How Do You Budget for Error and Calibrate a Temperature Sensor?

Detailed Explanation

A datasheet tolerance describes the sensor element in isolation, under the manufacturer's test conditions — it does not describe the accuracy of the finished measurement your product actually reports. Building an honest error budget means accounting for every error source the interface circuit and operating environment add on top of that baseline tolerance. This page covers the error sources common across the temperature-sensing technologies covered in this cluster — see What Is a Thermistor?, PT100 and RTD temperature measurement, and Thermocouples and cold junction compensation for the sensor-specific interface circuits these error sources apply to.

Self-Heating Error

Any resistive sensor measured by passing a current through it — an RTD or an NTC thermistor — dissipates a small amount of power in the sensing element itself (I²R), raising its actual temperature slightly above the medium it's measuring. The size of this effect depends on:

• Excitation current — a lower measurement current reduces self-heating but also reduces signal amplitude, trading measurement error for signal-to-noise ratio.
• The sensor's resistance at the excitation current used.
• Thermal coupling to the measured medium — a sensor in still air self-heats considerably more than the same sensor immersed in a flowing liquid, because liquid conducts the dissipated heat away far more effectively. Manufacturers typically specify a self-heating coefficient in °C per mW of dissipated power for one or more reference media (still air, stirred water) — always check which reference medium a quoted self-heating figure assumes before applying it to your own application.

Digital sensor ICs experience a related but distinct effect: the IC's own internal power dissipation (its logic, any onboard signal conditioning, and heat from nearby components on the same PCB) can raise the die temperature above the true ambient temperature it's meant to report, independent of any measurement excitation current. This is a PCB thermal design problem — component placement and copper spreading — rather than a sensor-excitation problem, but it produces the same practical symptom: a temperature reading that's biased above the true value.

Other Error Sources in the Budget

• Sensor tolerance — the manufacturer's stated accuracy class (for example, PT100 Class B is typically ±0.3°C at 0°C, with the tolerance band widening at temperature extremes) — the starting point for the budget, not the whole answer.
• Lead resistance — in a 2-wire RTD connection, the resistance of the wiring itself adds directly to the measured resistance and is indistinguishable from a genuine temperature change; 3-wire or 4-wire (Kelvin) connections remove or compensate for this term.
• ADC reference and quantisation error — the analog-to-digital converter's reference voltage accuracy and drift, plus its quantisation step size, both contribute directly to the final reported temperature's resolution and accuracy. See What Is an ADC? for reference-voltage and quantisation error fundamentals.
• Linearisation residual — a thermistor's Steinhart-Hart fit, or a lower-order polynomial approximation, leaves a small residual error across the fitted range that doesn't fully vanish even with a good curve fit.
• Cold junction error (thermocouples only) — the accuracy of the cold junction temperature measurement itself directly limits the accuracy of the final compensated reading; a thermocouple circuit is only as accurate as its cold junction measurement, on top of the thermocouple's own tolerance class.

Combining Error Sources

Once every significant error source is identified and sized for the actual operating conditions (not just the datasheet's test conditions), they need to be combined into a single total-error figure:

• Worst-case (arithmetic sum) — add every error source's magnitude directly. This guarantees the stated accuracy under the least favourable combination of conditions, which is the right approach when a design has a contractual or safety-critical accuracy guarantee to meet, but it is conservative because it assumes every source is simultaneously at its worst value in the same direction.
• Root-sum-square (RSS) — take the square root of the sum of each error source's square. This gives a more realistic statistical estimate when error sources are genuinely independent and random rather than correlated, and is the more common choice for general engineering accuracy estimates rather than a hard guarantee.

Choose the combination method deliberately based on what the accuracy figure needs to guarantee, and document which method was used — an accuracy figure without a stated combination method is easy to misinterpret as more (or less) conservative than it actually is.

Calibration Methods

Once the error budget establishes what accuracy is achievable, calibration confirms the actual circuit meets it:

• Fixed-point (ice point) reference — a well-prepared ice-water slurry sits extremely close to 0°C and provides a cheap, practical single-point check, either on the bench during development or in the field as a periodic sanity check. It validates accuracy at exactly one point, not across the sensor's full range. Laboratory metrology uses more precise fixed points still — the triple point of water (0.01°C, defined as part of the International Temperature Scale of 1990 (ITS-90)) is a common laboratory-grade reference point, though it requires specialised triple-point cells rather than a simple ice bath.
• Comparison (bath or dry-block) calibration — the sensor under test and a reference-grade sensor are both placed in a temperature-controlled bath or dry-block calibrator and compared across multiple points spanning the operating range, producing a calibration curve or correction table rather than a single-point offset.
• Manufacturer or factory calibration certificate — for sensors purchased with an accompanying calibration certificate, verify the certificate's traceability chain and stated uncertainty are adequate for your application before relying on it as the sole accuracy evidence.

Traceability

For a measurement to be traceable, the calibration reference used must itself be calibrated against a reference one step closer to a national or international standard, in an unbroken chain, with a stated uncertainty at every step. In Australia, laboratories accredited by NATA (the National Association of Testing Authorities) to ISO/IEC 17025 provide this traceable calibration service; a NATA-endorsed certificate is the standard evidence required when a measurement needs to withstand regulatory, contractual, or quality-system scrutiny — for example, in process industries, medical device manufacturing, or metrology-dependent product testing. A simple ice-point check or an uncalibrated bench comparison, while useful engineering practice, does not provide this formal traceability.

Practical Examples

An industrial process monitoring product uses a PT100 with a 4-wire connection and a MAX31865 interface IC. During design, the engineer builds an error budget combining the PT100's Class B tolerance, the MAX31865's reference resistor tolerance, and the RTD's self-heating coefficient at the chosen 1 mA excitation current in the still-air enclosure the sensor will be mounted in — arriving at a worst-case total error figure that's reported to the customer as the product's guaranteed accuracy, rather than simply quoting the PT100's own Class B tolerance as if it were the finished product's accuracy.

A calibration laboratory servicing food-safety-critical refrigeration equipment sends its reference thermometers for annual NATA-traceable calibration, maintaining an unbroken traceability chain back to national temperature standards — a requirement of the food safety quality system the equipment must comply with, not an optional engineering nicety.

Design Considerations

• Size self-heating for the actual mounting condition, not the manufacturer's reference medium. A self-heating coefficient quoted for stirred water will significantly understate the error for the same sensor mounted in still air — check which reference condition a quoted figure assumes.
• Choose the excitation current as a genuine trade-off, not a default. Higher excitation current improves signal-to-noise ratio but increases self-heating error; lower current reduces self-heating but can push signal amplitude closer to the ADC's noise floor.
• Decide worst-case vs RSS error combination deliberately, and document the choice. An accuracy specification without a stated combination method is easy for a customer or reviewer to misinterpret.
• Match the calibration method to what the accuracy claim needs to support. An ice-point sanity check is adequate for routine engineering validation; a contractual, safety, or regulatory accuracy claim needs a NATA-traceable (or equivalent) calibration certificate with a stated uncertainty. Zeus Design designs precision temperature measurement systems, including the error budget and calibration pathway, for industrial and process-critical products.

Common Mistakes

• Quoting the sensor's datasheet tolerance as the finished product's accuracy, ignoring self-heating, ADC error, lead resistance, and linearisation residual that the actual circuit adds on top.
• Applying a self-heating coefficient quoted for one media (e.g. stirred water) to a design mounted in a different media (e.g. still air), understating the real self-heating error by a significant margin.
• Treating a single ice-point check as equivalent to a full traceable calibration. An ice-point check validates one point and provides no formal traceability chain or documented uncertainty — it is a useful sanity check, not a substitute for accredited calibration when one is actually required.
• Combining error sources inconsistently or without documenting the method, making an accuracy specification impossible for a reviewer or customer to correctly interpret as either a conservative guarantee or a statistical estimate.
• Ignoring PCB-level self-heating on digital sensor ICs. Placing a digital temperature sensor IC next to a heat-generating component (a voltage regulator, a high-current trace, a processor) biases its reading above true ambient regardless of how accurate the sensor itself is.