Sensor Architecture, Excitation & Signal Conditioning

A bare sensing element rarely produces a trustworthy engineering value. The complete measurement channel is measurand → transducer → excitation and analogue front end → anti-alias filter → ADC → calibration and diagnostics. Each boundary can add loading, drift, noise, delay, saturation, or ambiguity, so channel accuracy cannot be inferred from the sensor data sheet alone.

Begin with an error-and-range budget #

Work backward from the required decision. A controller that must distinguish 0.2 °C changes does not merely need an ADC with nominally small LSBs: sensor tolerance, lead resistance, self-heating, reference drift, amplifier offset and noise, filter settling, quantization, and calibration residual must jointly fit the budget.

For independent standard uncertainties, combine them by root-sum-square,

$$u_c=\sqrt{u_1^2+u_2^2+\cdots+u_n^2},$$

but bounded worst-case errors, common reference errors, and correlated temperature drift must not be hidden inside that assumption.

Excitation is part of the measurement #

Passive resistive and reactive sensors require applied energy. Constant-voltage excitation makes a bridge naturally ratiometric when the same reference feeds the ADC. Constant-current excitation makes voltage proportional to resistance, but current creates self-heating. AC excitation and synchronous demodulation can move a weak signal away from DC offset and $1/f$ noise, yet add phase and settling constraints.

The analogue front end has four jobs #

1. Present the correct impedance to the sensor.
2. Translate differential or current output into the ADC input range.
3. Reject out-of-band energy before sampling.
4. Survive wiring faults, common-mode transients, and ESD without corrupting normal measurements.

Gain should place the largest credible signal below clipping with room for offset, tolerance, and transients. A 2 mV/V load cell at 5 V yields only 10 mV full scale; gain 250 maps it to 2.5 V, but a 2 mV bridge offset becomes 0.5 V after amplification and consumes one fifth of the range.

Conversion and calibration #

The ADC reference, input common-mode range, source impedance, acquisition time, integral nonlinearity, and effective noise all matter. Calibration identifies coefficients of the assembled channel, not merely the element. At minimum verify zero, span, monotonicity, overload recovery, warm-up drift, and a point not used to fit the calibration.

Diagnostic reasoning #

Probe the raw sensor nodes, conditioned analogue output, ADC code, and engineering-unit result separately. Without those observability points, software compensation can conceal a physical channel fault.