Does the "shape" of an FM bandpass filter matter?

[SoftwareSamurai]

Since I'm not an analog EE expert, I'd like to ask anyone here who might be...

I'm designing a bandpass filter that goes in front of a tank circuit for decoding an FM signal. Considering there are tons of filter designs to choose from, each with their own unique frequency response shape, I'm wondering:

Given an FM signal input, does the "shape" of a bandpass filter's frequency response have any significant affect on the decoded signal?

I'm talking about the band pass area specifically. Does it need to be flat? Or can it have up to a +/- 1.5 db ripple and still work effectively for the tank circuit to lock onto the signal? What if it has a "dip" in the middle of the bandpass (like two camel humps)?

My educated guess is that as long as the lower and upper cutoff frequencies are within spec for the desired FM signal, and as long as whatever ripple is "inside" the bandpass is no more than +/- 1.5 db, the tank circuit will lock onto the carrier and the decoded signal will be just fine. Am I missing anything here?

[amspire]

My educated guess is that what is important is reasonably smooth group delay, otherwise you will get distortion.

I guess it depends on the actual demodulator you are using, but the flatness of the passband should not matter too much. I would think 3db flatness would be OK.

Also you obviously need to meet the stop band requirements to prevent interference.

So I would guess that you don't want designs like the Tchebychev filter with lousy delay characteristics.  Bessel has flat group delay, but is lousy in the stop band. So the best compromise may well be the humble Butterworth filter which doesn't have a flat delay, but it is pretty smooth at least.

If there are any RF designers, it would be great to hear what they say.

Richard.

[vk6zgo]

I think you have the wrong end of the stick with the "tank circuit".

A "tank"circuit is Radio people's slang for a resonant LC circuit.
It normally refers to the circuit which is the plate load circuit for a vacuum tube amplifier.
A resonant LC circuit doesn't "lock" onto anything,it just resonates,that's all!

So what is your "decoder" circuit,a PLL,a quadrature detector,a ratio detector,Foster Seeley discriminator,or what?
If you are feeding a resonant LC circuit with a bandpass filter,you will find that the resonant circuit  will be the dominant influence in determining the bandwidth & shape of the resulting passband.

VK6ZGO

[SoftwareSamurai]

I think you have the wrong end of the stick with the "tank circuit".

Forgive me for not being clear.

The chip I'm using has three stages to it. The first stage is an amp for an antenna. The second stage is a limiter. The third stage is an FM demux.

The BPF goes between stages 1 & 2. The tank goes between stages 2 & 3.

Although the tank does have a resonant f, it's recommended that a BPF be used to help isolate the carrier f that you wish to demux. (The tank's Q is understandably low. Otherwise it would be difficult to, what I've probably inaccurately called, "lock on" to the carrier, or more precisely, feed the FM demux the proper resonant feedback in order to properly demodulate the FM signal.)

[Zero999]

You're not getting very helpful replies because you've not given enough information.

What IC are you using? Can you post the schematic or a datasheet?

[SoftwareSamurai]

What IC are you using? Can you post the schematic or a datasheet?

It's the TSH511.

So does the attenuation "shape" of the bandpass filter matter when it's used to feed an FM demux tank circuit?
Or is the group delay "shape" more important in this case?

[amspire]

The reason why I voted for Group Delay, is imagine you have a sawtooth waveform encoded as FM.

Also imaging the filter group delay had  spikes at, say, 5Kz and 10kHz that doubled the group delay.

Then as the increasing frequency of the leading edge of the sawtooth passes through 5kHz, is gets delayed a bit compared to other frequencies causing a kink in the decoded waveform. Same thing would happen at 10kHz.

Is it enough of an effect to worry about? Unless I designed a filter and did the maths, I have no idea.

As far as amplitude, the decoder includes a limiter which will greatly reduce any AM modulation in the signal caused by either varying RF signal strength, or frequency response issues of the filter.

Richard

[vk6zgo]

A very clever IC!

Are you using  a carrier on 2.3 MHz AND another on 2.8 MHz,as per the  suggested circuit?

Are you modulating the original FM carrier/s with audio,or data of some kind?

The width of your filters will depend on your maximum deviation.

For audio, reasonably flat amplitude response should be all you need.
Phase Response & Group Delay are normally not a problem in that type of service,BUT if you are using both channels,it is going to be difficult to get a nice flat response across the bandwidth you need ,& the steep skirts you need to avoid  crosstalk from the other channel,remembering that both carriers are present at the filter input.

LC filters that are sharp enough to avoid the other carrier may be too narrow at the required frequency.

From the App Note,I would guess that the filters normally used are ceramic or,maybe even SAW filters rather than LC for this reason.

If you are only using one channel,the requirements are considerably less onerous.
Why not build up a couple of filters & try them?

VK6ZGO

[amspire]

It looks like group delay itself is not a source for distortion but group delay symetry is more important then amplitude symmetry in FM for low distortion.

It also is not at all easy to get both amplitude symmetry and group delay symmetry.

See page 510 of this preview:

http://dc316.4shared.com/doc/FDlmsKyd/preview.html

Agilent Genesys S/Filter seems to be a great program for designing filters with both amplitude and group delay symmetry.

http://cp.literature.agilent.com/litweb/pdf/genesys200801/examples/filters/sfilter/response_symmetry_example.htm

A common way to design a filter is to design a low pass filter and transform it to a bandpass filter. What the link above says is that none of the common lowpass to bandpass transforms give good symmetry.

Here is the problem. (Sorry to get technical). When you transform a low pass filter to a bandpass filter using simple transforms, it is symmetrical if plotted on a logarithmic frequency axis.  So lets take an example of a really wide band filter.  Center frequency is 1MHz and the upper -3db point is 2MHz. The other -3db point with simple transforms is 500KHz. Now the FM deviation around 1MHz is linear, but the filter is not - 1Mhz width in one direction and 500KHz in the other.

Same happens with group delay symmetry.  So you need a different kind of transform that gives symmetry. That is what S/FILTER can do.

Now how good a filter do you actually need? You may not need great distortion figures, or you may be able to allow for a generous filter bandwidth versus the FM deviation.

If you can find a good filter design, but it is the wrong frequency, it is possible to rescale the design to the frequency you want.  Might be the easiest way.

Or just buy an IF filter module.

Richard.

[SoftwareSamurai]

Are you using  a carrier on 2.3 MHz AND another on 2.8 MHz,as per the  suggested circuit?

Those, and more.

Are you modulating the original FM carrier/s with audio,or data of some kind?

No extra modulation. Just what the companion transmitter (TSH512) chip does.

From the App Note,I would guess that the filters normally used are ceramic or,maybe even SAW filters rather than LC for this reason.

Well a simple LC filter wouldn't have a high enough Q, IMHO.
(Although the ST engineers claim a low Q is fine since the "tuning" range of the tank is usually limited to begin with.)

Agilent Genesys S/Filter seems to be a great program for designing filters with both amplitude and group delay symmetry.

Sure! Only a mere $10,000 software package! If you get everything, it's more like $30,000! (Where's the animated smiley icon that coughs up money?)
Heck, if I could afford that, I'd just hire a good analog EE to design the filters for me - or even finish my project while I paint my house!

But seriously - Genesys does look like a very good software package. And I did manage to sweet-talk my way into an eval license for it. So I've got 30 days to play with it and get some good designs out of it. (Although I did manage to crash it within 10 minutes of starting it for the first time. Hmmm...)

What's really fun is to mock up the same filter in Genesys and LTSpice, and try to get them to agree with each other.

Now how good a filter do you actually need? You may not need great distortion figures, or you may be able to allow for a generous filter bandwidth versus the FM deviation.

That...is a very good question!

My hunch is that all I really need is a filter with a nearly-flat delay group inside the FM deviation band, and at least a -3db drop outside the passband. So yes, the passband can probably be rather large in comparison, as long as the cutoffs filter out the carriers above and below the one it's passing through.

Or just buy an IF filter module.

I've looked - I can't find any manufacturer that has them in the 2-3 MHz range.

(Well, technically I did find one, but they want me to buy full reels - 1500 @ $3.50 ea - of BPFs for each frequency. Can't afford that!)

I'm seriously tempted to just put in a bandpass opamp and call it a day. But I'm trying to keep the complete circuit's power requirement as low as possible, and I have a slight fear it would introduce a little noise into the circuit. (I have nothing solid to base that fear on - I just know that opamps can't possibly be noiseless.)

[amspire]

Agilent Genesys S/Filter seems to be a great program for designing filters with both amplitude and group delay symmetry.

Sure! Only a mere $10,000 software package! If you get everything, it's more like $30,000! (Where's the animated smiley icon that coughs up money?)
Heck, if I could afford that, I'd just hire a good analog EE to design the filters for me - or even finish my project while I paint my house!

But seriously - Genesys does look like a very good software package. And I did manage to sweet-talk my way into an eval license for it. So I've got 30 days to play with it and get some good designs out of it. (Although I did manage to crash it within 10 minutes of starting it for the first time. Hmmm...)

What's really fun is to mock up the same filter in Genesys and LTSpice, and try to get them to agree with each other.

Shamelessly using a $10,000 package for free !

I love it !   I didn't dare lookup the price myself.

The S/Filter page did mentioned the transforms they are using, but just using the actual program to design the filter directly is great stuff.

Or just buy an IF filter module.

I've looked - I can't find any manufacturer that has them in the 2-3 MHz range.

(Well, technically I did find one, but they want me to buy full reels - 1500 @ $3.50 ea - of BPFs for each frequency. Can't afford that!)

I'm seriously tempted to just put in a bandpass opamp and call it a day. But I'm trying to keep the complete circuit's power requirement as low as possible, and I have a slight fear it would introduce a little noise into the circuit. (I have nothing solid to base that fear on - I just know that opamps can't possibly be noiseless.)

You could, but the opamp filters will have the same non-symmetry, unless you come up with the right magic numbers.

LC parts do have some big advantages - never overload, no noise, and no gain-bandwidth-slewrate  problems at RF frequencies.

Richard

[amspire]

This is a pretty impressive free filter design package I stumbled across.

Impressive Help file too.

http://www.aade.com/filter32/download.htm#download

Richard

[SoftwareSamurai]

This is a pretty impressive free filter design package I stumbled across.
http://www.aade.com/filter32/download.htm#download

Yes, I've already tried that software. It's mostly working. And by that I mean there's at least one "node" pattern that's broken (that I know of) - it doesn't calculate the S21 correctly when used. I emailed the guy and asked him if he knows if all of the node patterns are working correctly. His reply: "I don't know".   

[SoftwareSamurai]

On a related note (spice software): After playing around with Genesys for a few days, I had one of those facepalm/epiphany moments.

While I was comparing the graph results from Genesys and LTSpice, it suddenly dawned on me that I've been looking at the wrong "visible trace" in LTSpice all this time. It took me a day staring at the output graphs of the two software packages before I finally noticed why they didn't seem to quite match up, given the exact same filter design.

Probably like most people using LTSpice, I rigged the input and output impedances, then filled in my filter design, but ignorantly clicked on the voltage output of the output impedance "load" resistor to see the graph. This is technically incorrect. What I should have been looking at is the "S21" trace! Unfortunately, LTSpice doesn't automatically create the S21 parameter for viewing. Thankfully, Genesys automatically does!

So how do you get the S21 trace in LTSpice? Turns out it's a bit obscure, but can be done.

Let's say you're feeding the filter with a voltage source (V1). Naturally you've set it up to do "AC 1", entered a "series resistance" for the input impedance, and have a ".ac ..." spice directive to sweep the frequencies desired. If you have an output load resistor (Rload) for the output impedance, you need to add this spice directive...

".net I(Rload) V1"

...which at first seems to do absolutely nothing. However, that directive quietly creates a new visible trace, "S21(v1)", listed under the "Plot Settings -> Visible Traces" list. Select it, and bingo! There's your filter's correct response graph!

[amspire]

Great that you have got the two to match.

However, I don't understand why they would be different.

S21 is meant to be the forward voltage gain, so it should give exactly the same graph as V_out/V_in in LTSpice as long as the source and load resistances match exactly.

Puzzling.

Looking at your IC, the amplifier output impedance into the filter is 350 ohms which sounds pretty high? May have to add an emitter follower. If you actually design the filter for 350 ohms, then you will loose half your signal voltage from the amplifier.  The limiter input is 15K so I guess you will have to add a load resistor to suit the filter anyway. 

Richard.

[SoftwareSamurai]

S21 is meant to be the forward voltage gain, so it should give exactly the same graph as V_out/V_in in LTSpice as long as the source and load resistances match exactly.

In LTSpice, if the input and output impedances are the same, I would need to set the AC to 2 V in order for the voltage trace across the output impedance resistor (Rload) to "look" identical to S21. But when the impedances aren't the same, the difference between S21 and the voltage trace is only in db. (i.e. The voltage trace is higher or lower in db, but the curve looks the same.)

But now that I've discovered how to get S21 out of LTSpice, that doesn't matter anymore. (I wish things like that were documented better. Or perhaps I just need newer glasses...)

Looking at your IC, the amplifier output impedance into the filter is 350 ohms which sounds pretty high? May have to add an emitter follower. If you actually design the filter for 350 ohms, then you will loose half your signal voltage from the amplifier.  The limiter input is 15K so I guess you will have to add a load resistor to suit the filter anyway. 

Yes, as you can see, the input and output impedances I'm dealing with aren't the same. In fact, there's two stages where I can put filters: Between the LNA and the AMP, and between the AMP and the LIM.

Zout from the LNA is 200. Zin to the AMP is 10k.
Zout from the AMP is 350. Zin to the LIM is 15k.

Right now I'm toying with the idea of putting a low-pass filter between the LNA and AMP, and a high-pass filter between the AMP and LIM. Doing that would allow me to design simpler filters that won't load each other.

Q: How "flat" does the delay group need to be (within the FM modulation band) before I can stop worrying about it being a source for distortion?
e.g. If I've got it down to a slight curve that starts at 0.5us, curves down to 0.1us @ f_c, then ends at 0.5us, can I consider that essentially "distortion free"?

[amspire]

It I'd a bad idea to replace a bandpass filter with a lowpass and a highpass. The bandpass will out perform the two filters by a huge margin.


If you can design a bandpass filter with 350 input and output, that would work well.

Otherwise I would add a simple transistor emitter follower from the 350 ohm outputto get a much lower source impedances. Because the limiter input is high, you can lower it just with a parallel resistor.

Richard.

[SoftwareSamurai]

It I'd a bad idea to replace a bandpass filter with a lowpass and a highpass. The bandpass will out perform the two filters by a huge margin.

But if I'm putting a lowpass filter between the LNA and the AMP, then a highpass filter between the AMP and the LIM, why would that worse than a single bandpass filter? Having the AMP in between the LP and HP filters effectively isolates them, right?

By my count, I can do a lowpass filter with 2 components and a highpass filter with 2 components, but the best bandpass filter I can come up with would take 8 components. And from my LTSpice designs, putting LP and HP filters together like that produces a flatter group delay in the FM deviation band than the PB filter alone.

If you can design a bandpass filter with 350 input and output, that would work well.

So you're saying it's better to always have a matching input and output impedances for filters?

Considering the LNA has a 22 dB gain, and the AMP has a 20 dB gain, I'm not sure what improvement it makes to have matching input and output impedances. (Remember, I'm still new to analog EE!)

Otherwise I would add a simple transistor emitter follower from the 350 ohm output to get a much lower source impedances. Because the limiter input is high, you can lower it just with a parallel resistor.

So, drop the input and output impedances to, say, 50? I think I'll LTSpice it and see what difference that makes.

[amspire]

It I'd a bad idea to replace a bandpass filter with a lowpass and a highpass. The bandpass will out perform the two filters by a huge margin.

But if I'm putting a lowpass filter between the LNA and the AMP, then a highpass filter between the AMP and the LIM, why would that worse than a single bandpass filter? Having the AMP in between the LP and HP filters effectively isolates them, right?

By my count, I can do a lowpass filter with 2 components and a highpass filter with 2 components, but the best bandpass filter I can come up with would take 8 components. And from my LTSpice designs, putting LP and HP filters together like that produces a flatter group delay in the FM deviation band than the PB filter alone.

Sure you can do a lowpass filter with 2 components, but is will roll off really slowly in the stopband. Normally one of the main purposes of an IF filter is to reject signals outside the passband (like nearby radio stations in a FM radio receiver) , and the attenuation in the stopband of a bandpass is way better then low and high pass filters.

If you are not worried about interference from nearby frequencies, then you  can do without the IF filter all together. It will work without an IF, as long as the primary signal the RF stage is picking up is the one you want to decode. Every stage you add degrades the signal a bit, so it is worth asking if you do need an IF.

If you can design a bandpass filter with 350 input and output, that would work well.

So you're saying it's better to always have a matching input and output impedances for filters?

Considering the LNA has a 22 dB gain, and the AMP has a 20 dB gain, I'm not sure what improvement it makes to have matching input and output impedances. (Remember, I'm still new to analog EE!)

Matching impedance at 2Mhz just doesn't matter, which is why you really don't need to worry about "S" parameters and reflected power. If you were designing at 2GHz then matching impedances is very important.  The only place impedances come into play is if you design a filter that expects a 50 ohm input impedance, and you attach it to a 0 ohm impedance instead, you will get a slight change in the filter performance.

Otherwise I would add a simple transistor emitter follower from the 350 ohm output to get a much lower source impedances. Because the limiter input is high, you can lower it just with a parallel resistor.

So, drop the input and output impedances to, say, 50? I think I'll LTSpice it and see what difference that makes.

You can try designing a 350 ohm filter.  If you are going to design a 50 ohm filter, then you will need an amplifier like a transistor emitter follower, other wise you will get major signal losses. A 350 ohm output into a 50 ohm filter will attenuate the signal by 8 times.

A 350 ohm output for the filter is fine, as you can add the 350 ohms across the output and connect it to the limiter.

Richard

[SoftwareSamurai]

Sure you can do a lowpass filter with 2 components, but is will roll off really slowly in the stopband. Normally one of the main purposes of an IF filter is to reject signals outside the passband (like nearby radio stations in a FM radio receiver) , and the attenuation in the stopband of a bandpass is way better then low and high pass filters.

Ah, yes, I see. Well I was following the suggested 500 kHz carrier separation as noted in the TSH511 doc, so my passband is nearly 500 kHz. (Even thought the FM deviation is only +/- 75 kHz.) You're totally right if I wanted the passband to be much smaller. However, in making the passband smaller, I fear getting it too small and having the tolerances of the components and the stray capacitance/inductance of the board shift/distort the filter's performance. (i.e. High Q bandpass filters really need tuneable components.) The STMicro engineers cautioned me on making the bandpass filter too tight, saying that I shouldn't worry too much about it since the next section of the TSH511 is the tank circuit, which can easily be designed with a narrower operational bandwidth (but still pretty large).

If you are not worried about interference from nearby frequencies, then you  can do without the IF filter all together. It will work without an IF, as long as the primary signal the RF stage is picking up is the one you want to decode. Every stage you add degrades the signal a bit, so it is worth asking if you do need an IF.

Yeah, I've thought about that too. The STMicro engineers made that same comment, but added that eliminating the BPF might allow stray signals to creep in. I'll have to experiment with a prototype board and see what happens if I don't have a BPF

Matching impedance at 2Mhz just doesn't matter, which is why you really don't need to worry about "S" parameters and reflected power. If you were designing at 2GHz then matching impedances is very important.  The only place impedances come into play is if you design a filter that expects a 50 ohm input impedance, and you attach it to a 0 ohm impedance instead, you will get a slight change in the filter performance.

Yes, I can see that altering the input and output impedances does change the filter's response a little. At 2 GHz I can totally understand how those small changes would have a huge impact.

I think for my case it's safe for me to just roll with the impedances of the chip and not worry about it. (But I'll keep this in mind if I ever have to deal with GHz signals!)

You can try designing a 350 ohm filter.  If you are going to design a 50 ohm filter, then you will need an amplifier like a transistor emitter follower, other wise you will get major signal losses. A 350 ohm output into a 50 ohm filter will attenuate the signal by 8 times.

A 350 ohm output for the filter is fine, as you can add the 350 ohms across the output and connect it to the limiter.

Hmmm. I LTSpice'ed my filter with various matched impedances, but I didn't like the group delay response inside the FM deviation band. When I go with 200 & 10k or 350 & 15k, the gd looks really nice and flat there.

[vk6zgo]

One problem with using 50 Ohm Z for low frequency filters is that the series resistance of the inductors becomes significant,so that a real filter won't look like your modelled version.
You might have to wind up a sample inductor,measure R,then insert that into your model.
Another thing you can do is put impedance conversion sections at the input & output of the filter,&  design it for a somewhat higher impedance.
This of course,makes the whole thing more complex,& larger.

I had occasion to design & test a 50 Ohm filter for a bit lower in frequency (around 530kHz),& it was a nightmare.
The idea was to allow a Broadcast station to use another station "off air" as a standby program source,without interference from the station's local transmitter.
A filter designed from the standard tables (No modelling software back then),looked good,but the series R bugbear raised its ugly head,& I had to rethink the whole thing.
Empirical design lead to a filter which did the job,but reduced one sideband of the required signal more than I would have liked.
In retrospect,I would have designed it from the tables,but for a higher Z,& used impedance conversion sections.

VK6ZGO

[SoftwareSamurai]

One problem with using 50 Ohm Z for low frequency filters is that the series resistance of the inductors becomes significant,so that a real filter won't look like your modelled version.

Good point. That's why I'm sticking to known, manufactured inductors, so I can add their spec'ed resistance into the LTSpice model. (I considered adding their spec'ed capacitance too, but that's so small it doesn't seem to make too much difference.)