The Fault That Was Upstream of Everything Anyone Checked

Why this matters

Four failures in twenty-six months, four correct diagnoses, four correct repairs, and the building kept eating equipment. Nobody in that sequence did poor work. Every tech measured what was in front of them and found what was in front of them, and the thing that was killing the equipment sat upstream of every point anyone put a probe on. This is a reconstruction from the paperwork, because the fault was invisible at the time and obvious in the record.

The four tickets

A small commercial bakery, 208Y/120 V three-phase service, one main panel and two subpanels.

• Month 0. Walk-in cooler condensing unit down. Compressor failed. Replaced.
• Month 9. Dough mixer will not start, then trips. Motor failed. Replaced, and the old motor sent out. The report came back as thermal overload with no manufacturing defect, which everybody read as a closed loop.
• Month 18. Rooftop unit not cooling. Compressor failed, and the contactor was pitted badly enough to be replaced with it.
• Month 26. Control transformer in the mixer's starter enclosure failed, taking the control circuit with it.

Four different components, three different trades, two different shops. Each ticket is defensible standing alone. Read together they are a pattern, and the pattern is the diagnosis.

What the record did not say

Before the pattern, the gap. Across four visits and four written tickets, the supply was recorded exactly once, and that entry read as a single number with no context: a line-to-line voltage, no indication of which pair of legs, no note of what was running at the time, and no time of day.

That single deficiency is why the fault survived twenty-six months. A reading without its conditions cannot be compared to a later reading, so four visits produced no series, only four disconnected impressions. A sibling card in this library covers what to record about a supply so the next visit is shorter, and this case is the argument for it.

The one pattern that survived

Sort the four failures by what they had in common rather than by what failed.

Three of the four were three-phase motor loads: the cooler compressor, the mixer motor, the rooftop compressor. The fourth, the control transformer, was fed line to line from two of the three legs. In the same twenty-six months, not one single-phase 120 V load in the building failed: not a light, not a receptacle load, not a piece of countertop equipment.

That discrimination is the whole clue. A supply problem that hits three-phase loads and a line-to-line load, while leaving line-to-neutral loads alone, is not a general voltage problem. It is a problem with one leg, and the loads that suffer are the ones that see that leg against the others.

Motors are the loads that show it first because of how they respond. A voltage imbalance across the three legs produces a current imbalance in the motor that is several times larger, and the extra current is heat in the windings. The commonly cited multiplier is in the range of six to ten times, stated for a motor operating near its rated load; a lightly loaded motor sits below that band. The consequence is that a modest imbalance a meter would report as unremarkable becomes a large thermal penalty inside the machine, and the machine dies of heat months later with nothing wrong with it internally that a bench test can name.

Why every previous tech saw a healthy supply

This is the part that explains four honest misses.

A high-resistance connection in one leg drops voltage only in proportion to the current through it. With the building idle, the current is small, the drop across the bad connection is small, and the three legs read nearly equal. With the building in full production, the current is at its maximum and so is the drop, so the same connection produces its worst imbalance exactly when nobody is measuring.

Both ends of that range matter. At no load, this fault is invisible and the supply measures clean. At full production load, it is at its worst. Every one of the four prior visits happened in the morning, before production, because that is when a bakery can afford downtime. The supply was measured, once, at the one time of day it looked fine.

So the measurement plan for the return visit was set by the fault's own behavior: measure under the building's real peak load, at the main, and again at each subpanel.

Taking the readings

Everything below is measurement on energized conductors, which runs through the gate at 29 CFR 1910.333(a)(1): parts are de-energized before work unless the employer can demonstrate that de-energizing introduces additional or increased hazards or is infeasible. A load-dependent voltage measurement cannot be taken dead, which makes it one of the genuine cases, so it is performed as energized work under the employer's electrical safety program with boundaries and PPE per NFPA 70E-2021, 130.5 and 130.7 in the edition that program adopts. Anything requiring contact rather than measurement is locked out per 29 CFR 1910.333(b)(2), not 29 CFR 1910.147, which excludes electric utilization work at 1910.147(a)(1)(ii)(C), and proved dead with live-dead-live at NFPA 70E-2021, 120.5. Nothing on the line side of the main disconnect gets touched at all, because there is no overcurrent protection between it and the transformer.

The load condition itself is arranged, not induced: ask the customer to run their normal peak, and do not switch equipment on beyond its ordinary duty to manufacture a bigger number. Rotating machinery is running during all of this, so leads, sleeves and the meter's own cable stay clear of shafts and belts.

Idle, 6 a.m., before production. Line to line at the main: 207, 208, 206 V. Average 207, largest deviation 1 V, which is about 0.5 percent unbalance. Percent voltage unbalance here is the largest deviation from the average of the three line-to-line voltages divided by that average. Half a percent is a clean supply, and this is the reading everyone before had taken.

Peak, 1 p.m., full production. Line to line at the main: 205, 206, 193 V. Average 201.3, largest deviation 8.3 V, which is 4.1 percent unbalance. Same building, same meter, same terminals, four hours apart.

Current at the mixer motor, measured at the same time: 9.8, 10.1 and 14.2 A on the three legs. Average 11.4 A, largest deviation 2.8 A, which is about 25 percent current unbalance. Against 4.1 percent voltage unbalance that is a multiplier of roughly six, at the low end of the six-to-ten band, which is consistent with a motor that is loaded but not at its rated load. The rule of thumb held, and it held in the direction the mechanism predicts.

What that imbalance costs the motor in continuous rating comes from the derating curve in NEMA MG 1, in the edition the motor manufacturer's application data references, and it reaches you through that data rather than on its own authority. Two things about that curve are worth carrying without looking it up: the penalty rises steeply rather than gently as unbalance grows past about one percent, and continuous operation above five percent is not recommended at all. At 4.1 percent this motor was being asked to run derated by a substantial margin, on a load that had not been reduced by anybody.

Localizing it

With the imbalance confirmed under load, the same measurement at each subpanel showed the same low leg by the same amount, which places the cause upstream of both subpanels rather than in either one.

An infrared scan of the main's terminations under that same peak load showed one lug materially hotter than the other two. That comparison is the valid form of the measurement: three lugs of the same geometry and the same surface finish, at the same instant, on the same load, compared against each other. An absolute surface temperature read off a single lug would not have meant much, because emissivity and geometry both move the number.

The connection was then confirmed dead. With the service de-energized and proved dead, the lug and its conductor showed the discoloration and the oxide that go with a joint that has been running hot and cycling for a long time.

The correction and the proof

The repair itself was ordinary: replace the affected termination and conductor, then tighten every termination in that enclosure to the manufacturer's stated torque with a calibrated tool. Recent code editions require exactly that where a torque value is provided, at NEC 110.14(D) in the editions that carry it, and on this job it is not a formality, because a joint tightened by feel is how this one started.

Where any part of the work reaches the line side of the service, it is coordinated with the serving utility rather than performed on your own schedule; a sibling card covers that boundary and who owns which piece of hardware.

The proof came from repeating the bad measurement, not the good one. Idle readings would have looked fine before the repair and after it, and proved nothing. Back at 1 p.m., under the same production load as the failing reading: 206, 207, 205 V, average 206, largest deviation 1 V, about 0.5 percent. Same conditions, same terminals, same time of day, a different result.

That is the discipline the whole case turns on. A confirmation reading has to be taken under the conditions that produced the failing reading, or it is a different measurement wearing the same units.

What this changes about the next repeat-failure call

The reusable finding is not "check for voltage imbalance." It is the question that got there: when a building has eaten several unrelated components, stop diagnosing the component and ask what the failures share.

Sort the failures by which part of the supply they see, not by what they are. Loads on one leg, loads across two legs, three-phase loads and line-to-neutral loads are four separate populations, and which populations are failing tells you where to look before you take a single reading. In this building three of the four failures were three-phase, one was line to line, and none were line to neutral, and that alone pointed at one leg twenty-six months before anybody measured under load.

Then measure when the fault is awake. A supply that misbehaves in proportion to current is measured at peak current, and if the only time you can get on site is the quiet hour, say so on the ticket rather than recording a clean reading as though it settled the question.

References

• 29 CFR 1910.333(a)(1) and 1910.333(b)(2); 29 CFR 1910.147(a)(1)(ii)(C) for the carve-out; 29 CFR 1926.417 for the construction counterpart
• NFPA 70E-2021, 120.5, 130.5 and 130.7, as adopted by your employer's electrical safety program
• NEMA MG 1, in the edition referenced by the motor manufacturer's application data, for the definition of percent voltage unbalance and the derating curve
• NFPA 70 (National Electrical Code), 110.14(D) in the editions that carry it, for terminations tightened to the manufacturer's specified torque
• See related: What a Utility Owns and What You Own; What to Record About a Supply So the Next Visit Is Shorter