the A/D input channels of an energy meter. &#2013266080;Warning the document
is
> 31 pages long, only click on the link if you have a broadband internet > connection. &#2013266080;http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf.
&#2013266080;My
> understanding is you only need an anti aliasing filter if you want to > reconstruct the signal or find the frequency content of a signal. > Neither criteria apply to an energy meter. &#2013266080; Furthermore on page 6
the
> document says the sampling frequency is 4096 samples per second which > gives a Nyquist frequency of 2048 cps. &#2013266080; But the anti aliasing filter > is a single pole filter with the *+3db point at approximately 5000 > hz. > > Hence I ask did the designers make a minor mistake or am I missing > something?
One of the authors' names sounded very familier - with a bit of luck, you might get answers from one of the authors. Until then: This is a meter for a 3-phase system, that works on (presumably) 50 Hz or 60 Hz networks; maybe even at 400 Hz. These 3-phase systems often run high-power stuff, like motors and pumps, that are nonlinear and induce harmonics on the power grid. With that in mind, anti-alias filters are certainly needed in these kinds of systems. As for the details, the phrase "The anti-aliasing circuitry consisting of R26, R28, C20 and C22 follows the burden resistor." indicate to me that all these components are external to the IC, and that the values of these components are computed after a value for the burden resistor has been chosen. Rune
On Mon, 27 Jul 2009 13:44:52 -0700, hrh1818 wrote:

> The following document says on page 5 an anti aliasing filter is used on > the A/D input channels of an energy meter. Warning the document is 31 > pages long, only click on the link if you have a broadband internet > connection. http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf. My > understanding is you only need an anti aliasing filter if you want to > reconstruct the signal or find the frequency content of a signal. > Neither criteria apply to an energy meter. Furthermore on page 6 the > document says the sampling frequency is 4096 samples per second which > gives a Nyquist frequency of 2048 cps. But the anti aliasing filter is > a single pole filter with the *+3db point at approximately 5000 hz. > > Hence I ask did the designers make a minor mistake or am I missing > something?
You want to use an anti-aliasing filter if there is energy at frequencies that will alias down and give you trouble, or if your boss says you have to have one. In the case of an watt-hour meter like this one could conceivably have signal energy at frequencies that are disparate in continuous-time but which alias down to the same frequency in discrete- time, thereby causing an erroneous reading. Of course, one can always have a boss who thinks that _any_ sampled system has to have an anti-alias filter, lest the earth stop spinning. -- www.wescottdesign.com
Tim Wescott wrote:
> On Mon, 27 Jul 2009 13:44:52 -0700, hrh1818 wrote: > >> The following document says on page 5 an anti aliasing filter is used on >> the A/D input channels of an energy meter. Warning the document is 31 >> pages long, only click on the link if you have a broadband internet >> connection. http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf. My >> understanding is you only need an anti aliasing filter if you want to >> reconstruct the signal or find the frequency content of a signal. >> Neither criteria apply to an energy meter. Furthermore on page 6 the >> document says the sampling frequency is 4096 samples per second which >> gives a Nyquist frequency of 2048 cps. But the anti aliasing filter is >> a single pole filter with the *+3db point at approximately 5000 hz. >> >> Hence I ask did the designers make a minor mistake or am I missing >> something? > > You want to use an anti-aliasing filter if there is energy at frequencies > that will alias down and give you trouble, or if your boss says you have > to have one. In the case of an watt-hour meter like this one could > conceivably have signal energy at frequencies that are disparate in > continuous-time but which alias down to the same frequency in discrete- > time, thereby causing an erroneous reading. > > Of course, one can always have a boss who thinks that _any_ sampled > system has to have an anti-alias filter, lest the earth stop spinning.
Frequencies won't necessarily be at a meaningful phase after aliasing and can corrupt the reading. Jerry -- Engineering is the art of making what you want from things you can get. &macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;&macr;
On Jul 27, 4:44&#2013266080;pm, hrh1818 <hr...@att.net> wrote:
> The following document says on page 5 an anti aliasing filter is used > on the A/D input channels of an energy meter. &#2013266080;Warning the document
is
> 31 pages long, only click on the link if you have a broadband internet > connection. &#2013266080;http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf.
&#2013266080;My
> understanding is you only need an anti aliasing filter if you want to > reconstruct the signal or find the frequency content of a signal. > Neither criteria apply to an energy meter. &#2013266080; Furthermore on page 6
the
> document says the sampling frequency is 4096 samples per second which > gives a Nyquist frequency of 2048 cps. &#2013266080; But the anti aliasing filter > is a single pole filter with the *+3db point at approximately 5000 > hz.
this is the analog filter preceding the A/D, right? just curious, what is the order of the filter (or the slope of its transition band)?
> Hence I ask did the designers make a minor mistake or am I missing > something?
in the line of Tim's observation, it would only be "right" if it was known that there would be little energy in the A/D input spectrum between about 2 and 5 kHz. if the order of the filters is high and the transition band steep, it seems like it should be little more costly to bump the cutoff frequency down a little and put in a better anti-aliasing filter. seems to me, there is essentially no specific anti-aliasing filter. just an anti-interference filter to block stuff above 5 kHz. even though it's undersampled, the result of undersampling is folded spectrum, but if the phases in the folded spectrum are sorta random, this energy or power meter measures the energy in the baseband up to 5 kHz, even if it is sampling at less than that. i dunno. might be crude, but that might be what they're doing. still seems dum. r b-j r b-j
On 28 Jul, 00:57, Tim Wescott <t...@seemywebsite.com> wrote:

> > Hence I ask did the designers make a minor mistake or am I missing > > something? > > You want to use an anti-aliasing filter if there is energy at frequencies > that will alias down and give you trouble,
I've seen these kinds of readings in use, in systems where there were lots of harmonics on the line. Users do want anti-alias filters at the input. I suppose vendors want such filters to be configurable by users. Rune
>The following document says on page 5 an anti aliasing filter is used >on the A/D input channels of an energy meter. Warning the document is >31 pages long, only click on the link if you have a broadband internet >connection. http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf. My >understanding is you only need an anti aliasing filter if you want to >reconstruct the signal or find the frequency content of a signal. >Neither criteria apply to an energy meter. Furthermore on page 6 the >document says the sampling frequency is 4096 samples per second which >gives a Nyquist frequency of 2048 cps. But the anti aliasing filter >is a single pole filter with the *+3db point at approximately 5000 >hz. > >Hence I ask did the designers make a minor mistake or am I missing >something? > >Howard
This is a good question, which deserves an answer that drags on a bit..... The term anti-aliasing is being incorrectly applied here. Its one of those situations where one person started using an inappropriate term, and everyone else copied. The filter is really there to suppress wideband noise, which could badly mess up the operation of the sigma delta converter. If you look at competing solutions, they all use a similar circuit. For example in http://www.analog.com/static/imported-files/data_sheets/ADE7758.pdf figure 34 you will see the same time constant used with a slightly different RC configuration. Their circuit will be more fussy about board layout, which is why the TI one is a little different. If you look at applications information for other energy measurement devices you will see very similar single pole RC filters, with time constants not a million miles away. So, why do all these devices come with a recommendation to use this kind of filter? Well, its a compromise that works well. Aliases in the current and voltage waveforms matter far less than others have suggested. What matters very much is wider band noise energy. We need to suppress that, and do so without upsetting performance in the band of interest. The band of interest varies somewhat between power utilities, but they generally expect to capture at least the first 20 harmonics well. The gain and phase of the voltage and current samplers need to be very well controlled across the band of interest. In audio, 3dB is not a big deal. In comms, 0.2dB is not a big deal. In energy measurement we look for accuracies of 0.1% or better, which is about 0.009dB. So, really flat gain response is required. In audio a few degrees of phase is seldom a big deal. In comms a degree of phase is seldom a big deal. In energy measurement, at 0.5PF, 0.03 degrees of phase error gives a 0.1% measurement error. Really well controlled phase is required. You can get energy in the gigahertz range on a power line in an industrial area. Industrial systems using big arcs, or huge amounts of RF for heating, produce huge amounts of wideband energy on the power lines in the surrounding area. That could really mess up the operation of energy measurement devices, or perhaps even damage them. Really wideband noise needs separately filtering from each of the differential inputs, because the differential performance of the circuit starts to break down so far from its operating band. This is the purpose of the two small capacitors. The bigger capacitor filters medium frequency noise, where the differential functionality of the input does perform well. On a perfect board only two 33nF capacitors to ground would be needed. However, the three capacitor design is much more tolerant. Imbalance between capacitor values (which can be quite wide tolerance) will not inject unbalanced amounts of ground plane noise at the differential inputs. Capacitors not grounded at exactly the same point on the ground plane will also not inject inject differential amounts of ground plane noise. The differential RC filter fits the need for accurate phase very nicely. If the filter is well balanced the phase shifts in the two legs cancel out. Such balance is naturally achieved in the current sensor circuit, as long as the component tolerances aren't too wide. In the voltage sensor circuit we need to take care, because the impedance into one leg is zero (a ground connection) and the impedance into the other leg is about 1k ohms (the voltage divider). If we don't tweaks the resistors to get close to a zero phase shift we get two bad consequences. One is a phase shift that varies with frequency, so higher harmonics are not handled correctly. The other is a phase shift that varies with temperature. Capacitors generally have a horrible temperature characteristic, and in a consumer grade meter we don't want to use exotic devices. If there is a substantial phase shift, there will also be a substantial variation with temperature and the meter's accuracy will vary considerably with temperature for poor power factor loads. Others raised the issue that not suppressing frequencies >Fs/2 will lead to measurement error. This is not really true. There is nothing inherently out of band, which it is appropriate to filter away. Applying an anti-alias filter has as much potential to corrupt the results as not applying one. Let's look at the real measurement requirement.... In essence there should be a pure sine wave voltage signal, and a current signal somewhere between pure sine wave and massively polluted with harmonics. In practice, the utility's cables have some impedance, so the voltage signal may have some harmonic content. In the real world a typical residential area voltage signal is quite pure, and an industrial area signal may have 10% harmonic content. 15% would be pretty bad for even a heavy industrial area. The current waveform for an incandescent lamp load will be a pure sine wave, and the waveform with a half wave load will be horrible. >40% harmonic content is common for a current signal. If the voltage signal is pure it won't matter how much harmonic content there is in the current signal, or which harmonics they are. It also won't matter if the harmonics are aliased. They just won't correlate with a pure voltage signal, and affect the energy measurement result. The interesting case is when the voltage signal has harmonic content, as this may correlate with harmonic content in the current signal. If there are actual harmonic components beyond Fs/2 (rather than interference or noise), and they are filtered away by a true anti-alias filter, they won't contribute to the energy measurement as they should. So, anti-alias filtering corrupts the result. If these harmonics are not filtered away, they will fold back into the band. They will fold back similarly in both the voltage and current signal, will correlate, and will contribute to the energy measurement. Unfortunately, they won't have the correct relative phases, so they will contribute at the wrong effective power factor. Does it matter very much? The lower harmonics are always far bigger in amplitude, so the potential error is not that big, whatever we do about the aliases. The bottom line is we are really no worse off processing the aliases than we would be if we had an efficient anti-alias filter. So, if we try to produce a real anti-alias filter we have some very tough gain and phase constraints to meet if we are not to significantly affect the most important harmonics - the fundamental, and 3rd to 11th. If we ignore the aliases, which start at the 40th harmonic for a 50Hz supply and 4096 per second sampling, they aren't usually big enough to make a whole lot of difference. Rune mentioned 400Hz power. I should point out that the MSP430F47197 is designed for 50Hz and 60Hz energy measurement in high volume light consumer applications. 400Hz energy meters demand a high sampling rate, and are quite specialised products. I don't know of any single chip solutions targeted at that application. I hope that sheds some light on the issues. Steve
On Jul 28, 4:24=A0am, "steveu" <ste...@coppice.org> wrote:
> >The following document says on page 5 an anti aliasing filter is used > >on the A/D input channels of an energy meter. =A0Warning the document is > >31 pages long, only click on the link if you have a broadband internet > >connection. =A0http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf. =A0My > >understanding is you only need an anti aliasing filter if you want to > >reconstruct the signal or find the frequency content of a signal. > >Neither criteria apply to an energy meter. =A0 Furthermore on page 6 the > >document says the sampling frequency is 4096 samples per second which > >gives a Nyquist frequency of 2048 cps. =A0 But the anti aliasing filter > >is a single pole filter with the *+3db point at approximately 5000 > >hz. > > >Hence I ask did the designers make a minor mistake or am I missing > >something? > > >Howard > > This is a good question, which deserves an answer that drags on a > bit..... > > The term anti-aliasing is being incorrectly applied here. Its one of thos=
e
> situations where one person started using an inappropriate term, and > everyone else copied. The filter is really there to suppress wideband > noise, which could badly mess up the operation of the sigma delta > converter. > > If you look at competing solutions, they all use a similar circuit. For > example inhttp://www.analog.com/static/imported-files/data_sheets/ADE7758=
.pdffigure
> 34 you will see the same time constant used with a slightly different RC > configuration. Their circuit will be more fussy about board layout, which > is why the TI one is a little different. If you look at applications > information for other energy measurement devices you will see very simila=
r
> single pole RC filters, with time constants not a million miles away. > > So, why do all these devices come with a recommendation to use this kind > of filter? Well, its a compromise that works well. Aliases in the current > and voltage waveforms matter far less than others have suggested. What > matters very much is wider band noise energy. We need to suppress that, a=
nd
> do so without upsetting performance in the band of interest. The band of > interest varies somewhat between power utilities, but they generally expe=
ct
> to capture at least the first 20 harmonics well. > > The gain and phase of the voltage and current samplers need to be very > well controlled across the band of interest. In audio, 3dB is not a big > deal. In comms, 0.2dB is not a big deal. In energy measurement we look fo=
r
> accuracies of 0.1% or better, which is about 0.009dB. So, really flat gai=
n
> response is required. In audio a few degrees of phase is seldom a big dea=
l.
> In comms a degree of phase is seldom a big deal. In energy measurement, a=
t
> 0.5PF, 0.03 degrees of phase error gives a 0.1% measurement error. Really > well controlled phase is required. > > You can get energy in the gigahertz range on a power line in an industria=
l
> area. Industrial systems using big arcs, or huge amounts of RF for heatin=
g,
> produce huge amounts of wideband energy on the power lines in the > surrounding area. That could really mess up the operation of energy > measurement devices, or perhaps even damage them. Really wideband noise > needs separately filtering from each of the differential inputs, because > the differential performance of the circuit starts to break down so far > from its operating band. This is the purpose of the two small capacitors. > The bigger capacitor filters medium frequency noise, where the differenti=
al
> functionality of the input does perform well. On a perfect board only two > 33nF capacitors to ground would be needed. However, the three capacitor > design is much more tolerant. Imbalance between capacitor values (which c=
an
> be quite wide tolerance) will not inject unbalanced amounts of ground pla=
ne
> noise at the differential inputs. Capacitors not grounded at exactly the > same point on the ground plane will also not inject inject differential > amounts of ground plane noise. > > The differential RC filter fits the need for accurate phase very nicely. > If the filter is well balanced the phase shifts in the two legs cancel ou=
t.
> Such balance is naturally achieved in the current sensor circuit, as long > as the component tolerances aren't too wide. In the voltage sensor circui=
t
> we need to take care, because the impedance into one leg is zero (a groun=
d
> connection) and the impedance into the other leg is about 1k ohms (the > voltage divider). If we don't tweaks the resistors to get close to a zero > phase shift we get two bad consequences. One is a phase shift that varies > with frequency, so higher harmonics are not handled correctly. The other =
is
> a phase shift that varies with temperature. Capacitors generally have a > horrible temperature characteristic, and in a consumer grade meter we don=
't
> want to use exotic devices. If there is a substantial phase shift, there > will also be a substantial variation with temperature and the meter's > accuracy will vary considerably with temperature for poor power factor > loads. > > Others raised the issue that not suppressing frequencies >Fs/2 will lead > to measurement error. This is not really true. There is nothing inherentl=
y
> out of band, which it is appropriate to filter away. Applying an anti-ali=
as
> filter has as much potential to corrupt the results as not applying one. > Let's look at the real measurement requirement.... > > In essence there should be a pure sine wave voltage signal, and a current > signal somewhere between pure sine wave and massively polluted with > harmonics. In practice, the utility's cables have some impedance, so the > voltage signal may have some harmonic content. In the real world a typica=
l
> residential area voltage signal is quite pure, and an industrial area > signal may have 10% harmonic content. 15% would be pretty bad for even a > heavy industrial area. The current waveform for an incandescent lamp load > will be a pure sine wave, and the waveform with a half wave load will be > horrible. >40% harmonic content is common for a current signal. > > If the voltage signal is pure it won't matter how much harmonic content > there is in the current signal, or which harmonics they are. It also won'=
t
> matter if the harmonics are aliased. They just won't correlate with a pur=
e
> voltage signal, and affect the energy measurement result. The interesting > case is when the voltage signal has harmonic content, as this may correla=
te
> with harmonic content in the current signal. If there are actual harmonic > components beyond Fs/2 (rather than interference or noise), and they are > filtered away by a true anti-alias filter, they won't contribute to the > energy measurement as they should. So, anti-alias filtering corrupts the > result. If these harmonics are not filtered away, they will fold back int=
o
> the band. They will fold back similarly in both the voltage and current > signal, will correlate, and will contribute to the energy measurement. > Unfortunately, they won't have the correct relative phases, so they will > contribute at the wrong effective power factor. Does it matter very much? > The lower harmonics are always far bigger in amplitude, so the potential > error is not that big, whatever we do about the aliases. The bottom line =
is
> we are really no worse off processing the aliases than we would be if we > had an efficient anti-alias filter. > > So, if we try to produce a real anti-alias filter we have some very tough > gain and phase constraints to meet if we are not to significantly affect > the most important harmonics - the fundamental, and 3rd to 11th. If we > ignore the aliases, which start at the 40th harmonic for a 50Hz supply an=
d
> 4096 per second sampling, they aren't usually big enough to make a whole > lot of difference. > > Rune mentioned 400Hz power. I should point out that the MSP430F47197 is > designed for 50Hz and 60Hz energy measurement in high volume light consum=
er
> applications. 400Hz energy meters demand a high sampling rate, and are > quite specialised products. I don't know of any single chip solutions > targeted at that application. > > I hope that sheds some light on the issues. > > Steve
Steve, thank you for taking the time for providing a detailed response to my question. I conclude the filters should really be called power line noise rejection filters instead of of anti aliasing filters. Your response is highly appreciated. Howard
hrh1818 wrote:

> The following document says on page 5 an anti aliasing filter is used > on the A/D input channels of an energy meter. Warning the document is > 31 pages long, only click on the link if you have a broadband internet > connection. http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf. My > understanding is you only need an anti aliasing filter if you want to > reconstruct the signal or find the frequency content of a signal. > Neither criteria apply to an energy meter. Furthermore on page 6 the > document says the sampling frequency is 4096 samples per second which > gives a Nyquist frequency of 2048 cps. But the anti aliasing filter > is a single pole filter with the *+3db point at approximately 5000 > hz. > > Hence I ask did the designers make a minor mistake or am I missing > something?
Defenitly yes! You should have a look how a sigma delta ADC is working! The OSR (output sample rate) of 4096 samples per second is not the sample rate of the modulator. Simply spoken you have a 1 bit ADC with a samplerate of a multiple (which is called the oversampling ratio) which is typically *32, *64, *128, *256 ... So your analog sampling rate is much higher. I have not looked in detail on the descripion of the MSP (slau056h.pdf), but the maximum modulator frequency (= analog sampling rate) is about 1 MHz. For this frequency you need the analog anti-aliasing filter!. Converting the 1bit-datastreem (1Mbit/sec) into a slower 16-bit datastreem is done via a digital low pass filter. This filter should removes the frequencys above the digital OSR / 2 (4096 Hz / 2). The digital filter of a sigma delta ADC is what defines most of the performance of a sigma delta ADC. The passband is not flat and the attanuation in the stopband may be not big enough. The passband ripple is typically between +-0.1 .. 0.005 db and this counts if you want to make precise measurments. Hope this helps to understand the principle.
> > Howard
-- Freundliche Gr&#2013266172;sse -- Regards Friedrich
On Jul 29, 2:09&#2013266080;am, Friedrich Seuhs <friedrich.se...@hasos.com> wrote:
> hrh1818 wrote: > > The following document says on page 5 an anti aliasing filter is used > > on the A/D input channels of an energy meter. &#2013266080;Warning the document
is
> > 31 pages long, only click on the link if you have a broadband internet > > connection. &#2013266080;http://focus.ti.com/lit/an/slaa409a/slaa409a.pdf.
&#2013266080;My
> > understanding is you only need an anti aliasing filter if you want to > > reconstruct the signal or find the frequency content of a signal. > > Neither criteria apply to an energy meter. &#2013266080; Furthermore on page 6
the
> > document says the sampling frequency is 4096 samples per second which > > gives a Nyquist frequency of 2048 cps. &#2013266080; But the anti aliasing
filter
> > is a single pole filter with the *+3db point at approximately 5000 > > hz. > > > Hence I ask did the designers make a minor mistake or am I missing > > something? > > Defenitly yes! > > You should have a look how a sigma delta ADC is working! > > The OSR (output sample rate) of 4096 samples per second is not the sample > rate of the modulator. Simply spoken you have a 1 bit ADC with a samplerate > of a multiple (which is called the oversampling ratio) which is typically > *32, *64, *128, *256 ... So your analog sampling rate is much higher. > I have not looked in detail on the descripion of the MSP (slau056h.pdf), but > the maximum modulator frequency (= analog sampling rate) is about 1 MHz. > For this frequency you need the analog anti-aliasing filter!. Converting > the 1bit-datastreem (1Mbit/sec) into a slower 16-bit datastreem is done via > a digital low pass filter. This filter should removes the frequencys above > the digital OSR / 2 (4096 Hz / 2). > The digital filter of a sigma delta ADC is what defines most of the > performance of a sigma delta ADC. The passband is not flat and the > attanuation in the stopband may be not big enough. The passband ripple is > typically between +-0.1 .. 0.005 db and this counts if you want to make > precise measurments. > > Hope this helps to understand the principle. > > > > > Howard > > -- > Freundliche Gr&#2013266172;sse -- Regards > Friedrich
Friedrich, your response is very helpful in helping me understand why an an anti aliasing filter is used in this application. However, your use of the term OSR confuses me. The information I found uses the term OSR to mean over sampling ratio not output sampling rate. For example check the graph on page 54 of http://focus.ti.com/lit/ds/slas626a/slas626a.pdf. Warning this document is 82 pages long. Notice in the graph the best SINAD is at a high OSR. Whereas if OSR meant output sampling rate then the best SINAD would be at a low output sample rate. Howard