Science/Engineering/Technical forums

Peter, the "fine wine" preamp shown elsewhere on this board is a matched filter for this signal. It's not obvious, till you note turn on and off times for the transistors I chose, and the bias point. Many years went into evolving that tech, on and off, and it's simple elegance. The tricks are sort of non obvious. That's why some peoples' attempts to dupe it with what they thought were "better" parts failed...it's also stable for gain 1, so you can add capacity across the NFB resistance for shaping. It doesn't need to be our tiny smd board to work. Joe Jarski laid out that board so it would fit inside most detector heads for noise reasons, and it matches 52 ohm coax.

Been putting a little more thought into this, in particular after working with Phillipp's design. I think you really do need a way to prevent the peak detector from "re opening" when there is a signal present over threshold, so as not to catch the last part of some pulse and think you got the peak when all you got was the middle of the decay. On further study, some simplification seems possible, and some actual parts suggest themselves, so here goes. This could be done as a shield for arduino uno or chipkit uno32 (probably preferred, its A/D is far faster and it's just generally a better machine). Either would have both the required voltages available (5 and 3.3), though there might be some slight difference in how you'd adjust things, as for example, the arduino has a 5v input range where the chipkit board has 3.3v, and the arduino needs much closer attention to drift in hold mode as it's just plain a lot slower.

Here's the circuit I'm now looking at - and thinking of a digikey order to actually build (along with some nice fiber optics parts Peter suggested to me). I did some searching around on digikey for reasonable parts for this. There's just so much stuff out there now, there's no guarantee I found the best ones, just the first that seemed good enough and available.
For the transistors - easy - 2n3904 and 2n3906 should do, it's not picky. Though you might find a better NPN with less BE and CE leakage in reverse bias. Dunno. The point of the PNP is that by varying its emitter current via a rheostat, one can vary the base current (eg emitter current/Hfe = base current) and perhaps cancel some of the other leakages. FWIW, I'm using PDI for "peak detector input" and PDO for "peak detector output" on the schematic here.
Ch is not yet determined - the fet I'm looking at has about 100pf capacity anyway, and that might do it. If it turns out more is needed, we put some in on the bench while testing.
So, parts:
The diode is for Vbe compensation, and could be any garden variety silicon, or the BE junction of say, the same type NPN used for the diode/current amp. Adjusting how much current it sees will make the output of the opamp want to be at 1 Vbe above ground, so the peak detector sits at ground with no signal input. Again, the current into this is TBD, but will be pretty small, eg the resistor will be large in value.
For the opamp, I liked an AD8615 since they have most of the right properties - speed, low offset, low bias current, and will take 5v - we need a little more here so as to be sure to be able to make 3.3v after the Vbe of the NPN.
For the comparators...gheesh are there a lot out there most of which stink. This one looked good. BU7252G2CT (that's for an soic 8, there's a smaller one too). This one has a little built in hysteresis, and really low input current. Should be quick enough for this.
The or gate is an NC7532 single unit.
Flipflop is a SN74VLC1G74 (TI) part. Nothing special.
FETs - irf7530, a dual. Low gate threshold. Dunno if this is the best that can be found. At least this circuit isn't overly sensitive to charge injection or miller capacity (in a bad way, anyway).

Now, here's how this should work. At first, we assume no signal, noise under whatever we've set the threshold comparator for (I show R values for about 0-120mv, the setting will likely be at the lower end of that).
Disch is high from the uP, as is /reset enable, we're just waiting for a peak to come along.

When one does break threshold, then start back down, we'll get the flipflop set and an interrupt to the uP. This also shorts the opamp output (that's why there's a series resistor there to protect the opamp) and the uP should have been sitting in "sample" mode all this time - just tracking whatever the input was. When it sees the interrupt, it should start the a/d conversion, possibly waiting a little if it has a long acquisition time in its own S/H. When that's done, it increments a bin in the spectrum, and sets /reset enable low, so the next time the input goes back below threshold again, the flip flop allows another peak to be detected. The uP should hold /reset enable true (low) until this happens - either via reading its interrupt pin as a digital input (which most can) or tying another input pin to the same signal. At this point we are ready for step 1 again, repeat as required along with the usual housekeeping.

I personally like using USB to talk to a PC for this kind of thing - you get a free power supply even if it is a little noisy. For most things, you can either filter it out, or for 3.3v, there are low dropout regulators (usually on the uP board already) that clean things up.

Of course, to this old engineer, what stinks about modern times is that nearly all these parts are surface mount - at least I was able to skip the sort that don't even have visible pins and have to be reflowed or you can't use them at all. That means I get to lay out a PCB and make it before I can even start to play with it. Sometimes I miss the "good old days" because in some ways, they really were. If all this stuff was through-hole, I'd have built this in less time than it took to type this post.

Thanks John and Peter. And especially Phillipp for lighting a fire under my butt to do something here. Peter, the "fine wine" preamp, should it be used, is a matched filter both in F and T domain - fast onset/slow decay as is, designed just for that reason.
Many who have tried to dupe it using "better" parts found out you lose that - I chose carefully there in terms of charge carrier lifetimes and topology, among other things. I'd be interested to hear from John how you get rid of the "post pulse hangover" from charge picked up by the coupling cap during the pulse - with a finite load it will, and with no load, it's got a kV of dc on it...you can never really kill off all deadtime or pulse pileup, it's just not in the math of a Poisson distribution to be able to.

And any rate, here the current schiz, and I'm almost done laying out a board to go on a chipkit UNO 32 (arduino-like, but with real a/d and horsepower). It has some small issues - tempco and leakage aren't well defined, and in some cases, controlled opposite of the right way in the interest of simplicity. It looks kind of like this in my pcb layout tool. I "can't even" eagle - when the tutorial asks me to click icons or use menus that aren't there - in the bit bin it goes. I use traxedit, an old DOS program, in dosemu over linux. Fast as poop, and no crap about things like "you can't just put down an sot23-5 without telling me what the part number is so I can run spice or something against it". It just draws, well, and prints out stuff I can make boards from. There's now a public domain version (Easytrax I think) that's about the same, but I pirated a version back in the 70's...of the real deal. Not even sure which is better, but it sure is fast, having been designed for under 640k of ram, and 80186-class cpus. The NPN and PNP transistors cancel one another's Vbe a little, but it's all wrong here - couldn't make it perfect and simple both. Note the PNP will have the bigger Vbe with a zero height pulse, while the NPN will be at its lowest. And opposite for a big one. Life. That's what we have PC's for - to remove non-linearities like this.

I also designed a regular arduino uno pulse generator that makes 5v negative going pulses, around 5us (the fastest you can do with a 16 mhz processor and consecutive lines of "digitalWrite" in that platform - sometimes they are longer due to the "arduino opsys" taking a timer interrupt. But the cool thing there - and what plagued some other designs - is that it's random in timing with a poisson distribution, on purpose. It makes my thing, when just digitizing a pot (rather than the real front end) show any flaws - there are none so far. I'll publish the sketches for both somewhere in the software sections when I'm done...real close now.

It makes real pretty pictures off just a pot as the signal. Here's a log plot of one I made - you can even see the 1-2 count stuff where I moved the pot a couple of times during the test run, in the middle of millions of counts where I didn't, and there's no averaging going on at all here - this is raw, just logged amplitude. Linear, you can't see that, but you can get a better feel for the major peak height difference, so here's the same data linear. More when I've got more. I did get all these parts. Some have values "to be determined" as is always the real case. I used 2.7k resistors for example, to prevent 5v stuff from killing the 3.3v stuff...and some other values that "seemed to be right" in other projects. This needs cooperation from the UNO32 to work right - and only re-open the peak detector in a quiet time when there's no pulse present, as determined by the threshold pot. The UN"O is sitting in sample mode at all times (I had to write my own version of "analogRead" to get that) between triggers, so my peak holder doesn't have to be that great, and I have its interrupt prioritized so the time to get to conversion is deterministic. I may not actually need a Ch at all, since the fets I used have a few hundred pf anyway...we don't have to sweat charge injection or any of that other stuff in a traditional sample/hold either - if you pay attention to what happens when - any glitches happen when we're not looking (on purpose).
Slick ain't more complexity, it's subtlety.

Cool stuff!!

I have some questions about your schematic, though: which type of flip-flop are you using (I didn't find a datasheet for the SN74LVC1G75, as shown in your scan)? Also, is the "P"-input you're using the data- or preset input?

I think I understand everything up to the transistor pair for Vbe-compensation / hold capacitor charging. I'm afraid I don't quite understand the role of U2: I thought it might be for pulling down the common base of the Ts in order to switch to "hold" and trigger the A/D interrupt at the same time -- but wouldn't it discharge Ch, too? Am I missing something here?

Looking forward to your first real data /spectrum!

Philipp

Philipp

the dual Jfet is the reset ie the ch holds until the value is read then reset and ch is discharged via U2 and ch cannot get any pre charge while the reset is taking place then all is ready for the next pulse

You're both really sharp-eyed, and FWIW, there's a mistake on the schematic (but not on the board layout) which I might get around to etching today, we'll see. Finished it last night. I'm going to make that whole board into a library component and tile them so I get a few out of a sheet of laser printer transparency - that stuff isn't cheap, nor is it perfect edge to edge. And in a couple cases (not yet shown here) I pushed the design rules really hard and not all the prints might come out without an accidental short. Usually, the easiest place to fix that is after exposing and developing the board, but before etching, with a razor blade and a stereo microscope, which is plan-a - I usually do that inspection anyway. I found the one (only?) good use of ferric chloride in doing this. Wipe it on the developed board, and all the actually-exposed copper turns really dark. Places where a tiny amount of resist clung on (too thin to see otherwise), don't. Makes it easy to see exposure-development mistakes and fix them - and repeat as necessary with a new wipe test. I usually do this at 30x or so magnification and no caffeine but perhaps one beer and a couple hand steadying tricks.

Having done this before, I'm well aware that there will likely be some changes required before I have one I could sell or send off in good conscience. That's just how this usually works out. Hardware often needs "editing" just like software. It's just hard.

The fet gates (they are enhancement mode n type IG mosfets, not jfets) aren't supposed to be tied together - the right hand one gets it's own pin on the UNO32 and separate control. There's a good reason for that, which I'll get into. We want to turn that one off before we let the other one turn off, so we can't get into a high-current fight with the NPN transistor.

The PNP transistor gives me 1 Vbe more or less off ground when the opamp is putting out zero volts (it's a rail to rail fast opamp). It will be on the high side due to the low value emitter resistor (lots of current, Vbe is the log of current), and worst at near-ground when the emitter current is highest - a fast constant current source seemed like a futile adding of complexity there. The Vbe of the NPN will also vary, but mostly be at the low end once Ch is charged up, since no current will be flowing at that point, but the intent is to cancel out and make tiny pulses near ground possible to digitize at all. Since these are all SMD over groundplane, they should track temperature pretty well at least.

The result of that pair will be a bit of non-linearity in the results, as the larger pulses won't be perfectly as much larger as a small pulse would be, due to the current variance through the PNP. I can live with that - and NaI has it's own nonlinearities, so for a pro-grade thing, we'd have to do some heavy math lifting in the PC (preferably) to make it all come out accurate in keV. For what I want, and since I own a plethora of calibration sources, I probably won't bother unless it's really bad.
It should be deterministic, as I've prioritized the interrupt I'm using so the usual arduino-type opsys stuff they hide from you can't mess with my timing as much (it takes timer interrupts among other things).

I think the flip flop got mis transcribed as to the number. It's a TI part and some of the numbers are letters (5 and S are hard to tell apart in my sloppy handwriting, as are 1 and l). It only has a /reset, there is no /Q output (hence the change in the drawing and I'll have to change the software too). The digikey part number is 296-17617-1-ND, It's just a d type flop with a couple pins missing from a 7474 and only a "single" at that, in an sot23-6 package. Teeny stuff. Hate to admit it, I kinda miss the days when you could proto this stuff on perfboard...but most of these parts aren't even available in through-hole versions.

What is supposed to happen here:
Assume we are initially "at rest" and the signal is under the threshold set by the pot on one of the comparators. A pulse comes in, and shows up as a positive pulse at the opamp output (and yes, I know the gain control will interact a bit with the shaping - simple and perfect rarely occur together). Once we go "over the top", and the PDI (pulse detector input) signal drops below the PDO signal - we set the flop - but only if PDI is still above threshold, since that comparator feeds the D input. There might be some error on really tiny pulses - we'd miss them. That's life, I personally am not interested in the extreme low end of energy, and were I to be, well, that's what the gain control is for (losing high energies to clipping at the other end if you crank it up).

In real life, you use a different NaI for each end of the range anyway. My "Gallon jug" NaI's are horrible at the low energy end, but great in the megavolt range. I also have one with about a 2mm thick NaI head that is fantastic at the 10's of kev range, but of course, can't handle the big stuff well at all - it mostly scatters out. I have them all wired to take a negative power supply input, with the phototube anode floating, so they can be connected to this directly, no coupling capacitor required, they are an almost perfect current source. You can't pull this off with just any detector - some have noisy leakage from the photo cathode to the world..but with these, you can. For your basic 2 wire bicron, you have to use positive HV, a series resistor load (I use from 10k to 50k there, on the high side, but I like loud signals), and a coupling cap - a few hundred to 1000 pf, to this board. In either case, power supply regulation is pretty important, as most phototubes when run around their specs, are super sensitive to it - gain goes all over the place. I sometimes defeat that issue by simply running them on the hot side, voltage wise, where this effect is less pronounced. I've never figured out why everyone seems to get on the steep part of the error curve on purpose. I could be missing something, but these tricks are what made the really pretty spectra on that thread with the URSA II pro MCA we borrowed. And those showed us that no matter what - there's not much point in having more than 1k bins on an NaI - the lines are all a few bins wide. We wound up using a lot of inter-bin smoothing in the URSA software to get good looking plots - which wound up with around 500-100 effective bins, just prettier/smoother looking. I'll probably just blow this spectrum straight out to gnuplot. It should be good enough for what I need right now. I'll probably use perl as the duct-tape to make that go. You can open a USB serial device "by-id" with the serial port module in perl, so it doesn't get confused with other serial dongles. And in general, doesn't care about what baud rate you specify at either end. In our case, we'll be sending single characters to the unit to control it, and once in awhile getting a flood of 32 bit numbers back as ascii.
Very similar to the way Phillipp did his, but I might use some different characters for commands (since I do have some already established protocols for my other uP data aq here), and probably fewer, since I just want spectra - I don't need to look at single bins (which you can still do anyway once you have all 1k of them).

OK, setting the flop creates an interrupt to the uP. The uP responds by flipping the s/h internal to hold and starting the a/d acquisition. I rewrote their init code to leave the a/d always in sample mode on channel 0 so this could work. Setting the flop already pulled down the base of the NPN - a nice safe place to turn off the input vs shorting out the opamp and so on (the reverse emitter base diode of the PNP just barely can hold off the 5v, but barely counts here). Once the a/d conversion is begun, the uP will turn on the other half of the fet and discharge the main Ch - these are relatively large fets (ampere ratings) and have fairly high capacity already - we may not even need an explicit Ch here. Since all this happens either before or after the internal uP sample/hold is closed - we don't worry about charge injection issues, or glitches - all the noisy junk happens when we are not watching. That's on purpose.

Now that we have our number, and have incremented the appropriate bin in our spectrum, we turn off the Ch fet (the cap is presumably discharged by now, if we find otherwise we can add some delay in the code), and apply the reset enable signal (low true). The comparator and or gate ensures that the peak detector will not actually open again (allow the flip flop to reset) until the signal in question has fallen below the threshold (0-120mv with the values shown here, roughly). Everything returns to step one, waiting for another pulse at that point. The uP is responsible for holding that reset enable signal until it sees the flip flop reset - it can read that pin as well as use it for an interrupt, and it must do this or we might get "stuck" in a state we can't recover from.

The top half of all this runs off 5v, so we can always be sure of reaching 3.3v output. The bottom half runs on 3.3v, since that's what the uP wants - I used 2.7k resistors on signals driven from the 5v side to the 3.3v side to limit current. The UNO32 also has input protection stuff, but some of that you'd rather not put to work as it can actually pull up the 3.3v supply if you hit it hard enough. This should still work at all for pulses larger than 3.3v out of the opamp, but all those should stack up in the very last bin as they often do in a "real" MCA - for one thing, cosmic rays - you always have with you. We won't get an accurate value for that kind of thing, but we'll know it happened. For really low intensity sources, we'll probably have to add an option to not plot the last couple of bins so auto-scaling doesn't hide our desired signal. Of course, I will have a choice of log or linear plotting, since it's so easy in this system.

This will, of course, have some dead time. For most signals, that's a pretty large problem - you get funny results sampling, say, an IF strip looking for the PDF of it, unless (under)sampling is "truly random" - beats, aliases and so on (we had to do this to make an automated signal to noise measuring device for radios in the...outfit I can't talk about). Here, it's not such an issue assuming we get our randomness via the source radiation itself - if that turns out to be somewhat coherent, we'd have a problem. Since I have 32 bit everything here - we could likely just wait longer and average all that out, though. It's going to take a long time to overflow.

I strongly considered trying to pull this off with a garden variety arduino UNO. The specs just aren't there for this, the a/d is too slow, and so is the rest. The chipkit UNO32 does cost more...but it's orders of magnitude better in all ways. And I like PICs better than Atmel...even if this one does have a MIPS core - it still has their really good IO behavior, which is most of the reason I started with PICs before they even discovered that having interrupts would be nice. So for a couple times the price, we get orders of magnitude in a/d speed - mhz vs khz - and core speed (80 mhz vs 16), which we can put to good use here. Given that two consecutive lines of digitalWrite on the regular arduino can make at minimum a 5 uS pulse...something is a bit off there anyway - it should be tons faster than that at 16 real mhz. Maybe the atmel is more than one clock per instruction? More likely, the simplistic arduino IDE with tons of hidden library code full of #defines (to support every board under the sun), is the real problem there. And it's hard to find that stuff - long searches using grep in directories not mentioned in the dox...and so on. (Glad I'm on linux here, or even that would have been hard).

We'll see - I should have this in testing in at most a week, probably much sooner. I've got another project going on in parallel just now - fiber optics comm for a servo I want to have floating up at 50kv or so and tune a capacitor from down here at ground levels. It's taking me a little effort to test with this unfamiliar stuff. I thought of using WiFi for that (I have WF32 boards) but it's in a very noisy place, and near 100's watts RF as is...probably won't work with WiFi.

For giggles, here's what I'm about to etch, though what I did was tile 3 boards on what I'm really going to print out.
As you can see, I really pushed the design rules near the lower right corner with that ugly long 3.3v power track etc. Life.
Also, I don't do plated through holes - else I'd have put more on the blue (bottom) layer. When you see a light orange pad, that's top-only. The darker ones go through and are drilled. I'll be doing that and soldering WW wire through to get my grounds for things. I will also have to solder some parts in odd places on the ground plane side (like the main connector to the UNO) to get those signals. Once this is truly working...I'll send off gerbers to AP Circuits in Canada, who I've been using for PCBs for decades when I want more than one, they are good, fast, and cheap - all 3 (rare as that is). But not worth it for a onesie. Their proto-1 level uses solder for the etchant masking, and it's put on pretty thick. I find more often than not I can simply use mostly the existing solder and some Kester 952 flux to stick down the parts. I might then make another under-microscope pass and add a little solder to some spots, but there's enough there already to get the parts into position, hold them with a toothpick, and tap them with the Metcal super nice soldering iron to hold them in place while I "really" solder them later if required. I found their more-expensive proto service, one that does solder mask and uses gold-plated tracks nohwere near as good for this type home construction. The solder mask is so thick the pins don't always reach the board! No solder means you really have to solder each pin some way and you can't just heat one at each corner to lock the part in place. More money for more work isn't my way of doing things! Edit: Here's a link to a slightly different PCB drafting program (the one they actually did give away free for awhile, it only lacks auto-routing, which stank anyway, IIRC). I run it here under linux via dosemu. dosemu makes a .dosemu hidden directory in your home dir, which then has a drive_C directory under that - it's just like DOS, I added ".." to the path so I can put a directory for each new board under the "Trax" or easytrax directory I install the basic CAD stuff into, and still type the commands from that dir and have them run. Works a charm for simply laying out PCBs. Has no schematic capture junk, just patterns for parts, and it's easy to make new ones and add them to the libraries, as it is also easy to add apertures to the .apt files that define how the gerber will come out. This predates windows by a bunch, but you'll be surprised how much MS copied from this...left click = enter, right click = escape, the first letter of every command is its shortcut (without <alt>), and you can do this all keyboard only (recommended, it's hard to draw straight lines otherwise). The price is right: http://www.lupinesystems.com/easytrax/
I was able to set mine up so the little arrow keys move 1/10th as fast as the ones on the number pad - very handy. This is easy and quick to learn if you don't like Eagle, or that junk that forces you to endlessly buy libraries for every new component that comes out so it can do schematic capture (this doesn't do that at all). You can define your own track and pad sizes, what layer they go on - and it has a status bar, years before any other GUI opsys had one...so you know what mode you're in.
Like I said, those familiar with history will be astonished how much of now was copied from this thing....which came years earlier. I'll try installing this alongside the version I'm using and report somewhere here what the differences might be. IIRC, this one does metric (as well as imperial), mine does not but at one mil resolution, it's not a biggie.

Lots of interruptions lately, but progress anyway! Looks like this new pre-sensitized PCB stuff is way more sensitive. I used my usual exposure setup (which I should document more fully elsewhere) and time - and it developed super fast, leaving the odd missing resist spot. The laser printer will eat traxedit files for a 300dpi HP BW printer, but doesn't hit black as hard as I'd like. Would be handy to be able to add red on top - this is short wavelength (mostly) photo-resist. Now to measure a bunch more stuff and select the right size drills for the holes this needs. It will be somewhat of a kludge - some parts have a pin or two that goes through, and a pin or two that says top-only, as I used a straight "fill" for the groundplane and don't have plated through holes in my technique (I could, but it's not cheap or easy, so for onesies, we just solder a wire through a hole or whatever). This will at least reveal any design errors so I can make the inevitable next version.

Cool!

Thanks for the clarification on functionality, now it makes sense (even) to me!

I'm looking forward to the first spectra!

Keep us updated,
Philipp

I'm looking forward to the first spectra!

You and me both. It's on the list, but the list is long - I'm also in the middle of remoting the fusor - major task, just getting vid streaming going over the network now, getting ready for real winter and a few other things, so I get to play with this when I have a little time. Looks like for the moment I have to drop back to the pot and random interrupt timing and figure out if it's the chipkit board resetting the held output or something glitching on mine that a tek 475 can't see. Or pull the really good scope off the fusor, or use the one upstairs, all of which are major hassles with wiring and support to get it all going at once. I will get to it...but it had to drop a level on the priority list due to other pressing things - like burying some cable to the next building over so I don't get fried, building an addition on it so I can move to live there instead, etc etc. Been busy lately.