Ultrasonic Flowmeter (Clamp-on/Wet): AFE, TDC & Temp Compensation
← Back to: Industrial Sensing & Process Control
Core idea
Ultrasonic flowmeter accuracy is decided by a measurable evidence chain: transducer/acoustic path → Tx/Rx AFE integrity → Δt/Δφ extraction → temperature compensation and versioned calibration in NVM. This page shows what to measure first and how to isolate each error source so clamp-on and wet designs can stay stable from the bench to long-term field drift.
What This Page Solves
This page reduces “ultrasonic flowmeter” to a single engineering objective: extract a reliable time difference (Δt) and/or phase difference (Δφ) between upstream and downstream acoustic propagation, then keep that result stable across temperature, medium changes, and long-term drift—without drifting into protocols or full-system topics.
Two sensing families are covered (clamp-on and wet), but only in terms of how they reshape the error budget and the signal chain evidence you must collect:
- Model certainty: wet meters typically have a more predictable acoustic path, while clamp-on meters add pipe-wall and coupling-layer terms that change the effective path length and dispersion. The result is often a wider or multi-peak correlation response and more frequent phase ambiguity.
- Signal quality: wet paths are dominated by medium attenuation and flow-induced scattering; clamp-on paths more often face mode conversion and multipath through the pipe wall, which turns “clean waveforms” into unreliable Δt/Δφ if confidence is not tracked.
- Drift character: wet drift is frequently medium- and temperature-driven; clamp-on drift is frequently compounded by coupling repeatability and wall temperature gradients—making compensation and calibration storage central, not optional.
The page is organized as a measurement pipeline: Tx burst generation → Rx AFE integrity → Δt/Δφ extraction → clock/TDC limits → temperature & medium compensation → NVM calibration fields. Every later “debug” statement must map back to a measurable evidence field (waveform, confidence score, or stored calibration term).
Measurement Physics Refresher
The goal is not a textbook summary. The goal is to define what is observable and how the acoustic path turns those observables into flow—so later chapters can tie every failure mode to a measurable evidence field.
Two primary observables: Δt and Δφ
- Transit-time (Δt): measure upstream and downstream time-of-flight (TOF) and compute Δt = TOF_up − TOF_down. In many practical designs, Δt is much smaller than absolute TOF, which forces strict control of timing jitter, interpolation error, and correlation confidence.
- Phase shift (Δφ): measure phase relative to a reference (or between directions) and form a phase difference that can be highly sensitive to small delays under stable narrowband conditions. The trade-off is phase ambiguity (2π wrapping) and higher susceptibility to multipath and frequency-dependent group delay.
Clamp-on vs wet: the error geometry changes
- Wet: the dominant uncertainty is usually the medium (sound velocity vs temperature/composition) and flow-induced scattering. Path geometry is often more repeatable, so a single dominant arrival is common.
- Clamp-on: the path includes pipe wall and coupling layer, which can introduce dispersion, mode conversion, and multipath. Correlation peaks can become wider or multi-peak, and phase can become non-linear vs frequency, requiring explicit confidence metrics and more conservative filtering choices.
Evidence fields (what later chapters must point back to)
To keep the page auditable, evidence is grouped into three layers. Later chapters must reference these fields explicitly (waveforms, derived metrics, or health flags) rather than introducing new “invisible” causes.
- Layer 1 — Primary observables: TOF_up, TOF_down, φ_up/φ_down (or φ_rx relative to φ_tx), plus signal-quality fields such as SNR, burst amplitude, and ring-down duration.
- Layer 2 — Derived metrics: Δt, Δφ, correlation peak location and peak width, main-peak / side-peak ratio.
- Layer 3 — Confidence & health: correlation confidence score, phase unwrap error count, saturation flags, AGC/VGA state, and timing-quality counters (clock lock, jitter estimate).
These observables directly impose engineering constraints on the electronics: stable timebase (jitter budget), controlled group delay (filtering and AFE), and explicit confidence tracking (to detect multipath and drift).
Transducer & Acoustic Path
Many failures that look like “electronics instability” originate upstream: the transducer and acoustic path determine whether a single dominant arrival exists and whether its frequency response remains stable enough for Δt/Δφ extraction.
Transducer parameters that directly shape observables
- Center frequency & bandwidth: bandwidth controls correlation sharpness (time resolution) while center frequency sets phase sensitivity. Narrowband operation can benefit phase-based methods but is more sensitive to dispersion and multipath; wider bandwidth can stabilize correlation peaks but increases AFE and sampling demands.
- Q factor and ring-down: high-Q transducers ring longer, increasing overlap between direct and reflected components. This widens correlation peaks and raises phase ambiguity. Low-Q transducers reduce ring-down but can push the link into an SNR-limited regime.
- Drive headroom and repeatability: the goal is stable acoustic energy delivery, not maximum amplitude. Over-driving can amplify secondary paths; under-driving collapses SNR and confidence.
Clamp-on: coupling and pipe wall turn geometry into an error term
- Coupling-layer loss and frequency dependence: coupling materials add attenuation and can reshape the effective passband. A frequency region that looks “good” on one pipe may become SNR-poor on another.
- Pipe diameter/material effects: the pipe wall can support additional modes and reflections. The measurable result is often a wider peak (or multiple peaks) and a phase response that becomes non-linear across frequency.
Wet: medium-driven randomness dominates
- Direct coupling vs liquid attenuation: a more predictable geometric path is common, but attenuation and scattering vary with medium properties and temperature.
- Bubbles and turbulence: these introduce stochastic amplitude fading and arrival-time jitter, raising phase noise and lowering correlation confidence even when waveforms appear visually “clean”.
Transmit Path Design
The transmit chain is a metrology reference: it must deliver repeatable acoustic excitation with controlled timing. “Louder” is not the target; stable edges and a stable envelope are what keep Δt/Δφ estimators from drifting or becoming ambiguous.
Burst strategy: choose a waveform to stabilize confidence
- Short burst / single pulse: minimal complexity and clean timing references, but limited processing gain. When SNR is marginal or multipath is strong, peak confidence can collapse.
- Coded burst: trades implementation complexity for processing gain, sharpening correlation confidence in noisy paths. It also increases sensitivity to timing consistency across the code sequence.
- Chirp: sweeps frequency to find stable passbands and raise SNR, but makes group delay and filtering consistency more critical.
Driver and matching: keep the burst envelope stable
- Half-bridge / H-bridge drive: select for controlled output impedance and edge quality under varying transducer load. The driver must avoid edge-induced ringing that contaminates the timing reference.
- Impedance matching: poor matching can cause burst amplitude and start phase to drift with temperature and mounting variation, turning a transmit-side instability into an apparent receive-side or algorithmic problem.
Timing stability: jitter becomes measurement noise
- Edge jitter: unstable edges broaden correlation peaks and increase arrival-time uncertainty.
- Envelope stability: amplitude modulation over temperature or supply drift changes SNR and bias, especially near thresholds where confidence flips.
- Trigger-to-burst latency consistency: the delay between a digital trigger and the actual acoustic launch must remain stable to preserve Δt/Δφ repeatability.
Receive AFE Architecture
The receive front-end is not a “make it louder” chain. It is a metrology chain whose job is to preserve the dominant arrival’s timing and phase information. Noise, non-linearity, group delay distortion, and gain dynamics can silently convert a stable acoustic path into unstable Δt/Δφ measurements.
LNA: noise density versus input capacitance
- Input capacitance is a system variable: transducer capacitance plus PCB parasitics and protection capacitance shape effective bandwidth and noise gain. As Cin rises, it becomes harder to keep low input-referred noise while maintaining a phase-stable passband.
- Clamp-on sensitivity: clamp-on paths often operate closer to an SNR cliff due to coupling loss and multipath. A modest increase in input-referred noise can push correlation confidence below a threshold, causing peak instability and apparent “random” timing errors.
TIA / VGA trade-offs: preserve information, not just amplitude
- TIA-oriented designs: can provide a controlled input node for capacitive sources, but compensation and bandwidth choices impact phase response. Poorly controlled phase response manifests as measurement jitter downstream.
- VGA / variable gain designs: recover dynamic range across different pipes and media, but gain transitions and settling behavior create time-varying amplitude/phase effects. If the gain is still settling inside the measurement window, Δt/Δφ becomes biased and less repeatable.
Band-pass filtering: Q defines group-delay risk
- Q too high: group delay becomes steep and frequency-sensitive. When the received spectrum shifts (temperature, coupling, pipe modes), the effective delay and phase distort, increasing phase noise and widening correlation peaks.
- Q too low: wideband noise raises the correlation floor. The peak may remain visible, but its confidence drops, leading to peak selection instability in multipath conditions.
AGC side effects: gain dynamics can look like physics
- Correlation window sensitivity: gain motion reshapes the waveform within the correlation window, widening peaks or favoring a secondary arrival.
- Phase window sensitivity: gain-dependent phase response and settling artifacts raise phase noise and increase unwrap errors, especially when clamp-on multipath mixes arrivals.
Time & Phase Extraction Paths
Δt and Δφ are produced by two parallel hardware-verifiable paths that start from the same received waveform. Each path has distinct sensitivities to noise, bandwidth, sampling rate, group delay, and frequency drift. Treating these outputs as “pure algorithm results” hides the evidence required for debugging and validation.
Path A: Cross-correlation / TOF
- Signal chain: ADC samples → buffering/windowing → correlation engine → peak location (TOF) → Δt from upstream/downstream.
- Primary sensitivities: SNR and bandwidth define peak sharpness; sampling rate and interpolation define timing granularity. Clamp-on multipath can create multiple peaks, reducing confidence and causing peak selection instability.
- Evidence symptom: correlation peak width and peak-ratio trends indicate whether the chain is noise-limited, bandwidth-limited, or multipath-limited.
Path B: Phase / Frequency domain
- Signal chain: band-limited Rx → IQ/phase detector → phase difference → unwrap/validation → Δφ (and optional delay equivalent).
- Primary sensitivities: temperature-driven spectrum shifts, frequency offset, and AFE group delay nonlinearity raise phase noise and bias. Multipath mixes arrivals and increases unwrap errors.
- Evidence symptom: phase noise and unwrap error count separate clock/AFE stability issues from path ambiguity issues.
TDC vs ADC-Based Timing
Picosecond-class timing is not magic. It is the result of a closed budget: a stable timebase, a controlled interpolation method, and a waveform whose shape remains repeatable inside the measurement window. When any of these breaks, “high resolution” turns into unstable Δt.
TDC: fine quantization with calibration obligations
- Delay-line TDC: converts an edge/arrival event into a fine time code using a chain of delay elements. The practical cost is PVT drift; bin widths change with temperature and supply, so calibration or background tracking is required to keep the code meaningful.
- Vernier TDC: uses two slightly different delay chains to create a smaller effective time step. The cost is higher complexity and stronger sensitivity to mismatch drift; excellent nominal resolution can degrade into higher jitter if calibration is weak.
ADC oversampling + interpolation: stability-first timing
- Core idea: higher sampling density and stable filtering make the correlation peak (or edge model) smoother, enabling interpolation beyond raw sample spacing. The method is only as good as waveform repeatability—group-delay distortion or gain motion can move the “shape” and bias the fitted time.
- Engineering cost: power and bandwidth shift into ADC, buffering, and compute. In return, the method can be robust when the AFE preserves shape and the path supports a stable dominant arrival.
Clock jitter budget: the hard ceiling
- Reference jitter: sets the base limit. If the reference is noisy, finer TDC bins cannot translate into stable Δt.
- PLL phase noise: trading frequency flexibility for added phase noise affects both TDC event timing and ADC sampling aperture.
- Distribution consistency: clock gating, domain crossings, and trigger routing create additional uncertainty if not tightly controlled.
Decision rule: when TDC is required vs when ADC is enough
- TDC tends to be required when Δt is extremely small and must be resolved in a short window with limited averaging, or when ADC sampling rate/power cannot be increased to support stable interpolation.
- ADC interpolation tends to be enough (and sometimes more stable) when the AFE preserves waveform shape (controlled group delay), correlation confidence is strong, and the system can trade data/compute for averaging under stochastic media noise.
Temperature & Medium Compensation
A stable Δt/Δφ measurement does not guarantee accurate flow. The dominant error often comes from sound-speed modeling: c varies with temperature and medium, and clamp-on systems add a unique risk—wall temperature can diverge from fluid temperature. Compensation must therefore be model-driven and evidence-auditable.
Sound speed is the hidden scale factor
- c(T, medium): sound speed changes with temperature and medium properties. A small modeling error becomes a systematic flow bias because Δt is converted using c as a scale factor.
- Where it enters: the conversion from TOF/Δt/Δφ to velocity uses c explicitly or implicitly; the same Δt implies different flow when c changes.
Clamp-on: wall vs fluid temperature mismatch
- Mismatch mechanism: temperature sensors are often mounted on the pipe wall. The wall temperature can lag or differ from fluid temperature due to insulation, ambient airflow, or transient process changes.
- Resulting failure mode: the model receives the wrong temperature input, selects an incorrect sound-speed value, and produces a coherent flow bias that drifts with ambient or process conditions.
Model strategy: real-time sensing → characterization → segmented calibration
- Real-time temperature sensing: the key risks are offset and thermal time constant. A “correct” sensor with the wrong offset or slow response feeds incorrect c(T) during transients.
- Characterization curves: sound-speed behavior should be represented as a LUT or curve derived from characterization of the target medium and configuration.
- Segmented calibration: different temperature ranges or configurations can require segmented models to avoid boundary drift. Each segment should be traceable and versioned.
Calibration, NVM & Drift Control
“Accurate once” is not the goal. Long-term accuracy requires a parameter system: what gets calibrated, where it is stored, how it is versioned and validated, and how drift is detected early before it becomes a field failure.
Factory calibration vs field calibration
- Factory calibration: establishes a controlled baseline (zero-flow offset, scaling constants, temperature coefficients) with stable references. It is designed for repeatability and traceability.
- Field calibration: absorbs configuration differences (pipe/material class, clamp-on geometry factors, coupling state) that cannot be predicted at the factory. It must be guarded to prevent short-term noise or unstable conditions from being written as permanent truth.
What belongs in NVM: layered parameters
- Long-term constants: zero-flow offset, temperature coefficients, and timing/scale constants (when applicable). These turn raw Δt/Δφ into stable flow under a known baseline.
- Installation-specific parameters: configuration identifiers and geometry factors required to remain valid after pipe changes or clamp-on rework. They must be editable without overwriting factory baselines.
- Health & trend snapshots: low-rate summaries that reveal aging before failure. A stable flow chain with a drifting zero-flow offset indicates electronic drift; a shrinking confidence margin indicates acoustic aging or coupling degradation.
Drift sources and what they look like
- AFE offset drift: tends to appear as slow zero-flow bias movement across temperature or time, even when peak shape and SNR remain stable.
- Transducer aging / coupling loss: tends to reduce SNR and widen correlation peaks, increasing jitter at high flow and raising the required gain.
- Configuration changes: pipe/material changes shift acoustic modes and invalidate installation parameters; the correct response is a controlled update, not silent parameter reuse.
Error Sources & Debug Playbook
Debugging should start from symptoms, not theories. Each failure mode below is organized as: symptom → first two captures (fields/waveforms) → most likely subsystem → first fix. This keeps triage fast and evidence-driven.
Symptom A: zero-flow is not zero
- First capture (2): zero-flow offset trend; phase noise (or correlation peak width if TOF path is primary).
- Most likely subsystem: AFE offset drift or calibration baseline mismatch; temperature input offset/mismatch affecting sound-speed scaling.
- First fix: verify baseline parameters (version/CRC) and re-establish a valid zero-flow reference under stable thermal conditions.
Symptom B: high-flow jitter increases
- First capture (2): correlation peak width (and peak ratio if available); gain settling/AGC state (or a gain trend proxy).
- Most likely subsystem: multipath/turbulence creating multi-peak ambiguity; band-pass/group-delay distortion; timing chain jitter budget limit.
- First fix: validate whether jitter is confidence-limited (peak width) or timebase-limited (clock status); then stabilize AFE dynamics before tuning extraction thresholds.
Symptom C: drift after temperature change
- First capture (2): temperature sensor offset; sound-velocity LUT index/segment ID.
- Most likely subsystem: clamp-on wall-vs-fluid temperature mismatch; model segmentation/selection issue; untracked temperature dependence in timing scale (if TDC calibration is weak).
- First fix: confirm the temperature input is representative (offset and response), then confirm model segment selection is stable and traceable.
Symptom D: clamp-on works on one pipe, fails on another
- First capture (2): peak width / confidence proxy; installation parameter version/ID (pipe class, geometry ID).
- Most likely subsystem: acoustic mode change in the new pipe; filter center/Q mismatch causing group-delay warp; stale installation parameters reused without a controlled update.
- First fix: treat the pipe change as a configuration change: update installation parameters through a guarded process and re-validate confidence margins.
Validation & Test Signals
Validation must be bench-reproducible. The goal is to verify the full Δt/Δφ chain using controlled stimulus (delay, attenuation, echo), then quantify repeatability σ and drift versus temperature/time so the same evidence can be reused for calibration policy and debug triage.
1) Simulated acoustic path: make uncertainty controllable
- Delay injection (TOF surrogate): create a known time shift between a “Tx reference” and a synthetic “Rx arrival” so the extraction path can be verified without real fluid. The same setup enables linearity sweeps and step-response checks.
- Attenuation / SNR shaping: apply programmable amplitude reduction and noise shaping to emulate weak coupling or lossy media. A stable chain should degrade gracefully: confidence falls first (peak width rises) before bias explodes.
- Echo / multi-peak surrogate: sum a delayed, lower-amplitude copy of the main pulse/burst to emulate reflections. This validates the “confidence and ambiguity detection” behavior used by the debug playbook.
2) Pseudo-flow injection: sweep Δt and Δφ on purpose
- Δt sweep (TOF path): step or ramp the injected delay. Check that output Δt tracks the injected value with stable σ across the operating range.
- Δφ sweep (phase path): inject a small controlled phase ramp (or equivalent frequency offset) and verify phase unwrap stability and phase-noise growth under attenuation/AGC.
3) Environmental stress: quantify drift instead of arguing about it
Temperature chamber
Hold a fixed injected delay/phase. Measure repeatability σ at each temperature point, then record drift vs temp as the change in mean output. This directly validates compensation inputs and calibration stability.
Vibration / handling stress
Keep stimulus fixed and track confidence proxies (peak width/ratio) and output σ. If σ expands while stimulus is unchanged, the chain is sensitivity-limited (signal integrity) rather than model-limited.
4) Evidence fields for bench and production
- Repeatability σ: standard deviation under fixed stimulus (Δt σ, phase σ, or flow σ depending on the reporting layer).
- Drift vs temperature: mean output change across temperature at fixed stimulus.
- Drift vs time: long hold test (e.g., hours/days) summarized as maximum drift and slope.
- Parameter traceability: calibration version + checksum/CRC stored with test logs to avoid “unknown parameter state” investigations.
Example MPNs for a practical stimulus + logging bench
The MPNs below are reference building blocks used in many lab/production fixtures for timing injection, amplitude shaping, sensing, and traceable logging. Equivalent parts are acceptable if they meet bandwidth, jitter, and resolution targets.
| Test function | MPN examples | Why it fits the fixture |
|---|---|---|
| Time-of-flight delay injection |
DS1023-50 (Maxim/Analog) ·
TDC-GPX2 (ams OSRAM) ·
TDC7200 (TI)
|
Programmable delay line or TDC-grade timing reference to create/measure known sub-ns timing changes; supports Δt sweep and repeatability σ. |
| Low-jitter clocking (fixture timebase) |
SiT9365 (SiTime) ·
LTC6957 (Analog Devices)
|
Stable stimulus timing requires a clean timebase; improves test-to-test repeatability and prevents clock noise from dominating σ. |
| Synthetic burst / waveform generation |
AD9106 (Analog Devices) ·
AD9837 (Analog Devices)
|
Generates repeatable ultrasonic-like bursts or phase-continuous tones for Δφ validation; supports controlled ramps/steps. |
| Amplitude control / attenuation shaping |
AD8336 (Analog Devices) ·
AD5260 (Analog Devices) ·
DS1803 (Maxim/Analog)
|
Creates controlled SNR/attenuation profiles to test confidence behavior and AFE settling sensitivity without changing the stimulus truth. |
| Summing for echo / multi-peak surrogate |
ADA4899-1 (Analog Devices) ·
OPA828 (TI)
|
High-speed op-amp summer can combine a main waveform and delayed/attenuated copy to emulate reflections with predictable ratios. |
| Analog switching / routing |
ADG704 (Analog Devices) ·
TMUX1102 (TI)
|
Enables stimulus path selection (single/echo/noise) and supports production fixtures where a single instrument feeds multiple channels. |
| Stimulus/response digitization (fixture capture) |
AD9234 (Analog Devices) ·
ADS8354 (TI)
|
Captures stimulus/response snapshots for audit logs and σ computation; supports verifying “what was injected” versus “what was reported.” |
| Temperature sensing (fixture reference) |
TMP117 (TI) ·
MAX31865 (Analog Devices)
|
Provides stable temperature reference for drift-vs-temp characterization; RTD interface improves traceability for chamber runs. |
| Nonvolatile log / parameter store (fixture or DUT) |
MB85RS64V (Fujitsu FRAM) ·
24LC256 (Microchip EEPROM)
|
Supports versioned records (stimulus settings, σ, drift curves) with checksum/CRC; FRAM is robust for frequent writes in production. |
| Controller for automation + logging |
STM32H743 (ST) ·
ESP32-S3 (Espressif)
|
Drives sweeps (delay/attenuation/phase), computes σ and drift summaries, and stamps logs with version/CRC for audit-ready results. |
Request a Quote
FAQs
Each answer follows the same rule: a short diagnosis, two measurements to capture first, then one first fix. MPNs are shown as concrete reference blocks (AFE/clock/TDC/temp/NVM) only.
1
Zero-flow still shows non-zero flow — offset drift or grounding loop?
Most “zero not zero” cases are either true AFE offset drift or external injection that looks like signal. A drifted baseline
will move slowly over time/temperature; an injected artifact is often environment-correlated. A low-noise front-end
(e.g., OPA828 or ADA4899-1) reduces ambiguity but does not replace trend validation.
- What to measure first (2): zero-flow offset trend; input-referred noise / baseline at fixed gain.
- First fix (1): verify calibration version+CRC, then re-establish a valid zero reference under stable thermal conditions.
2
Works in lab, fails after clamp-on install — acoustic coupling or temp mismatch?
A post-install failure is usually confidence collapse (coupling/path mode change) or temperature representativeness (wall vs fluid).
Coupling loss shows up as SNR vs frequency degrading and peak width widening; temperature mismatch shows as stable SNR but systematic bias with temperature steps.
Use a stable sensor like TMP117 to separate “reads correctly” from “represents the medium.”
- What to measure first (2): SNR vs frequency (or peak width proxy); temperature offset/response (lag) during a step.
- First fix (1): lock down a repeatable coupling state (as a parameter ID) before tuning any compensation curve.
3
Correlation peak jumps under high flow — SNR collapse or ADC jitter?
Peak “jumping” is often an ambiguity problem: multipath/turbulence widens the peak and creates competing maxima before clock jitter becomes dominant.
If the ADC/timebase is the limiter, the error looks more uniform and less tied to amplitude changes. A clean capture chain
(e.g., AD9234 class) helps, but peak width is the primary truth signal.
- What to measure first (2): correlation peak width (and peak ratio if available); AFE gain settling / overload recovery behavior.
- First fix (1): stabilize Rx confidence (band-pass + dynamic gain behavior) before tightening correlation thresholds.
4
Phase path stable but TOF noisy — bandwidth or clock jitter?
When phase is stable but TOF is noisy, the TOF estimator is usually confidence-limited (bandwidth/SNR) or timebase-limited (jitter budget).
If peak width grows with attenuation, it is bandwidth/SNR. If peak width stays narrow but TOF σ grows, suspect reference clock/PLL noise.
Low-jitter sources like SiT9365 (or LTC6957) expose the true floor.
- What to measure first (2): correlation peak width; clock jitter/PLL phase-noise status (or a proxy counter).
- First fix (1): enforce a clean timing budget before increasing sampling rate or changing correlation windows.
5
Good accuracy at 25 °C, large error at 60 °C — compensation curve or sensor lag?
Large hot-side error is typically either wrong model segmentation (LUT/curve mismatch) or temperature input that lags the medium.
If the LUT segment index changes unexpectedly near a boundary, model/curve is the culprit. If the sensor reads “correct” but responds too slowly, the model is fed stale temperature.
A reference-grade sensor (TMP117) helps validate lag vs curve.
- What to measure first (2): temperature response (time constant) during a step; sound-velocity LUT index/segment ID.
- First fix (1): stabilize segment selection and only then refit compensation coefficients.
6
Flow spikes near a VFD motor — CM noise or AFE saturation?
Near VFDs, spikes are often front-end overload and recovery, not “mystery algorithm events.” Common-mode injection can push the Rx chain into saturation, then the recovery tail biases TOF/phase estimators.
Look for clipped waveforms, gain state jumps, or long settling time. A robust VGA path (e.g., AD8336) must still be proven stable under overload.
- What to measure first (2): Rx waveform clipping/settling time; correlation peak confidence (width/ratio) during the spike.
- First fix (1): prevent or shorten overload recovery before adjusting any extraction thresholds.
7
Clamp-on recalibration needed after pipe change — geometry or sound-velocity LUT?
A pipe change is a configuration change: acoustic modes and coupling efficiency shift, and any installation-specific parameter set can become invalid.
If confidence collapses (SNR/peak width changes) the issue is mainly acoustic/geometry. If confidence is stable but bias changes, suspect model/LUT or stale install parameters.
Store install IDs and versions in NVM (e.g., MB85RS64V FRAM).
- What to measure first (2): SNR vs frequency (or peak width); installation parameter version/ID and LUT segment.
- First fix (1): perform a guarded install-parameter update (versioned) before touching compensation coefficients.
8
TDC gives ps resolution but worse stability — reference clock noise?
Picosecond resolution does not guarantee stability; the stability floor is set by reference clock/PLL noise, power integrity, and temperature dependence.
A TDC such as TDC7200 or TDC-GPX2 can resolve fine steps, but the measured σ will still grow if the timebase jitters.
Validate with repeatability σ under fixed injected delay before trusting “ps” marketing numbers.
- What to measure first (2): repeatability σ at fixed stimulus; reference clock jitter / phase-noise proxy.
- First fix (1): improve timebase quality and calibration policy before chasing finer interpolation.
9
Long-term drift after months — transducer aging or AFE offset creep?
Month-scale drift splits cleanly into two patterns: baseline creep (electronic offset drift) or confidence erosion (acoustic aging/coupling degradation).
If zero-flow offset trend moves while confidence proxies stay stable, suspect AFE baseline and its stored calibration. If confidence steadily degrades (peak width rising, gain trending up), aging is likely.
Keep trend records versioned in NVM (e.g., MB85RS64V).
- What to measure first (2): zero-flow offset trend; confidence proxy trend (peak width/ratio or gain trend).
- First fix (1): gate any field recalibration behind trend evidence and parameter version control.
10
High flow accuracy OK, low flow poor — correlation window or zero-offset?
Low flow is where zero-offset and estimator confidence dominate. If the bias is similar in sign and magnitude across runs, zero-offset is the prime suspect. If results vary with SNR or peak width, the correlation window/threshold logic is limiting.
A stable front-end and capture chain (e.g., AD9234 class) helps, but low-flow quality is ultimately a “baseline + confidence” problem.
- What to measure first (2): current zero-offset value + trend; correlation peak width at low flow.
- First fix (1): correct baseline (versioned) before tightening windows or thresholds.
11
Temperature sensor reads correctly but flow still off — placement or model error?
“Reads correctly” can still be wrong for compensation if the sensor does not represent the medium temperature (thermal gradients, wall vs fluid).
If dynamic changes show error while steady-state looks fine, placement/thermal path is likely. If error persists even at steady-state with stable input, the model/LUT is wrong.
Use a low-drift sensor like TMP117 to validate response and offset.
- What to measure first (2): sensor response time constant during a thermal step; LUT segment/index stability.
- First fix (1): validate representativeness (response + offset) before refitting compensation curves.
12
Wet type stable, clamp-on unstable — SNR limit or mounting repeatability?
Wet meters often have a more consistent acoustic path; clamp-on adds coupling and path-mode variability that first appears as confidence loss.
If instability correlates with reduced SNR/peak confidence (peak width rising), it is SNR-limited. If SNR is acceptable but results vary run-to-run after rework, repeatability is the limiter and must be handled as a versioned installation parameter.
Log trend snapshots to NVM (MB85RS64V).
- What to measure first (2): SNR vs frequency (or peak width); run-to-run repeatability σ under fixed stimulus.
- First fix (1): treat mounting state as a controlled parameter set (versioned), then tune estimation thresholds.
