What changes in the AFE (core differences)

• Thermocouple front-end: must implement CJC and minimize thermal gradients at the connector/isothermal zone. Input integrity is about microvolt-level errors (thermal EMF, leakage, bias).
• RTD front-end: must provide excitation (typically constant current or ratiometric reference) and lead-resistance compensation (2/3/4-wire). Performance is often limited by excitation stability and self-heating.
• Shared blocks: low-drift amplification, 24-bit ΣΔ conversion, and digital filtering for mains rejection exist in both, but are tuned differently (gain, data rate, settling strategy).

Selection boundary (decision axes)

• Connector thermal gradients: large gradients or unstable airflow near terminals tends to hurt thermocouples unless an isothermal design is practical.
• Cable and pickup: long runs in a noisy environment require a plan for mains pickup and input bias/leakage; thermocouples are especially sensitive to microvolt-scale coupling.
• Channel count & scanning: high channel count with MUX scanning amplifies settling and filter latency. RTD 3-wire assumptions can fail if lead resistances drift asymmetrically.
• Power budget at the sensor: RTD excitation causes I²R heating; if the sensor is thermally insulated, even small currents can create measurable error.
• Range and robustness: thermocouples tolerate high temperature extremes; RTDs often win on accuracy and repeatability when self-heating and lead resistance are controlled.

Practical “must-prove” items (per sensor)

• Thermocouple: quantify CJC error sensitivity to a small terminal gradient; verify open-wire detection does not shift the measured microvolt level.
• RTD: quantify temperature shift versus excitation current (self-heating); verify 3-wire compensation holds across expected lead mismatch and connector aging.

Recommended order (and why it works)

1. Light ESD clamp near the connector: prevents large events from reaching the amplifier input. Choose parts whose leakage and leakage temperature drift do not create a measurable DC offset at the input.
2. Input RC right after the clamp: limits HF pickup and sharp edges before they hit the amplifier/PGA. The RC must be sized so its settling time is compatible with the target data rate (and with any channel scanning).
3. Series limiting (where needed): ensures the amplifier/ADC never sees a destructive or nonlinear input. Keep series resistance consistent and thermally stable to avoid creating a temperature-dependent gain/offset error.

Bias current × source impedance = hidden temperature error

• Thermocouple: the signal is on the order of tens of µV/°C. Any bias current flowing through cable resistance, input resistors, or protection leakage paths can generate a comparable µV-level offset. That offset becomes a fake “temperature shift” after linearization.
• RTD: the AFE measures voltage across a sensor driven by excitation. Input bias effects are often smaller than in thermocouples, but lead resistance and filter/series resistors can still create ratio errors and settling artifacts after switching.
• Design rule: treat every parasitic current path (bias + leakage) as a DC error source that can drift with temperature.

Differential vs single-ended (input topology boundary)

• Differential input preserves the sensor’s pair as a unit and is typically preferred for long cables and noisy environments, because the AFE can reject common-mode pickup before conversion.
• Single-ended input references one side to a local node. This can be acceptable for short, controlled wiring, but can convert cable pickup and reference shifts into apparent sensor changes.
• Practical test: if swapping cables or touching the shield/connector changes the reading, the input topology and bias paths are not controlled well enough.

Thermal EMF: the “unintended thermocouple” problem

Thermocouple-level signals are so small that dissimilar-metal junctions (connector pins, relays, screws, mixed alloys) can generate comparable voltages when a thermal gradient exists. The input stage must therefore minimize temperature gradients across junction pairs and keep sensitive junctions inside a locally uniform temperature region (the next section explains how that is created).

Input-stage checklist (actionable)

• Pick protection devices by leakage and leakage drift, not only by clamp voltage.
• Place the first clamp close to the connector, then apply RC filtering before the amplifier/PGA input.
• Keep differential paths symmetric (same components, same copper length, similar thermal environment).
• Use stable resistor technologies in the signal path to reduce temperature-dependent ratio errors.
• Assume contamination and humidity exist: design so small leakage currents do not dominate the error budget.

What CJC must achieve (the true requirement)

• Correct physical point: the CJC sensor must represent the temperature of the cold-junction node (where thermocouple metals transition to copper/terminal metals).
• Stable local temperature: the cold-junction node must be in a region with minimal temperature gradient across junction pairs.
• Separated error sources: CJC offset, CJC noise, and gradient-induced error must be measured and mitigated independently.

CJC sensor choice (AFE-relevant criteria)

• Drift & repeatability: long-term stability matters more than absolute “headline accuracy” when field recalibration is rare.
• Noise vs response time: lower noise often needs longer averaging; select a time constant that matches the temperature dynamics.
• Thermal coupling: package thermal resistance and mounting method decide whether the sensor measures the cold-junction node or merely the PCB’s average temperature.

Gradient control (the most common root cause)

• Gradient creators: terminal blocks, copper pours, screws, and airflow can create a temperature slope right across the cold-junction node.
• Design actions: short thermal paths, symmetric layout, keep hot components away, and add thermal mass where needed to reduce sensitivity to airflow and transient heating.
• Metal junction management: keep dissimilar-metal junction pairs inside the same isothermal zone to prevent thermal EMF from drifting with gradients.

Three CJC error buckets (separate them on purpose)

• CJC offset: a steady temperature bias from sensor calibration or poor thermal contact.
• CJC noise: short-term jitter that averages down with longer filter windows.
• Gradient error: drift caused by spatial temperature differences across the cold junction; often depends on airflow and nearby heat sources.

CJC checklist (actionable)

• Define the cold-junction node explicitly (terminal + metal transition), then place the CJC sensor to represent that node.
• Create an isothermal zone: use symmetric copper, avoid local heat sources, and add thermal mass if airflow exists.
• Validate gradient sensitivity: apply a small local airflow or nearby heat step and measure the resulting temperature error.
• Keep the mechanical stack-up stable: screw pressure and terminal contact should not create one-sided thermal gradients.

Excitation choices: constant current vs constant voltage

• Constant current: the RTD voltage is proportional to resistance (V = I·R), which is simple to linearize. The design must provide enough compliance voltage to cover RTD + leads + protection/filter drops, and must control current noise and drift.
• Constant voltage: the RTD current changes with temperature (I = V/R), so power and self-heating behavior varies with R(T). Lead resistance and contact changes can have stronger impact on the effective measurement.
• Practical preference: instrument-grade RTD AFEs typically favor constant current combined with ratiometric readout to cancel excitation drift.

2-wire: lead resistance directly becomes temperature error

In 2-wire wiring, both lead resistances add to the measured resistance: R_meas = R_RTD + 2R_L. The temperature error is approximately ΔT ≈ (2R_L)/(dR/dT). For Pt100, dR/dT ≈ 0.385 Ω/°C, so even small lead changes can create visible °C shifts.

• When it can work: short, stable leads; moderate accuracy targets; predictable terminals.
• Common failure: lead/contact resistance changes with temperature, vibration, or oxidation and appears as a sensor change.

3-wire: cancellation works only if lead resistances match

3-wire compensation relies on the assumption that two lead resistances are equal (or track each other): R_L1 ≈ R_L2. The AFE measures in a way that makes those two resistances enter the result symmetrically, so they cancel (often via switching or differential sensing).

4-wire (Kelvin): separate force and sense for best accuracy

4-wire wiring drives current on a dedicated pair (Force+ / Force−) and measures voltage on a separate pair (Sense+ / Sense−). Because the sense lines carry nearly zero current, lead drops contribute minimally to the measured voltage. This is the most reliable option for long leads, low-resistance RTDs, tight accuracy targets, and unstable connectors.

Ratiometric measurement: cancel excitation drift the right way

Use the same excitation and the same measurement path to read the RTD and a stable reference resistor R_REF. With constant current excitation, the ratio cancels current amplitude drift: V_RTD/V_REF = R_RTD/R_REF. This removes excitation drift from the error budget, but it does not fix lead mismatch (3-wire) or self-heating.

Self-heating: set the excitation current limit (engineering steps)

1. Allocate allowable self-heating contribution ΔT_allow inside the total temperature error budget.
2. Estimate worst-case thermal resistance to ambient (or use a conservative self-heating coefficient) for the RTD installation.
3. Compute RTD power under excitation: P = I²R (constant current) or P = V²/R (constant voltage).
4. Adjust excitation so the predicted temperature rise stays below ΔT_allow, then verify by measurement with a step in excitation level.
5. If multiplexing channels, account for duty cycle: average power decreases, but switching requires enough settling time for each channel.

What “24-bit” should mean in temperature terms

• Resolution: code width, not a guarantee of stable readings.
• Noise-free bits / ENOB: determines real stability at the chosen bandwidth.
• Temperature resolution: the useful metric is output RMS noise in °C: °C_RMS ≈ (input-noise)/(sensor sensitivity) (µV/°C for thermocouples, Ω/°C for RTDs).

How ΣΔ filtering creates strong mains rejection

• Digital filter shapes the response: sinc/FIR filters can create deep notches at 50 Hz and/or 60 Hz.
• Notch depth depends on output rate: deeper rejection typically requires a longer effective averaging window and a lower output data rate.
• NPLC intuition: longer integration windows average more of the mains cycle, lowering residual ripple at the output.

Trade-offs that must be declared

• Update rate vs mains rejection: higher update rates reduce averaging and often reduce notch depth.
• Filter group delay: strong filtering increases delay; after a step input, the output needs time to settle to a valid value.
• Multi-channel scanning: channel switching requires enough settling time; otherwise readings include filter memory (“ghosting”).
• Over-filtering: can hide real temperature changes and slow detection of genuine events (still within temperature acquisition context).

Acceptance definition (simple and measurable)

• Define the final data rate: specify output samples/second and filter mode.
• Inject mains interference: apply a known 50 Hz or 60 Hz component and measure residual output ripple at that configuration.
• Convert to °C: report residual in µV_RMS/Ω_RMS and in °C_RMS.
• Specify settling: after a channel switch or step input, define the waiting time until output is valid.

Step-by-step conversion: input-referred → °C

1. Lock the configuration: output data rate and digital filter mode define bandwidth and mains rejection.
2. Pick sensitivity at the operating point: Thermocouple uses S(T)=dV/dT (µV/°C). RTD uses dR/dT (Ω/°C), and with constant-current excitation, dV/dT = I_EXC·dR/dT.
3. Express each contributor as input-referred: (µV_RMS or Ω_RMS for noise, or µV/Ω for bias terms).
4. Convert to temperature: Thermocouple: °C_RMS ≈ V_RMS/S(T). RTD (resistance domain): °C_RMS ≈ R_RMS/(dR/dT). RTD (voltage domain): °C_RMS ≈ V_RMS/(I_EXC·dR/dT).
5. Combine correctly: RSS for uncorrelated noise terms; keep systematic terms as allocated limits and sum by worst-case policy.

Source layers (and what they become)

• Sensor & cable: source resistance noise; pickup that shows up as measured input-referred ripple (often near 50/60 Hz and harmonics).
• Amplifier/PGA: voltage noise e_n, current noise i_n×R_SOURCE, and 1/f that dominates low-frequency stability.
• ADC: input-referred noise at the chosen OSR/filter; “24-bit” does not imply noise-free temperature resolution.
• Reference & excitation: drift/noise maps to gain error or ratiometric residual; treat as systematic unless proven random within the bandwidth.
• Thermal error terms: thermal EMF, CJC offset/noise, gradient error, and RTD self-heating (often dominant in °C even when electrical noise is small).

Minimum auditable budget table (structure)

A usable budget must record the domain, the conversion, and the verification method. The list below is a “table template” that can be copied line-by-line during implementation.

The budget is complete when the dominant term is identified and tied to a concrete mitigation action.

Dominant-term diagnosis (why budgets matter)

• If CJC gradient or thermal EMF dominates, increasing ADC bits will not improve °C accuracy.
• If mains residual dominates, the fix is filter window/data rate and input hygiene, not a “lower-noise ADC”.
• If RTD self-heating dominates, reduce excitation or use duty-cycled scanning and validate the rise empirically.

What MUX/relays/solid-state switches change (AFE-level)

• Leakage paths: temperature- and contamination-dependent leakage can create input bias errors, especially at high impedance nodes.
• Charge injection / step disturbance: switching injects a transient that must settle before the channel is valid.
• Thermal EMF at junctions: more contacts and dissimilar metals increase the chance of µV-level offsets that drift with gradients.

Settling has three components (each must be budgeted)

1. Input RC + source impedance: cable + protection/filter form a time constant that differs by sensor and wiring.
2. PGA/amplifier settling: gain or input changes can require extra time beyond the simple RC constant.
3. ΣΔ filter group delay: strong mains rejection increases filter memory; after a channel switch, output needs multiple cycles to flush prior history.

Practical scanning recipe (repeatable and auditable)

1. Switch channel (keep routing consistent; avoid unnecessary gain changes).
2. Wait analog settling (RC + amplifier) until the residual is below the allocated °C error.
3. Discard N output samples to clear digital filter memory (N depends on filter mode and output rate).
4. Capture M valid samples inside a defined window; optionally average to reduce random noise.
5. Group channels by similar source impedance and required filter settings to share one dwell policy.

Built-in self-check channels (catch drift vs real temperature)

• Shorted-input (TC) check: a periodic shorted channel monitors AFE offset/µV drift and highlights thermal EMF issues.
• Reference resistor (RTD) check: a fixed R_REF channel monitors excitation/ratio stability and long-term gain behavior.
• Scan-order sanity: schedule check channels at stable intervals to detect “scan-induced” artifacts early.

Thermocouples: segmented polynomials with CJC correctly applied

• Keep “type” as configuration: firmware selects one coefficient set (and its valid range), not a universal table.
• Segmented polynomial discipline: the implementation must define how segments are selected and must avoid discontinuities at segment boundaries.
• CJC is an EMF add, not a temperature add: the robust workflow is: convert T_cold to V_cold (equivalent thermocouple EMF), add V_hot=V_meas+V_cold, then invert EMF to T_hot.
• Range awareness: if the measured EMF or inferred temperature is outside the configured range, output overrange rather than a misleading number.

RTDs: CVD mapping and coefficient realism

The Callendar–Van Dusen (CVD) model maps resistance to temperature with a small set of coefficients. Firmware should treat the RTD nominal and coefficient set (including α variants) as configuration, not a hard-coded assumption, because coefficient mismatch becomes a systematic temperature error.

• Inversion method: for R→T, use a controlled method (iteration or a bounded LUT) that is stable across the intended temperature range.
• Coefficient integrity: store coefficient set IDs and apply the same set during calibration and normal operation.
• Do not over-model: keep the model simple; use calibration to remove repeatable offset/gain rather than fitting high-order curves.

Calibration strategy (separate what can be separated)

• End-to-end (2-point): removes repeatable offset and gain error over a defined range with minimal complexity.
• Multi-point: improves residuals only if the thermal condition is stable and points are well-distributed; avoid high-order fits.
• Calibrate CJC separately: treat CJC offset as its own term; do not “hide” it inside the thermocouple end-to-end fit.
• Per-channel trims: use channel-level offset/gain trims to enforce channel-to-channel consistency in scanned systems.

Avoid “calibrating noise” (hard rules)

1. Only calibrate in stable thermal conditions: require a low temperature-change-rate window before capturing calibration points.
2. Average long enough: the averaging window must match the final filter/data rate (otherwise mains residual is fitted as offset).
3. Use bounded fits: prefer 2-point or low-order correction; high-order fits easily absorb short-term noise and worsen long-term accuracy.
4. Freeze configuration: store which filter mode/data rate was used during calibration and warn if operation differs.

Calibration data hygiene (auditable firmware behavior)

• Metadata: coefficient set ID, calibration version, temperature range label, and measurement configuration (filter/data rate).
• Runtime reporting: expose which calibration set is active and whether it is valid for the current mode.
• Fail-safe: if calibration is invalid or missing, output a clear quality status rather than silently using defaults.

Thermocouple checks (channel-level)

• Open circuit: use a controlled bias/burnout current so a floating input moves to a recognizable extreme; then assert OPEN.
• Reversed polarity: detect sign inconsistency and trend inconsistency (temperature increases but EMF decreases); assert REVERSED.
• Overrange / clipping: if input protection or ADC rails, assert OVERRANGE and suppress misleading temperature output.

RTD checks (open/short and 3-wire failure cues)

• Open / short: compare inferred resistance against physically plausible bounds for the configured RTD and range; assert OPEN or SHORT.
• 3-wire compensation failure: if lead mismatch breaks the cancellation assumption, readings can drift with terminal temperature or show step-like jumps; assert LEAD_MISMATCH or degrade quality.
• Overrange: if the inferred R(T) exceeds configured range, assert OVERRANGE instead of extrapolating beyond validity.

Plausibility checks (reduce false positives)

• Rate limit: apply a maximum dT/dt bound per channel; if exceeded, mark SUSPECT rather than immediately declaring a hard fault.
• Neighbor consistency (only when justified): compare to physically adjacent channels of the same sensor class; use as a quality hint, not a sole fault trigger.
• Instability detector: if short-term variance exceeds a threshold at fixed filter settings, set UNSTABLE quality.

Define the diagnostic interface (what to output)

Temperature is trustworthy only when fault_flags are clear and quality_bits indicate OK.

How to run the checklist (audit-friendly)

1. Freeze the mode: data rate + filter + notch + range/excitation define the measurement bandwidth and latency.
2. Log raw + temperature: capture raw code (or input-referred value) and final °C output in the same stream.
3. Discard settling: for scan tests, discard initial samples after each switch before computing metrics.
4. Attach metadata: store calibration ID, coefficient set ID, and measurement mode in the test report.

Test 1 — Noise & resolution (short-term stability)

• Setup: thermocouple input short (low-thermal-EMF shorting fixture); RTD simulated by a precision resistance source.
• Example lab items (equivalents OK): IET Labs RS-200 decade resistance substituter; Keysight 34465A/34470A, Keithley DMM6500, or Fluke 8846A for cross-checks.
• Procedure: after warm-up, log N seconds of output at the final filter/data-rate; compute °C_RMS and °C_PP.
• Metric: °C_RMS (and optionally Allan deviation at τ=1–10 s for stability profiling).
• Pass criteria: °C_RMS ≤ allocated random-noise budget for this mode (criteria must be stated together with filter/data rate).

Test 2 — 50/60 Hz rejection (notch effectiveness)

• Setup: inject a controlled 50 Hz / 60 Hz interference component through a repeatable coupling network.
• Example lab items (equivalents OK): Keysight 33500B series or RIGOL DG4000 series function generator; Vishay RN55/RN60 metal film resistors; shielded twisted-pair such as Belden 1800F.
• Procedure: capture with notch OFF and notch ON at the same output data rate; compute residual ripple in °C and the rejection in dB.
• Metric: Rejection(dB) = 20·log10(Injected / Residual) and residual ripple (°C_PP or °C_RMS).
• Pass criteria: residual at 50/60 Hz ≤ allocated mains-residual budget (or rejection ≥ target dB) for the selected mode.

Test 3 — CJC gradient sensitivity (airflow & local heating)

• Setup: hold the thermocouple input at a stable point (short or simulator), then apply controlled airflow or localized heating near the cold-junction region.
• Example lab items (equivalents OK): Noctua NF-A4x10 fan (repeatable airflow); 3M 8810 thermal pad or 3M 986 thermal tape (mounting/isothermal experiments); aluminum/copper isothermal block (simple machined fixture).
• Procedure: run a baseline, apply a defined disturbance step (fan on/off or mild local heat), log the CJC reading and final temperature output.
• Metric: CJC-induced temperature error (°C step response, °C drift per disturbance condition).
• Pass criteria: worst-case CJC error under the defined disturbance ≤ allocated CJC/gradient budget.

Test 4 — RTD self-heating curve (excitation current sweep)

• Setup: a real RTD probe in a stable thermal environment (ice bath, thermal block, or controlled ambient) while sweeping excitation current.
• Procedure: step excitation current across intended operating points; wait for thermal steady state at each step; log temperature output.
• Metric: ΔT_self-heat(I) = T(I) − T(I_low) and slope vs I² (for model sanity).
• Pass criteria: ΔT_self-heat at nominal excitation ≤ allocated self-heating budget; if not, reduce I or duty-cycle measurement.

Test 5 — Scan settling (MUX + analog settle + ΣΔ latency)

• Setup: two stable inputs (A and B) alternated by scan order; evaluate dwell time and discard-sample policy.
• Example lab items (equivalents OK): Omega CL543B thermocouple simulator or Fluke 714B calibrator; IET Labs RS-200 for RTD simulation; switching examples: ADI ADG1208/ADG1219 (low-leakage analog switches) or Pickering reed relay modules (low-thermal-EMF switching class).
• Procedure: switch A↔B; for each dwell setting, compute error vs time-after-switch (or sample index) until it converges.
• Metric: settling curve: error(t_wait) or error(k); minimal dwell/discard to meet the allocated error.
• Pass criteria: at target channel count & update rate, the required dwell/discard still meets the allocated settling error budget.

Test 6 — Warm-up & drift (prove stability over time)

• Setup: stable input (short/simulator/decade box) and a simple logger (PC or microcontroller capture).
• Procedure: log from power-on through warm-up (e.g., 1–2 hours), then extend to long duration (e.g., 24 hours) at a lower sample rate.
• Metric: time-to-spec (minutes to enter ±threshold band), and drift (°C over 24 h, or °C_PP).
• Pass criteria: warm-up time ≤ limit and 24 h drift ≤ allocated drift budget for the intended environment.

Minimal report template (copy/paste)

Lab BOM examples (part numbers to start fast)

• TC / temperature sources: Omega CL543B (TC simulator class), Fluke 714B (temperature calibrator class)
• RTD / resistance simulation: IET Labs RS-200 (decade resistance substituter class)
• Cross-check DMMs: Keysight 34465A/34470A, Keithley DMM6500, Fluke 8846A
• 50/60 injection: Keysight 33500B series or RIGOL DG4000 series; Vishay RN55/RN60 resistors; Belden 1800F shielded pair
• Scan switching (examples): ADI ADG1208, ADI ADG1219; Pickering reed relay modules (low-thermal-EMF switching class)
• CJC gradient fixtures: Noctua NF-A4x10 fan; 3M 8810 thermal pad; 3M 986 thermal tape; simple aluminum/copper isothermal block

Equivalent instruments and fixtures are acceptable as long as the setup is repeatable and the injected stimulus is quantified.