Filters, Snubbers and EMI Mitigation

The overshoot:

Fifty volts from a hundred nanohenries — which is a few centimetres of busbar. This is why converter layout is a design activity rather than a packaging afterthought.

Double the stray inductance and the overshoot doubles:

The last row is a real failure mode: a converter that works on the bench with a short lead destroys devices when the capacitor is moved. The energy involved is trivial; the voltage is not.

Why faster devices make it worse. The overshoot scales with the switching speed:

A SiC MOSFET switching the same current in 20 ns produces five times the overshoot — 250 V instead of 50 — from the same layout. Wide-bandgap devices therefore demand low-inductance packaging, and much of their packaging development has been about reducing \(L_s\) rather than improving the die.

The RCD clamp. Rather than slowing the device, divert the inductive energy into a capacitor:

The clamp diode conducts as soon as the device voltage exceeds the capacitor's, dumping the inductive energy into it and limiting the excursion. Note that a smaller \(\Delta V\) needs a quadratically larger capacitor — halving it to 15 V would need four times the capacitance.

The resistor must discharge the capacitor before the next switching event:

The stored energy is dissipated once per switching cycle, so the resistor's rating is proportional to frequency — a 100 kHz converter would need 12.5 W. That is the clamp's real cost, and it is why reducing \(L_s\) is always preferable to clamping the consequences.

The hierarchy of remedies, best first:

The first two are what a good design does; the last three are what a compromised one needs. Set 4 designed an RC snubber for a thyristor's \(dv/dt\) withstand; this is the same idea applied to a different parasitic, and the same conclusion applies — a snubber is insurance, not a substitute for layout.

The mechanism. The switching node swings the full input voltage at every transition, and its capacitance to the earthed heatsink carries a current spike:

Sixty milliamps into the earth conductor from a converter that may only be delivering a few amps. The current is brief, but a spectrum analyser does not care about duration — it cares about the amplitude in each frequency bin.

What the LISN measures. The line impedance stabilisation network presents a defined 50 Ω to each line, which for common-mode current is two in parallel:

The LISN exists so that measurements are repeatable — without it, the emission depends on whatever mains impedance happens to be present. Its 50 Ω is a standard, not a physical property of the supply.

Against the limit:

This is the number that surprises people meeting EMC for the first time. The converter is not marginally over — it is over by a factor of two thousand, and no amount of tidying will close that gap. A filter is mandatory, not optional.

Where the emission sits in the spectrum. The conducted band runs from 150 kHz to 30 MHz:

A converter switching below 150 kHz keeps its fundamental out of the measured band — which is exactly why so many designs sit at 100–140 kHz. The harmonics still fall inside, but they are smaller, and the filter's job becomes much easier. Choosing the switching frequency is an EMC decision as much as a magnetics one.

Common mode versus differential mode. Two distinct mechanisms need two distinct filters:

Differential-mode noise is a circuit problem with a circuit solution — more input capacitance, as Set 18 sized. Common-mode noise depends on stray capacitances that are hard to measure and harder to change, which is why it dominates most EMC failures and why the 67.5 dB above is a common-mode figure.

The Y capacitance is capped first, because it sets a safety limit:

That is the absolute limit for a protectively earthed appliance, and prudent designs stay far below it — typically 2.2 nF per line, giving 4.4 nF total. The leakage flows through the earth conductor and is what an RCD sees, so several such devices on one circuit can cause nuisance tripping.

Try a single stage. Working back from the required attenuation:

Three hundred and forty millihenries. A common-mode choke of that inductance would be enormous, and its self-resonant frequency — set by its own winding capacitance — would fall well below 200 kHz, at which point it stops behaving as an inductor at all. The single stage is not just impractical, it is self-defeating.

Two stages share the work, and because attenuation is logarithmic, sharing is very effective:

Seven millihenries per stage — an entirely ordinary common-mode choke, available as a standard part. Splitting the filter reduced the required inductance by a factor of nearly fifty, because the corner frequency only had to rise sevenfold and inductance goes as \(1/f_c^2\).

The complete filter:

The fourth row is a useful economy: a common-mode choke's windings are never perfectly coupled, and the resulting leakage inductance — typically 1–3% of the CM value, so 70 to 200 µH here — acts as a differential-mode inductor. With the X capacitor it forms the DM filter at no extra cost.

What makes real filters underperform:

The third row is the most common failure in practice. A filter that measures perfectly on a network analyser can achieve nothing in a product if its input and output tracks run near each other, because the noise couples around it. Physical separation of the filter's two sides matters as much as its component values.

And the last row echoes Set 19. The filter is an \(LC\) network on the converter's input:

A switching converter presents a negative input resistance — drawing constant power means drawing more current as the voltage falls — so an undamped input filter can oscillate. The Middlebrook criterion above is checked in the same way Set 23 checked a loop's phase margin, and the cure is a damping network across the X capacitor.

The principle. A series \(LC\) branch placed in shunt presents near-zero impedance at its resonance, so harmonic current at that frequency flows into it rather than into the supply:

And below resonance the branch is capacitive, so it also supplies reactive power at the fundamental. One component does two jobs, which is why passive filters remain common despite active alternatives.

Size the capacitor from the reactive requirement:

Then the inductor from the tuning frequency:

Why detune. Consider what happens to the tuning frequency as components drift:

Every mechanism moves the tuning frequency upward. A filter tuned exactly to 250 Hz will therefore drift above the fifth harmonic in service — and above resonance the branch is inductive, so it no longer absorbs the harmonic. Tuning to 235 Hz means the drift moves it towards 250 Hz rather than away.

The danger of tuning above. An inductive branch in parallel with the supply's capacitance forms a parallel resonance:

And a parallel resonance amplifies rather than absorbs. This is exactly the mechanism Set 21 found at 269 Hz — a capacitor bank resonating with the supply inductance close to the fifth harmonic, magnifying it instead of removing it. A mistuned filter can make a harmonic problem substantially worse than having no filter at all.

The full design in practice:

The capacitor rating deserves care. At the fundamental the branch is capacitive, and the series inductor's voltage adds to the capacitor's rather than subtracting — so the capacitor sees more than the line voltage, typically 5–10% more. Rating it at the line voltage is a common and expensive error.

The cable is a transmission line. At a 100 ns edge the relevant wavelength is comparable with the cable length, so lumped analysis fails:

The impedance mismatch. A motor's surge impedance is far higher than a cable's:

An open circuit reflects a voltage wave with the same sign, so the incident and reflected waves add at the motor terminals. The machine sees twice what the inverter sent.

When it happens. The reflection returns to the inverter after the round trip, and only matters if the edge is still rising:

Almost twelve hundred volts on the winding of a 415 V motor, thousands of times per second. Standard motor insulation is not qualified for that, and the failure mode is progressive partial discharge in the first turn of the first coil — where the whole step appears, because the winding is also a transmission line.

The critical cable length:

Seven and a half metres is shorter than almost any real installation, so this is the normal case rather than an edge case. And faster devices shorten it further: a SiC drive with 20 ns edges has a critical length of 1.5 m.

The remedies:

The first row is standard practice. A \(dv/dt\) filter that stretches the edge to about 1 µs makes the round trip short by comparison, and the reflection arrives while the edge is still rising — so the two waves never fully add. It costs a small inductor and capacitor per phase and a few tenths of a per cent in efficiency.

What this joins. The cable is now doing three unwanted things at once:

All three come from the same source — a fast edge on a long conductor — and all three are reduced by the same remedy. A \(dv/dt\) filter at the inverter output simultaneously protects the insulation, reduces the bearing current and lowers the radiated emission.

Collect the problems. Each was identified in its own set and left open:

Seven distinct problems from one drive, and no single measure addresses more than two.

Layer 1 — the device and its loop. Fix what can be fixed for free:

The third row is the layer's characteristic trade. Slowing the gate drive reduces every emission problem in this list at the cost of switching loss — a knob that trades Set 2's efficiency directly against this set's EMC.

Layer 2 — the converter's own terminals:

The third row deserves emphasis: a DC link choke takes the input power factor from 0.44 to about 0.9 and roughly halves the input THD, for one inductor. It is the single most cost-effective power quality measure available to a drive.

Layer 3 — the installation:

The first row is worth more than its cost suggests. A symmetric shielded cable bonded 360° at both ends confines the common-mode current to a loop that stays inside the shield — so it never enters the building's earth system and never radiates. A cable terminated by a pigtail achieves almost none of this.

Layer 4 — the system:

The third row is often skipped and is where Set 21's 269 Hz resonance came from — an existing power factor correction bank that becomes a resonant circuit as soon as a harmonic source is connected. That interaction cannot be found by testing the drive alone.

What each layer costs and buys:

The pattern is the same one that ran through Set 24 and Set 23: the measures taken earliest and closest to the source are the cheapest and most effective. Layer 3 in particular costs almost nothing if specified before installation and is extremely expensive to retrofit — replacing a buried cable is not a filter change.