How to Tell Whether You Have a Power Quality Problem

Why this matters

"Dirty power" is where a diagnosis goes to die. It is unfalsifiable, it sounds technical, it puts the blame on the utility, and it sells filters and conditioners that fix nothing because the fault was a lug in a panel forty feet away. Most complaints that arrive under that heading resolve to something ordinary: a shared feeder that sags when a large motor starts, a high-resistance connection, a control transformer sized without regard to inrush, a neutral carrying more than it was meant to. All of those are findable before anyone rents an instrument.

The work is correlation, not measurement. Answer five questions in order, because each one narrows what the next has to consider, and instrument only when the first four have failed to resolve it.

Question 1: what exactly failed, in the equipment's own words

Start with the equipment's fault record rather than the customer's description, because they are rarely the same thing. "It keeps resetting" covers an undervoltage trip, an overcurrent trip, a ground fault, an overtemperature trip and a control power interruption, and those are five different investigations. A drive, a controller and most modern appliances keep a fault log with a code and often a timestamp and a snapshot of conditions at the trip.

That record tells you which electrical quantity to look at, and it eliminates the largest category in one step: an undervoltage code puts you in a voltage investigation, an overcurrent or ground-fault code takes the supply off the table and sends you at the load. Skipping this question is how techs end up recording perfectly good voltage data for a fault that was never about voltage.

Write down the code, the count of events, and the window they occurred in. "Seven events in thirty days" is a measurable thing a fix can be tested against; "it keeps happening" is not.

Question 2: how many things failed, and what do they share

Scope locates the boundary of the problem, and the boundary is usually the answer.

Walk the electrical tree and mark what failed and what did not: one device, several devices on one panel, everything on one phase, everything on the service. The useful outcomes are the negative ones. If equipment on another panel fed from the same switchboard rode through the same moment, nothing upstream of that switchboard can be the cause, and you have just eliminated the utility, the service and the transformer without a single reading. If everything on one phase misbehaves and the other two are clean, you are in an unbalance or single-leg investigation instead.

Do this on paper. A one-line sketch of what is fed from where, with the failures marked on it, is worth more than any instrument in this procedure.

Question 3: when did it happen, and what else happened at that moment

This is the question that most often ends the job.

Get timestamps from the equipment's fault log and set them against what else the building does: large motor starts, elevator calls, welding, a chiller staging, a shift change, a lighting schedule, a generator test. If you cannot get timestamps, ask for a manual log for a week with time of day written down at each event, which costs the customer nothing and is more useful than a day of monitoring.

Correlation gives three outcomes and each points somewhere different. Events tracking a specific machine's start point at a shared impedance between it and the affected loads. Events tracking a schedule point at load distribution or a switching operation. Events with no internal correlation point outward, at the utility or the weather, and that is the case where instrumenting is genuinely justified.

Question 4: what the free measurements say at the equipment

Now measure, and measure at the affected equipment's terminals rather than at the panel, because everything between the two is a suspect.

These readings are taken on energized conductors, so 29 CFR 1910.333(a)(1) applies: de-energize before working on or near exposed live parts unless the employer can demonstrate that de-energizing introduces additional or increased hazards or is infeasible due to equipment design or operational limitations, and a reading that only exists under the fault condition is the ordinary case for the second clause. Work under the employer's energized-work program with boundaries and PPE selected under NFPA 70E-2021 130.5 and 130.7, in the edition your employer's program or your authority having jurisdiction has adopted, with leads and instruments rated for the measurement category and voltage of the point being tested.

Take four things: voltage at rest and during the suspect condition, percent voltage unbalance across the three line-to-line pairs on a three-phase system, current on each phase, and neutral current with neutral-to-ground voltage where there is a neutral. Use a true-RMS instrument that can capture a MINIMUM rather than an average, because a sag lasting a few cycles does not appear in an averaged reading at all.

If reproducing the condition means starting a machine, clear it first: confirm nobody is at the driven end, guards are in place, and valves are in their running position so a pump or compressor is not started against a closed path. Respect the equipment's own restart limits - a compressor has a minimum off time between starts in its literature, and repeat-starting it to collect data damages it, which turns your diagnostic into the next work order.

Question 5: whether to instrument, and what to ask it for

Instrument when questions 1 through 4 have left you with an event you cannot reproduce, no internal correlation, or a disagreement between the measured behaviour and what the installation should do. Do not instrument to confirm something you already found.

When you do, specify it rather than "putting a meter on it": monitoring at the point of use rather than only at the service, capturing minimum and maximum with the DURATION of each excursion, timestamped so it can be aligned against the building's schedule, over a window long enough to include several events. An instrument reporting one-second averages cannot see a four-cycle sag, and a report full of clean averages is the most expensive way to learn nothing.

Worked case: seven drive faults in thirty days

Question 1. The drive's fault log shows DC bus undervoltage, seven events in thirty days, each timestamped, with the bus voltage at trip recorded in the snapshot. That is a voltage investigation, and the equipment has told us its own threshold in its own units.

Question 2. Two drives on panel B fault. Three drives on panel A never do. Both panels are fed from the same 480 V switchboard. Everything upstream of that switchboard is therefore eliminated, because panel A rode through every one of the same moments.

Question 3. Six of the seven timestamps fall within a minute of the rooftop compressor starting, which the customer's own runtime log confirms. The seventh does not correspond to anything anyone can identify, and it stays open rather than being folded into the pattern. The compressor is fed from panel B.

Question 4. At panel B with the compressor off, 478 V. During a compressor start, the captured minimum is 441 V. That is 441 divided by 480, or 91.9 percent, an 8.1 percent sag. At panel A during the same start, the captured minimum is 468 V, or 97.5 percent, a 2.5 percent sag.

Read the difference. Panel A's 2.5 percent is the sag developed in the shared path, the service and switchboard. Panel B sees 8.1 percent, so 5.6 percentage points are being developed in panel B's own feeder, which is the only part of the path the two panels do not share. And the drive's recorded bus voltage at each trip corresponds to a line voltage between the two measured minima, which is why panel A's drives never faulted on the identical event.

Is 5.6 percent what that feeder should do? The compressor nameplate gives a locked-rotor current of 210 A. The feeder is 180 feet of 1/0 copper, 0.122 ohm per 1000 feet at 75 degrees C from the conductor properties table in the adopted NEC. Three-phase drop is 1.732 times 210 times 0.122 times 180/1000: 1.732 times 210 is 363.7, times 0.122 is 44.4, times 0.180 is 8.0 V. Against 480 V that is 1.7 percent.

The feeder should produce 1.7 percent and is producing 5.6 percent, more than three times the calculation. A conductor cannot do that. Something in that path is adding resistance the conductor table does not know about, which means a connection.

Finding it. An infrared scan of the panel B feeder terminations under load, taken from outside the restricted approach boundary or through an infrared window rather than by opening a door and reaching in, shows one line-side lug materially hotter than its two siblings.

The repair. De-energize, lock and tag under 29 CFR 1910.333(b)(2) - the standard that owns electrical work, since 1910.147 excludes exposure to electrical hazards from work on conductors and equipment in electric utilization installations at (a)(1)(ii)(C), with 29 CFR 1926.417 as the construction counterpart - and prove dead live-dead-live per NFPA 70E-2021 120.5. Inspect the conductor end and the lug: a termination that has run hot enough to discolour usually has an annealed or oxidised conductor end and a damaged lug, so cut the conductor back to clean copper and replace the lug rather than re-torquing damaged parts. Land it to the manufacturer's stated torque with a calibrated torque tool, because that value is the equipment maker's and no amount of feel substitutes for it.

What the whole exercise took. No monitoring rental, no filter, no utility involvement: a clamp meter, a multimeter with minimum capture and an infrared camera. The decisive step was question 2, which eliminated everything upstream in ten minutes of walking the one-line.

What this procedure cannot see

Three things it will miss, so know when to hand off:

• Sub-cycle transients. Surges from switching and lightning last microseconds. Nothing in this procedure captures them, and they need a transient-capable instrument.
• A problem that is genuinely at the point of common coupling with the utility. When question 3 produces no internal correlation, this procedure has done its job by telling you to look outward, and the utility owns the measurement at that point.
• Slow, cumulative harm. Harmonic heating in a transformer or a neutral does not produce an event to correlate against, so nothing in a fault-log-driven procedure finds it. That is a thermal survey and a loading review, and the neutral and system-effects cards own it.

How to verify you fixed it

Verify against the count from question 1, not against a good reading. Repeat the question 4 measurements at panel B during a compressor start and confirm the sag has come back to the neighbourhood of the calculated 1.7 percent rather than merely improving. Then leave the fault counter alone and check it after a window at least as long as the original thirty days.

Two outcomes are acceptable and they are not the same. Zero faults in thirty days closes the job. One fault in thirty days is consistent with the seventh event that never correlated with anything, which was never claimed to be fixed, and it needs its own investigation starting again at question 1 rather than a second visit to the same lug. Recording that distinction in advance is what stops a successful repair from being written up as a failure.

References

• 29 CFR 1910.333(a)(1) and 1910.333(b)(2), general industry electrical safety-related work practices, with the 1910.147(a)(1)(ii)(C) exclusion; 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 or your authority having jurisdiction
• NEC Chapter 9 conductor properties table, in the edition your authority having jurisdiction has adopted
• Equipment manufacturer literature for fault codes, undervoltage thresholds, minimum restart intervals and termination torque values
• See related: What Voltage Drop Actually Costs the Equipment; The Motor That Ran Hot on a Healthy-Looking Supply; What Non-Linear Loads Do to a System