Acoustic / Vibration Edge Node: Mic/TIA + ΣΔ ADC Triggers
← Back to: IoT & Edge Computing
Center Idea
An acoustic/vibration edge node is a low-power sensing system that keeps listening locally, extracts lightweight features, and triggers event capture and reporting only when defined signal conditions are met. It is engineered around a low-noise analog front end, ΣΔ ADC + decimation latency control, robust trigger rules, and evidence buffers that keep field detection reliable without burning always-on power.
H2-1. Definition & Boundary: What is an Acoustic/Vibration Edge Node?
An acoustic/vibration edge node is an ultra-low-power sensing device that continuously monitors analog motion or sound signals, runs local detection, and only captures/logs data when an event trigger fires. The goal is to preserve energy and storage while still retaining high diagnostic value around real-world events.
One-sentence definition (SEO-friendly)
A low-power acoustic/vibration edge node continuously senses signals, extracts lightweight DSP features, detects events locally, and records evidence only when triggers occur.
Sense (always-on): keep a minimal power domain sampling and producing stable features.
Decide (local DSP): convert features into a trigger decision using thresholds + state logic, not raw intuition.
Act (event evidence): capture a short pre/post window, attach metadata, and then log/notify.
Acoustic vs Vibration (same architecture, different front-end constraints)
Both share the same “low-noise chain + trigger logic,” but differ at the input model. Acoustic inputs often behave like voltage sources (analog MEMS mic) or arrive digitally (PDM). Vibration inputs are frequently charge/current sources (piezo), pushing the front-end toward TIA behavior, bias/leakage control, and fast recovery from impulsive shocks.
Scope Guard What this page owns vs does not cover
Owns
- Signal chain: sensor interface → low-noise AFE (LNA/TIA) → ΣΔ ADC → decimation → edge DSP features → trigger.
- Event-driven evidence: ring buffers, pre/post capture windows, minimal metadata for diagnostics.
- Always-on strategy: power domains, duty-cycling choices that avoid missed events.
Does NOT
- Wireless protocol stacks, gateway/cloud pipelines, or backhaul architectures.
- Industrial PdM specifics (IEPE constant-current, multi-channel synchronous DAQ, plant deployment).
- Vision pipelines (image sensors/ISP/NPU) or full ML model training.
Figure F1 — System boundary: always-on sensing → local trigger → event capture
H2-2. Use Cases & Event Taxonomy: What events are we trying to catch?
Trigger design is only “easy” after events are translated into engineering constraints. Different events demand different time windows, frequency focus, dynamic range, and false-alarm tolerance. This section defines an event taxonomy and maps each class to required signal traits and chain implications.
Why classify events?
Without a taxonomy, sampling rates, filter splits (analog vs digital), trigger thresholds, and buffer lengths become guesswork. With a taxonomy, each design decision is traceable: event class → signal traits → features → trigger logic → capture window.
Event classes (expressed as signal behavior, not “application stories”)
- Impulse / knock: short, high-crest-factor, often wideband; stresses front-end recovery and latency.
- Sustained loudness: energy stays high for longer; demands stable RMS/envelope metrics and robust debouncing.
- Band-limited rise: energy rises in a specific band; best detected with bandpower features.
- Resonance drift: peak frequency or amplitude shifts; requires peak tracking rather than raw thresholds.
- Timbre anomaly (lightweight): spectrum shape changes; can be approximated with simple spectral shape metrics.
Four constraint dimensions that directly shape the chain
Time scale (ms → minutes)
Frequency focus (band)
Dynamic range (noise floor → peak)
False / miss tolerance
These dimensions decide windowing and hop sizes, analog HPF/LPF placement, ΣΔ decimation latency budgets, and trigger protections (hysteresis, debounce, cooldown).
| Event class | Signal traits | Trigger features | Chain implications (what must be designed) |
|---|---|---|---|
| Impulse / knock | Short duration, high peak, wideband energy, high crest factor; can saturate AFE; fast recovery matters. | Peak, peak-count, short-time energy; optionally bandpower for “impact band.” | AFE: headroom + fast clipping recovery; DSP: short windows, low latency; Trigger: hysteresis + hold-off; Capture: strong pre-trigger to preserve onset. |
| Sustained loudness | Energy stays elevated; sensitive to drift and background variation (wind/handling/structure). | RMS, envelope, duration-above-threshold; rate-of-change gating. | Filters: stable HPF/LPF split; DSP: longer windows, smoothing; Trigger: debounce + cooldown; Capture: shorter pre-trigger, longer post-trigger. |
| Band-limited rise | Energy grows in a band while total RMS may look normal; noise outside band should be rejected. | Bandpower, band ratios, narrowband envelope. | DSP: band filters after decimation; ADC: ensure band is within passband; Trigger: multi-condition gating (bandpower + duration). |
| Resonance drift | Spectral peak shifts or changes amplitude; can be gradual; may require baseline tracking. | Peak frequency tracking, peak amplitude trend, spectral centroid. | DSP: peak detection per frame; Trigger: compare to baseline with hysteresis; Capture: store features + short raw snippets to diagnose shifts. |
| Timbre anomaly (lightweight) | Spectrum shape changes (not just level); robust to overall gain variation is useful. | Spectral flatness/centroid proxies, band ratios, feature vectors (small). | DSP: stable normalization; Trigger: threshold on feature distance; Capture: save features + metadata; keep compute bounded for always-on. |
Practical rule: the trigger point is not the event start. Decimation + windowing introduces detection latency.
A pre-trigger ring buffer is what preserves the onset of fast events.
Figure F2 — Event timeline, detection latency, and pre/post trigger capture windows
H2-3. Sensor & Interface Choices (Practical, Not Catalog)
Input selection becomes reliable when based on source model rather than part numbers. The front-end topology is determined by whether the sensor behaves like a voltage source, a charge/current source, or a digital stream.
Three input models that decide the entire front end
- Voltage-like (analog): analog MEMS microphone, analog MEMS accelerometer outputs. Typical path: LNA + analog conditioning.
- Charge/current-like (analog): piezo vibration pickups (often high impedance). Typical path: TIA / charge amplifier to make gain and bandwidth controllable.
- Digital stream: PDM/I²S microphones, digital accelerometers. Typical path: digital domain (filter/decimation/features), minimal analog AFE.
Microphones: analog vs digital (interface only, not bus deep-dive)
- Analog mic / analog MEMS: behaves like a voltage source; front-end priorities are input-referred noise, RFI/EMI tolerance, and stable biasing.
- PDM/I²S mic: arrives as a bitstream/frames; priorities shift to always-on clocking cost, data movement cost, and stable digital filtering/feature extraction.
Vibration sensing: piezo vs MEMS accelerometer
- Piezo (charge/current source): input capacitance and leakage paths can dominate behavior; TIA/charge amplification keeps gain and corner frequencies predictable.
- MEMS accelerometer: can be analog or digital; analog output is often voltage-like but installation and structure coupling can shift the effective spectrum.
Installation and coupling influence the system transfer function (resonances, bandwidth, structure-borne noise). For edge triggers, treat mounting as a signal-path element:
changes here can move energy between bands and reshape trigger features.
If input is charge/current-like → choose TIA path; manage Cin/leakage/bias and overload recovery.
If input is voltage-like analog → choose LNA path; prioritize input-referred noise, RFI robustness, and stable biasing.
If input is digital stream → go direct to DSP; control always-on clock/data cost and ensure feature stability.
Figure F3 — Input models and the matching front-end path (LNA vs TIA vs digital)
H2-4. Low-Noise Analog Front-End: LNA vs TIA, Biasing, and Filtering
In always-on trigger systems, the analog front-end determines the noise floor and the post-overload recovery time. Trigger logic cannot compensate for drift, leakage, or slow recovery after impulsive events. The objective is a front-end that is predictable in gain and bandwidth, robust to overload, and diagnosable with clear test points.
Why TIA exists (beyond “amplification”)
A TIA/charge amplifier turns a charge/current-like sensor into a controllable voltage output. This makes gain and corner frequencies designable and testable, rather than being dominated by sensor capacitance, wiring capacitance, and leakage paths.
AFE has three jobs (each must be verifiable)
Gain
Bandwidth
Overload recovery
- Gain: map typical events into ADC range without frequent saturation; leave margin for rare spikes.
- Bandwidth: pass only what carries event information; block out-of-band noise that inflates features and false alarms.
- Overload recovery: return to a valid detection state quickly after shocks; slow recovery causes missed onsets and repeated false triggers.
Biasing & leakage: why “small currents” become real signals
- Input capacitance (Cin): sensor + cable + ESD + routing; shifts poles and stability needs in TIA paths.
- Leakage and bias currents: ESD devices, protection networks, contamination; create slow drifts that look like low-frequency motion.
- Reference integrity: bias/reference noise leaks into the passband and becomes trigger energy in envelope/bandpower features.
Analog filtering split vs ΣΔ digital filtering (role separation)
- Analog HPF: remove DC/very-low-frequency components that waste headroom and prolong recovery.
- Analog LPF: block high-frequency interference and reduce out-of-band energy entering the ADC front-end.
- ΣΔ decimation filtering: refine band shaping and feature-friendly passbands after conversion.
Design checklist (diagnosable, field-friendly)
- Noise contributors: identify dominant terms (source noise, amplifier noise, R/C thermal noise) in the event band.
- Leakage paths: enumerate ESD/protection/bias routes; verify drift with long captures and temperature changes.
- Overload behavior: test clipping and recovery; confirm features do not “ring” into repeated triggers after impact.
- Test points: provide measurable nodes (AFE output, ADC input, injected stimulus) to separate analog vs digital causes.
Figure F4 — Noise & drift contributors: LNA vs TIA and their impact on triggers
H2-5. ΣΔ ADC & Decimation: Why it’s common in audio/vibration nodes
Sigma-delta conversion is popular in acoustic/vibration edge nodes because it achieves low in-band noise and strong anti-alias behavior with modest analog complexity, then shapes bandwidth and noise in the digital domain via decimation filtering. The trade-off that matters most for triggers is group delay, which can shift detection timing and pre-buffer needs.
Three building blocks (system meaning, not textbook)
- Oversampling (OSR): spreads quantization noise over a wider band; more freedom to reduce noise in the target band after filtering.
- Noise shaping: pushes a large fraction of quantization noise toward higher frequencies, lowering the noise seen inside the event band.
- Decimation filter: sets the final bandwidth and sample rate, provides steep digital anti-aliasing, and defines group delay.
What parameters mean for an event trigger
OSR
In-band noise
SNR / DR
Group delay
- OSR ↑ can reduce in-band noise, but often increases compute, power, or delay depending on the decimation design.
- In-band noise ↓ stabilizes threshold-based features (envelope, bandpower), reducing false alarms for weak events.
- Group delay ↑ shifts the trigger later; without sufficient pre-buffer, the onset can be missed even when detection is correct.
| Parameter | Typical change | System symptom | Compensation action |
|---|---|---|---|
| OSR | Higher oversampling | Cleaner in-band floor; sometimes higher latency | Validate group delay; size ring buffer for onset capture |
| Decimation bandwidth | Narrower passband | Less integrated noise; may suppress some event energy | Align passband with event taxonomy; avoid “too narrow” filtering |
| In-band noise | Lower noise after decim | Lower false-trigger rate for weak signals | Revisit AFE gain so the ADC is not underutilized |
| Group delay | Longer impulse response | Trigger reacts late; onset features clipped | Increase pre-buffer; align windowing to delayed features |
Figure F5 — ΣΔ + decimation pipeline and where group delay shifts trigger timing
H2-6. Noise & Dynamic Range Budget (From “What you need” back to design)
A trigger can only be as reliable as the separation between noise floor and the minimum event level, while still keeping enough headroom to avoid clipping during rare impacts. This section provides a practical budgeting workflow that starts from event needs and ends at front-end gain, ADC range, and margin checks.
Four “levels” to define before choosing gain or range
- Noise floor: the baseline energy seen by features in the chosen bandwidth.
- Minimum target event: the smallest event that must reliably trigger.
- Full-scale / headroom: the maximum expected amplitude without saturation and long recovery.
- Margin: allowance for installation variance, temperature drift, and environment changes.
Noise budget “buckets” (design handles)
- Sensor noise: sets a lower bound; cannot be corrected by adding gain.
- AFE noise: LNA/TIA input-referred noise and bias/leakage effects; drives false alarms if dominant.
- Conversion noise: quantization and in-band noise after ΣΔ + decimation; improves with OSR and bandwidth choices.
- Bandwidth effect: wider bandwidth integrates more noise; narrow enough to exclude irrelevant bands but not so narrow it cuts event energy.
Common field mistakes (symptom → cause)
- “More gain is always better”: clipping becomes frequent, recovery slows, and impacts turn into repeated triggers.
- “Threshold alone defines reliability”: thresholds are relative; if the noise floor shifts, the same threshold yields different false/miss rates.
- “Average noise looks fine”: impulsive overload or interference can create non-linear artifacts that inflate features after the event.
1) Define the event feature (peak, RMS, bandpower) and the minimum trigger level in the target band.
2) Choose bandwidth to match event taxonomy; narrower bandwidth lowers integrated noise but risks cutting event energy.
3) Set a noise-floor target so the minimum event sits comfortably above baseline with margin.
4) Allocate budget across sensor, AFE, and ADC in-band noise; the dominant bucket defines what must be improved.
5) Pick gain and ADC range to avoid frequent clipping while keeping adequate resolution for weak events.
6) Validate in practice: baseline noise capture, overload/recovery test, and trigger timing vs pre-buffer.
Budgeting is iterative: if the AFE dominates noise, return to front-end choices (gain, bias, leakage paths). If conversion noise dominates, revisit OSR, decimation bandwidth, and in-band noise targets.
Figure F6 — Dynamic range picture: noise floor, target event, full-scale, and margin
H2-7. Edge DSP Pipeline: Feature Extraction for Triggers (Not Full ML)
Trigger-oriented DSP is built to answer a narrow question: did an event happen, and does it match a simple signal trait (impulse, sustained energy, band-limited rise, resonance shift). The goal is low cost, low latency, and stable fixed-point behavior—not classification at scale.
A practical trigger DSP pipeline (minimal but complete)
- Pre-conditioning: DC removal / simple smoothing / overload flags to keep features stable.
- Band shaping: one band or a small set of bands that match the event taxonomy.
- Framing: window + hop define time resolution, latency, and compute.
- Feature set: low-cost metrics (RMS, envelope, bandpower, peak/resonance) designed for gating.
- Temporal consistency: smoothing + duration rules to suppress short spikes.
- Outputs: features and flags feed a trigger rule system (hysteresis, debounce, cooldown).
Framing choices that directly affect trigger quality
- Window length: longer windows average noise better but add latency and can blur impulse onsets.
- Hop size (overlap): smaller hops react faster and reduce misses, but increase compute and power.
- Feature timing: window + group delay define when a feature becomes “available” for a decision.
Fixed-point stability (common failure modes)
- Overflow hot spots: squared sums (RMS/bandpower) and accumulators; use scaling and saturating arithmetic.
- Noise-gate region: near-threshold jitter can flip features; clamp or freeze updates below a baseline gate.
- Consistent Q-format: avoid frequent re-scaling across blocks; define a small set of safe ranges.
| Feature | Compute cost | Added latency | Best for events | Pitfalls / notes |
|---|---|---|---|---|
| Peak | Low | Low | Impacts / knocks | Can be polluted by clipping recovery; pair with cooldown + overload flag. |
| RMS | Low | Window-limited | Sustained noise rise | Depends on bandwidth; baseline drift needs adaptive floor or hysteresis. |
| Envelope | Low | Low–Medium | Friction / continuous energy | Near-threshold jitter; use hysteresis + debounce (N frames). |
| Bandpower | Low–Medium | Window-limited | Band-limited anomalies | Wrong band causes misses; multi-band gating reduces sensitivity to environment. |
| Peak / resonance bin | Medium | Medium | Resonance shift | Installation changes can move peaks; use “trend + duration” not a single frame. |
| Spectral centroid | Medium | Medium | Timbre changes | Sensitive to broadband noise; combine with bandpower and duration rules. |
| Zero-crossing rate | Low | Low | Coarse HF content changes | Unstable at low SNR; gate it with minimum energy (RMS/bandpower). |
Figure F7 — Framing (window + hop), feature outputs, and the trigger decision point
H2-8. Event Triggers: Thresholds, Hysteresis, Debounce, and False Alarms
A threshold alone does not form a reliable trigger. Robust event detection is a rule system that combines hysteresis, debounce, duration checks, multi-feature gating, and cooldown behavior to control false alarms and “repeat triggers.”
Core mechanisms (what each one fixes)
- Hysteresis: prevents rapid on/off toggling when features hover near the threshold.
- Debounce / duration: requires N consecutive frames (or minimum time) to reject short spikes.
- Hold-off / cooldown: suppresses repeat triggers during overload recovery or ringing.
- Multi-condition gating: combines features (e.g., bandpower + envelope + duration) to reduce single-metric fragility.
Common false-alarm sources (symptom → mitigation knob)
- Clipping recovery / overload tail: use overload flags, longer cooldown, and ignore features during recovery.
- Mains hum / wind-like low frequency: align bands to the target event; add duration and bandpower gating.
- Structural resonance / ringing: require multi-frame persistence and add post-trigger hold-off.
- Power ripple coupling: avoid ripple-dominant bands; gate by minimum energy and stability checks.
Single-metric vs multi-metric: practical recipes
- Impulsive events: Peak + short duration + strong cooldown (+ overload guard).
- Sustained noise rise: Envelope/RMS + minimum duration + mild cooldown.
- Band-limited anomalies: Bandpower + envelope + duration (optionally multi-band consistency).
- Resonance shift: Peak-bin / centroid trend + persistence; avoid single-frame decisions.
A robust tuning order is: choose feature bandwidths first, then set thresholds relative to the noise floor, then add hysteresis and debounce, and finally tune cooldown to suppress repeat triggers during recovery.
Figure F8 — Trigger state machine template with key parameters (thresholds, debounce, cooldown)
IDLE: low-power monitoring; optional baseline tracking for features that drift with environment.
ARM: feature approaches threshold; enable multi-condition checks and debounce counters.
TRIGGER: latch on confirmed conditions; freeze sensitive adaptive updates if needed.
CAPTURE: commit pre/post buffers and timestamps; apply hold-off rules while recording.
COOLDOWN: suppress retriggers during overload recovery or resonance tails; return to IDLE when stable.
H2-9. Always-On Power Architecture: Duty Cycling Without Missing Events
“Always-on” does not mean full-rate sampling and full-rate compute at all times. A practical edge node stays vigilant by keeping a tiny always-on domain awake for low-cost gating, while the main domain wakes only when evidence crosses a reliable trigger boundary.
Split the system into domains (what stays on vs what wakes)
- Always-on domain (AON): low-power clock/counter + simple feature gating (energy / bandpower / envelope) and wake logic.
- Main domain: higher-rate capture, richer features, packaging, and storage operations during short “event windows.”
- Domain gating: keep the minimum set alive (time base + gate feature + wake path); power down heavy compute and fast storage paths until needed.
AON gate
Main wake
Domain off
Short active bursts
Coarse-to-fine listening (hierarchical monitoring)
- Level-0 (sentinel): low-rate, low-cost features to detect “something changed.”
- Level-1 (confirm): tighter band checks + duration rules to reduce false wake-ups.
- Level-2 (evidence): main domain wakes for high-rate capture + pre/post buffers and event metadata.
The “do not miss events” guarantee is achieved by a fast sentinel plus pre-trigger buffering (capture can include what happened before the trigger decision).
Sampling rate vs feature rate (two different dials)
- Sampling rate is set by bandwidth and anti-alias needs.
- Feature update rate is set by acceptable trigger latency and hop/window choices.
- Power wins often come from reducing how often expensive features run, not from reducing the ADC rate alone.
A reliable tuning order is: define a low-cost Level-0 gate, add Level-1 confirmation to suppress false wake-ups, and only then size Level-2 capture windows (pre/post) to preserve evidence.
| State | Typical activity | Current | Duty (how often / how long) | Design knob / risk |
|---|---|---|---|---|
| Sleep | Most blocks off | Isleep | Dsleep ≈ 1 − (others) | Leakage, clock strategy, wake sources |
| Listen | AON gate features | Ilisten | Dlisten (continuous, low) | Feature rate, band choice, baseline drift |
| Capture | High-rate window | Icap | Dcap = fevent · Tcap | Pre/post size, false alarms inflate duty |
| Report | Package / queue | Irep | Drep = fevent · Trep | Keep abstract: “exists”; do not bind to protocols |
| Average | Ī = Isleep·Dsleep + Ilisten·Dlisten + Icap·Dcap + Irep·Drep Lower false alarms → smaller Dcap/Drep → longer life. | |||
Level-0: cheapest gate feature (energy / bandpower) with conservative hysteresis.
Level-1: confirm with duration + band consistency before waking the main domain.
Level-2: main domain capture with pre/post buffers + event metadata for evidence integrity.
Figure F9 — State timeline and energy steps (sleep → listen → trigger → capture → report)
H2-10. Data Capture & Timestamping: Buffers, Pre/Post Trigger, Storage Integrity
A trigger is only valuable if it preserves evidence. A practical node captures a short pre/post window from a continuous ring buffer, stores raw snippets plus feature summaries, and attaches metadata and timestamps so every event stays comparable and debuggable.
Ring buffer → pre-trigger evidence (no missed onsets)
- Continuous overwrite: the buffer always contains the most recent history window (Tpre).
- Trigger = slicing: on trigger, freeze (now − Tpre) to (now + Tpost) into an event record.
- Edge cases: handle wrap-around and overlapping triggers using cooldown or merge rules.
Two data channels: raw vs features (storage-aware)
- Raw snippets: short segments for forensic replay (what actually happened).
- Feature summary: compact descriptors over longer time windows for indexing and trending.
- Why both: features accelerate triage; raw protects against false interpretations.
Timestamping (local-only, protocol-free)
- Tick counter: stable relative timing inside the captured window.
- RTC: absolute time across events; attach quality flags if needed.
- Do not expand: network time sync is out of scope for this page.
Storage integrity (trustworthy evidence)
- Atomic commit: write payload first, then write a commit marker to confirm validity.
- CRC: detect corruption and reject partial records after power loss.
- Recovery rule: on boot, scan for the last valid commit and discard trailing partial writes.
| Record part | What it contains | Why it matters | Common fields (examples) |
|---|---|---|---|
| Raw segment | Pre/post waveform snippets | Forensic replay and debugging | Tpre, Tpost, sample rate, clip flags |
| Feature summary | RMS/envelope/bandpower stats | Indexing, trending, fast triage | peak, duration, band IDs, confidence gates |
| Event metadata | Context and provenance | Comparability across time and builds | thresholds, rule version, firmware, temp/voltage |
| Timestamps | RTC + tick references | Align events and sequence them | rtc_time, tick0, tick_rate, quality flags |
| Integrity | CRC + commit marker | Detect partial writes/corruption | crc32, commit, record length |
Figure F10 — Ring buffer slicing (pre/post) + raw/features/metadata event packet
Acoustic / Vibration Edge Node — Validation & Parts Pointers
Two practical chapters for field-proofing event triggers and selecting core silicon (with multi-vendor example MPNs). Single-column layout and mobile-safe inline SVG figures.
H2-11
Engineering closure: detectability • latency • battery
Validation & Debug Playbook: How to prove it works in the field
Field readiness is not “it triggers in the lab.” It requires three metrics to pass together: capture quality (miss/false rate), real-time response (trigger latency), and energy (average current vs. event rate). Any one failure mode can break the product experience.
1) KPI definitions that prevent “argument-by-log”
| Metric | Engineering definition (testable) | What it drives (design knobs) |
|---|---|---|
| Detection / Miss rate | Given a labeled event window, does the final trigger fire within an allowed time bound? | Gain/headroom, feature choice, threshold margin, multi-condition gating, mounting/structure coupling |
| False alarm rate | Triggers per hour (or per shift) when no target event exists; report by environment segment. | Hysteresis/debounce/hold-off, clip recovery handling, hum/wind rejection, ripple coupling fixes |
| Trigger latency | Event onset → trigger decision timestamp; separate Level-0 gate vs Level-1 confirm delays. | Decimation group delay, window hop, confirm duration, pre/post buffer policy |
| Battery / Ī | Average current over a representative duty cycle including capture + logging spikes. | Always-on domain, tiered listening (coarse→fine), write batching, false alarm suppression |
2) A minimal, repeatable validation loop (concept-level)
- Controlled excitation: use a repeatable sound/vibration source to sweep thresholds and quantify misses/latency.
- Probe-point isolation: validate at AFE out → ADC out → feature out → trigger state.
- Clipping & recovery: force saturation (impulse/high SPL/shock). Measure recovery time and post-clip false alarms.
Field debug priority:
start with front-end saturation/recovery and power/ground coupling,
then tune threshold/window/state machine, and only then revisit sensor mounting/structure coupling.
This order removes “hardware contamination” before parameter optimization.
3) Symptom → root-cause pattern matching (fast triage)
| Symptom | Likely causes (tag) | Verification action | Fix direction |
|---|---|---|---|
| False alarms spike | [AFE] clip recovery tail [PWR] ripple injection [ALG] no hold-off |
Log “clip flag” / max code density; correlate with supply ripple and state transitions | Add post-clip blanking, tighten hysteresis/debounce, improve analog filtering & grounding |
| Missed short impulses | [ALG] gate too slow [DSP] long hop [MECH] weak coupling |
Check Level-0 gate latency; compare raw vs features; verify mounting torque/adhesive | Coarse→fine tiering, smaller hop, increase pre-trigger ring depth, fix mounting path |
| Late triggers | [ADC] group delay [DSP] confirm duration |
Measure end-to-end delay by injected step/impulse; separate decimation vs DSP | Lower-latency decimation mode, shorter confirm, keep longer pre-buffer to preserve evidence |
| Battery drains early | [SYS] frequent capture [LOG] write spikes [FA] high false rate |
Break down Ī by state: sleep/listen/trigger/capture/log; count false alarms per hour | Reduce false rate, batch writes, store feature summaries more often than raw waveforms |
Figure F11 — Field debug flow: symptom → probe points → knobs
H2-12
Selection criteria + example MPNs (multi-vendor)
Parts / IC Selection Pointers (with example MPNs)
The goal is to select by system consequences (noise floor, recovery, latency, logging integrity), not by catalog browsing. The MPNs below are examples to anchor the selection dimensions; final choice should be validated against availability, lifecycle, temperature range, and the exact sensor/interface constraints.
MPN usage note: Each table lists multiple vendors to reduce lock-in risk. If one part is unavailable, the selection criteria still holds.
A) AFE (LNA/TIA/Op-Amp) — must / nice / red flags + MPN examples
| Must-have | Nice-to-have | Red flags | Example MPNs (not exhaustive) |
|---|---|---|---|
|
Low input noise over target band Stable with expected Cin / sensor model Recovery behavior understood (post-clip) |
EMI-robust input stage Rail-to-rail I/O as needed Power down / fast wake for duty modes |
“Low noise” but slow recovery → false alarms Marginal phase margin with sensor capacitance Hidden distortion near rails/common-mode limits |
TI: OPA380 (TIA-style front end), OPA140/OPA145 (low bias), OPA1678/OPA1662 (audio LNA class) ADI: ADA4522-2 (low drift), ADA4807-2 (fast/low noise class) ST: TSV79x / TSV99x families (low power op-amp class) |
Tip: for charge/current-output sensors, prioritize stability with input capacitance and recovery behavior before chasing the last nV/√Hz.
B) ΣΔ ADC — why it fits audio/vibration + MPN examples
| Must-have | Nice-to-have | Red flags | Example MPNs |
|---|---|---|---|
|
Band-limited noise/DR aligned to event thresholds Group delay (decimation) acceptable for triggers Interface supports streaming + DMA reliably |
Switchable modes (listen vs capture) Integrated front-end options (mic/line levels) Low-EMI continuous-time front-end |
Great SNR but long delay → late triggers Mode transitions that glitch the baseline Clocking/PLL constraints ignored until late |
TI: PCM1863 (2ch), PCM1865 (4ch) audio ADC class; ADS127L01 / ADS127L11 (precision ΣΔ class) ADI: ADAU1978 (4ch ΣΔ audio ADC class) Cirrus: CS53Lxx families (audio ADC class) |
C) MCU / Low-Power DSP — always-on decisions + MPN examples
| Must-have | Nice-to-have | Red flags | Example MPNs |
|---|---|---|---|
|
Low-power always-on domain or deep sleep with fast wake Efficient MAC/FFT path for feature extraction DMA + ring buffer friendly streaming |
Hardware timestamps / low-jitter timers Sufficient SRAM for pre/post buffers Low-leakage retention options |
Wake latency too long for short impulses DMA limitations cause dropped samples Logging path spikes current excessively |
ST: STM32U5 series (ultra-low-power MCU class) Ambiq: Apollo4 Plus (ULP Cortex-M4F SoC class) NXP: LPC55S6x (M33 + DSP-ish workloads class) TI: MSP430FR5994 (ULP + FRAM logging class) |
D) Storage & RTC — logging integrity + MPN examples
| Must-have | Nice-to-have | Red flags | Example MPNs |
|---|---|---|---|
|
Power-fail safe write strategy support (atomic record / CRC) Endurance aligned to event rate (esp. frequent short events) RTC drift observable or compensable |
FRAM for “write-heavy” metadata RTC with timestamp/event input pin Backup switchover/trickle charger options |
NOR-only logging without wear strategy → early failures RTC drift ignored → thresholds drift with temperature/time No integrity tags → “evidence not trustworthy” |
FRAM: Infineon FM24CL64B (I²C FRAM class) SPI NOR: Winbond W25Q64JV (serial NOR flash class), Macronix MX25R64xx (low-power NOR class) RTC: Micro Crystal RV-3028-C7 (ultra-low-power RTC class), ADI/Maxim DS3231 (TCXO RTC class) |
Figure F12 — Concept selection matrix (where each block matters most)
Practical reading tip: start from the strongest pain point (false alarms / misses / latency / battery) and trace left-to-right using Figure F12 to pick the most impactful component constraints first.
Request a Quote
FAQs — Acoustic / Vibration Edge Node
Practical troubleshooting questions for low-noise AFE + ΣΔ ADC decimation + edge DSP triggers, always-on power, evidence capture, and field validation.
H2-13
FAQs ×12
User-readable answers + structured data (FAQPage JSON-LD)
Figure F13 — FAQ coverage map (pain points → questions → chapter knobs)
1 Why do false alarms increase after switching to a TIA with the same sensor? What three checks come first?
A TIA can amplify non-event artifacts: bias/leakage paths, input capacitance effects, and post-clipping recovery tails. First checks: (1) confirm whether the front end clips during impulses and how long it takes to recover, (2) verify stability with the sensor’s effective capacitance and cabling, and (3) inspect trigger state-machine settings (hysteresis, debounce, hold-off) that may now be too sensitive.
Maps: H2-4H2-8H2-11
2 How should trigger thresholds be set to avoid “daytime triggers, nighttime misses”? How to handle temperature drift and changing ambient noise?
Fixed thresholds fail when the noise floor and baseline drift with temperature and environment. Use a threshold tied to a measured baseline (noise statistics or envelope floor) and apply slow adaptation with bounds. Track temperature alongside triggers to identify drift correlation. Validate by segmenting data into environment slices (quiet/noisy/hot/cold) and comparing miss/false rates per slice before locking the policy.
Maps: H2-6H2-8H2-11
3 Does ΣΔ ADC decimation delay make triggers slower? How to compensate with pre-trigger buffering?
Yes—decimation filters introduce group delay, which shifts the trigger decision later than the physical event. The fix is not “faster thresholds,” but evidence management: keep a ring buffer that always stores a short history, then cut a window that includes pre-trigger samples. Measure delay end-to-end (event onset to decision) and size the pre-buffer to exceed the worst-case delay plus window/hop latency.
Maps: H2-5H2-10
4 Why do low-frequency wind noise or structural sway destabilize triggers? Should the high-pass filter be analog or digital?
Low-frequency energy can push the AFE toward saturation, shift baselines, and inflate envelope/RMS features, causing runaway false alarms. An analog HPF prevents headroom loss before the ADC, while digital HPF cleans residual drift after conversion. A robust split is: analog HPF to protect the front end (anti-saturation), digital HPF plus duration gating to prevent “slow motion” from satisfying trigger rules.
Maps: H2-4H2-5H2-8
5 What “fake trigger” symptoms appear when front-end saturation recovery is too slow, and how can it be measured?
Slow recovery often looks like repeated triggers after a single impulse: the waveform baseline creeps back, or ringing crosses thresholds multiple times. Measure it by forcing a controlled over-range event (impulse/high SPL/shock), logging AFE/ADC peak codes, and marking trigger timestamps. If triggers cluster during the recovery tail rather than the event core, add post-clip blanking/hold-off and revisit gain/headroom and stability margins.
Maps: H2-4H2-11
6 RMS trigger or bandpower trigger—how to choose, and what are the common pitfalls?
RMS is cheap and works for broad loudness/energy changes, but it is vulnerable to low-frequency drift and slow sway that inflate energy without a real event. Bandpower is more selective and fits “specific band rises” or resonance changes, but it depends on correct filtering, windowing, and stable frequency response. A common safe pattern is bandpower + duration gating, with RMS used only as a coarse pre-gate.
Maps: H2-7H2-8
7 How should frame and hop be chosen to balance latency and detection rate?
Larger frames stabilize features and reduce false alarms, but add latency and can smear short impulses. Smaller frames respond faster, but become noisy and require stronger debounce/confirm logic. Treat frame as “feature stability,” hop as “decision cadence,” and confirm duration as “false-alarm guard.” Validate by sweeping frame/hop and measuring miss/false and decision latency together; then lock a configuration that meets KPI targets across environments.
Maps: H2-7H2-8
8 Always-on power will not meet budget—what sampling or compute actions should be cut first for maximum impact?
Start with actions that scale continuously with time: reduce always-on sample rate (coarse listening), lower feature update rate, and avoid frequent spectral transforms unless strictly needed. Prefer a tiered strategy: simple envelope/bandpower at low rate for Level-0 gating, then short bursts of high-rate capture and heavier features only after gating. Finally, reduce logging frequency by storing summaries more often than raw waveforms.
Maps: H2-9
9 After a trigger, should raw samples be stored or only features? How to trade storage endurance versus diagnosability?
Raw snippets enable offline re-checking and root-cause analysis, but write-heavy logging can shorten flash life and raise power spikes. Features plus event metadata are lighter and scalable, but may miss evidence for edge cases. A practical compromise is “rare raw, frequent summary”: store compact features and metadata for every event, and store short raw windows only for selected events (first-in-shift, high severity, or anomalies). Use batching and integrity tags.
Maps: H2-10H2-12
10 Why can a node be “accurate in the lab but not in the field”? Is it usually mounting/structure coupling or power coupling?
Field variance is commonly a combination. Mounting changes coupling and resonance, reshaping the spectrum and amplitude so thresholds no longer match. Power coupling injects ripple and digital noise into the analog chain, creating phantom energy in features. Triage efficiently: check for clipping and recovery artifacts, then correlate triggers with supply ripple and state transitions, and finally verify mounting repeatability (torque/adhesive/contact) using the same controlled stimulus across setups.
Maps: H2-3H2-11
11 Can the same hardware run both acoustic and vibration triggers? What conflicts appear when sharing the signal chain?
Conflicts typically appear in headroom, filtering, and rule tuning. Acoustic and vibration events can differ in dominant bands, amplitude distribution, and impulse content. Shared gain and HPF settings may protect one modality while degrading the other. Shared trigger thresholds often fail because baseline and noise statistics differ. The safest approach is minimal separation: per-channel calibration/thresholds, modality-specific features (or bands), and independent state-machine parameters while sharing capture/log infrastructure.
Maps: H2-2H2-4H2-8
12 Are more complex trigger rules always better? How to avoid rule stacking that becomes unmaintainable and unverifiable?
Complexity can reduce false alarms, but it often destroys testability. Prefer a structured template: Level-0 simple gate, Level-1 confirmation, then cooldown. Every added condition should have a measurable benefit (false-alarm reduction vs missed events) and should be versioned in logs. Avoid hidden coupling by limiting the number of conditions and keeping parameters interpretable (threshold, duration, hysteresis, hold-off). Use the validation checklist to prove each rule’s contribution.
Maps: H2-8H2-11
