Why a thermocouple reading is off by a few degrees

A thermocouple board works on the bench, then reads several degrees high once it sits in an enclosure next to a regulator. Or it tracks the oven well at one setpoint and badly at another. The probe is usually fine. The error comes from the board, the connector, or the math between them.

What a thermocouple actually measures

A thermocouple does not report temperature. It produces a small voltage that depends on the difference between the hot junction at the probe tip and the junction where the thermocouple wire meets copper, on your board. That second point is the cold junction.

To get an absolute temperature, the firmware or the amplifier has to know the temperature at the cold junction and add it back in. This is cold-junction compensation. Every error in that second temperature goes straight into the result, one for one.

Two things follow. The reading is only as good as your knowledge of the board temperature at the connector. And the thermocouple curve is not a straight line, so the conversion has to use the published reference tables or polynomial coefficients for your thermocouple type. The NIST ITS-90 thermocouple database is the standard source, and Microchip states that the MCP9600’s correction coefficients come from it.

Where the error comes from

The cold-junction sensor is not at the cold junction. Most thermocouple interface chips measure their own die temperature. That is correct only if the die and the connector terminals are at the same temperature. Put the connector at the board edge and the chip beside a voltage regulator, and they will not be. Texas Instruments’ thermocouple application note makes the same point in general terms: the cold-junction terminals should be isothermal, the sensor should sit as close to them as it can, and a large mass of metal helps. It also warns that air currents and heat from nearby components degrade the compensation.

The enclosure changes the gradient. On an open bench the board sits in room air. Inside a case, a power stage, a display backlight or a cellular modem warms one side of the board. The gradient across the board is new, and your calibration from the bench does not include it. This is why the symptom often shows up only after the product is assembled.

Extra junctions. Any place where two dissimilar metals meet is a junction that makes its own voltage. A connector with ordinary copper-alloy contacts, a solder joint, or a terminal block can add one. Thermocouple-grade connectors use matching alloys for this reason. If you use a standard connector, the two contacts need to be at the same temperature, which brings you back to isothermal layout.

Wire type and extension. Thermocouple extension wire is chosen for the type of thermocouple. Copper hookup wire in the middle of the run moves the cold junction to wherever the splice is, which may be somewhere the board knows nothing about.

Noise and mains pickup. The signal is tiny, and a long probe lead is an antenna. Interface chips like the MAX31856 list built-in rejection for 50 Hz and 60 Hz line frequency. Choose the setting that matches where the product is used. For an amplifier-plus-ADC design, the filtering is your job, so budget for it in the schematic.

Choosing the front end

There are three common routes.

1. A dedicated digitizer with an on-chip cold-junction sensor and linearization, such as the MAX31856 or the Microchip MCP9600 family. The chip returns a temperature over SPI or I2C. This is the shortest path and the one we recommend first. You still own the placement problem.
2. An analog amplifier with built-in cold-junction compensation, such as the AD8495, feeding your own ADC. Analog Devices’ CN0354 circuit note shows this approach with an isothermal terminal block, and adds a dummy thermocouple to cancel the voltage from any temperature difference between the block and the compensation circuit. It is a useful idea if the connector has to sit away from the chip.
3. A general-purpose delta-sigma ADC reading the thermocouple voltage plus a separate thermistor or RTD at the connector. It is the most flexible and the most work. You write the compensation and the conversion yourself, and you decide where the sensor goes.

Check the variants before you commit. Microchip’s MCP960X datasheet lists integrated open-circuit and short-circuit detection on the MCP9601 variants. The base MCP9600 is a different part with different fault behavior. Read the datasheet for the exact part number on your bill of materials, not for the family. The MAX31856 datasheet lists open-thermocouple detection and over- and under-temperature fault flags. Confirm in the register map what each flag means for your wiring.

Layout rules that matter

These are the layout checks we run, and none of them depend on a particular chip.

• Put the connector and the interface chip close together, on the same side, with copper between them to help them track each other. Keep them away from regulators, inductors, motor drivers and any part that dissipates heat.
• Route the two thermocouple traces as a tight pair with the same length and the same copper. Keep them over a solid ground region and away from switching nodes. Do not put a via on one trace and not the other.
• Keep heat sources on the far side of the board and consider slots in the board to cut conduction toward the cold-junction area. Whether slots help depends on the board and the enclosure, so treat them as something to test.
• Add a footprint for the input filter, even if you leave it unpopulated at first. Changing it after layout is much cheaper than adding it afterward.
• Plan the grounding of the probe. A grounded-junction probe ties the sensor tip to the machine it is measuring. If that machine is at a different potential than your board, current flows through the thermocouple. Check the input common-mode range of the chip, and decide whether the design needs isolation. This is a datasheet question for the specific part, and a safety question for the product.

Checks before you trust the number

1. Shorted-input test. Short the thermocouple inputs at the connector, or plug in a short length of the right thermocouple wire. The reading should equal the board temperature at the connector. A second thermometer taped to the connector terminals tells you the truth.
2. Warm-up test. Run the product at full load in its enclosure until it settles. Watch the reading with the probe held at a fixed, known temperature. If it drifts, the gradient is the cause.
3. Ice-point check. Put the probe in a well-made ice bath and compare. It checks the whole chain at one point, including the thermocouple type setting in firmware.
4. A second point. Check at a higher temperature against a trusted reference. If the first point passes and the second fails, suspect the type setting, the conversion coefficients, or the wire.
5. Fault injection. Disconnect the probe. The product should report a fault, not the last good value or the board temperature. Do this in production too, as part of the test.

The last check belongs in your production test as well as your lab notes. A fixture that reads a known reference through the real connector catches the assembly defects that no schematic review will, such as a missing filter part or a reversed pair.

How we approach it

We pick the front end from the thermocouple type, the accuracy you need and where the connector has to sit, then place the sensor and connector together and plan the checks above into the first prototype. If you send us the schematic, the enclosure drawing, the probe type and the temperature range, we can tell you where the cold junction will end up and what is likely to push it off.

Sources used for the technical claims above: the NIST ITS-90 database, TI application note SLOA204, Analog Devices CN0354, the Microchip MCP960X datasheet and the MAX31856 datasheet.