Does an Inverter's Own Switching Overshoot Corrupt Its Current Sensors?

Fast silicon-carbide transistors switch in nanoseconds. Every time one turns off, the stray inductance of the circuit makes the voltage overshoot and then ring, like a struck bell, at tens to hundreds of megahertz. Engineers usually treat this ringing as an insulation problem, because the peak voltage can damage motor windings, or as an emissions problem, because it can radiate into nearby radios. A third possibility has not been measured. The inverter's own current and voltage sensors are built around operational amplifiers (op-amps, the small analog chips that condition sensor signals). Op-amps are known to act as accidental rectifiers when they are hit with high-frequency energy: they turn an AC disturbance into a steady DC error. If the inverter's self-generated ringing reaches those op-amps, the controller would read a current or voltage that is slightly but systematically wrong, and the error would change with every switching event.

This note proposes to measure that effect on a silicon-carbide inverter, model it, and test which mitigations actually remove it. The main claim is that the sensor error depends on the energy of the ringing within the op-amp's rectification band, and not on the peak overshoot voltage that insulation studies track. If that holds, then "clamp the peak," the standard fix for insulation, and "damp the ringing" become separate design choices with different consequences. Designers would need to choose between them on purpose rather than treat them as a single overshoot problem.

Where This Sits

The closest prior result is ReThink. It showed that radio-frequency energy at or above 1 GHz rectifies inside the sensor op-amps of commercial inverters and a TI C2000 control board. With a 20 W transmitter at 150 cm, the authors caused false trips, drove the DC bus above 500 V, and cut photovoltaic output from 35.6 kW to 2 kW. That paper supplies the mechanism we test. It studies only an external attacker, though, and leaves open whether the inverter's own switching transients cause the same kind of error. We study the self-inflicted case. The source is near-field coupling of turn-off ringing, which is much stronger than a radiated attack at 1.5 m but sits at much lower frequencies. Whether rectification still matters at those frequencies is exactly what we do not know.

Two papers establish that switching transients do reach secondary circuits.

  • Liu et al. measured a 10 kV static var generator with transients of 1012 V/µs. They found that the 600–1000 pF capacitance between the module baseplate and its heatsink carries conducted common-mode noise, and that radiated noise dominates from 30 to 100 MHz. That work is about emission and slow IGBTs; it never asks what happens to the module's own sensing.
  • Tian et al. used a partial-element circuit model of a 1000 kV platform. It showed primary transients inducing up to 1146 V common-mode on secondary measurement cables, with most energy below 20 MHz. That work is passive high-voltage switchgear, not a converter's control loop.

Together these motivate the coupling paths, but neither measures a closed-loop sensing error.

Two lines of work constrain the hypothesis.

  • Lusuardi et al. showed that partial-discharge inception in winding insulation is independent of rise time from 7 to 150 ns. It depends only on peak voltage. This is the property our hypothesis is set against: we predict that sensor corruption depends on ringing spectrum, not peak. Peak voltage and ringing energy can be varied independently, so the two outcomes can be separated.
  • Zhao et al. clamped a 1.6 kA fault-interruption spike from 475 V to 295 V with an RC-TVS (transient voltage suppressor) snubber. They also noted that the clamp branch's own inductance creates secondary di/dt, and they left EMC (electromagnetic compatibility) testing out entirely. That suggests clamping a peak can create new ringing.

The modulation papers optimize common-mode voltage without looking at sensing integrity. Jin et al. capped it at $V_{dc}/6$ with a model-predictive controller, and Peng et al. did similar work with hybrid T-type modulation.

Core Claim

The DC error $\Delta y$ that a sensor op-amp adds during one switching event should follow, for modest ringing amplitudes,

$$\Delta y \approx \sum_k \kappa(f_k)\, A_k^2 \, \frac{\tau_k}{T_s},$$

where:

  • $A_k$ and $f_k$ are the amplitude and frequency of the $k$-th ringing component coupled onto the sensor input;
  • $\tau_k$ is how long that component lasts before it dies away;
  • $T_s$ is the switching period;
  • $\kappa(f)$ is the op-amp's rectification sensitivity, which is negligible within its bandwidth and grows above it.

The square law comes from the second-order nonlinearity of a differential input stage, the standard small-signal rectification picture. A large-signal regime would move toward linear dependence, and we will report which regime the testbed sits in.

Two predictions follow and can be tested. First, peak overshoot alone should predict the error poorly once ringing frequency and damping are varied independently of peak. Second, because ringing differs between turn-on and turn-off and with the sign of the phase current, the error should be correlated with the operating point. If so, it will not average out over a fundamental cycle, and it will show up as a current-dependent bias in the control loop.

Research Program

The testbed has three parts:

  • a three-phase silicon-carbide inverter controlled by a TI C2000 board, the same family ReThink attacked;
  • its stock Hall-effect and shunt-amplifier current sensing;
  • an optional 3–10 m cable to an inductive motor-like load, so that reflected-wave overshoot can be added as described by Pastura et al..

The key design step is giving the experiment independent control over four parameters that normal operation keeps tied together:

  • Edge speed is set by gate resistance.
  • Peak overshoot is set by DC-link voltage and by a TVS clamp.
  • Ringing frequency is set by deliberately inserting small, calibrated loop inductances and moving the decoupling capacitor.
  • Damping is set by an RC snubber whose resistance is swept.

A fractional-factorial sweep over these four parameters lets peak and ringing energy be separated statistically.

Ground truth for the true current comes from a coaxial shunt read by a battery-powered, optically isolated digitizer, so that the reference does not share the controller's ground. We isolate the sensor's DC error in three ways:

  1. Comparing the C2000's sampled current against this reference at matched instants.
  2. Running each condition at fixed DC current from a bench supply, with switching enabled and then disabled.
  3. Injecting the recorded ringing waveform into the sensor input with an arbitrary waveform generator and amplifier, with the power stage off, to confirm the cause directly.

In parallel, we measure $\kappa(f)$ for each sensor chip by sweeping a continuous-wave injection from 10 MHz to 1 GHz. The fitted model above then predicts the error from the measured ringing spectrum with no further tuning.

The closed-loop part runs the inverter under its own current control and records four outcomes: steady-state current error, current total harmonic distortion, DC-bus regulation error, and false protection trips. We also test whether the bias pushes the operating point in a consistent direction, the self-inflicted version of the bus-overvoltage failure in ReThink.

Baselines

The naive reference point is the datasheet assumption that the sensor error is its rated offset and gain error, with zero contribution from switching. Any measured switching-correlated bias larger than the rated offset already refutes it.

The strongest existing predictive approach is the one in the Modified SVR Prediction Method of Transient EMI. It predicts secondary-side interference from the primary transient overvoltage peak and reports a Spearman rank correlation above 0.94 and a precision of 98.6% on a 220 kV gas-insulated switchgear dataset of 1,600 tests. We will reproduce its method on our data: a support vector regression from peak overshoot (plus cable length) to measured sensor error, using the same training splits and the same hyperparameter search budget as our spectral model. Beating it means two things. Our model, built from the op-amp's rectification sensitivity and the ringing spectrum, must achieve clearly lower held-out error on conditions where peak and ringing energy were varied independently. It must also keep its rank correlation on those conditions when the peak-only model falls below the level that paper reports. If peak-only regression stays above 0.94 even on the decoupled conditions, our central claim is wrong.

For mitigation, the standard practice baseline is a first-order RC filter at the sensor input combined with ADC sampling synchronized to the center of the PWM pulse, the moment furthest from switching edges. The strongest competing hardware fix is a larger gate resistance. We will run every mitigation at matched inverter efficiency where possible and report efficiency alongside sensor error, because a fix that costs several points of efficiency is not equivalent to one that does not.

The key ablation separates clamping from damping. A TVS clamp lowers the peak at roughly fixed ringing energy, and an RC snubber lowers ringing energy at roughly fixed peak. Our hypothesis predicts that the snubber reduces sensor error and the clamp does not, possibly making it worse given the secondary di/dt that Zhao et al. observed.

Why not just…?

Why not just sample the ADC at the PWM center, away from the edges? This is the most serious threat. If the rectified error decays faster than half a switching period, center sampling may already hide it. We do not know the recovery time of these op-amps after a burst of high-frequency energy, and no reported result was found on Lacuna for it. We will measure the error with the sampling instant swept across the switching period. If the error disappears at the PWM center for all conditions, we will narrow the claim. The remaining contribution would then be a quantitative safe-sampling window and a warning for controllers that sample asynchronously or oversample, such as model-predictive schemes.

Why not just put an RC filter at the op-amp input? Rectification happens inside the input stage, and a filter only helps if the energy enters through the pin it filters. If the energy instead couples through PCB trace capacitance or the supply and reference pins, as ReThink describes for its attack, the filter does not help. The direct-injection experiment, run separately into the signal pin, the supply pin, and the reference pin, settles where the energy enters and therefore whether the filter is enough.

Why not just raise the gate resistance? Raising the gate resistance would likely reduce the error, but at a cost in switching loss, and the claim is about which property of the edge matters. If our model holds, a well-tuned snubber should reach the same sensor error with much smaller efficiency loss. The efficiency-matched comparison tests this directly.

Why not just shield the control board? Liu et al. showed that the dominant common-mode path below 30 MHz is conducted through the baseplate capacitance, which a radiation shield does not intercept. We will still test a grounded enclosure over the control board, to measure how much of the error is radiated and how much is conducted.

Decisive Checks

The idea is supported if three things hold. First, switching-correlated sensor error exceeds the rated offset under ordinary operating conditions. Second, the error tracks ringing energy weighted by the measured rectification sensitivity, while the peak-only regression fails on the decoupled conditions. Third, in closed-loop operation the error produces a measurable, current-dependent control bias or false trips that persist with synchronous sampling.

The idea is refuted if the error is within the datasheet offset, or if it vanishes under center sampling across the whole parameter sweep. A refutation would still be useful: it would bound the self-inflicted risk and show that the hazard ReThink found is specific to external attack.