The platform is the easy part to compare: mass, endurance, IP rating, climb capability. The payload is where deployments succeed or fail quietly, because a sensor that is one specification grade too weak produces a system that appears to work and misses the defects you bought it to find. This guide covers the four payload families that carry most industrial inspection duty — thermal, acoustic, gas and lidar — and the numbers worth insisting on.
1. Thermal (LWIR) imaging
Thermal is the highest-value payload in most industrial inspections and the most commonly under-specified. Buying decisions usually come down to pixel count, which is the wrong first question.
| Specification | Why it matters | What to ask for |
|---|---|---|
| Resolution (e.g. 160×120 vs 320×240 vs 640×512) | Determines the smallest object that resolves as more than one pixel at your standoff distance. At 5 m standoff, a 160×120 sensor cannot resolve a cable-joint hot spot that a 320×240 sensor can. | Specify resolution together with standoff distance and target size. Ask the vendor to calculate the spot-size ratio for your actual asset. |
| NETD (noise-equivalent temperature difference) | The smallest temperature difference the sensor can detect. A poor NETD hides the early-stage faults where the value is highest. | Prefer ≤50 mK. Below 40 mK is better and increasingly common. |
| Absolute accuracy and emissivity handling | Absolute readings are only meaningful if emissivity is set correctly per surface. Shiny metal has low, variable emissivity and will read wrong. | Ask how emissivity is set — fixed, per-asset profile, or user-selected at run time. For shiny targets, ask whether the vendor compensates or simply flags the limitation. |
| Thermal–visible registration | You need to know which physical object the hot pixel belongs to. | Insist on a fixed geometric relationship and a documented parallax correction. Loose "dual camera" claims are not enough. |
| Window material and cleaning interval | Contamination shifts apparent temperature gradually, with no hard failure. | A defined maintenance interval, and where possible a reference target in the scene to detect drift. |
The single most useful commissioning test: point the robot at a known reference — a calibrated blackbody if available, or a controlled heat source — at each of your working standoff distances, and record the offset. Then repeat it monthly. That curve is your data-quality certificate for every subsequent report.
2. Acoustic and ultrasonic sensing
Partial discharge in switchgear, gas leaks, and mechanical bearing faults all show up acoustically before they show up thermally. Acoustic payloads are less standardised than thermal and need care.
- Ultrasonic (40 kHz class) for partial discharge and pressurised-gas leak detection. Requires a directional horn or parabolic collector, and the detection is highly directional — the mounting geometry and the robot's approach path define what you can find. Ask for the detection cone in degrees, and check it against your asset layout.
- Audible-band microphone arrays for mechanical anomaly. Useful for bearing and gearbox faults on rotating equipment. Value depends almost entirely on the analysis pipeline behind it — a microphone alone produces a recording, not a diagnosis. Ask what the classifier does and what its false-positive rate is on your equipment class.
- An acoustic camera (beamforming array) is a different purchase. It localises the source in an image, which is far more useful for maintenance action. It is also heavier and hungrier; confirm the platform's payload and power headroom before specifying it.
One caution we repeat to every customer: acoustic detection in a plant environment competes with plant noise, and plant noise is not stationary. A system validated in a quiet test bay may behave differently at full production. Validation must happen under operating conditions.
3. Gas detection
Gas is the payload with the most regulatory weight and the least tolerance for specification shortcuts.
- Sensor technology matters more than the gas list. Catalytic bead (LEL), infrared (CO₂, CH₄), and electrochemical (H₂S, CO, O₂) sensors have different cross-sensitivities, lifetimes and drift behaviours. Ask which technology is used for each channel, not just which gases are covered.
- Response time (T90) must be specified. If the robot walks past a leak at 1 m/s and the sensor's T90 is 30 seconds, you have driven 30 m past the source before the alarm. Match T90 to your operating speed, or slow the robot down at gas-sampling points.
- Sampling method. Diffusive sensors read only what reaches the sensor; pumped sampling with a probe can reach voids, sumps and enclosed spaces. The two are not interchangeable in a refinery or a mine.
- In hazardous areas, the gas sensor is inside the certification boundary. In an ATEX Zone 1 deployment the sensor must be certified as part of the assembly, not added afterwards as a peripheral. This constrains choice significantly and should be settled before you compare vendors.
- Calibration interval and field-replaceability. Ask how a sensor is replaced on site, and how long the platform is out of service while it happens.
4. Lidar and the navigation payload
Lidar is usually considered navigation rather than inspection, but in most deployments it also produces the geometric deliverable — clearance measurement, volumetric change, structural deviation. Treat it as a dual-purpose payload when specifying.
- Channels and range. More channels means denser vertical coverage, which matters for close-range asset geometry more than for navigation.
- Mounting and vibration isolation. On a legged platform the Lidar is the sensor most sensitive to gait-induced vibration. Insist on mechanical isolation, and ask what the mounting's resonant frequency is.
- Calibration to the thermal and visible cameras. The extrinsic calibration between Lidar and thermal is what lets you overlay a hot spot onto a 3D asset model. Ask for the calibration procedure and the re-calibration interval; on a platform that flexes, this drifts.
The power and mass budget nobody plans for
Each additional payload costs mass, power and cooling, and the costs compound. A useful rule when reading a datasheet: a payload that consumes X watts typically costs more than X in endurance terms once you account for additional conversion losses, and every kilogram of payload costs climb performance and reduces the safety margin on stairs and obstacles. Before finalising a payload list, ask the vendor for the endurance figure with the full payload fitted, at the operating temperature you will actually run at — not the bare-platform number, and not a lab-temperature number.
Acceptance testing: make the vendor prove detection, not specification
Specifications describe the sensor. Acceptance tests must describe the detection. The difference is the whole of your procurement risk. A workable acceptance protocol:
- Create known defects. A controlled heat source at a defined temperature rise and standoff distance; a calibrated gas release at a defined leak rate; a bearing with a known fault signature. Real, physical, measurable.
- Run the robot along the route at operating speed and conditions. Not a static bench test.
- Measure detection rate and false-positive rate separately. Both numbers, over at least three runs. A vendor who only offers you a detection rate is showing you half the picture.
- Record the pose where each detection occurred. This validates the localization and the sensor in one test, and it catches the "confident but wrong" failure mode described in our GPS-denied deployment guide.
- Test the edge cases explicitly. Wet versus dry surfaces for thermal emissivity; full plant noise for acoustic; the background gas level for gas sensors.
Write these tests into the contract as acceptance criteria, not into a commissioning week as a formality. The structure for that is in how to write an inspection robot specification.
Putting the payload list together
A practical sequence that avoids over-specification:
- List the defect classes you must detect, from your own inspection records — not from a vendor's capability sheet.
- For each defect, identify what physical quantity changes when it occurs: temperature, sound, gas concentration, geometry.
- Pick one sensor family per physical quantity. Do not add a second sensor for the same quantity unless it is genuinely redundant in a different failure mode.
- Specify the detection numbers for the worst-case asset: smallest target, longest standoff, worst environmental conditions.
- Now add up the mass and power, and check the endurance number with everything fitted.
- Only then compare vendors.
Most over-specified systems we see were built by starting with a vendor's payload catalogue and working backwards to the defect list. It is a reliable way to spend budget on sensors that produce data no one acts on.
Our pages on oil & gas, power grid, mining, rail and nuclear inspection describe the payload configurations we build for each environment. If you send us your defect list and operating conditions, we will tell you which sensors are actually required — including the cases where the answer is fewer than you expected.