Hot-plug transients, and why your input capacitors keep dying

A board works on the bench for weeks. It goes into the field. Input capacitors start failing — not all of them, not predictably, and never while anyone is watching.

The supply is within rating. The parts are from a reputable manufacturer. Everything measures fine afterwards. This one is worth understanding properly because the mechanism is invisible unless you deliberately look for it, and because the usual fixes address the wrong thing.

What actually happens

Connecting a live supply to a board with bulk input capacitance is a step function into an LC circuit you did not design on purpose.

The inductance is your cable, your connector and your PCB traces. The capacitance is your input bulk. Neither is optional, and together they resonate.

An ideal undamped LC circuit driven by a voltage step overshoots to twice the applied voltage. Real wiring has resistance, so the ring is damped — but on a typical board with a meter of cable you still see 1.5× to 1.9×.

On a 24 V system that is a spike between 36 V and 45 V, lasting a few microseconds.

Why the intermittency misleads people

The peak depends on how far into the mating cycle contact is made, cable length, how discharged the bulk is, and source impedance. That produces exactly the pattern that gets a design fault misclassified as a component quality problem:

• Fails on some connections, not all
• Never reproduces on the bench, because the bench cable is short
• Never appears in the field logs, because it is microseconds long
• The replaced part works fine — until it does not

If a failure only appears on some power-ups and never during operation, look at the connection transient before you look at anything else.

Measuring it

You need a scope at the connector, not at the regulator input, and you need to catch a single event.

1. Probe directly at the input connector, ground spring rather than a clip lead — a 10 cm ground lead will ring on its own and confuse the measurement.
2. Set the timebase to microseconds, not milliseconds.
3. Trigger single-shot, slightly above nominal.
4. Plug it in. Repeat twenty times: you are sampling a distribution, and the first connection is unlikely to be the worst one.
5. Try a longer cable. More inductance, higher peak — and if that changes the result, you have confirmed the mechanism.

Four fixes, in order of preference

1. Derate for the transient, not the rail. The most common root cause is a capacitor selected against nominal voltage. On a 24 V input, 50 V parts rather than 25 V. This costs almost nothing and removes the failure outright.

2. Inrush limiting. An NTC thermistor, or an active soft-start using an ideal-diode or hot-swap controller such as the LM74700-Q1. Controlling the rate of rise stops the ring from forming rather than surviving it. Necessary on anything with substantial bulk capacitance.

3. Damping. A small series resistance, or a resistor-capacitor snubber across the input, drops the Q of the circuit. Cheap, effective, and costs a little efficiency.

4. A TVS as a backstop. Fast enough and rated for the energy. It limits the peak; it does not prevent the ring. Fit it as a safety net after the first three, never instead of them.

The mechanical half nobody mentions

There is a second failure mode that looks identical and is not electrical at all.

Large ceramic capacitors — 1206, 1210 and up — are brittle, and they crack under board flex. Enclosure assembly, a connector being levered, thermal cycling, vibration or impact will all do it. A cracked MLCC often fails short, which then looks exactly like an overvoltage failure.

Soft-termination MLCCs solve it. They include a compliant conductive polymer layer between the ceramic and the solder joint that absorbs board movement. They cost marginally more and they are the right default on any board that is not sitting still in a rack.

On a robot that gets hit repeatedly, both mechanisms were present simultaneously — electrical overstress from the ring, mechanical cracking from impact. Fixing only one would have looked like a partial fix and been misread as a different problem.

The checklist

Before releasing any board with a hot-pluggable input:

• Input capacitors rated for at least 1.5× nominal, preferably 2×
• Every part’s absolute-maximum rating checked against the transient, not the rail
• Inrush control present if bulk capacitance is more than a few tens of microfarads
• Soft-termination parts for anything 1206 or larger in a non-static environment
• The transient actually measured at the connector, on the longest cable the product will ship with

Five minutes at design review. Weeks of field failures otherwise.

---

This came out of a competition robot that kept eating capacitors — described in full in designing for repeated high-energy impact. If your product is failing intermittently in the field and nobody can reproduce it on a desk, that is a conversation worth having.