Teardown: Agilent E3631A Power Supply (Picture Heavy)

[vl400]

I have had this supply for awhile now and always planned on doing a teardown when replacing the front connectors but still have not got around to it Had a chance today to pull it apart and take some photos along with some performance screen caps.

This is a triple output (0 - 6V @ 5A, 0 - 25V and -25 - 0V @ 1A) supply with some fairly reasonable accuracy, noise and ripple specs.











The top board contains the processor, front panel control and +-25V section.











Processor







While the bottom board does the 6V section, HPIB and serial











Front panel board is fairly basic as the processor does the work, so didnt pull it apart any further. Each terminal has its own sense wire. The mounting is nice with tab towards the middle/left that clips it in





And it has a proper power switch 






And a few quick tests

No load output on @ 6V


1R Load on 6V


No load output @ 25V


1R load on 25V


10R load on 25V


Weird delayed start that happened every so often

[Kibi]

Nice teardown. Thanks for sharing.
Those meters are pretty much spot on!

[ElektroQuark]

I don't like at all how they have blocked the pressure vents of the capacitors.

[Ronnie]

Agilent really makes quality products.
Very clear scope screen shots!   

[vl400]

Yeah other than the cap vents being blocked I thought it was well made and thought out. Its not be cal'd since new in the early 2000's 

[gerrysweeney]

I recently done a teardown and repair of one of these power supplies - I just found this post so I thought I would share my efforts too, text and picture heavy.   

http://gerrysweeney.com/agileent-e3631a-power-supply-teardown-repair/

[free_electron]

-edit- one of these days i'm going to invent a non-dyslectic keyboard... ( or learn better typing ). polishing right now ..

A bit more information on this power supply, its design and the how and why, of agilent ( this was designed when they were still hp ) doing the things they did in this design.

First off : those caps do not have their vents blocked. It is silicone isolation to prevent possible shorting of the bare top metal of these caps to the heatsink of the 6 volts section underneath when the lid is closed. Not all supplies have this. Originally there were capacitors that were shorter, only later models use longer caps so they solved it that way. Nevermind , was thinking of another supply in the same series.
Here it is purely mechanical stability.

Second. The A/D is indeed built with discrete components and .. an asic. They use a conversion technique called multislope III which is a charge balancing convertor. It is similar to a dual slope convertor. They charge a capacitor in an integrator circuit during a fixed amount of time from a reference voltage. They then use the, to be measured voltage, to discharge back to zero and measure time again. The ratio gives you the weighted ratio of Vin versus Vref.

The problem with this is that conversion time changes with input voltage. So, what does multislope do? If they detect that they are 'far off' they 'add' or 'substract' reference current. So they can speed up the discharge of the cap. An algorithm switches reference currents around coming from a -10 volt and +10 volts reference through 30 k resistors. They use a simple 4051 to do the switching. A high precision opamp for the integrator and a really good comparator for zero detection.

The asic controls the current switches and holds the counters to measure the times.
So , the whole a/d system does not work with 'volts' but with 'time'. Time that is derived from a charge current derived from a voltage reference.

The beauty of this converter is that all errors are self canceling. Only the precision of the reference is important. If you want more digits out of the convertor .. simply measure time more accurately... This done by increasing clockspeed and having counters with more bits. The absolute clockspeed doesnt even matter as it is a ratio between two numbers. If you get more leading numbers before division you get more effective numers post division.

Now, the machines were designed at the same time als the 34401 multimeter and 34970 switch ... Which uses the same A/D technology and the same asic. So for them , this was a freebie ... And if that display tells you 5.784 volts: that last digit is spot on because internally there is one more full digit ... They have 5 1/2 digits in this design...

You can't really measure anything in this a/d convertor when it comes to troubleshooting ( i've seen 3631's with a bad convertors .. Typically the integrator chip that is fried, not a cheap opamp, its like 8$...) because everything is current controlled. So poking around with a scope probe or voltmeter tells you nothing. The multislope system mechanism keeps all voltage in the convertor at 0 by injecting or extracting currents. so the only thing you see is 'ground' everywhere and you think it's broken. Nope. working perfectly fine. If you don't see '0 volts' somewhere : that's the problem spot.

The A/D does not scan the parameters. It only measures the channel that is beeing displayed on the screen, or temporarily switches over if a request is made through gpib.  The power supply control loops are fully analog.

Core processor is an 80196 with external memory and sram. Again, same beast as in the other machines from that era 33120 generator, 34401 meter , 34970 switch,  basically any machine with vfd's, so they have tons of experience with this thing. Proven design , and for them , very short design cycle as it is paste and repeat ... With some software changes.

Now, the dac is another interesting beastie ...
They use a 16 bit dac to do the setting of all voltages and currents. The dac has a multiplexer on the output followed by analog memories built around tl074 and capacitors . They continously replenisch the caps by scanning and reprogramming the dac. Why not 6 of these DAC's? Because when these supplies were designed one dac costed more than making 6 analog memories with caps and opamps as unity gain followers...

Right. More interesting tidbits.
The cpu and a/d system share a common 'ground' with the +/- 20 output. The 6 volt ground is fully isolated from chassis ground, as is the 20 volt section. Only the gpib and serial port sits at chassis ground. The supplies float in respect to chassis ground, and each other.
So , how do we get the info in and out ? Like i said the DAC is expensive and we already have a superduper a/d... We are not going to replicate that and send it across a digital optical link... That would be too expensive... But wait ! We are HP ! And we make optocouplers... Even analog ones ! Lets do that !

The analog coupler uses an led and two matched photodiodes. Two simple opamps make a system that drives the led , measures emission using one of the photodiodes and makes a servo system. The other photodiode is a mirror... Simply amplify it and done ! Cheap, elegant and very very clever !

Now ,the communication bit.. We find the classical TMS9914 ieee488 controller driven by an 8051 as well as standard rs232 port. These talk through regular digital optocouplers with the 80188.

The front panel has its own 8051 as well. That thing scans keyboard, rotary encoder and the VFD (through high voltage buffers) and talks serially with the 80196 via the asic that ontains a second uart... )

This may all seem like overkill but it isnt... There is again something clever going on.
The VFD requires an ac voltage to drive the filament and a relatively high dc voltage to drive the segments. Problem is that this requires some voltage shifting . The scanning, and high voltage switching, also creates electrical noise. If you blink a digits intensity you see the noise spectrum change. The display is scanned in a matrix.. So this would overload the 80188  , especially when gpib commands would come in you would see a noticable flashing of the display. We are HP.. We can't have that.

So , we give the frontpanel its own processor to keep the scanning noise away from the a/d , offload the main cpu and... Wait for it ... We can do some trickery with the power supplies for the VFD ! The whole frontpanel runs at -18 volts in respect to the ground of the processor. We simply use 4 opamps to level shift the control signals between domains. This negative voltage allows us trickery with the display filament so we only need one additional winding on the mains transformer, which alread has a lot of windings..  4 opamps is cheaper than an extra winding !

So this may all look like it is 'overengineered' but in reality it is not. For them it was very cheap and they actually cut costs. The asic they had, the display system too, the analog optos too . They worked really hard to make this as cost efficient as possible.

Now the supplies themselves. These are fully analog control loops , because even HP would not dare attempt a digital control loop as they know that is always too slow.
The control is very simple. Two opamps do the work. One works as overcurrent detector and can pull the output of the voltage servo downwards. I don't recall of the top of my head the schematic but i believe they use two simple diodes to do that, sort of an analog 'and'

The driving of the power stage is a bit odd too. The make a constant current source with an lm317 and inject this in the base of the power transistor. All the regulation opamps do is deviate this current.
This allows the supply to react very fast , as opposed of having a weak opamp try to drive a relatively large 'load' and be slow at it. And you do NOT want an additional gainstage behind the opamp as this makes the loop unstable. So .. current source it is !

There is a couple of other tricks in that system. They employ a pair of LM339 or LM393 comparators to monitor the delta between iset/ireal and vset/vreal.  Depending on the logic combination they can detect cv/cc (if we are entering current limit they show CC on the display) but also when the regulator totally messes up ('unreg' is dsiplayed then).
 
To turn a channel 'off' they do not simply program to zero...i forgot the exact trick , i'd have to look back at schematics. I know they had a software bug in firmare 1.1 and 1.2 where the supply would give -0.3 volts at 30 milliampere if you turned a channel off. Later software fixed that but you can't swap software easily. It's flash rom but not uploadable, you need to pop the chip out and make sure you have the right firmware for the display. These things have been in production so long that at one point the high voltage display controller went obsolete. They did a revision but the scanning is slightly different and the 80196 needs to know that . A 1.4 firmware is bug fixed with old display, the 2.xx firmware is for new display. Agilent publishes repair notes that you can download for free with all this stuff in it. They also publisch service manuals with a full description and explanation of all the techniques used.

Wanna really learn electronics ? Go read and study all Service manuals from HP /agilent equipment and look at the schematics.. You will be amazed at what kind of a goldmine you are sitting on!

[Bored@Work]

Now, the dac is another interesting beastie ...
They use a 16 bit dac to do the setting of all voltages and currents. The dac has a multiplexer on the output followed by analog memories built around tl074 and capacitors
They continously replenisch the caps by scanning and reprogramming the dac.

Sample and holds synchronized with the multiplexer's output.

I have recently seen that in a DIY powers supply. The S&Hs were driven with a PWM signal from an AVR CPU via a multiplexer. I didn't get the reason for this, since DACs are easy to get and use. But now that you mention it, the developer might just have copied an old design.

[free_electron]

Correct. Analog sample and holds that run in sync with a demux behind the dac.

Now, there is another thing i have not explained in my long 'post' above....
Calibration ! There is not a single potmeter in sight.... So how do you get all the nonlinearities and offsets out of this system...

Well this supply is self calibrating ! Another piece of genius from HP. The firmware has a calibration algorithm on board you can run this yourself. The password is 003631. Hold down one of the buttons on the frontpanel while powering up to enable cal mode ( forgot which button, bottom row, third from left i believe ) and use dial to enter password

WARNING ! DO NOT RUN THIS IF YOU DO NOT HAVE AN APPROPRIATE PRECISION VOLTMETER AND LOAD .

YOU NEED A 5-1/2 digit meter you can trust , both for voltage and current , and you need a dc load that goes up to 5 ampere.
Dave's load is ok , but you need an ampere meter with 5-1/2 digit... that can do 6 ampere...

Here is how calibration runs : the supply will make something it claims is 0.5 volts and ask you to measure the output. You simply enter the reading from you multimeter. It then does what it thinks is 5.5 volts and asks you to enter what you see on the multimeter. It goes on with the dual channel making 0.5,-0.5,19.5,-19.5.

And then it will ask you to draw 0.5 amp and 4.9 amp and 0.1 amp and 0.9 amp on the supplies. Each time you simply enter the number you see on your 'golden' multimeter.

And then .. Magic ! Using all the 'time' constants it has now accumulated from the multislope converter ( which measures averything as 'time') they run a vey clever bit of code that yields offset and gain for everything ! The calibrate the a/d , the linear optos , and any other offset or gain anywhere in the whole supply. This set of parameters is stored in a table and that's it.

The drawback is that you cannot 'measure absolute' in this system, which makes troubleshooting a bit hard sometimes, you can only see if something goes up or down but not absolute, only relative. In essence the A/D and the entire electronics is tuned to itself, all derived from a single 10 volts reference voltage that is converted to time on an integrator. This means that ,whenever you replace ANYTHING , even a simple resistor, the calibration needs to be run.

They don't even need accurate current sources.. They only need to be long term stable , the absolute precision is irellevant. Who cares if it is 1.078943 mA or 0.9983245 .. As long as it is long term stable we are ok. It is all tuned out during calibration. It's all just 'times' on an integrator. You can use cheap analog muxes with unknown rds on , your resistors can be 5% .. It all doesnt matter. You don't care about only one thing : long term stability of all parameters. The offset and gains, and tolerances are all calibrated out.

This multislope system is a very clever A/D. You can breadboard this. Someone on another forum ( i am active on another electronics forum too ...) wouldn't believe it could be that easy. He took two LF411 opamps out of the 'pile' , a temp stable capacitor for the integrator , put it on a breadboard in dead spider construction , hooked it to an FPGA he had laying around, wrote half a page of verliog code to make the counters... He used a simple voltage reference and he got 6 digit first try ! The damn thing was spot on compared to a 34401. There are some 'volt-nuts' on that forum that have super stable chemical references. So these guys got together , tweaked it a bit and compared the results to a solartron 8-1/2 digit... and they got 7 digits .. with two crap opamps , a capacitor and some counters breadboard in an afternoon ... beat that ... i don't know if they went beyond that. there was talk of trying it with ad711 and ad706 like the agilent uses , but i don't know if they ever did that.

I have seen HP do an automatic calibration on the supplies. ( they used to come once a year to do VOSCAL, their on site calibration service. If you have enough equipment they make house-calls. when i was at the factory we had enough stuff. the design center i work now has not enough to merit voscal )

They hook up two 3458 8-1/2 digit dmms , a shunt and a programmable load. They plug in gpib in the back and hit enter ... 10 seconds later .. A fully calibrated E3631 that is reliable down to 4 trailing digits...

Beat that with you 'pic' or 'avr' driving an lm317 with fancypants lcd display.

You'll have a very VERY hard time making a supply with the same specs (current / voltage / load response, readout accuracy and precision , floating outputs , gpib and serial control ) for even double the price of what the supply goes for new off the shelf at agilent.
The isolated floating outputs alone will kill you in cost.. Youll need one cpu per channel, send it through optos , have another earth referenced cpu and you wont even come close to the precision of control and readback... You need like 18 bits or more...

Agilent is the absolute king in this kind of stuff. Their designs are incredibly clever. Optimized in all aspects, some not immediately obvious. They do not 'overengineer'.. They do it 'right'. Apply the correct technique , cut costs but do not sacrifice performance or capabilities.

Another example is the 34401 multimeter. You can study that service manual for a whole week. The explanation on the converter alone is very very interesting.

Same with the 33120 and 33150 / 33250 signal generators. The DAC , filters and output stage is pure genius ! They have also a self calibrating system in there. Not a trimmer in sight. They can even calibrate the output impedance of the amplifier .

Those siglent , rigol and others dont even come close ! And they are 15 years more 'modern'...

[A Hellene]

For anyone interested, this is the full service manual with schematics (dated April 2000) for the Agilent E3631A, in a 6.25MB .PDF file.


-George

[free_electron]

some more 'walkthrough'



from left to right :
-5 volt regulator in TO220 with 5 pins ( reset output for CPU )
- cable to the 6 volt supply. carries +/- 15 volt for analog opto servos and the volatage current set and readback signals.
- all the 'analog' drab between cable and the white 220nf caps is the multislope III convertor. U4 ( close to heatsink) is the input multiplexer. U5 ( under cable) is the referenc eucrrent multiplexer. the analog devices chip just above the connecotr i a buffer opamp making the + and - 10 volts reference from a chip that sits just left of the heatsink (U2).
- the 220nf capacitors and U9 + another under the cable ( hard to see ) are the 6 'analog memories ). U 13 is the demux for scanning them and the wide body analog devices chip is the DAC
next we have the sram and flash eprom.
-the big fat chips to the right are (bottom) the asic holding the digital part of the multislope converte, an extra uart for the frontpanel and the glue logic for the 80196 (that sits above it)


the business end ... bottom left ASI , bottom right 80188 .. ehh 80196 .. where di i get this 80188 from ? brainfart. need to go edit this ..
flash rom ( firmware 2.1 , so new display panel ) and 256k sram. the AD1851 is the 16 bit dac , u13 does the demux , r22 r24 r27 r25 r26 and the 220nf caps are the 'analog storage'


the A2211 opto on the left ( there is another one just outside the picture ) is for gpib/rs232. the next two are for CV/CC and UNREG detection (U13 does that : lm339) the HCNR200 are the analog optos. 4 in total. 2 outgoing , two returning : Vset, ISet, Vmeasured , IMesured
U14 is the 'current generator for the powertransistor. Q2 is the 'sink' that pulls current away from the source to regulate. U21 does the real work ... U20 U22 and U23 ( and the other one between the two left analog couplers U19) are the opto servo's. have nothing to do with supply regulation loop.

Here's a tip : wanna make a really good 'analog' power supply ? go strip that section out and replace the analog optos with potmeters. it'll outperform the 5 billion stupid lm317 "wrongly implemented , half understood, can't regulate to zero and doesnt have adjustable current limit" schematics floating around on the interwebs...

[Mechatrommer]

i wish i could read all free_electron posts. genius! from what i read so far, valuable reference for later.

[free_electron]

learn dutch i got over 15000 posts on the dutch electronics forum ... i really need to get a life

[T4P]

Thanks free_electron i learnt a shitload just by reading all these ...
We are witnessing a shitload of thought going into the power supplies
(A accomplished engineer is one that makes something better without making it more expensive)

[free_electron]

(A accomplished engineer is one that makes something better without making it more expensive)

and that is exactyl what Agilent is good at ! it didn;t cost them anything extra, they cut corners left and right and come up with a product with unbeatable performance at its pricepoint. You open the box and simply keel over at the staggering number of components in the design... but if you do the pricecalc it turns out it is cheaper to slap in 6 opamps , 6 resistors and six capacitors + a multiplexer under software control , than 6 DAC's....

They optimized the snot out of this thing withouth sacrificing performance.
HP : High Performance , Heavy Product, High Price ... That is when 'HP' was still 'HP' and Bill and Dave were still in charge...

If you really want to see their 'magic' at work : service manual of a 34401 ! Or if you can find it : a 3485.. That beast has been in production since the dinosaur age ... and is still the 'gold standard' when it comes to multimeters. Heck these are the machines that today are still used to 'calibrate' the calibrators ... and its a 30 year old design ! ( mid 80's )

That multislope technique is simply brilliant ( go read the patents on it .. this is why you don't find it on the market as a chip ... it is heavily patented and they don't licence ...).
You don't care about anything 'absolute', not even your reference.  The only key factor is long-term stability and low-tempco. Anything else is automatically nulled by the converter itself.
So you end up with a converter that needs no manual adjustment and can easily be replicated. Just assemble it with cheap off the shelf parts. The only key part is the integrator opamp ( low bias current on the inputs and stable ) ,the integrator capacitor (  agood quality foil or silver/mica cap ) and a precision comparator to find '0 charge'.

The system is autozeroing as well. the input stage can switch between 4 'signals' - a positive reference, a negative reference , 0 , and the to be measured signal.
Autonulling algorithm : switch to zero , wait a fixed amount of time , reset counters , switch to + ref , wait a fixed time , switch back to zero wait same fixed time ,  store residual count value.
the residual count + the count of the fixed time give you the offset in the convertor. done.
Measuring algorithm : zero the integrator , charge with fixed time from reference , switch to input , measure time needed to hit 'zero'.

Take all numbers ( residual count of autozero , charge time and measuring time as well as the 'reference times' stored during calibration ) and apply simple math ..
Calibration parameters are stored as 'numbers'. you don;t care about precision of anything. who cares if my reference voltage is 5.000001 or 4.78634.
I know that , if i apply exactly 5 volts at the input the 'time' for that is x this 'time' is my 'reference' ( i nothing in the machine has been changed since it was last calibrated)
So they store a 'time value' for what is'believed to be the internal '10 volts reference' as well as for the '-10 volts reference. And that's it.
Since it is a balanced system ( charge up , discarge back to zero ) even your crystal is allowed to drift ! the absolute frequency of the crystal doesnt even matter. It's all time ratios. If the crystal ticks a bit slower they charge longer . Charging is a fixed number of ticks.

need more digits ? sample faster in the same time window. Simply jack up the clock and charge count value. A matter of more flipflops and a larger number in the digital comparator that decide 'end of charge' cycle. So it scales beautifully with microscopically added cost...

So calibration is a matter of applying a known reference, storing 2 numbers and the machine can do the rest all by itself. For production and initial alignment this is a tremendous cost saver. you can calibrate the machine on all ranges , all units in under a minute. Fully automatic. Simply store 2 numbers for each possible configuration. They even store a different set of numbers if they switch between front and back terminals ... so contact resistances in relays and switches is even acommodated for. As for the cal setup : you only need a stable reference.. done. so that too is cheap.

The base principle (dualslope) is very old. even the icl7106 uses that ( as does the icl7035 and many other chips found in meters ) . the problem is that dualslope is slow and easuring time is dependent on the to be measured voltage. Multislope speeds that up by injecting known currents ( which are again stored as a 'time-ratio' ). What these current really are in terms of microampere is not important. al it matters is that you know : if i turn on this mosfet : i deduct this much 'time' for every clocktick i leave it 'on' from the count value.
When i hit zero charge ( the comparator toggles ) i know how many 'counts' are in the convertor , how many i 'forced' off and my autozero residual. Apply simple math...
Multislope II runs continuously and attempts to keep charge on the capacitor at 0. the rest is simply counters going up and down and looking at the delta's between current and previous you know where you are... so it is incredibly fast too. ( that's why some of these meters can do 10000 readings a second or more at high resolution with great repeatibility)

It is an amazing idea. Elegant , simple, and brilliant.

[Bored@Work]

That multislope technique is simply brilliant ( go read the patents on it .. this is why you don't find it on the market as a chip ... it is heavily patented and they don't licence ...).

Patents don't last forever. The stuff from the '80th and early '90th should have expired.

It is probably more because it is simpler and easer for One Hung Low companies to grap some cheap off the shelf IC from the '70th, copy the example circuit from the datasheet, remove everything they think isn't necessary (filter caps) or they can't afford (precision resistors) and mass produce it with unskilled labor.

[free_electron]

patents do expire , but the we are at generation 3 or 4 now. and every generation has been renewed. so it's been under patent quite a while ...

[tinhead]

remove everything they think isn't necessary (filter caps) or they can't afford (precision resistors)
and mass produce it with unskilled labor.

you made my day!

[vl400]

Thanks for the really good info free_electron, will never look at the supply the same again!

[codeboy2k]

Now, the dac is another interesting beastie ...
They use a 16 bit dac to do the setting of all voltages and currents. The dac has a multiplexer on the output followed by analog memories built around tl074 and capacitors
They continously replenisch the caps by scanning and reprogramming the dac.

Sample and holds synchronized with the multiplexer's output.

I have recently seen that in a DIY powers supply. The S&Hs were driven with a PWM signal from an AVR CPU via a multiplexer. I didn't get the reason for this, since DACs are easy to get and use. But now that you mention it, the developer might just have copied an old design.

The Rigol DS1052 does this too, using a single DAC to refresh 4-5 holding caps per channel, to set some levels , i.e. offset, gain, trigger level, etc
EEVBlog user A Hellene once posted his hand drawn schematics here on the forum. I can't seem to find the thread anymore.

I've not seen it anywhere else before I saw it first in the Rigol, and I thought it was pretty slick then; I haven't read the HP schematics but I will now.
It seems like HP did it first.

[A Hellene]

I can't seem to find the thread anymore.

Here is it.


-George

[codeboy2k]

I can't seem to find the thread anymore.

Here is it.
-George

Thanks George! Beautiful hand drawn schematics!

[A Hellene]

You are welcome.

[saturation]

Most excellent commentary free electron, kudos from me too.  I thought I was reading the 'theory of operation' section of the HP DMMs!

One quickie though, in the bold, you mean 3458a, yes?  I had to check the Agilent website as this DMM did not ring a bell.
If so, yes, 3458a was the culmination of the line of DMM beginning with the 3455a.

Thanks for the really good info free_electron, will never look at the supply the same again!

If you really want to see their 'magic' at work : service manual of a 34401 ! Or if you can find it : a 3485.. That beast has been in production since the dinosaur age ... and is still the 'gold standard' when it comes to multimeters. Heck these are the machines that today are still used to 'calibrate' the calibrators ... and its a 30 year old design ! ( mid 80's )

[free_electron]

eh, yes 3458A . i have a lysdectic syldectic dylsectic dyslectic memory

HP service manuals should be mandatory reading lecture in EE courses... They also used to have something called 'bench briefs' ( no not underwear ). the whole collection is somewhere on the web. There they published inteview with their designers and highlighted the techniques used. extremely interesting reading material.

When i moved into my new townhome last year i got strange looks because i would sometimes sit at the pool on weekends and read these big HP service manual binders with fold out schematics... Now they're used to the nutty guy because they know where to go when theri pc or tv don't work ..

[saturation]

Yes, agree, nearly all of them provide enough information to a skilled user to keep the product going far beyond its service life.  I haven't seen a recent design Agilent device being as thorough, except the carry overs from HP such as the 3458a DMM or the 36xx analog PSU.

eh, yes 3458A . i have a lysdectic syldectic dylsectic dyslectic memory

HP service manuals should be mandatory reading lecture in EE courses... They also used to have something called 'bench briefs' ( no not underwear ). the whole collection is somewhere on the web. There they published inteview with their designers and highlighted the techniques used. extremely interesting reading material.

When i moved into my new townhome last year i got strange looks because i would sometimes sit at the pool on weekends and read these big HP service manual binders with fold out schematics... Now they're used to the nutty guy because they know where to go when theri pc or tv don't work ..

[gerrysweeney]

@free_electron, great writeup on the workings of the E3631A PSU, thanks for post in.  Despite having repaired my own I did not get close to that level of detail and a few things fell into place as I read your posts here. I have to say using the PSU its very very good, as you say HP know a thing or two about this sort of stuff.  None the less I am still going to have a go a building my own, its my way of learning more about analog electronics than I currently know

[codeboy2k]

If you really want to see their 'magic' at work : service manual of a 34401 ! Or if you can find it : a 3485.. That beast has been in production since the dinosaur age ... and is still the 'gold standard' when it comes to multimeters. Heck these are the machines that today are still used to 'calibrate' the calibrators ... and its a 30 year old design ! ( mid 80's )

(I know you mistyped 3485)..

Is this the manual you are talking about, for the 3458A, or is there another service manual for that DMM?  I've never owned a 3458A so I don't know, but I would still love to read a good read.   A seller on eBay has 4 of these for $20 each.

http://www.ebay.com/itm/Agilent-3458A-Assembly-Level-Repair-Manual-/320603678046

[free_electron]

You can download the assembly level repair manual for free from agilent. All it tells you is what module is bad and how to swap the module. No schematics....
You need the CLIP... ( that is what hp calls their internal service manuals. component Level Information Package..) i got clips from some scopes. But the 3458 i have not found yet. They do not release these publicly. This is for their repair centers only. You need to know someone who works there and can make scans / copies.... Afaik nobody has ever seen the clip for the 3458...

[alm]

Component level information packages were intended for customers, not just for service centers. They were a compromise between the old way of providing component level service manuals for free, and the new trend of providing only assembly level information. They charged a fair amount of money for them to recover the costs. I'm surprised you haven't been able to find one, Agilent Find-A-Part lists p/n 03458-90033 as currently orderable for about $150.

There are also some HP Journal articles (probably in one of the 1989 issues) describing several of the key circuits and techniques in the 3458A. If you're interested in this kind of thing, reading the HP Journal articles about its predecessors like the 3456A and 3455A might also be informative. The HP Journal articles often described important circuits like A/D converters.

[codeboy2k]

You can download the assembly level repair manual for free from agilent. All it tells you is what module is bad and how to swap the module. No schematics....
You need the CLIP... ( that is what hp calls their internal service manuals. component Level Information Package..) i got clips from some scopes. But the 3458 i have not found yet. They do not release these publicly. This is for their repair centers only. You need to know someone who works there and can make scans / copies.... Afaik nobody has ever seen the clip for the 3458...

Ah ok. Thanks for the info; good that I checked before shelling out $20. It's not that much money, but I don't like to get scammed either.
The seller might have downloaded the PDF and is making prints at Kinkos and selling those.

[casinada]

Free_Electron,
you can download the HP_3458_CLIP.pdf  from the KO4BB repository manuals site.
 

[A Hellene]

Excellent piece of information on the HP_3458_CLIP.pdf. The filemane worked miracles in the search engines.
Thank you!


-George

[free_electron]

I just downloaded it. The first 20 pages or so are detailed explanations of how the A/d works in that machine. It explains multislope in greater detail than i did. Like how they avoid saturationg the integrator by already starting to inject current during the first phase. They aslo swap the slopes. A normal dual slope charges first from reference then discharges from input. They flip that around because then they shorten time by playing with substraction during slope 1 ( called runup , slope two is called rundown ). They essentially digitize 4 digits while the capacitor is still charging.. The remainder 4 digits are got during discarge.

This machine is totally bonkers... All discrete fets everywhere. Amazing how they get this thing to do 81/2 digits. And it was made in the early 80's and is today still in production.. Unmodified. And still the gold standard.. Highly recommended reading material

[Christe4nM]

@free_electron: those 20 pages of detailed explanation, which file are you talking about? I downloaded the CLIP, the service manual and a HP Journal from 1989, but I cannot find it. Am I overlooking something here? Also, may I ask at which forum I can read about that multislope tryout on a breadboard?

[Stephen Hill]

Where did you get your screwdriver from and (more importantly) can you get them in other colours?

[free_electron]

The clip that A.Hellene pointed at is incomplete... The real clip is 204 pages ..... Not 66.
I will host and place a link.

The other forum is in dutch .. ( flemisch) ... So unless you can read that.. Bummer , but here is the post : http://www.circuitsonline.net/forum/view/94488/10/circuitsonline+multislope , somewhere halfway that page. We were discussing hp34401 versus the new fluke ( the linux running thingie 8xxx series. I had done a teardown on a 'loaner'. Versus keithley 200x versus some others. We were discussing converter techniques. One thing led to another....

this was the first build ..


And then the volt- heads took over ...
http://www.circuitsonline.net/forum/view/95584

[WorldPowerLabs]

Very nice post -- appreciate all your effort...

I used to use these power supplies.  In my experience, they are extremely durable and quite trouble-free.

I'd also like to know the source for the screwdriver...

[Stephen Hill]

Found a few Screwdrivers which look very similar:

http://www.flightboxrc.co.uk/acatalog/Pliers_and_Wrenches.html

http://www.redrc.net/2012/03/eds-metric-inch-combo-drivers/

http://www.arrowmax-rc.com/index.php?main_page=product_info&products_id=346

[T4P]

The coloured ones are common in the hobby/rc channels

[krivx]

The other forum is in dutch .. ( flemisch) ... So unless you can read that.. Bummer , but here is the post : http://www.circuitsonline.net/forum/view/94488/10/circuitsonline+multislope

Google does an OK job for anyone else who wants to follow along http://translate.googleusercontent.com/translate_c?hl=en&ie=UTF8&prev=_t&rurl=translate.google.com&sl=auto&tl=en&twu=1&u=http://www.circuitsonline.net/forum/view/94488/1&usg=ALkJrhh2quwO04ziKdxEDyJBOB_V5bPJUA

I would lobe to see the complete version of that CLIP...

[vl400]

Where did you get your screwdriver from and (more importantly) can you get them in other colours?

They are just cheapies from ebay (search something like precision screwdrivers), the picture showed them as black but arrived pink

[codeboy2k]

The other forum is in dutch .. ( flemisch) ... So unless you can read that.. Bummer , but here is the post : http://www.circuitsonline.net/forum/view/94488/10/circuitsonline+multislope

Google does an OK job for anyone else who wants to follow along http://translate.googleusercontent.com/translate_c?hl=en&ie=UTF8&prev=_t&rurl=translate.google.com&sl=auto&tl=en&twu=1&u=http://www.circuitsonline.net/forum/view/94488/1&usg=ALkJrhh2quwO04ziKdxEDyJBOB_V5bPJUA

I would lobe to see the complete version of that CLIP...

me too. Waiting for free_electrons link ... thanks in advance!
After I read it I am going to build me a 6.5 digit multimeter just for fun 

[free_electron]

Kurts manuals has the correct one.

http://128.238.9.201/~kurt/manuals/manuals/HP%20Agilent/HP%203458A%20Component-Level%20Information%20Packet.pdf

[A Hellene]

Fabulous site!
Thank you, free_electron!


-George

[gerrysweeney]

There is a pretty detailed explanation of the multislope integrating converter on Wikipedia here. http://en.wikipedia.org/wiki/Integrating_ADC, looks like the HP description so I guess the detail comes from the patent.

@free_electron  You said in an earlier post, "You'll have a very VERY hard time making a supply with the same specs (current / voltage / load response, readout accuracy and precision , floating outputs , gpib and serial control ) for even double the price of what the supply goes for new off the shelf at agilent. The isolated floating outputs alone will kill you in cost.. Youll need one cpu per channel, send it through optos, have another earth referenced cpu and you wont even come close to the precision of control and readback... You need like 18 bits or more..."

I will take that as a challenge then I am not so convinced about your assessment, its defeatist - its easy to give up and tell yourself you can never do better, it’s much harder to actually try and do better and of course be prepared to accept success or failure as an outcome - that my friend is called progress - and the human spirit is a master of making progress  - so thank you for setting me the target......  I am not for one minute supposing that I can match the R&D capability, knowledge and knowhow that HP have out of my home office here in London with a small hobby workbench (that ironically is stocked with some pretty decent HP/Agilent kit), but between 1992 and now a great deal has changed in the world. 

Firstly, components and availability of the same has been transformed, accessibility to information, data sheets and other designs are at everyone’s fingertips because of the internet, contrast that with 20 years ago when building a datebook/sheet library was the most important thing to do as an electronics engineer.   Secondly, components are much better and more easily available. When HP designed their stuff in the late 80's early 90's high precision op-amps DAC's and ADC's where ridiculously expensive - not the case anymore.  Digital electronics and processing speed are so far advanced today;  a "silly PIC or Atmel" as you referred to them in your earlier post will outperform the 80's style CPU/RAM/ROM  found in the E3631A a hundred-fold in an embedded design and will cost you less than $2 in a one off quantity.  Setting all of those advantages aside, much more is generally known about the theory of operation around regulators, the pitfalls and analogue system design, so even a novice like me can understand most of what is going on.    I have some alternative ideas about getting the resolution I need for the DAC's using software and Delta Sigma modulation, that’s the advantage of having one dedicated 10mips ($2) controller per PSU channel , your assessment for result ion of 18 bits is incorrect by the way; to match the range of the Agilent 3631A I only need 13 bits resolution including headroom for calibration.

I don't really mind if I fail or succeed, but I will try, and I will share with anyone else who is interested what I learn from trying.  I don't know everything there is to know about analogue electronics, actually far from it, my knowhow is more in the digital domain, so if you or anyone else is able to offer any suggestions or advice as I make progress, (so long as it’s not too defeatist that is it would be great fully received!  I am only on Part 3 of my attempts and have already hit some problems as I expected to but have overcome them and learned a lot along the way.   

Gerry
http://gerrysweeney.com/

 

[free_electron]

You will fail...
Here is why :
1) the transformer you need will bomb your budget...
2) on the 18 bit: that os for the meter part, not the dac part ... This thing measures 5 1/2 digit. 16 bit gets you. A max of 65535... So you need 17 bit assuming you pick up no noise.... Make it 18 or 19 then....
3) how are you going to implement gpib ?
4) heatsinks are custom jobs...
5) you will need multiple cpus . Since the channels are isolated. So you need an outguard cpu for the comms , and at least two more ( one for 6 volt, one for the +/- 22 volt channel )
6) display ? I don't want to see a stupid lcd ... Needs to be a vfd, starburst preferred...

That's what i mean with ' you will have a hard time duplicating it... ' at that price. It is not economically feasible. Just the cost of the parts alone is bombing the project...
As a hobbyist you cant build anything that comes close to these. If you want to shave it down like using an lcd , forfaiting gpib , forfaiting isolated channels , forfaiting the readout precision and accuracy, forfaiting tracking capability and a bunch of stuff. Yeah you could make something. But it is stilll going to be more expensive thana wingpangpong with the same specs as what you built.

Of course thereis still the thrill of trying and that's what we do it for. On the other hand... You can pick these supplies up , used but fully working, for as low as 300$...  Then it becomes a real hard nut to crack...


As for 'advancements in regulation' ... There is none. Same circuitry is still in use. These are proven regulator designs. Been in use for 40 years. Can't really improve them.
By the way, the precision opamps are today still 8 $ a pop ...

[gerrysweeney]

Well despite your pessimism I am not going to give up just yet

1. My aim is to create something modular so you can use the modules to make a single, two, three or four channel PSU, just keep adding the modules and the transformers.  I am designing the PSU module to work from a single secondary winding so you can use low-cost off the shelf transformers - remember, just like you can pick up these PSU's for $300 (which by the way if you know of any would buy a couple more at that price), as a DIY'er you can also pick up transformers cheap.   If CATII/III isolation is important then you can always use one transformer in the box.

2) Yes of course you are right I was not thinking of the meter, the ADC chip will be the most expensive part by far, I think a decent dual ADC for this resolution will be $6-8 a piece. I have not done much with high resolution ADC's, its something I need to play around with, I was thinking of something like a 16bit decade counter chip in conjunction with one of the the MCU's 16-bit timers would give me a 24bit counter, I fancy having a go at a dual or multi-slope implementation in software.  Failing that, there are off the shelf ADC's that are high enough resolution for not so many $'s a pop.

3) This is one area where I don't see much point to be honest.  GPIB is an old standard, it belongs in the dark ages of the centronics parallel port and there would appear to be less and less support for it as serial protocols like RS232/485, USB and in particular ethernet seem to be much more current, especially since SCPI is supported over these.   That said I am well aware that GPIB is still in widespread use for test automation etc, I don;t know that much about it so I can't really comment on it apart from it can not be so hard to implement, its really just a parallel bus of some form.  This is certainly not on my list of priorities but if I get some time I may look into it as an academic exercise.  My thoughts are with RS232 (simplicity) and Ethernet (TCP/IP) because anyone can plug in and make use of it.

4) Enclosures are a whole different problem, something that would be great to solve.  However, what I am trying to achieve is the electronics, for s DIY build the enclosure and heatsinks would be left up to the builder to accommodate.

5) Yes, they are cheaper than most other IC's nowadays so not a problem, in fact anywhere I can replace a chip(s) with a CPU it will save time and money (GPIB and their custom chips spring to mind, I think a decent PIC could possibly run a 1Mbit parallel buss).

6) I too like VFD's very much, but as you say they are expensive so not accessible to everyone. Of course I am trying to make something that is modular so there are options. For those of us that could get a VFD for cheap (I recently bought some 64x256 graphics ones for £5 each) then its an option, for others a "stupid LCD" might be acceptable....

Apart from the fact that I am not actually trying to go into production and compete with HP, the point was really to create something that is every bit as good (spec wise) but on a DIY budget - maybe I can and maybe I can't but I am still going to try

The precision op amp used in the HP slope detector is about $5 at a one off price from one of the most expensive R&D sources, they go down by 50% at a quantity of 100 (http://uk.farnell.com/analog-devices/op2177arz/op-amp-dual-precision-smd-2177/dp/9994238?whydiditmatch=rel_3&matchedProduct=ad706&matchedProduct=ad706&whydiditmatch=rel_3) so not too bad, but there are plenty of other more modern devices that are even better and cheaper.

I was not thinking about advancements in regulation specifically, they are what they are, what has advanced is the components you can use to create the thing and all of out control circuitry around it.

You obviously have a very strong conviction that this can't be done which is fine but I am still going to try, there is nothing like finding out for ones self!

[SeanB]

Just remember a 16 bit DC accurate ADC is not the same as an audio 16 bit ADC. The non audio device has much higher requirements with regards to reference noise, stability and monotonicity as well as needing an order of magnitude more stable internals. Audio devices are fast and reasonably accurate, but at DC you need it to be a lot better than audio, where it is hard to detect that the bottom 2 bits or so are not really usable due to noise and offsets. The ADC will not be a $2 part at all, probably will have a name like AD or TI, NSC and will cost $ 20 or more in volume, and will be only available in limited package types and accuracy specs.

[gerrysweeney]

Hi Sean,

I was looking at ADC's designed specially for instrumentation, these are the sorts of thing I was looking at http://www.farnell.com/datasheets/62412.pdf they are about £8 ($12) as one off's so not too bad.  Not entirely sure a differential input is OK for what I wan't but there are single input ones too.  I am going to try a DIY slope converter for academic reasons but if push comes to shove there are plenty of these types of devices available.   

Gerry

[free_electron]

You obviously have a very strong conviction that this can't be done.

what i'm saying is that you can't design an identical supply at that price. Identical in terms of electrical specs, performance , io and communication.

The linear section of these supplies os nothing special. I guess you can replicate that for 10$ without mechanics or transformer. After all its only two opamps, a current source and a fat transistor.
Slap on a dual channel 16 bit dac to set cvoltage and current, under control of a cpu over an optical rs232 link. I would look at analog devices aduc8xx series cpus. Cheap and they have abuilt in 24 bit A/d ... You can drive a local lcd display and rotary encoder. Display voltage and current. Push encoder to toggle between current and voltage. The moment you turn , the set values are displayed. One second after last turn it switches to real values. Use one extra button for channel pn / off.

Three optos doe the work two for rs232 one for global on/ off.
The chasssi processor needs two uarts or one uart and ethernet.
Chassis processor controls global on/ off and powerup sequencing and does interfacing.
You can use the intel 9 bit uart mechanism where the ninth bit signifies address/ command.
Very easy to make a distrbuted system that way

[gerrysweeney]

@free_electorn - I agree as a one off thats not possible, but that was never what I was trying to do. I  would say though, that if one had enough conviction it could be done, you could compete with HP, you just need to design something well (like they have), do some reasonable volume (like they do), make it in china (like they do) - after all, Agilent would appear to be able to do it and make a margin doing it - or perhaps they can't; that could go some way towards explaining why HP are in a financial hole

I will have a look at the aduc8xx series chips, I was not aware of them.  That said, I know PIC's well and have all the tools, code libraries etc so that should be much faster for me to get up and running and debug - they don;t have a 24-bit A2D built in though, but as I said, I do like the idea of creating a dual or multi-slope ADC just to learn more about the approach.   Keep an eye on my blog for progress, you won't find the first few parts of much interest, its just me coming up with my own variation of a CV/CC controlled regulator circuit as I get more familiar with the nuances of regulator design.

Thanks for the posts, certainly food for thought and an excellent source of motivation

Gerry

[free_electron]

@free_electorn - I agree as a one off thats not possible, but that was never what I was trying to do.

even in moderate volume it will be hard.. You'r ebritish. Thurlby Thandard ring a bell ? They make good power supplies. I got multipleof their QM series ( three channels gpib programable. I also have xantrexes ( which are TTI designs , not real xantrex / sorensen )  shitload of electronics inside...

Afaik the agilent supplies do not come form china but from their penang - malaysia plant. Agilent has had an assembly plant in malaysia since the 70's.

i would not try tocome up with my own regiulator circuit. anything that can be done has been done on the linear front.... the best and easiest regulator is a twin opamp where the current monitor pulls down the output of the voltage opamp to provide the foldback. bulletprrof , rock solid ,simple.
if you need some speed in this : drive the base of the powertransisotr using a current source.
Go get the manual of the hp supply and study the 6 volts section... it's hard to make it better than that. it is simplified to the extreme and has stellar impulse response and load regulation. use tht as the basis of your design and scale it up in voltage.

other similar designs are the e3610 series. that regulator method ha been around since the late 70's and it is still being used any almost any supply out there. simply because it is hard toimprove on it.

you could throw in a preregulator using two thyristors in a bridge to bring down internal dissipation. agilent uses that in the e344x series i believe.
other machiens to study : 6624 quad channel machines... the regulator there is very clever ...

[gerrysweeney]

@free_electron - yes I heard of TT, they always reminded me of low-end professional gear, I am sure they have gotten better of late although, but being a british company they are probably not at the bleeding edge of the design spectrum - us brits don't seem to do engineering/manufacturing as well as the rest of the world.

In terms of the regulator circuit, you are absolutely right, I am not trying to invent anything new, but part of this exercise for me is learning what I can about how these well proven designs actually work, what problems they have and how you solve them; you can get a sense of whats right or wrong by looking at other working designs but for me at least, I only truly understand once I have experienced the problems and solved them myself.  What you describe is pretty much what I have ended up with, a pass device, a couple of op amps and a pre-regulator switching device.  I am a couple of steps ahead of what I have published in my blog so far. 

I spent a fair amount of time pouring over the 6v regulator circuit on the 3136A when I was repairing my own, the regulator its self is fairly simple as you say apart form one thing - the use of multiple bias supplies floating around the main regulator supply - a mains transformer to get all those windings would be a custom job and expensive, I might be misguided but I think I can design that need out which in turn makes things simpler.

I am using a FET as a switch for the pre-regulator, simpler and keeps the BOM down (as I use the same FET type as the pass device too), I have not yet put in the analog feedback loop for the pre-regulator, so the controller currently drives it (and yes I know it won't be fast enough and I will be changing it

I did have a constant current source in the driver stage but removed it again, I need to do measurements although just from playing around I think it might be fast enough already, but as I say I can't confirm that.

I will take a look at the other designs you mention although by and large I am there with a working regulator thats good for 50 watts and can be configured for various ranges 0-6/5A, 0-30/1A and some inbetween.  I am using single supply op amps and I think this is why I have a 2mv offset problem which I have yet to tackle.  So nothing new, but I have been through the pain of making it work, getting it stable and understanding the regulator circuit behaviour - big step for me and my own knowledge.  I have also tried to keep the BOM down from the off, so simplicity where possible...

Cheers,

Gerry

[T4P]

Actually if they were made in china it would only last 2 years no?
But they are made in malaysia because the stuff they make in malaysia is pretty good (hakko made the 936 currently makes the triple-8 there as with all others)


Or you can also buy a 120W switching supply from aliexpress and have a postregulator to filter out the noise and also another 24W switching supply as the 5V supply line

As for the post regulator you probably know where i got the idea from ... DJ !

[gerrysweeney]

@DaveXRQ @free_electron - My mistake, I knew they were made in the far east, now I know specifically they are made in malaysia - thanks for the correction.

@DaveXRQ - I have seen the switch mode/post regulator approach before, its a nice idea. If I had anything like the knowledge needed to build a reliable switch mode I would probably give it a go, but alas.....

Gerry

[ben_r_]

Very glad I found this thread! There is some crazy good info in here! I just picked up a used E3631A for $450 and am awaiting delivery. FWIW I called Agilent and found out they charge $146 for a one-off calibration. I might do it seeing as how the one I bought hasnt been calibrated in almost ten years, but I havent decided yet. We'll see when it comes and after I do a bit of measuring of my own.

[boris.t]

Hello.
I have problem. My 3631 say me 550 error for MEAS:VOLT? and MEAS:CURR? over GPIB.
Error 550: "Command not allowed in local. You should always execute the SYSTem:REMote command before sending other commands over the RS-232 interface."
But STST:REM not allowed for GPIB communication.
I have no idea how, but it does not work. Hope to heal it.
Where I can find firmware files?
Sorry. It is Keysight IO Libraries problem, solved by installing previous version.