What “PdM Edge Node” means (device-side only)

Two field profiles (design priorities differ)

• Industrial / always-powered (24 V / PoE / wired): prioritize robustness against ground shift, supply noise, and field transients; keep strong evidence logs for troubleshooting.
• Remote / duty-cycled (battery + short captures): prioritize repeatable wake-up timing, capture window discipline (pre/post), and configuration traceability across firmware updates.

Rule of thumb: Always-powered nodes fail “quietly” (noise/ground coupling). Duty-cycled nodes fail “silently” (missed events due to settling/trigger/capture window).

Three questions every node must answer (turn into measurable outputs)

• What to measure? Define frequency band, amplitude range, and event shape (steady trend vs shock/impulse).
• How to trust sampling? Define sampling rate & anti-alias margin, phase alignment targets, and minimum self-check evidence (bias/clip/drop counters).
• How to make uplink usable? Separate trend vs event payloads; attach sequence, configuration hash, and error counters for auditability.

Module responsibilities (each with “field evidence”)

• Sensing & mounting: mechanical coupling changes show up as band-dependent amplitude shifts; track mounting state as metadata when possible.
• IEPE excitation: constant-current stability and compliance headroom; field evidence: bias voltage, clip/saturation counters, cable drop checks.
• AFE conditioning: coupling corner, gain/DR, anti-alias; field evidence: sweep response and alias signatures.
• Synchronous sampling: no drop/overrun and known timebase; field evidence: sample-seq continuity and timestamp monotonicity.
• Edge features: explainable metrics (RMS/peak/crest/kurtosis/envelope/FFT bands); field evidence: repeatability under controlled stimulation.
• Uplink + evidence: payload structured for diagnosis; field evidence: seq gaps, retry stats, and config version alignment.

Start with a measurable target (band + amplitude + event shape)

• Low-frequency trend (structural vibration, imbalance): longer windows, stable RMS/peak trends, tolerance to minor phase variation.
• High-frequency shock/impulse (bearing/gear early fault): short windows, strong need for headroom (no clipping) and disciplined pre/post capture.

Common pitfall: “Looks fine in steady state” while short impulses clip or alias, making envelope/crest-based indicators unreliable.

Translate sensor specs into node-side consequences (with evidence hooks)

• Sensitivity: sets ADC code utilization; too high clips on shocks, too low buries early faults in noise.
• Noise density: defines feature jitter floor; verify via quiet-run spectrum and metric repeatability.
• Max g / overload: protects event integrity; verify by clip counters and waveform flat-top signatures.
• Mounting coupling: mechanically filters high-frequency content; verify by controlled repeat tests after remounting.

When IEPE is the practical choice (clear boundary, device-side reasons)

• Long cables / harsh noise: constant-current excitation + defined bias point supports stable operation and easier field diagnostics.
• Need for evidence: bias voltage and compliance headroom are measurable; clipping risk can be monitored and logged.
• High-frequency reliability: common PdM ecosystem for industrial vibration; broad sensor options with known behavior under shock.

Scope note: This section stays at the node interface level (excitation + signal integrity). Network architecture and cloud analytics remain out of scope.

Design mapping: band → sampling/AA → DR/noise → features

• Sampling rate: choose with anti-alias margin (filter roll-off + transition band), not only Nyquist.
• Anti-alias: align cutoff to the target band; avoid “alias spikes” that mimic fault signatures.
• Dynamic range: reserve headroom for impulses; avoid “feature distortion” from rare events.
• Feature set: trend (RMS/peak/FFT bands) vs shock (crest/kurtosis/envelope + short FFT windows).

IEPE, simplified (what must be guaranteed)

• Excitation current affects noise robustness and cable drop (more current improves some noise margins but increases I × R drop).
• Compliance voltage must cover bias + signal swing + cable drop + protection drop + margin.

Compliance budget (turn “it works” into a measurable window)

• V_bias: sensor bias point after settling; measurable at the node input.
• V_signal(peak): peak swing during worst-case shock/impulse (the event that PdM cannot afford to lose).
• V_cable_drop: proportional to excitation current and cable/connector resistance; increases quickly with long runs or poor contacts.
• V_protect_drop: protection elements can add drop during large excursions; this directly steals headroom.
• Margin: required so the largest event remains linear (no flat-topping, no slow recovery).

Field symptoms → evidence (two common failure trees)

Engineering check trio (a minimal, repeatable workflow)

• 1) Measure static bias voltage: confirm it settles within the expected window and remains stable across temperature and cable changes.
• 2) Stress the max-event swing: apply or capture the largest expected shock; verify no clipping and fast recovery.
• 3) Estimate cable drop: compute/measure cable resistance and connectors; verify I × R drop does not consume margin at the worst event.

Practical rule: PdM credibility is decided by the rare event. A chain that is linear for steady vibration but clips on impulses will produce confident-looking, wrong features.

Canonical chain (device-side)

Key trade-offs (write as “decision + field evidence”)

• HPF corner: too high removes real low-frequency vibration (trend becomes artificially flat); too low allows drift/bias wander into metrics (trend drifts without mechanics).
• Anti-alias filtering: not “stronger is always better” — aggressive filters can add phase/group-delay behavior that harms impulse timing; insufficient filtering creates alias peaks that mimic faults.
• Input protection: protection parasitics can steal headroom and add distortion under large events; symptoms look like “hard bends” and slow recovery during shocks.

Stage-by-stage: what each block must prove

• AC coupling / HPF: proves low-frequency integrity; verify with low-band excitation and long-window stability checks.
• Gain (TIA/INA/PGA): proves noise floor vs headroom balance; verify that rare impulses never clip while quiet-run noise stays low.
• AA LPF: proves alias control; verify via sweep near Nyquist and by changing sample rate to see whether suspicious peaks “move” (alias signature).
• Protect: proves survivability without measurement damage; verify large-event behavior (no premature conduction causing flat-tops).
• Diff driver / Vcm: proves ADC compatibility; verify input common-mode and full-scale swing alignment under load.

Minimum validation set (fast acceptance tests)

• Sine (single tone): confirm linearity, no clipping, and reasonable harmonic content (gain + ADC range check).
• Sweep (band + near Nyquist): confirm amplitude/phase sanity and absence of alias peaks (AA + Fs check).
• Impact / impulse: confirm peak capture without flat-tops and fast recovery; confirm trigger and pre/post buffer produce stable features.

Acceptance goal: features remain stable across repeats, and any failure mode leaves a clear signature (clip counter, alias behavior, recovery time).

Sample rate is not only Nyquist

• Anti-alias margin: leave transition band between the highest useful band and Fs/2 so the AA filter can roll off without folding energy back as false peaks.
• Impulse time resolution: Δt = 1/Fs. If Δt is coarse, short shocks look “rounded,” peak/crest features drift, and triggers become inconsistent.
• Window discipline: feature windows are T = N/Fs. Trend windows and event windows must both remain meaningful and repeatable.

Dynamic range: noise floor vs maximum shock (avoid “steady looks fine, shock clips”)

• Noise-floor KPI: quiet-run spectrum floor and short-term feature variance (repeatability) indicate whether the chain is dominated by electrical noise.
• Headroom KPI: maximum event must stay below full-scale with margin; clipping and slow recovery contaminate the post-event window and break envelope/crest metrics.
• Bring-up check: step through gain/scale options and record: noise floor, clip counter, and recovery time under large impulses.

Why synchronous ADC matters: channel skew becomes phase error

• Multi-point comparability: two-end bearing measurements and multi-axis sensing rely on stable cross-channel phase relationships.
• Skew KPI: measure channel-to-channel time alignment and stability across temperature and run time (device-internal).
• Evidence: apply the same input to multiple channels and check correlation lag / phase difference repeatability.

Interface & sampling clock: practical conclusions + checks

• Clock jitter hurts high-frequency credibility: a “clean looking” low-frequency spectrum can hide high-frequency SNR loss if the sampling clock is noisy.
• Check 1 (HF sine): inject a near-top-band sine and observe SNR / noise floor; degradation that grows with frequency is a classic jitter signature.
• Check 2 (peak shape): as input frequency increases, watch for “fatter peaks” and rising skirts/noise floor around the tone.

Boundary reminder: this section covers device-internal ADC clock/reference integrity only (no network time algorithms).

What “timebase” means inside a PdM node

• XO/TCXO: sets the sampling cadence; temperature and aging shift frequency and phase behavior.
• PLL (optional): can multiply/clean clocks but adds its own lock/warm-up behavior that must be observable.
• RTC: anchors long-term scheduling and timestamps; must remain monotonic across sleep cycles and reboots.
• Warm-up stability: define a settle window after power-up before comparing “day-to-day” trend features.

“Same machine, different day” not comparable: split into three causes

Actionable rule: log installation events as metadata; otherwise electrical drift and mechanical coupling changes become indistinguishable.

Multi-point phase relationships drift: three device-side chains to check

• Sampling alignment chain: channel skew changes with temperature/supply → phase error grows at higher frequency.
• Trigger alignment chain: trigger latency/debounce varies → t0 alignment drifts inside pre/post buffers.
• Path/connection chain: cable/connector changes create delay steps and amplitude changes → correlation lag steps.

Evidence chain table (device-side only)

• Waveform phase / correlation: use correlation lag and phase difference repeatability to detect skew and trigger drift.
• Spectral peak stability: track peak position and peak width; frequency drift and “peak fattening” point to timebase issues.
• Timestamp continuity: enforce monotonic timestamps and sequence continuity; gaps are direct proof of capture/logging faults.
• Warm-up/temperature tags: correlate metrics with temperature and time-since-boot to separate warm-up effects from real machine changes.

Explicit boundary: do not attribute drift to network sync; this section is strictly internal (XO/PLL/RTC + timestamps).

Quality gate first: do not compute features on untrustworthy samples

• Signal integrity: clip/overrange, saturation count, recovery time indicators.
• Capture integrity: sample drops, FIFO overflow, timestamp/sequence gaps.
• Context integrity: gain/range state, temperature and supply summary for the record.

Rule: if quality is Bad, emit evidence (counters + snippet) but suppress “confident” conclusions.

Explainable time-domain features: what each one is good at

• RMS: stable energy trend for steady vibration; robust for long-run monitoring.
• Peak: highlights shocks and impacts; sensitive to clipping and bandwidth limits.
• Crest Factor (Peak/RMS): amplifies rare impulses riding on a steady baseline; collapses if peaks clip.
• Kurtosis: emphasizes sparse spikes; requires consistent windowing and installation metadata to stay comparable.

Common failure mode: clipping can make crest/kurtosis look “healthier” by flattening real spikes.

Frequency-domain features: node-side implementation mindset

• FFT band energy: map spectrum into a few fixed bands (low/mid/high or application bands) and trend band deltas.
• Peak tracking: track top peaks (frequency + amplitude) plus peak stability (drift/width) as credibility evidence.
• Envelope / sideband (minimal): band-limit → rectification or simple envelope → low-rate FFT/bands; keep compute bounded.

Credibility hint: suspicious peaks that shift when Fs changes are often alias artifacts (device-side check).

Event triggering: threshold + sliding window + debounce (with pre/post capture)

• Sliding window: short window for detection, longer window for confirmation to suppress noise spikes.
• Debounce: separate enter/exit thresholds or minimum duration to prevent “chatter.”
• Pre/Post capture: store a short snippet around t0 so the alert can be audited later.

Output recommendation: feature + confidence + minimal raw evidence

• Features: fixed fields (time + frequency) to keep results comparable across firmware versions.
• Confidence: derived from quality-gate status + repeatability checks (not an opaque probability).
• Evidence snippet: small raw segment (pre/post) for forensic review and model improvement later.
• Context: timestamp, config version, gain/range, temperature/supply summary, counter snapshot.

Goal: the same parameter set remains stable across temperature, supply variation, and installation changes (tracked as metadata).

Two-layer local storage: short raw snippets + long feature trends

• Ring buffer (snippets): store short raw waveform segments around events (pre/post) for forensic review.
• Trend store: store low-rate feature trends for long-term baselining and drift detection.
• Link key: join by event_id, timestamp and a sequence index so nothing is ambiguous.

Benefit: an alert can be re-checked even when backhaul is unreliable or the device reboots.

Minimum required event record fields (standardized, compact, reproducible)

• Time & ordering: monotonic timestamp + continuous sequence number (gap is direct evidence of capture/logging faults).
• Config identity: config version or hash, plus sampling/feature parameter set identifier.
• Measurement state: gain/range state, full-scale mapping, temperature and supply summary.
• Integrity snapshot: clip count, drop/overflow counters, timestamp gap count, sync/skew status (if available).

The three most useful log families in the field

• Sensor-chain health: bias/compliance indicators, saturation counters, sensor disconnect hints.
• Sampling-chain health: drop/overflow counters, clock anomaly hints, alignment/skew stability flags.
• Reporting-chain health: retry/latency summaries, sequence breaks, payload integrity checks (device-side).

Evidence Pack: make every alert “question-proof”

• Conclusion: severity + short feature-change summary (what changed, how much).
• Evidence: raw snippet reference + feature vector + confidence derived from quality gate.
• Context: config version, gain/range, temperature/supply summary, install metadata tag if available.
• Integrity: timestamp/sequence continuity + counter snapshot (clip/drop/overflow/gaps).

Non-negotiable: without evidence + integrity, an alert is not trustworthy, regardless of back-end analytics.

Two uplink modes: Trend vs Event (with built-in rate limits)

• Periodic trend uplink: low bandwidth and highly robust. Send compact feature summaries and deltas.
• Event uplink: alert + evidence. Send short pre/post snippets only under a strict quota.
• Degrade ladder: event+snippet → event-only → trend-only when quota/queue pressure rises.

Principle: when the field gets noisy (bursts of events), the node must stay alive and keep auditable records locally.

Payload schema: optimize for audit and reconciliation

• seq (continuity): exposes loss, duplication, or reordering; makes “what was missed” measurable.
• timestamp: aligns trend and event timelines; enables time-window audits.
• config_hash: binds the measurement to firmware and parameter set (repeatability).
• quality_flag + reason codes: indicates whether features should be trusted.
• counter snapshot: clip/drop/overflow/gap/retry counts to explain anomalies.

Reliability as diagnosable behavior (not a vague promise)

• Queue observability: queue depth, overflow count, last-success timestamp.
• Retry observability: retry count, backoff state, consecutive-fail counter.
• State code: emit a compact uplink state: Up / Degraded / Down with reason codes.

Outcome: a failed uplink is still informative if it produces consistent state + counters.

Ethernet-side evidence (strictly node-side)

• Link evidence: link up/down, link-flap count, PHY error counters (when available).
• Power coupling evidence: reboots/resets correlated with high-current uplink activity (e.g., PoE rail droop).
• Local timing: last-success timestamp + sequence continuity (no TSN/PTP algorithm discussion).

Three coupling paths that create “pseudo vibration”

• Current-source ripple: modulates IEPE bias/output, raising low-frequency floor and triggering false events.
• AFE reference / common-mode drift: appears as gain/bias drift, making day-to-day data incomparable.
• ADC Vref jitter / ground bounce: broadens peaks and lifts high-frequency noise floor, destabilizing band energy.

Key idea: “normal looking” time traces can still produce unstable crest/kurtosis/band energy because those metrics amplify subtle shifts.

Why features are more fragile than visual waveform checks

• Crest / kurtosis: highly sensitive to clipping, tiny spikes, and window inconsistency.
• Band energy / peak tracking: sensitive to noise-floor shifts and reference jitter (peaks widen or drift).
• Event state machine: sensitive to low-frequency drift; debounce can fail if the baseline moves.

Ground & shielding in long IEPE cabling: node-side symptoms only

• Ground loop / common-mode swing: injects mains components and slow baseline movement.
• Observable symptoms: low-frequency lift, 50/60 Hz + harmonics, unstable triggers, phase inconsistency across channels.
• Node-side evidence: rising quality flags, clip counts, baseline drift indicators, and repeatability failure across sessions.

Boundary: focus on symptoms and evidence at the node; avoid turning this into a general EMC guide.

Three must-measure points in the field (and what instability looks like)

• Probe 1 — current-source ripple: ripple correlates with low-frequency lift and false triggers.
• Probe 2 — AFE ref/common-mode: drift correlates with gain-equivalent drift and day-to-day mismatch.
• Probe 3 — ADC Vref/ground bounce: activity-dependent noise correlates with HF band jitter and peak broadening.

Minimal mitigations (node essentials only)

• Partition & timing: reduce coupling by separating analog/reference rails and avoiding heavy uplink bursts during capture windows.
• Reference stability: prioritize low-impedance reference return and stable common-mode biasing for the ADC driver.
• Verify by evidence: expect counter/quality improvements and reduced feature jitter under the same parameter set.