I strongly endorse the use of such a circuit.  I've been using them since my earliest work in power electronics (above 1kW), and would still be able to hold all the devices I've destroyed, in two hands.  That includes one big industrial module, which is the main reason they're probably not all in just the one hand.

I always used a comparator circuit, which provides more accurate measurement of Vce; this might be of interest for crude current monitoring (because Vce depends on Ic) or for use with MOSFETs (which benefit from exactly the same treatment, they just don't need to respond so fast -- thanks to the lower power density, 20-100us may be adequate time).  For detecting the rapid rise on leaving saturation, nothing fancy is needed at all, and indeed a few BJTs will do the job.

A note: you'll want a B-E resistor on the "sense" PNP, to prevent it from accidentally turning on.  Also mind the capacitance and reverse recovery of the sense diode: there's massive V and dV/dt on the other side of it.  The series base resistor helps, but I would consider a E-B clamp diode, and maybe RC filter as well.  Also check the relative values of collector and base resistors; I assume these values are just for a quick simulation, not an actual build, so just remember to check this later.  (The high collector current, ~50mA, is rather hungry for a gate driver; but that can come down substantially, no problem.)

Note that any filtering you add, slows the response.  You can afford a few microseconds, as long as it's still a modest fraction of the total switching waveform.  Too much, and obviously, there's just nothing to sense, you'll need a different method.

Back then, the conclusion seemed to be that desaturation detection was not fast enough to save an IGBT experiencing a huge overload in a DRSSTC.

It is a perfect counter to the situation it targets.  There are other situations that result in failure, however.  For example, plain old overheating -- whether from switching too slowly (large t_r, t_f), or too often (high Fsw), or drawing greater than rated load current (but still without going into desaturation).

To cover these, the basic method is to simply limit peak, average or other load current.  Monitor with a suitable sensor (current transformer, say; Hall effect sensors are pretty slow, but also adequate when Fsw is low) and shut it down if it exceeds nominal.

For more finesse, you'd unfortunately need a more complex method, perhaps a real-time measurement of device voltage and current, the product of which is fed into an equivalent thermal impedance model, to infer real junction temperature.  This is a PITA because analog multiplication is hard.

That SOA curve looks a bit wrong. Usually they're shown in datasheets like this:

As you can see the SOA is also a factor of time.

The highlighted region corresponds to the leftmost column; I suppose arguably, as long as it's under the DC curve, one could say time actually isn't a factor.  But that changes quickly at only little higher voltages!

Tim

FYI, gate capacitance is apparently quite trans-linear, the downsides being high resistance (for power devices anyway) and low voltage range (-Vgs(max) < Vgs < Vgs(th)).

Not sure if RF devices exhibit the same response; it may be specific to VDMOS and not LDMOS.  That is, I would expect LDMOS may have more Cin(Vgs) dependency than VDMOS (which has essentially only Cin(Vds) dependency).

Tim

Try more capacitance across the primary?  Idunno, it's a dumb oscillator (if it's the usual simple circuit), it can do whatever it wants, and transformers are complex networks with many characteristic frequencies.

Tim

The IGBTs turn on and off hundreds of times per burst, and there are hundreds of bursts per second in the example.  Depending on exactly what settings you've configured.

The rate at which the IGBT change state, from off to on or vice versa, is on the order of 200ns, one five-millionth of a second.

Tim

I am last in line in Denmark, long way ahead as they are currently at those born in 1951 and prior.

Upshot to the... *gestures vaguely at the entire last year plus*, we're pushing a hell of a drive, and vax is opening up to much of the public (varies by state).  Rollout has been slower elsewhere, probably as we finish up we'll send over a bunch of our leftovers?  Hang in there, stick with precautions, and keep the air sanitized with ozone.

(But not really... but, I do wonder, how effective are corona or spark discharge actually against pathogens, not sure I've seen a study on that?)

Corona doping is only 95% effective, see attached picture I took outside of the city power plant

Its taken in pitch dark at 30 second exposure and ISO 12800 in case anyone wondered...

Lovely shot!  Wonder if they have get any RFI complaints...

Tim, I hope it works out for you. You know, the South African variant B351 is 3 times worse than B117, since 351 = 3 * 117. The mutant B617, which is believed to be the cause of the steep rise in India is even 500 points worse that B117. 

That's like five whole times worse!

But seriously, hope they get it under control quick.

Tim

A lock-in amp is also a correlator, or very narrow bandpass filter.  Simply correlate it with any signal that is linearly dependent to the waveform.

I suppose given the sectors are... sectors, it has a more triangular waveform?  So, another triangle would be easy enough to generate also, and would correlate very nicely with it.

The simplest form is just inverting the signal every half cycle; instead of a rampy waveform you get pairs of peaks, which average above zero.  So, filter out the ripple and there's your detection.

Well, simplest in terms of implementation -- just use some analog switches or whatever and there you go.  Mathematically a sine wave would be simpler (pure frequency mixing, hetrodyning), but often this isn't as easy to do (needs an analog multiplier).

You can also consider orthogonal pairs, namely, quadrature -- this shouldn't tell you anything about the field (Q should always be ~0, and any error should just mean timing is off), but is interesting in general: for example, measurement of impedance (real and reactive), or receiving a radio signal without a perfectly locked (coherent) reference clock.

Tim

Why not just use a UC3842?

If discrete is desired, here's a seven10 transistor* model:



*TL431 is basically an improved transistor. Trust me, I know how to count.

Missing from this schematic is cross regulation (take equally weighted feedback from each output), but otherwise it works.  Scale up output in the usual way.  Control is peak current mode with frequency control as well, which is similar to pulse skipping at light loads, but smoother.  It may have weird chaotic modes at the extremes of supply voltage or under heavy load; this is just one of the realities of discrete designs, and also of peak current mode controls incidentally.  Note the diff pair sensing switch current: it wastes less power on this than the usual Vbe triggered reset circuit, at the expense of two additional transistors.

Tim

Heh... class C is typically 60-75%, with an upper limit of I forget what, maybe 80%?

Tubes in class D can do a bit over 80%, but may require quite high grid/screen power dissipation, or careful drive and load matching, to achieve.

Tim

Sure, but what for?

I think you'll have problems doing it at high frequency, for multiple reasons:
- Note the balancing capacitances in the first example
- The avalanche diodes in the second example also have capacitance; worse, it's nonlinear
- The stack is long, i.e., stray inductance is unavoidable; at the very least the load current per stack will have to remain within some limit

Preferably, the assembly will also be very compact, perhaps a hybrid, and potted in some kind of goop to keep corona away and allow much smaller clearances.  This isn't friendly to prototyping.

You can apparently already get SiC devices up to 10kV or so, which are either academic samples, limited production, or special qualified customer access (probably because ITAR and such).  I haven't seen anything over 1800V in the usual places (well, some MOSFETs up to 4kV, but not very big or fast).

Tim

Show board layout.

Also, that looks to be a #26 powdered iron core in the GDT, don't expect a great frequency range from it.  It should be hi-mu ferrite, for least loading of the drivers and extended LF range.

Tim