Thermal Imaging for Predictive Maintenance

It usually starts at a lug. A breaker feeds a motor control centre, the cable lands on a terminal, and over a few thousand thermal cycles the connection works itself a quarter-turn loose. Nothing trips. The panel door stays shut. But resistance at that joint has crept up, and resistance under current makes heat. By the time anyone smells the insulation, the bus bar behind it has been running far above its neighbours for weeks. Open the door with an infrared camera and the fault is not subtle: one lug glows while the two beside it, carrying the same current, sit cool. That contrast is the whole game.

Infrared thermography turns that invisible temperature pattern into a picture. Any object above absolute zero radiates energy, and the amount climbs with temperature. Our eyes work in a narrow slice of the spectrum, roughly 0.4 to 0.7 microns; the thermal band sits well outside it, generally taken to run between about 2 and 15 microns, per the U.S. Department of Energy's Operations & Maintenance Best Practices Guide. A thermal imager carries a detector sensitive to that band and builds a two-dimensional map of radiated energy across whatever it is pointed at. Where a loose or corroded connection drives up local resistance, the temperature rises with it, and the pattern shows long before the connection is hot enough to fail or start a fire. The camera does not measure the fault. It measures the symptom the fault leaves on the surface.

What makes that valuable is timing. The DOE guide frames maintenance as four broad postures: run it to failure, service it on a calendar, service it on measured condition, or build the whole program around reliability. Reactive work still dominates. According to an industry survey cited in the guide, the average plant's maintenance spend ran more than 55% reactive, with preventive near 31% and genuine predictive work around 12% (the rest filed under "other"). So most facilities are still buying parts after the part has already broken, on overtime, with the line down and no spare staged. Thermography is one of the cheapest ways to move spend off that curve. The instrument's portable, the inspection's non-contact, and a single technician can walk a switchroom in an afternoon.

Seeing the fault before it announces itself

The economic case for catching faults early isn't new, and it isn't small. According to independent industry surveys reported in the DOE guide, plants that stood up a functional predictive-maintenance program saw average savings above 30% to 40% against a run-to-failure baseline, and a further 8% to 12% against a calendar-based preventive one. The figures the guide cites for a mature program are blunt. It puts the return on investment at roughly tenfold, with maintenance costs down a quarter to a third, breakdowns cut by about three-quarters, downtime down by a third to nearly half, and production up by a fifth or more. Those are program-level numbers, not thermography alone, and the guide is careful to call them surveyed averages rather than guarantees. But thermography is usually the first technology a plant adds, precisely because it pays for itself on a single avoided switchgear failure. A camera that costs less than a mid-range vibration rig can find a fault that would otherwise take out a transformer.

Regulators have started to treat it as table stakes rather than a nice-to-have. In 2023 the National Fire Protection Association rewrote NFPA 70B, the Standard for Electrical Equipment Maintenance, converting it from a recommended practice into an enforceable standard. The language shifted from what maintenance practices "should" be to what they "shall" be. The 2023 edition also folds in the potential-failure-to-functional-failure curve, the P-F curve, as a method for setting maximum maintenance intervals, and it builds infrared inspection of energized electrical equipment into the maintenance program rather than leaving it optional. For a lot of plants that quietly turned thermography from a line item someone defended every budget cycle into a documented obligation.

The method that 70B leans on is comparison, not absolute temperature. As the standard's thermography provisions describe, you measure the temperature difference between similar components carrying similar load, and between a component and the surrounding air, and you document that delta, as FLIR's reading of the 70B requirements lays out. The reason is practical. The absolute number a camera reports depends on emissivity, distance, reflected ambient, and a dozen settings a hurried operator gets wrong. The difference between two phases of the same breaker cancels most of that error, because both legs sit in the same conditions. Three identical lugs, two cool and one running well above them, is a finding you can act on without trusting the camera's calibration to a degree. "How hot is it?" the DOE guide notes, is usually far less important than "how much hotter than it should be?"

Load is the catch most people miss. A connection only reveals a resistance fault when current is flowing through it, and heat scales with the square of that current. Scan a feeder at a tenth of its rating and a genuinely bad joint can read stone cold. The working rule for electrical surveys is to inspect under meaningful load, and many programs set a floor before they will trust a clean scan. That single constraint quietly dictates when you can do the work: you walk the switchroom when the plant's running hard, not during the weekend shutdown when it's convenient and everything is de-energized and cold.

Electrical connections are the headline application, but they're not the only one. The DOE guide lists the same camera finding a warm inboard motor bearing, an overloaded contact, a slipping or misaligned belt drive throwing off friction heat, and losses across steam systems. Steam traps are a quiet favourite: a trap stuck open bleeds live steam straight to condensate, and on a thermal image the downstream line lights up where it should be cool, so a walk down a steam header can pay for the camera in recovered energy alone. In the plants we instrument, the same logic extends to refractory and furnace walls, where a cold spot or a hot streak on the shell maps to thinning or failed lining behind it, and to tank and vessel levels, where the product line reads as a sharp thermal boundary on the outside of the wall. The mechanism is always the same. Something the eye can't see changes a surface temperature, and the camera reads the surface.

Where the picture lies

Thermography earns its reputation by being easy to point and shoot, and that is also where it betrays the careless. The first trap is emissivity. A camera assumes the surface in front of it radiates efficiently, and most do: painted enclosures, oxidized metal, rubber, electrical tape all sit high on the scale and behave. Bare polished metal does not. A shiny copper bus bar or a clean aluminium housing radiates poorly and mirrors the room instead, so a thermal image of it shows you the reflection of the technician, the lights, and the warm panel across the aisle rather than the bar's own temperature. The standard dodge is to put a patch of high-emissivity tape on the target and read that, or to read an adjacent painted surface, but the operator has to know to do it. Visible colour tells you nothing about how a surface behaves in the thermal band; a gleaming silver bus and a matte black one can read wildly differently at the same true temperature.

The second limit is line of sight. A thermal camera sees surfaces, and only surfaces. It cannot see through a closed metal cabinet, an insulating cover, or an oil film. A hot connection buried inside a sealed enclosure shows up only once enough heat conducts to the outer skin, by which point the margin you bought has shrunk. This is why energized inspection, with the panel open or fitted with an IR-transparent window, is part of the method and part of the hazard; it is also why thermography pairs so well with other techniques. A bearing degrading inside a gearbox often shows in vibration or in the oil long before its heat reaches the casing surface. The DOE guide treats infrared as one technology in a suite alongside vibration analysis, oil analysis, and ultrasonics, not a replacement for them, and that framing matters. The plants that get burned are the ones that buy a camera and decide they have "done" condition monitoring.

There's a safety cost to the line-of-sight problem that's easy to underrate. Reading a hot connection usually means looking at it energized, which on a switchgear lineup means an open door and an arc-flash boundary, the territory NFPA 70E governs. That's why infrared-transparent windows and viewports have spread through the industry: a permanently fitted IR window lets a thermographer scan the buswork behind a closed, latched door, so the survey happens without exposing anyone to an open energized panel. It costs something to fit them, and they only cover what they're aimed at, but they turn a hazardous annual ritual into a routine walk-past. That trade pays off on anything you'll scan for years.

Interpretation is the third, and it's where the discipline lives. A thermal image is full of hot spots that mean nothing: a motor is supposed to be warm, a transformer runs hot by design, sunlight on a wall reads like a fault. Turning a thermogram into a diagnosis, and a severity, is a trained judgement. The international standard for this work, ISO 18434-1, lays out general procedures for applying infrared thermography to machine condition monitoring, and its companion ISO 18434-2:2019 covers image interpretation and diagnostics directly: how to separate a real anomaly from a reflection or an expected gradient, and how to grade what you find. NFPA 70B is explicit that the inspection has to be carried out and read by a qualified thermographer, not handed to whoever happened to be free. The camera is cheap relative to the skill it takes to read it well.

That asymmetry is the part budgets get wrong. The hardware line gets the scrutiny, because it's a number on a quote, while the training and the procedure that make the hardware worth anything get treated as overhead. It's the wrong way round. A modern imager will resolve temperature differences far finer than any fault you care about, and the better detectors now pack enough pixels that a single frame covers a whole switchboard at standoff. None of that helps if the operator doesn't set emissivity, doesn't check the load, and can't tell an expected gradient from a developing fault. ISO 18434-1 reflects this by treating the program, the qualifications, and the recording discipline as part of the method, not an afterthought to buying a device. The plants that get value out of thermography are the ones that invest in the reading, not just the camera.

None of this means thermography is fragile. It means the limitations are known and avoidable. Where this fails is predictable, and the list is short: a de-energized panel, a mirror-bright surface read without correction, a fault sealed inside a casing that hasn't yet warmed the skin, a scan run at trivial load. Each of those is a procedure problem, not a physics problem, and each is why the standards spend more words on method than on hardware. A camera that won't lie to you is a camera operated by someone who knows where it can.

From a walk-around to a baseline

The shift that turns thermography from a periodic walk-around into real condition monitoring is trending. A single image tells you one lug is hot today. A baseline tells you it has been getting hotter for three inspections running, that the delta against its neighbours is widening, and roughly when it will cross the line into a planned outage rather than an unplanned one. That is the difference between finding a fault and forecasting it, and it is the difference the P-F curve in NFPA 70B is trying to formalize: the interval between the point a failure becomes detectable and the point the equipment actually fails is the window you are managing, and the more often you sample it the more of that window you keep.

Doing that across a plant means the images cannot live on the camera's memory card. They have to be tagged to the asset, time-stamped, stored with the load and ambient conditions at capture, and compared image to image over months. The DOE guide makes the same point in passing, recommending that maintenance technologies feed a common system sharing equipment lists, histories, and work orders rather than each tool keeping its own siloed record. That is the layer where infrared inspection stops being a clipboard exercise and becomes data: severity trended per asset, repeat offenders flagged, work orders raised automatically when a delta crosses a threshold the thermographer set. It is also the layer where the newer fixed-mount thermal sensors earn their place, watching a critical bearing or a furnace wall continuously instead of waiting for the quarterly walk. Tying handheld surveys and fixed thermal feeds into the same trended record, alongside vibration and process telemetry, is exactly the kind of work our edge telemetry and analytics platform is built to carry.

Continuous monitoring brings its own discipline problem, which is the alarm. A fixed thermal sensor watching a bearing will see the bearing get hot every time the load climbs or the ambient does, and an alarm that fires on absolute temperature will cry wolf until someone mutes it. The fix is the same comparison logic the handheld survey uses, pushed into software: alarm on the delta against a baseline at matching load, not on the raw reading, and the false alarms mostly vanish. Setting that threshold is judgement, not arithmetic. Set it tight and you drown the control room in noise; set it loose and you miss the slow drift you installed the sensor to catch. The thermographer's reading and the data engineer's pipeline have to meet on that threshold, because one that ignores load and ambient is worse than no sensor at all, and a sensor with no one watching its trend is just an expensive thermometer bolted to a machine. The point of putting infrared on a critical asset isn't the picture. It's the slope of the line the pictures draw over a year.

What thermography buys you, in the end, is lead time. The hot lug was always going to fail; the question was only whether you found it on a Tuesday afternoon with a spare on the shelf, or at 3 a.m. with the line down and a melted bus to replace. The physics has been understood for decades and the cameras keep getting cheaper and sharper. The standards now tell you how often to look and how to read what you see. So the open question for most plants is not whether the technology works. It is whether the image you captured last quarter is sitting in a database where this quarter's image can be laid against it, or whether it is still on a memory card in a drawer, proving only that on one afternoon, once, a lug was hot.