When the arc reaches something earthed, five parts of the machine feel it at once: the feedback, the pole frequencies, the pole structure, the mode the driver lands on and the primary current. Builders usually meet these one at a time, so they rarely add up to one picture.
The machine these numbers belong to has never been switched on, so none of them came from its own arc. Most come from our model. The one measurement used, davekni's 413 kHz, is real but was taken off another coil. Treat the figures as an example and work them out for your own machine. Where you can, measure the angle between the arc's voltage and current rather than R and C separately. The angle does not depend on how the model is written, and both values follow from it.
The event
arc -> earthed object
secondary loaded by
a resistance the coil
did not chooseA coil's own loading is designed: the arc you draw is the load you tuned for. A strike replaces it with a resistance set by what the arc touched, how wet it was and how far the current has to travel through it. That can be anything from a few hundred ohms to megohms, and you have no say in it.
Consequence one: the secondary stops reporting its own frequency
This one has been known longest, and it shaped every driver in use. Steve Ward, in his DRSSTC log book:
When a ground strike occurs, the secondary appears to jump to another harmonic, usually 3X or 5X the Fres of the coil… This causes all sorts of bad switching transitions… I believe this is why I've lost too many IGBTs in the past.
A secondary ringing at three or five times its resonance tells a driver nothing useful about when the current crosses zero. A driver that trusts it switches into a full bus at the wrong moments, and the damage lands on the bridge rather than the topload.
The fix is structural. Take the feedback off the primary current, and the loop can no longer hear what the secondary does in a strike. That is why the feedback comes off the primary at all. It usually gets presented as a design preference, but it came out of failure analysis.
Steve Ward puts the same collapse more compactly on the forum (HVF 117): a ground arc is such a low resistance beside the secondary's impedance that "the secondary no longer acts as an LC, but just an L with a parallel resistance, which can be boiled down to a single-resonant transformer."
The next step is our own reasoning. Power crosses the coupling as P = -M·Isec·dIpri/dt, so it is carried by the secondary current. When the strike collapses that current the primary has nothing left to deliver into, and its own current rings up instead. That is the fifth consequence below, and the overcurrent detector therefore only meets the event after it has happened.
Consequence two: it pulls the frequency further than an arc does
This one has a measurement behind it rather than a model. davekni, on his oversized QCW in HVF 2397, reports the upper pole at three points:
unloaded 485 kHz
end of a normal arc 440 -9.3 %
end of a ground 413 -14.8 %
strikeThe strike took it 27 kHz below where an ordinary arc on the same machine ended. The direction is what you would expect, since an earthed object is a bigger load than air, but there was no number for the size until this one. It also means a strike uses more detuning than the budget a coil is built with.
The pole equation itself was checked on the same coil, and the check is on the poles page.
Consequence three: the pole structure can come apart
Load a coupled pair hard enough and its two resonances merge into one near the primary's. The condition, the source and what it costs you belong to a property of coupled resonators rather than to this event. What matters for a strike is that it only does this across a band of resistances.
Damping peaks where the strike's resistance matches the secondary's reactance, 21.1 kΩ on this coil, and the structure comes apart across roughly 11.2 to 39.6 kΩ. Outside that band, in either direction, both poles remain. Too high a resistance barely loads the secondary, and too low a one stops damping it and acts as a short.
The band has no measurement under it. davekni's 413 kHz shows that a strike loads a coil heavily, but not where the edges are. No one has published a strike resistance measured against a frequency, so the numbers above are our model and nothing more.
The ordinary arc's resistance is better pinned down. The collapse is a resistance threshold, R < Z/k, which is 63.2 kΩ on this coil [derived, Z/k], with no capacitance in it. For a streamer in air, five published sources land within 3.3× of each other once they are put in one form, 77 to 258 kΩ at a metre and a half, and all of them leave the poles intact. In a strike that same term is set by whatever the arc happened to touch. The mechanism does not change, but you no longer know the resistance.
A sixth source disagrees with all five, and it is where the 50 kΩ used further down comes from. Anders Mikkelsen, revising his own series model because "a parallel one is more representative", gives the arc as 8 pF per meter in parallel with 75 kohm*meter, fitted to his measured 50 degrees near 340 kHz and 60 degrees near 460 (HVF 3140). The constant reproduces both of his angles. At 1.5 m it is 12 pF and 50 kΩ, whose reactance is 39.0 kΩ at 340 kHz and 28.8 at 460, so arctan(R/X_C) is 52 and 60 degrees [derived]. The 50 kΩ below is therefore his 75 kΩ·m over a metre and a half. It is fitted to measurements rather than assumed, and it is the harshest arc on file by a factor of 1.5 against the next one [derived, 77 over 50]. It also sits under the 63.2 kΩ above, so on his model an ordinary metre-and-a-half arc collapses this coil's pole structure by itself. Five sources say it does not and one says it does. The numbers below use the harsh end.
Consequence four: the strike that leaves the machine intact is the worse one
A strike that collapses the pole structure leaves a single frequency near the primary's 349.9 kHz, above the sword threshold, and nothing for the driver to fall onto. A milder strike that leaves both poles in place is the dangerous one. The driver recaptures onto the pole with the most gain, and which pole that is depends on how long the arc was when the strike landed. The gain figures give the lower pole +9.4 dB at half a metre and +1.9 at a metre and a half, a crossover at about 1.8 m, and -1.0 by two metres, because the loaded secondary passes the primary's 349.9 kHz somewhere in between. The lower pole therefore wins across the part of the ramp this coil actually uses, and the upper one past 1.8 m.
How long either pole rings is not established. Gain is what a recapturing loop follows, and the gain is published; ring times are not. That makes the mild strike, the one that does not even interrupt the bang, the one that can move the machine into the wrong mode and leave it there.
After a strike, then, check what frequency the machine came back on. That it came through the strike in one piece tells you very little.
Consequence five: the current rises, and not by enough
Damping raises the resistance reflected back into the bridge, and at a fixed bus a higher reflected resistance lowers the primary current. Loading a coil harder therefore pulls its current down at first, the opposite of what an overcurrent detector is built to catch.
On one of the two arc models that fall is already used up before the strike arrives. The turn, where more loading stops lowering the current and starts raising it, sits at 41.1 kΩ of total load [derived, the minimum of the primary-current expression]. An ordinary 1.5 m arc presents 50 kΩ on Anders Mikkelsen's parallel model, his 75 kΩ·m over a metre and a half (HVF 3140). On that model this coil runs its normal ramp 1.9 per cent above the bottom of that dip. A strike does not walk it down into a current minimum. It starts beside the minimum and goes up from there.
On the five sources that convert higher, and on our own 210 kΩ for the same metre and a half, the ordinary arc sits five times above the turn rather than beside it [derived, 210 over 41.1]. A strike then walks the current down through the minimum before taking it back up. Which of the two this machine does is not known, and the difference matters. On the harsh model there is no dip left to fall into. On the other five there is one to cross before the rise starts. The previous paragraph gives the harsh model's answer, and the table below is drawn on the harsh model throughout.
Somebody does report the fall, on a real machine. dr. kilovolt, describing the primary current envelope of his ferrite-cored SiC QCW: "During a ground arc the current drops. This is probably the opposite to a classic DRSSTC" (HVF 1073). That is the falling case, so at least one real coil sits far enough above the turn to have a dip to fall into.
His second sentence suggests a different split. The five against one above is a split between arc models. He is suggesting the sign may split between machine types, a QCW one way and a DRSSTC the other, which would come from where each topology puts its loaded impedance rather than from whose arc model you believe. He offers it as a guess, from one coil, so it is a hypothesis with a single data point. It does show that the falling case happens on real hardware.
The classic DRSSTC case he compares it with is on record too, from one coil at one scope setting, a frame apart.
The rise itself is small:
strike total load primary current [derived, every row]
none 50 kΩ 1.00x
50 kΩ 25.0 1.11
39.6 22.1 1.18
25 16.7 1.41
11.2 9.2 2.32
5 4.5 4.50
1 1.0 20.60Only the 25 kΩ row has ever been measured. The one published measurement of a strike channel, in HVF 117, has the load going from capacitive to resistive and bottoming out around 25 kΩ as the channel establishes. Nothing below that was measured, and every other row in the table is the model's curve. At that one measured point the current is already 1.41 times the normal ramp. A threshold set with a working margin of half again over the normal ramp is not crossed until the channel reaches 22.1 kΩ [derived], which is below the one strike anybody has published.
It is tempting to stop there and say that through the current the strike draws, the detector is not late, it does not fire. Where the threshold is set decides it. Half again is crossed at a channel of 22.1 kΩ and a fifth over at 37.3 kΩ [derived from the same expression that produced the table], and the one published strike, at 25 kΩ, falls between them. It passes straight through the lower setting: at 1.20 the detector fires the moment the arc lands, at 1.30 times. It never reaches the higher one.
The result therefore flips inside the range a builder might reasonably set, and the single measurement sits in the gap. What the model shows is how sensitive the result is. Half again over the ramp is not reached by any strike anybody has published, a fifth over is reached at once, and the boundary between them runs through the one channel resistance that has been measured. By this route the burst can run to the end of its ramp in a mode nobody chose, and whether it does is set by the detector setting. A second route, faster and by a different mechanism, does reach the threshold, and it is below.
That table has a calculator behind it, on the page about the arc that reaches the object. Set the channel resistance and the detector's threshold, which is its second slider, and watch whether they ever meet. Moving the threshold across the band changes the verdict with nothing else about the machine touched.
Below about 3 kΩ the model stops describing anything, for the same reason as the band: the topload is being pulled towards earth and the secondary is no longer a quarter-wave resonator. The bottom two rows of the table are arithmetic, not physics.
The current in the table is the one the strike draws through the resonator in the steady state, and it stays under the threshold. A strike raises another current as well. In the first consequence the secondary jumps to a harmonic and the switching loses sync. Once the bridge fires at the wrong moments the primary is no longer part of a tuned tank. It is an inductance straight across the bus, and its current climbs at V = L·dI/dt with nothing resonant to hold it. From a working 160 A to a threshold half again over it, 240 A, that takes 3.1 µs [derived, t = L·ΔI/V]. It does not depend on which harmonic the lock jumped to. The step per half cycle falls as the frequency rises while the count of half cycles grows to match, so the frequency cancels out of t = L·ΔI/V.
That is fast but not instant, and close to the speed of the protection itself, which the overcurrent page calls slow, a few cycles: about 8.6 µs at the working frequency here. It is how a bridge that the steady-state table says should never trip gets reported tripping on a wall strike on a real 600 A machine (kechuang t/80422). The protection is catching the loss of lock, not the reflected-resistance current.
The 3.1 µs is fixed by the machine. Its 12.2 nF tank capacitor sets the primary inductance at 17 µH for the 349.9 kHz primary resonance, and across this machine's 440 V buck ceiling that inductance takes the current up at 25.9 A/µs [derived, V/L on 440 V and 16.96 µH]. It is still a lower bound, since t = L·ΔI/V assumes the bus stays applied one way, and a bridge switching out of phase averages less.
The detector's few cycles of delay cost something too. In the 8.6 µs it takes to answer, the current at 25.9 A/µs reaches about 382 A, more than half again over the 240 A threshold, before the protection acts. That is behind the overcurrent rule that a threshold has to sit below the destructive current rather than at it. On this machine a loss of lock overshoots the set point by about 1.6 times.
Which decides an argument on the next page over
Whether the detector fires at all depends on where it is set, as above. So the argument is about a strike hard enough to trip whatever threshold you chose, and about what the machine does for the rest of the ramp when nothing trips:
latching OCD
burst over, mode never
had a chance to change
pulse skip
drops a cycle, returns,
into a different buck
voltage and a different
pole structurePulse skip is the right choice for a long bang and the wrong one here. Read the page that owns that setting keeping in mind how rarely the detector gets a say.
There is indirect evidence that a burst carries on through a strike rather than being ended by it. davekni's 413 kHz was read at the end of the ramp that struck earth, so his flash ran to completion, and the rest of it ran in whatever mode the strike left behind. The choice between latching and pulse skip decides what that remainder looks like.
What to do about it
- Feedback off the primary. It is not optional, and consequence one is the reason.
- A ground plane or a target, so the strike lands where you chose. A strike you arranged has a resistance you roughly know, which removes the one truly uncontrolled term in all of this.
- Coupling high enough to keep the margin. At
k= 0.15 there is no margin to keep; at 0.4 to 0.5 there is. - Check the frequency after a strike, not the fuse. That is consequence four. A quiet detector does not mean nothing happened.
- Latching rather than pulse-skip overcurrent, if the machine draws to earth at all.
Where next
- DRSSTC tuning: the two poles, and which one the driver lands on, which is what the collapse above takes away.
- Inside the driver, where the primary feedback decision lives.
- Overcurrent is not a setting, which is what the detector is actually for.
The department diagram does not draw the strike. It happens outside the machine and reaches every box in it.