How should sensors be chosen for a hardware product?
Write the decision the reading will change. A threshold that warns a person, a log for later, or an act by the product. Each has a different tolerance for being wrong, and a different need for what happens when the sensor is absent. BrahmEarth, a BrahmWorks family of sensor nodes, is one place this discipline applies: the node is still enclosure, sensing path and electronics reviewed together. No performance figure for BrahmEarth is stated here. How stage evidence is handled is set out in Hardware Product Development Process Explained. The error stack below is invented teaching arithmetic, not a calibration and not a BrahmWorks result.
A retail payment terminal on a shop counter
The question defines the sensor, not the reverse
Name the quantity and the location. Air beside a door is not the temperature of goods. Current in a wire is not torque. A flag is not proof a part is seated unless the flag can move only when the part is seated. If you cannot draw the location, you are not ready to buy the part.
A flag is not proof a part is seated unless the flag can move only when the part is seated.
Write the range you will meet, including abuse. A sensor that saturates and still looks plausible will be believed. Prefer a failure that is obvious. Specify what the product shows when the element is disconnected, shorted or out of range. A blank that looks like zero is a lie. Do not leave it as the default.
Time matters as much as the value. How fast the quantity moves, how often you sample, and how you filter. A filter that calms a demo can hide the event you built the product to catch. Write the duration you must not miss, then choose the filter.
The product around the sensor is part of the measurement
Windows, membranes, magnets, air gaps, self-heating and a hand on the case are in the measurement. Review the element in the geometry you will ship. A result on an open board is allowed only if the note forbids a claim about the enclosed product.
Placement error is often larger than the part's own tolerance, and it is not on the catalogue page. Manufacturing adds glue thickness, a reversed polarity and a lens clipped the wrong way. The instruction and a polarity check are part of the sensing design. A constant typed into one prototype is not a calibration process.
Calibration can match a reference at defined conditions. It cannot turn a door-air sensor into a goods-core temperature. If the job needs the core, sense the core, or change the sentence so it admits it is about the air.
Claims that outrun the sensing
A single point is not a map. One node is not a field. Say what the sample does not see, in the way the number is presented, or people will generalise it.
An act based on a sensor needs a defined behaviour for absurd values and for silence. A pump or a motion that runs because a reading froze at a plausible value is a hazard by omission. Where the act can hurt someone, that path is a safety scope for qualified people. This article does not set it and reports no test.
Keep element, placement, filter and any calibration identity with the unit. Do not publish the teaching figures below as accuracy. You have no accuracy until you measure the configuration you ship, at the location you named.
Worked example: a cold-room door
The teaching product warns that a cold-room door has been left open, using air temperature near the door frame, not the temperature of the food. It is not a BrahmWorks project, not a BrahmEarth specification and not a tested product. The prototype question is whether a person can tell an open door from a sensor that has come unplugged, and whether the enclosure in the intended spot still couples to that air rather than to the wall.
In the build: the mount at the frame, the element in its window, an indication for "open" and a different indication for "no sensor", and a note that food temperature is not measured. Out of the build: a second room, a log of stock, an automatic defrost command, and any accuracy figure. If the wall conducts heat into the element so that an open door looks closed, the placement has failed. A sensor with more digits will fail in the same way.
Forbidden conclusions: food safety, energy saved, and any BrahmEarth performance. The example is here to separate the location from the part.
Labelled calculation: the term a better part does not fix
These contributions are assumed, in degrees, for one reading. They are not a calibration, not a specification and not a BrahmWorks result.
Assume the element's own offset is ±0.6. Assume placement, air at the frame versus the spot you actually care about, is ±1.4. Assume self-heating of the enclosed board is ±0.3. Assume, only so the sum is easy, that these are independent and that no other term exists.
The worst-case span is 0.6 plus 1.4 plus 0.3, which is 2.3 degrees, if every term leans the same way. The root-sum-square is the square root of 0.6 squared plus 1.4 squared plus 0.3 squared. That is the square root of 0.36 plus 1.96 plus 0.09, which is the square root of 2.41, which is about 1.55 degrees.
Placement is most of both sums. Replacing the element with one whose offset is ±0.2 changes the worst case to 1.9 and the root-sum-square only from 1.55 to about 1.44. You would have bought very little. Spend the next prototype on the mount and the wall, or change the sentence so the frame air is what you meant. Do not quote 1.55 degrees as accuracy. It is the arithmetic of assumptions that omit time, humidity and installation error.
Where the error usually sits
Checklist
- The quantity, the location and the decision are one paragraph.
- Disconnect, short and saturation cannot look like a normal reading.
- The event you must not miss is longer than the filter delay, or the filter is unjustified.
- Window, mount, sun and neighbouring heat are in the same review as the element.
- A finer part is not funded while placement dominates.
- Calibration has a reference, a condition and a revision, or it is not claimed.
- An act based on the reading has a behaviour for silence and absurd values.
- The record can find element, placement and firmware together. No BrahmEarth figure is invented.
Related questions
Should the part be chosen before the question?
No. Choose the location and the decision, then the part that can live there. A part chosen first will be defended with a claim written to match its page. That claim is how products measure the wrong thing accurately.
When is the enclosure part of the measurement?
When it changes what reaches the element: light, air, moisture, heat, a magnetic path, a pressure path. In those products the shell is not cosmetics around a finished circuit. Review it before the tool. A sensor node, including one in a family such as BrahmEarth, is this kind of product. Still no performance number.
What is a calibration record for?
To show which units were compared to which reference, under which conditions, at which revision. It is not a licence to extend the meaning of the measurement to a location you did not sense. Keep the record with the unit identity.
Can one sensor serve two claims?
Only if both claims are about the same quantity and location, and both can tolerate the same failures. A door-air sensor and a food-temperature claim are two products. Shipping one and writing the other is how a useful warning becomes a false assurance.
Review the location before the next part number
Bring the sentence, the mount, and the term in the error stack you have not measured. BrahmWorks can say whether the next build should move the sensor, change the claim, or only then change the component.