Sensor conditioning: strain gauges, thermocouples, IEPE and LVDT

The ceiling is set before the converter

It is tempting to judge an acquisition system by its converter: the number of bits, the sample rate, the datasheet figures. But by the time a signal reaches a converter its quality is already decided — by the sensor, its excitation, the cable, the amplification and the filtering. Everything after that point can only preserve or degrade what it was handed.

One arithmetic example settles the argument. A quarter-bridge strain gauge with a gauge factor of two, at a thousand microstrain, produces about five hundred microvolts per volt of excitation — some 2.5 mV at a five-volt supply. A sixteen-bit converter on a ±10 V input has a step of about 305 microvolts. The entire signal, at full scale of the measurement you cared about, is eight converter steps wide.

Nothing in the digital half of the chain can fix that. Only gain, applied cleanly and early, can — and that is what conditioning is.

The signal path, stage by stage

Read that path as a list of questions rather than a block diagram. Is the excitation stable enough, and does it heat the sensor? Is the bridge completed where it should be, and does the lead resistance enter the measurement? Is the gain set for the range the signal actually uses? Is the filter matched to the sample rate that channel was actually assigned? Is the input isolated, or does the airframe get a vote?

Strain gauges: the bridge, the excitation and the leads

A resistive gauge changes its resistance by a fraction of a percent, so it is always measured in a bridge: quarter, half or full, with the missing arms completed by precision resistors in the conditioner. The excitation must be stable, because the output is proportional to it — a one percent droop in the supply is a one percent error in every reading, and sharing one supply across too many gauges is how that droop appears.

The leads are the classic trap. In a two-wire quarter bridge the resistance of the cable sits in series with the gauge and is indistinguishable from strain, so the reading changes when the cable warms up in the sun. Three-wire and four-wire connections exist precisely to remove that term, and using them is not optional on an aircraft where the cable run is long and the temperature range is wide.

Then there is temperature acting on the gauge itself. Self-temperature-compensated gauges, matched to the coefficient of the material, remove most of it; a dummy gauge in an adjacent bridge arm removes more. Neither removes the need to know which one you used when the analysis disagrees with the model.

Thermocouples: the reference is the real subject

A thermocouple does not measure temperature. It produces a voltage that depends on the difference between its measuring junction and its reference junction — about forty-one microvolts per degree for a type K, which on that same ±10 V input is a seventh of a converter step per degree. So thermocouples need gain, and they need it close to the input.

The reference junction is where the accuracy really lives. Whatever error there is in knowing the temperature of that junction passes straight through into every reading, and it is usually larger than everything else in the channel combined. A conditioner with a poorly placed or poorly characterised cold-junction sensor produces beautifully repeatable readings that are all wrong by the same amount.

Three more practical points. The relationship is not linear, so the conversion is a polynomial or a table and it belongs in the calibration data rather than in a spreadsheet at the end. The junction is often electrically bonded to the structure, which makes isolation a requirement rather than a refinement. And an open circuit should be detected and reported, because a broken thermocouple does not read zero — it reads whatever the amplifier drifts to, which can be a perfectly plausible temperature.

IEPE accelerometers: constant current, and a low-frequency limit

An IEPE accelerometer contains its own amplifier, powered by a constant current — typically a few milliamps — sent up the same coaxial cable that brings the signal back. The signal rides on a bias voltage of several volts, and checking that bias is the fastest way to know whether the sensor is alive, shorted or open: three distinct and easily recognised values.

The consequence people forget is the low end. The coupling that separates the signal from the bias is a high-pass filter, usually somewhere around one hertz, and it is inside the measurement path. An IEPE channel cannot report a steady acceleration, and it attenuates slow events before you ever see them. If the phenomenon you are chasing lives below a few hertz, IEPE is the wrong sensor, and no processing afterwards will bring the missing part back.

At the top end, long cables and their capacitance limit the usable bandwidth for a given drive current, which is a real constraint on an aircraft where the accelerometer is thirty metres from the acquisition unit.

LVDT, resistance probes and the others

An LVDT measures displacement by transformer coupling, so it needs an AC excitation — a few kilohertz — and synchronous demodulation to recover both magnitude and direction. The carrier has to sit comfortably above the mechanical bandwidth of interest, an order of magnitude is the usual rule, and the demodulation has to be ratiometric so that a drift in the excitation cancels instead of appearing as movement.

Resistance temperature probes want a small, known excitation current and a four-wire connection, and they punish a large one: the current heats the element, and the reading tells you about the self-heating rather than the air. Potentiometric sensors are ratiometric by nature and should be measured that way. Charge-mode piezoelectric sensors need a charge amplifier and, unlike IEPE, tolerate high temperatures — which is exactly why they still exist.

The errors that survive into the report

• Ground loops. The airframe is a conductor with volts across it, and a non-isolated input measures that as well as your sensor. Isolate, or accept a common-mode signal you cannot remove afterwards.
• Shields grounded at both ends, which turns the screen into a loop and injects exactly what it was fitted to keep out.
• Thermal EMF at connectors. Dissimilar metals in a low-level circuit produce microvolts with temperature, which is negligible on a ±10 V signal and dominant on a thermocouple.
• Gain set for the theoretical range rather than the measured one, discarding bits before the first sample.
• A filter left at its factory default while the sample rate for that channel was changed on the flight line.
• And the most common: a sensor calibration certificate in the file, and no end-to-end check of the installed chain.

The remedy is the same in every case and it is not a document: inject a known value at the sensor end, in the flight configuration, and read what arrives in the file. Shunt calibration for bridges, a reference bath or a calibrator for thermocouples, a shaker or a known tilt for accelerometers, a mechanical gauge block for displacement. Whatever the method, the number you compare is the one that will be in the report.