Sensor arrays: the converter is almost never what limits your measurement

Signal levels for common laboratory sensors, the common-mode arithmetic that dominates them, why the multiplexer settles differently at high gain, and why the narrowest input range is free resolution.

Research rigs grow. A setup that starts with four thermocouples acquires a strain bridge, two pressure transmitters, a flow meter and a photodiode before the first paper is written, and all of it ends up on the same acquisition system. The reason the converter is rarely the problem is that the signals involved are already difficult before they reach it.

Quantify "small" before specifying anything

Swipe the table sideways to see all columns

Read the last column as a budget, not as an accuracy. The point is that the same 16-bit converter delivers 0.15 °C of resolution on a thermocouple and 0.003% of span on a transmitter, purely because the range was chosen differently. Choosing the narrowest range that contains the signal is the only gain stage in the system that costs nothing.

The common-mode arithmetic that dominates everything else

A thermocouple or a bridge sitting on a rig with a motor, a heater and a switched supply sees a large common-mode voltage: both of its wires move together with respect to the DAQ ground. Common-mode rejection ratio is quoted in decibels, and the conversion is worth doing once by hand.

A rejection ratio of 120 dB is a factor of a million. Against 20 V of common-mode, that leaves 20 µV of feedthrough — and the table above says one LSB on a thermocouple channel is 6 µV. In other words, at that rejection ratio the common-mode interference is already a few LSBs, which is to say it is visible in the data. Rejection of 100 dB would leave 200 µV, which is larger than the entire thermocouple signal at a modest temperature difference.

• Use differential inputs for every low-level channel. This is not optional — it is the mechanism by which common-mode rejection exists at all.
• Use shielded twisted pair, with the shield connected at one end only. A shield grounded at both ends is a conductor, and it will carry the ground-loop current you were trying to keep out.
• Give a floating source a defined reference with bias resistors, typically 10 kΩ to 100 kΩ from each input to ground.
• Keep the common-mode voltage itself inside the module's limit. On the higher-gain ranges that limit is tighter than on ±10 V, and exceeding it does not degrade gently.

The multiplexer settles differently at high gain

A multiplexed front end switches between channels, and every switch brings a small packet of charge with it. After switching, the amplifier and its input capacitance need time to settle to the new level before the converter takes its sample. Settling gets slower as gain rises, and the scan rate has to leave room for it.

The failure mode is distinctive and worth recognising: a channel reads slightly wrong, and reads *more* wrong the more channels are enabled, because each switch event has less time to settle. The diagnostic is to scan one known-good channel between the suspects. If the known channel drifts when it is surrounded by others and is clean when scanned alone, the problem is settling rather than the sensor.

In a mixed array, this is most likely to bite where a millivolt channel is scanned immediately after a channel carrying several volts. Grouping similar-magnitude signals together in the scan list is a cheap mitigation, and it costs nothing but a moment of thought when the channel list is being written.

Cold junctions: the error that adds directly to the reading

Thermocouple tables assume the reference junction sits at 0 °C. It never does. The terminals on your DAQ are at room temperature, and the thermocouple voltage you measure is the difference between the hot junction and those terminals. Correction means measuring the terminal temperature and adding back the EMF that temperature would produce.

Two numbers make the case for taking it seriously. First, the correction is not small: a reading of 4.096 mV taken naively is around 100 °C on a Type K curve, but with a 25 °C terminal temperature properly accounted for it is closer to 127 °C. Second, the tolerance on the sensor itself is measured in degrees: IEC 60584-1 Class 1 for Type K is ±1.5 °C or 0.4% of the reading, whichever is greater, so at 600 °C the wire alone can be 2.4 °C out.

Channel count and rate: what a large array actually needs

The reason to buy an 80-channel module is rarely speed. Temperature is a 1 Hz to 10 Hz measurement; static pressure and strain are not much faster. Spreading 250 kS/s across 80 channels leaves about 3 kS/s per channel, which is two orders of magnitude more than a thermocouple array requires and still leaves room for the occasional fast channel.

Swipe the table sideways to see all columns

Valves, relays and the digital half of the rig

Most laboratory automation is a handful of analog channels and a lot of switching: solenoid valves, pumps, shutter drivers, relay banks, status lamps. The USB-6509 puts 96 digital lines on one module with 24 mA of sink current per line, which is enough to drive an indicator panel or a bank of small relays directly rather than through a driver board. The analog modules sink considerably less, so a panel of lamps hung off a 16-channel analog module will not light.

Running unattended, and telling the truth about it

A USB acquisition device is not a deterministic controller. The host operating system schedules it, and a background update or a busy disk can delay a software-timed read by far more than the measurement interval. That is acceptable for logging and unacceptable for control.

• Use hardware-timed acquisition with onboard buffering rather than a loop that reads one sample at a time. The card's own clock then sets the timing, and host jitter affects only how often you collect the buffer.
• Assume the device will re-enumerate at some point over a long run. Log the gap and resume, rather than failing silently or losing the file.
• Dedicate one channel to a housekeeping signal — a fixed reference or a shorted input — so that a drifted channel is provable rather than suspected.
• Log a temperature channel on the structure or rig alongside the measurement. Thermal drift is the single most common explanation for a slow change in a long record.

Errors that look like sensor faults

• Reading a floating thermocouple single-ended and blaming the wire for the drift.
• Leaving the range at ±10 V for every channel, then concluding the converter is not good enough.
• Ignoring cold-junction correction and reading 100 °C where the answer is 124 °C.
• Grounding a cable shield at both ends and calling the resulting hum "environmental noise".
• Interleaving a millivolt channel with a multi-volt channel in the scan list and never suspecting settling.
• Trusting a linear sensitivity for a thermocouple across a wide temperature span.

Every figure in this note is either arithmetic from published resolution and sample-rate specifications, or a value stated by the standard it cites. Where a number depends on a particular module, confirm it against that module datasheet before design freeze.