high-side current sensor

For situations like this I tend to like using a simple moving average filter with the window being variable. If so inclined, could be made adaptive. First or second order biquads can also be effective. I usually chose direct form one implementations

I'm fond of making digital filters out of exponential lags or integrators, starting with a classic analog filter and essentially Spicing it inside the FPGA or uP. My fpga folk really don't like that for some reason. Not enough Zs or something.

They like filters that have a constant of integration in the output, and assume that nothing ever goes wrong to make it nonzero. A sinc3 filter terrifies me.

Recursive filters can exhibit limit cycles, though. For a simple example, if you put a step function into a one-pole IIR lowpass implemented in fixed point, with a decrement of 0.01 (say), it will get stuck somewhere around 50 units away from the 'true' value. That's DC, but it's still a limit cycle. Fancier filters can have AC limit cycles.

Cheers

Phil Hobbs

I still like an ordinary 36V dual or quad RRO op amp for this, running off +26ish and ground, with a good quad pack such as an ACASA1003S1003P100 (100k, 0.1% absolute, 0.05% relative), 37 cents.

That puts both inputs near +24V with the output near ground. There'll be a little bit of offset, but nothing major, and since you can switch off all the loads, you can measure it at POST. A little bit of intentional imbalance will prevent the output from really going to zero, so the op amp and ADC will both be on-scale at all times.

Cheers

Phil Hobbs

I could mux into a single official highside measurement chip. But we'd still need to control the mux without slowing down the scan much.

AC couple the mux selects? Or use a fast digital isolator, with highside power.

If I can actually buy the highside sensor chips for 80 cents each, we can just use 9 of them.

We plan to put a fuse on the +48 input to each of the power module boards, to avoid internal forest fires.

Here's the box concept:

If you do the exponential smoothing equation with K=0.01, in 32 bit scaled integer math, it works fine.

In really dinky machines, the trick is to do

Out = Out + (In-Out)/2^n

which is just a right shift.

Run the logic outputs into the emitters of a quad PNP via resistors, and use the same resistance on the high side. Hang the bases on the logic supply. Works better at 3.3 or 5 V than lower rails--the logic threshold is 30% to 70% of VDD.

Sure. Press on, next problem.

Good idea.

Phil Hobbs

torsdag den 13. januar 2022 kl. 17.49.33 UTC+1 skrev snipped-for-privacy@highlandsniptechnology.com:

$0.5803 if you buy 100

the AD8479 is about 10x the price of a part made for the purpose

But Vout - Vin doesn't decay all the way to zero. With lots of bits and intelligent scaling, you may not care, of course.

Sure.

Cheers

Phil Hobbs

Nice. We could buy 1000.

We could do the dual layout to have a stuffing option. That's life these days.

Something like LM331 is half the price, and doesn't tie up ADCs, because frequency counting is already logic-level info. If overhead counts, I've always liked frequency conversion because it allows the CPU to tailor the resolution and response, with minimal latency at each status check. Think of a motor-bearing analysis, accumulating the harmonics of the rumble of rotating machines, for days... a big digital accumulator and phase-locking works great with voltage-frequency conversion.

Yes, I suggested the related ZXCT1084 up thread a while back.

If you're in the mood to mux then you could take your 2xPNP current mirrors and parallel all the measuring collectors into one resistor/ADC and selectively enable each mirror with a HV open collector driver (thinking old school ULNxxxx or discretes) ?

piglet

Maybe you don't even need the open collector enables, a 1/8 or 1/9 drive should AC couple OK. I guess it depends whether a 60V capacitor is going to cost more than a 2N7002?

piglet

Me too. But the charm of the AD8479 is in the four closely matched resistors.

Getting precisely calibrated frequency generators depends on getting precisely calibrated reactive components as well as resistors, so it is piling Pelion upon Ossa.

Precision resistors aren't cheap, but you can buy them. Equally precise capacitors don't exist. Precision inductors are even less precise. You can get stable parts, and calibrate them, but it's a good idea to recalibrate frequently and automatically.

The frequency generation is driven by the sensed signal, but the AD8479 and its ilk use 100:1(?) pre-attenuation and that requires roughly 100x the precision of those matched components. Only the 3 -terminal current monitor chips like ZXCT1009F are actually amplifying straight off the sense resistor.

1% capacitors are plenty good enough (unless a 1 kW power supply cannot be derated to 990W plus/minus 10W). For the AD8479 gain/offset, you need resistors better than that, .01% or so.

It's mainly about aging, and for the resistors, also matching a quad. Either can be calibrated, and neither should drift much. There's some auto-zero instrument amps, just in case you want extra security.

Maybe use the idea with the resistive divider, but then don't use the ADC, use a slope converter for increased precision. That can be done with one cheap comparator, and a loop from the FPGA/Micro

Add a resistor with known drop, to calibrate

If the voltage range is not excessive you can use a differential input on the FPGA as a comparator. If it is excessive you can scale it down.

It's fun to design circuits, until some clown sells an IC that does it all for you cheap. Looks like we'll use INA280, which is available and costs 80 cents, according to TI.

Digikey wants $3. Digikey has become a scalper.