H2-1. What This Page Solves

This chapter sets a strict decision boundary for Level Radar (non-contact) versus Guided-Wave Radar (contact), and for Pulse versus FMCW versus TDR/FDR in level-measurement terms. The goal is to make selection evidence-driven—each choice is tied to a measurable field (SNR, confidence, drift, multi-echo behavior), not generic “better accuracy” claims.

Selection rules that can be proven (what to check first)

• Range and geometry: long range and open beam paths favor non-contact, while narrow tanks/nozzle constraints often favor guided-wave. Prove by whether a stable dominant peak exists across time (peak persistence) without heavy gain.
• Dielectric (εr) and weak echoes: low εr reduces surface reflection strength in non-contact systems. Prove by echo SNR staying low even when gain/AGC is high; confidence metric becomes intermittent.
• Foam, turbulence, vapor/condensation: foam and vapor can broaden or fragment surface echoes; condensation on radomes can detune and add near-field clutter. Prove by peak width growth, multi-peak count increase, and confidence fluctuation tied to environmental changes.
• Nozzle/stilling well/internals: hardware structures create fixed-range clutter and “phantom levels”. Prove by peaks that remain at constant range while the true surface changes.
• Maintenance tolerance: guided-wave probes may accumulate buildup, changing impedance and creating extra reflections; non-contact radomes may suffer condensation or coating. Prove by baseline drift + emergence of persistent secondary echoes.
• Required accuracy vs robustness: choose the waveform/engine that dominates the local error sources (timing jitter, phase noise, dispersion, slope nonlinearity). Prove by whether distance noise tracks clock jitter (pulse/TDR) or LO phase noise/slope errors (FMCW/FDR).

6-line “Choose this if…” (no deep math, all evidence-linked)

H2-2. System Architecture Map (Block-Level, Evidence-Driven)

This chapter defines a shared block diagram that the rest of the page will reference. The architecture is shown as two parallel chains—non-contact FMCW/pulse and guided-wave TDR/FDR—because they fail for different reasons. Every block includes an evidence tag: what to measure (waveform/spectrum), what to compute (metrics), and what to log (events) so troubleshooting does not devolve into guesswork.

Two system variants (why the block boundaries matter)

• Non-contact FMCW/Pulse chain: the hard part is preserving dynamic range in the presence of Tx-to-Rx leakage and installation clutter, then producing a stable confidence metric from the IF/baseband signal.
• Guided-wave TDR/FDR chain: the hard part is making reflections “parseable” (probe impedance uniformity, controlled excitation, stable time/frequency pickoff) while tracking long-term drift from buildup and temperature.

Evidence model: three layers that must agree

• Physical: leakage / IF spectrum / TDR trace
• Metric: SNR / ENOB / jitter / persistence / confidence
• System: reset reason / surge counter / cal version / temp bin

What to log/measure at key boundaries (minimum viable set)

• Tx boundary: Tx power or chirp amplitude; Tx-to-Rx leakage estimate; T/R isolation state.
• Rx analog boundary: noise floor, gain state, compression indicator; IF bandwidth and DC offset state.
• Clock boundary: clock jitter or PLL lock quality; timebase health; temperature at the timebase.
• Digitizer boundary: ADC ENOB/SNR at IF (FMCW) or TDC RMS/time-walk (pulse/TDR); clipping counters.
• Estimator boundary: echo SNR, peak width, peak persistence, confidence; multi-peak count and ranking.
• System boundary: internal temperature; calibration version/hash; brownout/reset reason; surge/ESD event counter.

H2-3. RF Front-End Design Priorities (LNA / Mixer / PA / Switching)

“Microwave front-end” in level sensing is defined by one constraint: the true echo is often small while the system is surrounded by near-field clutter, wide temperature swings, limited power budgets, and harsh EMC. This chapter maps front-end blocks to the four failure signatures that matter: phantom levels, blind zones, jumping readings, and temperature/EMC-triggered drift.

What the front-end must protect (non-negotiables)

• Dynamic range: prevent Tx leakage and structural clutter from consuming headroom before the weak echo is amplified.
• Baseline stability: keep near-zero IF/baseband energy stable; otherwise weak echoes disappear or become “fake peaks”.
• Repeatability over temperature: preserve gain, noise floor, and leakage behavior across hot/cold states.
• EMC immunity: avoid front-end modes that convert interference into DC/low-frequency baseline shifts.

Deep priorities (mapped to failure signatures)

1) Tx leakage becomes self-jamming. If Tx-to-Rx isolation is insufficient, the receiver sees a large “built-in interferer”. In FMCW, this often appears as a strong near-zero beat tone and an elevated baseline that reduces usable dynamic range. In pulse systems, leakage plus slow recovery increases the blind zone and masks close-range echoes.

• Symptom: strong fixed near-range peak, weak echoes disappear, or top/bottom level becomes unreadable.
• Prove with evidence: T/R isolation (dB), IF/baseband baseline level, recovery time trace.
• First fix direction: isolate/leakage path before adjusting thresholds or DSP.

2) Spurs can look like “phantom levels”. A spur is not random noise; it is a coherent tone that stays stable across frames. After FFT or correlation, it can form a peak that resembles a valid level—especially when installation clutter already creates fixed-range reflections. The key is distinguishing “moves with surface” versus “fixed in range and correlated with RF state”.

• Symptom: a stable peak persists at a constant distance even when the surface changes.
• Prove with evidence: spectrum snapshot (spur frequency), LO phase noise/spur plot, temperature/voltage correlation of peak amplitude.
• First fix direction: spur/root-cause in Synth/PLL/PA biasing before “peak selection rules”.

3) LNA compression hides weak echoes. A low noise figure improves sensitivity only if the receiver stays linear. With strong leakage or near-field clutter, the LNA/mixer can enter compression and “flatten” the information-bearing differences between echoes. This creates a deceptive output: it may look smooth, yet the correct peak is no longer separable.

• Symptom: weak-echo liquids become unreadable; reading becomes “confident but wrong”.
• Prove with evidence: gain/linearity sweep, compression indicator, noise floor versus gain state.
• First fix direction: restore headroom (attenuation, filtering, isolation) before increasing gain.

4) Mixer low-frequency noise and IF filtering decide close-in clutter behavior. Near-range clutter and leakage tend to concentrate energy close to DC/low IF. If the IF/baseband chain does not suppress these components appropriately, baseline becomes unstable and changes with temperature or interference, causing level jitter. IF filtering is not “for EMI” only—it shapes what the estimator will treat as plausible echoes.

• Symptom: reading jumps when nearby equipment switches; baseline shifts after EMC events.
• Prove with evidence: IF spectrum (baseline + skirts), flicker/1/f characteristics, event-correlated baseline shifts.
• First fix direction: stabilize IF/baseband baseline and reject close-in clutter before tuning detection logic.

Minimum evidence fields (store these before troubleshooting)

• Spectrum snapshot (spur + baseline)
• LO phase noise / spur state
• Rx gain & linearity indicators
• T/R isolation (dB)
• Leakage-induced baseline level
• Recovery time (pulse)

H2-4. Waveform Choice (Pulse vs FMCW vs TDR/FDR) — Practical Error Budget

Waveform choice should be driven by dominant, measurable error sources, not by generic claims about resolution. In level sensing, many deployments fail because the selected waveform cannot suppress the environment’s leading error source: timing jitter, phase noise, slope nonlinearity, dispersion, or multipath / multi-reflection.

Pulse radar (non-contact): resolution is not the main limiter

• Dominant error sources: Tx/Rx recovery (blind zone), timing jitter, near-field structural ringing.
• Practical symptom: close-range level is missing or unstable; distance noise tracks clock state.
• Quick proof fields: recovery-time capture, jitter-to-distance noise trend, near-range clutter persistence.

FMCW (non-contact): precision is set by phase noise + slope + IF chain

• Dominant error sources: chirp slope linearity, LO phase noise, IF/baseband baseline from leakage.
• Practical symptom: strong near-zero tone/baseline; weak echoes vanish; stable phantom peaks appear.
• Quick proof fields: IF spectrum baseline and skirts, phase-noise/spur state, confidence stability across frames.

TDR (guided-wave): time pickoff quality and probe physics dominate

• Dominant error sources: edge rise time, dispersion along probe, time-walk, multi-reflection ambiguity.
• Practical symptom: two or more persistent peaks; wrong peak gets selected when amplitude changes.
• Quick proof fields: TDR trace (multi-echo map), time-walk vs amplitude, baseline drift over time.

FDR / swept-frequency (guided-wave): separates close reflections if sweep is controlled

• Dominant error sources: sweep linearity, frequency-to-distance mapping stability, temperature drift of sweep/reference.
• Practical symptom: distances are repeatable in lab but shift with temperature or supply conditions.
• Quick proof fields: sweep residual/linearity check, reflection spectrum repeatability, calibration repeatability.

Dominant error source table (compact, evidence-linked)

H2-5. ADC vs TDC: Choosing the Measurement Engine

The measurement engine is the core fork: TDC (time pickoff) is strongest when the problem is “when did the edge/echo arrive?”, while high-speed ADC is strongest when the problem is “how to separate weak echoes from leakage/clutter in the sampled waveform/IF spectrum”. Picking the wrong engine forces unstable thresholds, fragile peak rules, and excessive gain—leading to phantom levels and drift.

TDC path deep dive: pickoff quality beats “spec resolution”

In pulse radar and guided-wave TDR, the chain is: edge formation → time pickoff → TDC quantization → distance estimate. A high-resolution TDC does not guarantee stable distance if the pickoff time shifts with amplitude or edge shape.

• Threshold pickoff: simple and fast, but strongly amplitude-dependent; time-walk can become a “fake level drift” when echo amplitude changes with temperature, foam, or coating.
• CFD (constant fraction): reduces amplitude dependence, but becomes sensitive to waveform shape and noise; requires stable edge shaping.
• Correlation / matched pickoff: improves stability when waveform is consistent, but depends on templates and can be challenged by multi-reflection structures.
• Edge shaping requirement: rise time, dispersion, and ringing determine whether the pickoff is single-valued; poor edge quality produces multiple plausible pickoff points.

ADC path deep dive: ENOB × jitter defines separability

In FMCW and sampled-IF systems, the chain is: IF/baseband formation → sampling → digital downconversion/FFT → peak extraction. Sampling rate alone is not the limiter; the key is usable dynamic range at the IF of interest and how clock jitter turns into noise in the frequency domain.

• Sampling rate vs IF: choose sampling strategy that preserves the beat/IF band with controlled aliasing and stable filtering; “faster” is not automatically “cleaner”.
• ENOB at IF: many converters lose ENOB as frequency rises; verify ENOB/SNR at the actual IF band, not at low-frequency datasheet conditions.
• Clock jitter penalty: jitter smears phase in sampled sinusoids and lifts the noise skirt; in FMCW it reduces peak prominence and can mimic “weak echo” conditions.
• Leakage baseline headroom: if leakage consumes ADC headroom, quantization and clipping dominate; peaks become unstable even if DSP is correct.

Hybrid engines (when neither pure approach is enough)

• Coarse ADC + fine TDC: ADC identifies the correct echo region in a multi-echo scene; TDC refines timing within that region to reduce jitter sensitivity.
• Oversampling + matched filter: improves pickoff stability by raising effective SNR, then uses correlation to reduce time-walk and ringing sensitivity.
• Digitized pulse with dual metrics: keep both pickoff time and peak-integral energy; degrade output if energy drops but pickoff remains unstable.

Minimum evidence fields (loggable and decision-ready)

• Clock jitter spec + lock state
• ADC ENOB / SNR @ IF band
• TDC RMS (time bins) vs temp
• Time-walk vs amplitude (slope)
• Usable dynamic range (dB)
• Leakage baseline headroom

H2-6. Signal Processing That Actually Moves the Needle (Not Generic DSP)

The DSP that matters in level sensing is the DSP that removes leakage baseline, rejects false echoes, survives tank internals, and produces a diagnostic confidence output. This chapter lists only the pipeline steps that directly reduce phantom levels, jumping readings, and long-term drift.

Non-contact radar DSP (FMCW / pulse): four steps that fix real pain

• 1) Leakage cancellation / baseline subtraction: Remove the strong near-zero component created by Tx-to-Rx coupling and close-in clutter so weak echoes recover headroom.
  Prove: baseline_dB drops and peak SNR increases without raising gain.
• 2) Windowing + FFT (beat extraction): Use stable windowing to control spectral leakage; reduces wide peaks and false side lobes that imitate a second level.
  Prove: sidelobe_ratio improves and peak_width narrows.
• 3) Peak picking with confidence metrics: Choose peaks by a scored model: SNR + peak width + multi-frame persistence + rank margin; this prevents “peak hopping”.
  Prove: persistence_N rises and rank_margin stays positive across disturbances.
• 4) Multipath / false echo suppression: Suppress fixed-range clutter from nozzles, internals, and ringing by learning a clutter mask and detecting “non-moving peaks”.
  Prove: fixed_peak_flags increase during clutter events while level output is gated/held.

Guided-wave DSP (TDR / FDR): four steps that make probes reliable

• 1) Reflection identification along the probe: Separate liquid surface reflection from known probe features (coupler, end reflection, connectors) using a calibrated feature map.
  Prove: probe_feature_align_err remains bounded.
• 2) Clutter learning for buildup/foam: Track slow baseline drift and the emergence of secondary peaks; adjust parsing rules without rewriting thresholds.
  Prove: baseline_drift_rate captures slow changes and secondary_peak_count trends.
• 3) Multi-echo parsing: Rank multiple discontinuities and keep a stable hypothesis; detect conflicts when two candidates are too close in score.
  Prove: echo_count and rank_margin predict when the output should degrade.
• 4) Stability gating (degrade/hold): Avoid control-system harm by freezing or degrading output when confidence collapses due to ambiguous echoes.
  Prove: quality_state transitions correlate with low SNR or conflict codes.

Confidence output definition (required outputs besides “level”)

A level sensor should output not only a numeric level, but also diagnostic fields that explain whether the reading is trustworthy and why. These fields enable reproducible validation, remote troubleshooting, and safe control-system behavior.

H2-7. Temperature Compensation & Calibration Strategy (Factory + Field)

Temperature drift in level sensing is rarely a single linear error. It typically mixes scale drift (slope/VF changes), offset drift (internal delays), and detectability drift (gain/noise-floor/baseline headroom). A usable strategy must be testable: each drift source maps to a calibration object, a verification plot, and loggable versioned coefficients.

Temperature impact map (what moves, and how it shows up)

Factory calibration (freeze what is stable and separable)

Factory calibration should capture repeatable device-specific behavior and store it as versioned coefficients. The goal is not “perfect absolute accuracy”; the goal is to separate scale vs offset vs detectability so field actions remain narrow and safe.

• FMCW sweep linearity correction: correct chirp slope nonlinearity using a residual map; validate with a sweep residual plot across temperature points.
• Gain/noise-floor table: store gain trims or a compensation table so echo_snr_dB stays comparable across temperature.
• Guided-wave VF calibration: calibrate probe velocity factor (VF) and a probe feature map (known reflections) to separate surface reflection from structural reflections.
• Internal delay calibration: calibrate fixed timing offsets (switching/coupler/internal routing) so the system has a stable zero reference.
• Coefficient versioning: store coeff_version + CRC; any runtime change must be traceable and rejectable.

Field calibration & self-test (safe, bounded, reversible)

Field calibration should avoid “deep rewriting” of factory coefficients. The safe approach is baseline learning + reference checks + self-test pulses, with sanity gating when confidence collapses.

• Reference targets / known conditions: use an “empty tank baseline” (non-contact) or known probe features (guided-wave) to verify offset alignment.
• Empty-tank baseline learning: learn fixed clutter signatures (nozzle/stilling well) and generate a clutter mask that is valid for that installation.
• Self-test pulse / loopback: inject a known test reflection/delay to track internal delay drift over temperature without needing a process stop.
• Sanity checks: if level changes exceed physical limits or conflict with confidence metrics, hold or degrade output and log reason codes.

Minimum evidence fields (must be logged to make drift testable)

• Temp sensor placement list
• coeff_version + CRC
• Drift-over-temp plots IDs
• cal_age + last_cal timestamp
• self_test_delay vs temp
• echo_snr_dB vs temp

H2-8. Antenna / Radome / Probe: The “Analog Physics” That Creates Digital Headaches

False readings are often created before any DSP runs. Mechanical and electromagnetic boundary conditions reshape echoes into fixed peaks, multipath clusters, baseline drift, and peak widening. This chapter links installation choices to measurable echo artifacts and a checklist that can be audited on site.

Non-contact (antenna + radome + nozzle): how installations manufacture false echoes

• Beamwidth vs nozzle/stilling-well constraints: nozzle “clips” the near-field, creating ringing and fixed-range reflections.
  Artifact: fixed peaks near the top; broadened peaks; blind-zone expansion.
• Radome dielectric loading + condensation: a wet radome changes effective permittivity and detunes the antenna match.
  Artifact: slow SNR decay; baseline_dB drift; range bias that correlates with humidity/temperature.
• Near-field reflectors (agitator, ladder, internals): coherent multipath produces stable peaks that do not move with level.
  Artifact: multi-echo scenes; rank_margin collapses; peak hopping under disturbance.
• Mounting angle / offset / ground reference: shifts main-lobe/side-lobe energy into different clutter features.
  Artifact: site-to-site inconsistency; persistent phantom peaks at different fixed ranges.

Guided-wave (probe): impedance reality drives reflection ambiguity

• Probe type and impedance uniformity (rod/cable/coax): non-uniform impedance creates multiple discontinuities along the probe.
  Artifact: multiple small “steps” on the trace; echo_count increases; surface peak becomes less dominant.
• Coating/buildup changes impedance: systematic impedance shifts create new reflections and change surface reflection amplitude.
  Artifact: baseline_drift_rate rises; secondary_peak_count trends upward; time-walk sensitivity increases.
• Mounting, sealing, and ground reference: unstable return paths alter coupling and feature map alignment.
  Artifact: probe_feature_align_err grows; re-installation changes readings.
• Mechanical motion/vibration: small geometry changes create time-varying reflections.
  Artifact: persistence_N drops; quality_state frequently degrades.

Installation risk checklist (tie each item to echo artifacts)

H2-9. EMC, Surge, Isolation—and Why Level Sensors Reboot at the Worst Time

EMC and surge events are not only “compliance costs”—they directly break measurement integrity. In level sensing, disturbances often enter the chain as noise-floor jumps, phantom peaks, baseline shifts, or brownout resets. The key is to link each symptom to a coupling path, then to a verifiable evidence field.

Failure chain A: EMC coupling → noise floor jump → phantom level

Conducted and radiated emissions can couple into the Rx chain, IF/baseband, ADC reference, or the clock/PLL. In FMCW this often appears as a lifted noise skirt and spurious beat components; in pulse/TDR it appears as unstable pickoff timing and widened peaks.

• Where it couples: LNA bias nodes, mixer/IF filters, ADC reference pins, clock traces, and digital switching edges near analog ground returns.
• What it looks like: noise_floor_dB step-up, peak_width expansion, phantom_peak_count increase, rank_margin collapse.
• Why it causes wrong level: peak ranking flips; the algorithm “locks” onto a stable but wrong fixed peak during bursts of interference.

Failure chain B: ESD/Surge → latch-up / baseline shift → corrupted calibration writes

Surge/ESD can create transient latch-up and bias shifts, but the most damaging case is when the event overlaps with a flash write. A single corrupted coefficient can turn a stable sensor into a drifting sensor that “looks calibrated” until temperature or clutter changes.

• Device-level risks: latch-up, IF offset shifts, ADC clipping bursts, memory bit flips during writes.
• Field symptom: reboot happens once, then long-term accuracy degrades or baseline_dB remains elevated.
• Proof requirement: coeff_crc mismatch, flash_write_fail counter, calibration_version jumps unexpectedly.

Failure chain C: long cables + ground potential differences → comm glitches → watchdog resets

Tanks and remote PLC cabinets can sit at different potentials. Without isolation, surge currents and common-mode noise flow through signal grounds, causing UART/RS-485 framing errors, bus contention, and MCU exceptions that trigger watchdog resets—often exactly during pump starts or lightning transients.

• Where isolation helps: isolate UART/RS-485 interfaces and provide isolated power to the sensor head when long cables or noisy grounds are unavoidable.
• What to watch: crc_error_count, seq_gap_count, framing_error_count, watchdog_reset_count.

Evidence fields (minimum logs to make EMC failures diagnosable)

• reset_reason (BOR/WDG/UVLO)
• rail_min_mV + droop_event_count
• emc_test_mode_flag + test step ID
• surge_event_counter
• coeff_version + coeff_crc
• noise_floor_dB / baseline_dB
• phantom_peak_count / rank_margin

Practical countermeasures (tie to the chains above)

H2-10. Isolated Links & System Integration (RS-485 / 4–20mA / Ethernet / IO-Link)

This is an integration chapter, not a protocol tutorial. The practical questions are: where isolation belongs, how to keep time and data integrity across that barrier, and how to make updates and calibration writes brownout-safe. The deliverable is a minimum viable telemetry contract that lets systems diagnose measurement vs link vs power failures remotely.

Where isolation belongs (sensor head vs comms board)

Data integrity across isolation (timestamp + CRC + sequence)

Across an isolation barrier, the common failures are not only bit flips. Packet loss, reordering, and replay-like behavior can make a control system believe the level jumped. A minimum integrity set should include: CRC, a sequence counter, and a monotonic timestamp (or sample counter).

• CRC: detect corruption; increments crc_error_count for diagnostics.
• Sequence counter: detect drops/reorder; track seq_gap_count and last_seq.
• Monotonic timestamp: prevent time going “backwards”; enable stable rate checks and jitter analysis.
• Degrade/hold rules: if seq gaps exceed a threshold or timestamp is non-monotonic, set quality_state=DEGRADED/HOLD and log reason codes.

Update & diagnostics safety (firmware + calibration data + brownout-safe commits)

Field updates and calibration writes must survive real-world power transients. A robust approach is: separate data regions, validate rails before commits, and only “activate” new versions after verified writes and CRC checks.

• Firmware safety: A/B slots or staged activation; never overwrite the only bootable image during field update.
• Calibration protection: split into factory read-only coefficients and field-learned coefficients; each with version + CRC.
• Brownout-safe writes: dual-copy commit (write new → verify CRC → bump version pointer) gated by rail_ok windows and droop detection.
• Diagnostics channel: keep a minimal readout path for reset_reason, coeff_crc, and rail_min_mV even after partial failures.

Deliverable: minimum viable telemetry schema (field names + meaning)

{
  "level_value": 3.42,
  "quality_state": "GOOD",
  "confidence_score": 0.92,
  "echo_snr_dB": 18.4,
  "baseline_dB": -42.0,
  "peak_index": 137,
  "multi_echo_count": 2,
  "internal_temp_C": 46.1,
  "coeff_version": 12,
  "coeff_crc": "0x8A31",
  "reset_reason": "NONE",
  "rail_min_mV": 4740,
  "surge_event_counter": 3,
  "crc_error_count": 0,
  "seq_gap_count": 0
}
      

H2-11. Validation & Troubleshooting Playbook (Evidence-First, Not Opinion-First)

This playbook forces an evidence-first workflow. Every symptom maps to exactly two captures (power + measurement) and two logs (reliability + measurement health). Root causes are ranked by probability and proven by those fields before any major tuning is attempted.

Standard capture kit (use the same 2 waveforms + 2 logs for all symptoms)

10–12 common symptoms (each with capture → causes → first fix)

Four-step triage (minimum decision order)

• 1) Power/reset first: confirm reset_reason and rail_min_mV before tuning measurement.
• 2) Measurement health: check echo_snr_dB, baseline_dB, multi_echo_count, confidence_score for ambiguity.
• 3) Installation physics: identify fixed peaks and near-field artifacts (nozzle/radome/probe).
• 4) Only then tune: gating/windows, peak rules, baseline learning limits, and temperature compensation tables.

H2-12. FAQs (Accordion) — Evidence-First Answers

Each answer stays short and always starts with what to measure first. Use the chapter links to jump to the deeper evidence chain and fix paths.