GDT keeps on killing IGBTs

[johnf]

My god
what a mess
you need to twist the transformer output wires together 3 or more twists per cm
this helps stopping picking up stray signals and or radiating signals into circuitry it shouldn't.

[davekni]

Thank you for the images.  I'm presuming the gate-drive schematic sketch with the 5.8-ohm resistors is replicated to drive 4 IGBTs total from two GDT outputs of your controller.  That schematic looks good.  If gate-drive undershoot voltage gets too high, adding a lower-value resistor in series with the diode can help, at the expense of less dead-time (slower turn-off).

As John said, twisted pairs of wires would be good for both the GDT leads and current-sense leads.  For scope monitoring, a coax cable would be even better.  Keep in mind the pairing of the GDT leads even when adding the resistor/diode sets.  The resistor and diode sets should be close to the IGBTs, with the other emitter lead of the GDT secondary in close proximity to each set.

I can't tell from the image if your 10-ohm CST load resistor is wire-wound or not.  Wire-wound resistors have relatively-high parasitic series inductance, which could be amplifying high-frequency signals (ie. spikes).

I like your litz-wire primary winding - thought I might be the only person using it.  Litz isn't very important for the leads, but is useful within the coil itself to reduce losses.  For the primary inductor and MMC leads and H-Bridge output wiring, twisted pairs aren't practical due to the relatively-high voltage.  Keeping the wiring as close to each other as practical (but still far enough to not arc over) is helpful.  Reduces stray fields to couple into control wiring (current sense, gate-drive, etc.) and reduces parasitic inductance (increases effective coupling factor to your Tesla secondary coil).

Concerning phase lead, I'm surprised there's no adjustment.  The FPGA is likely implementing some phase lead, but not sure how it would know how much is needed for your IGBTs and wiring etc.  It's hard to design an external phase-lead circuit without knowing the controller's input schematic.  I'd suggest getting the wiring as clean as possible (twisted pairs), then some more scoping to see if phase-lead is truly insufficient.  (Do you have two-channel scoping capability?  Then you could scope current and H-Bridge voltage simultaneously, making phasing easier to measure.)

[ritaismyconscience]

I can't tell from the image if your 10-ohm CST load resistor is wire-wound or not.  Wire-wound resistors have relatively-high parasitic series inductance, which could be amplifying high-frequency signals (ie. spikes).

I scraped away some of the ceramic and the inside looks black. I'm pretty sure my resistor is not wirewound. The resistor used by the CPLD board is SMD so it's probably thin film.



^What's inside.

Reduces stray fields to couple into control wiring (current sense, gate-drive, etc.) and reduces parasitic inductance (increases effective coupling factor to your Tesla secondary coil).

Right now, the primary connection wire is quite long, as I still have to tune it so I don't know where to cut it yet. Probably around 50cm extra.

I'd suggest getting the wiring as clean as possible (twisted pairs), then some more scoping to see if phase-lead is truly insufficient.  (Do you have two-channel scoping capability?  Then you could scope current and H-Bridge voltage simultaneously, making phasing easier to measure.)

I have the 1744A oscilloscope. As far as I know, there's two ways to display both traces, ALT and CHOP. CHOP is way too slow (250kHz), and ALT gets the traces from two different runs and puts them together. Perhaps I could press A and B to display A+B on the oscilloscope, then try to reconstruct A and B? I don't have a digital scope, which is a problem (also, I don't have proper oscilloscope probes). Someone who knows more about analog oscilloscopes might be able to help me here.

[davekni]

Chances are that the circuit behaves the same each time an enable pulse is received.  Presuming so, and that you have a stable trigger setting on the scope, then dual-channels should work fine.  Even chop works well if the scope's persistence is long enough to cover many trigger events.  Alternate is fine too.  If the circuit behaves differently each time, then that's something to chase first.  (I started engineering before digital scopes, and worked for a scope company, Tektronix, so am familiar with analog scopes.)

I understand your reason for long primary wires.  To mitigate the effect of the long wires, I'd suggest running the leads parallel to each other, close together, but at least 10mm apart to avoid arcing between them.  Use some cheap tape to tack the coil leads to a scrap of plastic (piece of PVC pipe for example) or other such insulating item.  That will minimize the excess lead inductance with its associated stray fields.  Placing the primary coil away from your drive circuitry, at least for this initial testing, will eliminate coupling from that coil as a possible issue.

Without scope probes, it's especially important to use coax (ideally) or at least twisted pair wiring from whatever you are measuring to your scope.  Route the scope connection(s) directly away from your circuitry, not adjacent any other wiring.  Gate drive and current-sense wiring needs to be twisted pairs too.

I can't see for sure from your pictures, but another issue that probably needs addressing is VBus inductance (power to the H-Bridge).  That needs to be extremely low, even lower than any of the other wiring.  Most designs have polypropylene capacitors directly at the IGBT terminals (high-side collectors to low-side emitters).  Using individual TO247-packaged IGBTs adds some challenges to low-inductance interconnect.  (My DRSSTC uses 10 TO247 parts in parallel for each switch, 40 total.  Designed a custom ECB just for the power connections - IGBTs and VBus capacitors.  I'd suggest at minimum some form of copper plane for VBus-, such as copper foil.  Or, better, use a piece of raw un-etched ECB material, with one side for VBus- and the other side for VBus+.

Finally, using TO247 IGBTs, make sure the gate-drive emitter connections are at the IGBT emitter leads, separate from the emitter power connection.  Higher-power IGBTs usually have two separate emitter terminals just for that purpose - one for power and one for gate-drive.  This is referred to as "Kelvin connection".

[ritaismyconscience]

I fixed the wiring today and I replaced the FB and the GDT wires with coax cable. Here's some pictures:



^ the blue wire is for the OCD, so twisted wire is probably good enough for this

I just wanted to make sure I was doing everything correctly before I insulate the wires.

Some more stuff about the board: The thick green and red wires go to the + and - of the capacitor bank. (My red wire is a bit long because I needed to connect it to the other side of the capacitor bank for current sharing. My capacitor bank is made out of 14 1000uF 200V capacitors arranged 2s7p to double the input voltage. There's also more decoupling capacitors on the back side, and I made sure to drill holes in the PCB to connect the copper on the two sides together with thick wire. I also reinforced the thinner parts of my board with copper strips and lots of solder. The giant blue capacitors are some snubber capacitors I got from China, they're 1uF 1200V.

Here's some pictures of the board layout
front:

back:


I made a couple minor changes to these, but that shouldn't affect anything that much.





The capacitor bank is made out of some chinese induction heater capacitors rated for 50kHz at 800VAC. I'm using 48 of these 300nF capacitors in a 12s4p configuration. The primary coil is made out of thicker litz wire



Here's the dummy load I'm using for testing.

No scope waveforms today, it's a bit too dark outside for testing.

[davekni]

Informative set of pictures!  Great to see that you do have a circuit board for the IGBT power interconnect.  (It's clear that you need a larger soldering iron for the heavy wires and copper planes.  I bought a cheap 150W iron with 10mm-base-diameter tip for such work.  Use it with a light-dimmer as it doesn't have temperature control.)  Added copper foil can help too if you run high duty cycles.  I do the same thing.  The induction-cooker capacitors work great for the MMC - again the same as mine (after other parts failed).

Good to see some local film capacitors for VBus.  The long wires to your bulk caps will eventually cause trouble at higher power.  (Mine were relatively short, but still required better pairing.  Otherwise the IGBTs see a voltage spike at the end of a burst, when the H-Bridge changes from using power to returning power to VBus, reversing current in the wiring to VBus bulk caps.)

The coax should help.  Yes, twisting should be plenty for the over-current connection.

What are the individual wires above the litz-wire primary coil?  Be careful they don't touch multiple windings, as the magnet-wire insulation won't handle the primary coil voltage.

It looks like you have some finer litz wire in some of the interconnect, but coarser in the primary winding.  Am I seeing that correctly?  Litz isn't important in the general wiring, but is useful within the coil.

Keep your wood dry!  Damp wood will form arc tracks and eventually burn at Tetsla primary voltages.

[ritaismyconscience]

I've conducted another test, this time at 75% the final bus voltage(which is about 250V)



1 div = 100A

There seems to be less spikes compared to last time.

To figure out when the transistors were switching, I came up with this:



It doesn't seem to work very well, it picks up a lot of noise. Also I still can't figure out how to display both traces on the oscilloscope, because when I turn the trigger level, this happens:

https://drive.google.com/file/d/1iQYbKvam7idLdfpl3MY4IfTEXZ9FejbX/view?usp=sharing

I also get weird traces like this:



^{1 div = 200A}
This doesn't make sense because there shouldn't be any more gate drive pulses after the current starts to go down, so there's probably something wrong with my setup.

Here's a picture of my entire setup:


I felt all the parts afterwards and all of them were pretty cold, but I have a cooling system for the IGBTs

(Also somewhat unrelated, but I guess if I finish this project I might build a larger version using more IGBTs. I can get more of these 75N60s for dirt cheap on chinese websites. You can buy them here for roughly 20 cents each)

[davekni]

The part number listed on that web page is "IGW75N60T", which looks like an Infineon part that does not include an internal emitter-collector diode.  The parts you have presumably include such a diode since they haven't all fried yet.  It would be worth a quick test with a meter to see if there's a diode internally that conducts when the collector is more negative than the emitter.  If not, that may explain your failures. Perhaps they don't fail instantly with excess reverse voltage.  (The shorter "75N60" part number includes both parts with and without internal diodes.)

I'm guessing that the $0.20 parts are counterfeit.  I've had that experience with a number of FET parts from China.  Some factory is making similar parts and marking them with a more reputable companies label and selling them as genuine items.  They may work reasonably, but often fail to meet all the specifications.  It is possible that they are genuine, and someone has a large stock to dispose of.  That would be more likely if they don't include internal diodes, as most designs don't use such parts any more.

A single turn on the gate-drive transformer makes a relatively-low voltage signal, so I wouldn't be too concerned about the noise level on it.  Turning on the scope's 50-ohm termination on that channel will likely help reduce the noise.  (Of course, don't use 50-ohm termination on higher-voltage signals, to avoid frying the scope's internal termination resistors.)

The second scope image in your last post is quite helpful.  The current is dropping because the controller isn't locking to the current waveform.  It's driving the gates at a fixed frequency that is a bit lower than your resonant frequency.  After a few cycles, the drive is out-of-phase with the resonant current, so forces the current back down.  So, there's some issue with current feedback.  You could try reversing current-sense polarity, although that may not matter for this driver.  Looking at the driver image, it appears that the current-sense feedback may be rectified even for the primary feedback, not just for the over-current.  The FPGA may be auto-detecting the phase, and monitoring only zero-crossings.  It also appears that this driver is self-oscillating.  I consider self-oscillating a good feature, but it does need to lock to current sense after a cycle or two.  For some reason yours isn't.  Perhaps scoping the current feedback voltage at the driver input terminals would provide a clue to the issue.

The scope images in your video are due to expanding the horizontal scale (faster sweep).  The scope is triggering a second time on each burst, so showing two overlapping time windows.  That's what "trigger hold-off" is there to solve.  Increase the hold-off (lower right knob) so the scope waits too long to catch another trace within the same burst.  (The other option is to go to a slower sweep, then use the 10x horizontal magnification feature to view a section of the longer sweep.)

[ritaismyconscience]

I'm currently using the transistors I bought from the seller listed above. The reason why they're so cheap is because they're pulled from old parts (that's also how I was able to get the capacitors for a couple cents each). I've taken a few of them apart and they do seem to have the correct die. Plus, I've tested them at 200A a while ago and they didn't fail.

Also, I'm pretty sure the polarity matters for the feedback. One of the pins is shorted to ground. The driver also seems to be working at the correct frequency, if you look at the second scope shot, the falling edge seems to always occur at the top of the sine wave. I think it's because I didn't set the trigger holdoff high enough, so the two traces got misaligned. Anyways, I'll test the setup later.

[davekni]

I'm not seeing the same thing you are in the second scope plot.  The first falling gate-drive edge at the left of the trace occurs just before the positive current peak.  By the sixth falling gate-drive edge, it's at the current zero-crossing after the positive peak.  It's shifted about 120 degrees in 5 cycles.

If the IGBTs are genuine Infineon IGW75N60T parts, then you do have an issue with missing free-wheeling diodes.  Definitely check for that.  I think many of the IGBTs with diodes use two separate die within one package.  I've never dissected any.  A few data sheets specifically mention two die.

[ritaismyconscience]

I'll try flipping the FB transformer and trying again.

[ritaismyconscience]

I found out more information about the driver.

When I flipped the FB transformer, the driver did not oscillate. It only oscillated when the FB transformer was the right way around.

Also, I did a low voltage test with the trigger holdoff set to max, and I got a different waveform than shown in the second picture. My waveform showed the current increasing when the driver was on, and when the driver turned off, the current went down quickly. I'll send a picture when I get more time to run tests at higher voltages.

[John123]

For what it's worth, I've had really bad luck using ethernet wire for gate drive transformers. I had much better results using single strand/core bell wire, not sure what its called in other countries but here in the UK its what the primary telephone/broadband network people use to wire up people's houses (full fiber and coax to the house isn't in most places yet).

It's like ethernet wire and the same thickness but the conductor is just one single core of copper. When wound onto a core it "hugs" the core much better and the insulation doesn't fight back and try to straighten itself out like the ethernet stuff, this allows the internal conductor to couple better.

[ritaismyconscience]

A bit more testing today:

Looks like I need to figure out how to add phase lead on this thing:

(250V on bus, 2us/div, 100A/div)

Same thing but at 5us/div


My controller has this interesting thing called "pulse skipping." This is what happens (170V on bus, 0.2ms/div, 200A/div):


The frequency seems to be right (you can see that there's about 1.5 small divisions = 600ns between the time the GDT switches and when the current switches). There seems to be phase lead, but it's the wrong way. It's probably hardcoded into the FPGA.


Maybe I'll have to flip the GDT connections?

[Mads Barnkob]

Download this guide and go to page 10, be sure to check for skew between your voltage and current measurements, before looking for problems in your inverter: https://highvoltageforum.net/index.php?topic=111.0

[ritaismyconscience]

Both of my traces are produced using transformers, the square wave is made by looping wire through the GDT and the sine wave is made with a 1:1000 current transformer. Is there anything that could cause delay using my current setup?

[Hydron]

The gate drive waveform isn't as important as the actual output waveform when it comes to worrying about phase lead. There will be delays in the IGBT switching that you're not seeing. Aim for having it switch just a tiny bit before the zero crossing of the current - it is better to turn off a little early than late (avoids any diode recovery issues).

[ritaismyconscience]

How would I measure the voltage safely wihtout a differential probe? Could I connect 9 1M resistors to drop the voltage?

[davekni]

Since I understand that you don't have any 10x scope probes, a capacitive divider would work better than resistive for 135kHz signals.  Run coax from your scope to a small piece of foil around one of the H-Bridge output wires (around an insulated part).  The foil to wire will make a small capacitor, combined with capacitance of the coax and scope input.  Not calibrated, but plenty useful for measuring timing and rough waveform shape.

Yes, it does appear to need significant lead compensation.  The lag you have now is actually a bit worse than it looks due to IGBT gate-to-output delay as Mads mentioned.  At least it does appear to be locking to your current feedback in your recent scope images.

Good luck figuring out how to adjust phase lead on that driver!

[ritaismyconscience]

Found this datasheet for my driver that I ran through Google Translate