System Boundary & Sensor Variants

This page treats contact sensing (door/window open–close and tamper) and glass-break detection (event-level) as a single endpoint category because they share the same long-life power, ULP MCU, and Thread/Zigbee radio. The sensor front-end and decision logic are intentionally separated: contact uses magnetic state; glass-break uses acoustic signatures. The goal is a design that is measurable, explainable, and robust against false alarms.

The selector below forces a clean split: pick the user-visible outcome first, then bind to a measurable evidence chain.

• VBAT droop during radio TX burst: measure VBAT with a scope at the moment of transmission; confirm UVLO margin and reset behavior.
• Raw sensor observables: contact edges/time-stamps (Reed) or analog readings (Hall); acoustic two-band envelopes (glass-break).
• Event log: store the last N events with timestamps and key scalars (edge count, peak envelope, noise floor) for field debug.

Event Physics & Signal Signatures

Sensor front-ends must be sized from what the real-world event looks like at the sensor pins. This section defines the minimum signatures needed to build an explainable detector: contact events are dominated by magnetic geometry and threshold crossings; glass-break events are dominated by a two-stage acoustic pattern (impact → shatter burst). Each signature is framed with a measurement plan so thresholds can be tuned without guessing.

• True open/close: a clean transition that crosses the threshold once, with stable settle after closure.
• Chatter near threshold: edge bursts (multiple crossings) caused by door vibration, loose magnet, or insufficient geometry margin.
• Geometry drift: the “closed” state sits too close to the threshold (metal frame, misalignment, door warp), making small motion look like an event.

• Stage A — impact energy: a lower-frequency rise from an удар/knock; common in false triggers (slam, thunder, door hit).
• Stage B — shatter burst: a short, high-frequency-rich burst; the most discriminative local evidence.
• Sequence gate: Stage B must follow Stage A within a defined time window; this filters random noise spikes.

Practical rule: if thresholds cannot be explained by the above shapes (edge bursts, two-band timing, baseline ratio), the design is not field-debuggable and will drift into false alarms.

Reference Architecture Block Diagram

A reliable endpoint is defined by a single evidence chain: battery → power tree → wake event → sampling window → local decision → radio burst → event log. This reference diagram prevents scope drift by forcing every design choice to map to a measurable node: VBAT droop, raw sensor observables, and time-stamped decisions.

Debug discipline: any field failure must map to one of the taps above (VBAT droop, raw signal shape, decision scalars, reset flags). If it cannot be measured, it cannot be stabilized.

Contact Sensing Path A: Reed Switch

A reed switch is the lowest-power contact sensor because it can behave like a near-zero-leakage state input. The engineering challenge is not the switch itself, but magnetic geometry margin and threshold chatter: small vibrations can produce edge bursts that look like multiple opens. A robust design turns those edges into measurable evidence and filters them without hiding real opens.

• Geometry margin: closed-state distance and misalignment window must sit far from the effective threshold.
• Edge-burst statistics: measure burst length and edge count under controlled vibration; tune debounce to cover typical bursts.
• TX-aligned checks: verify whether false edges cluster at radio transmit moments (VBAT droop or EMI injection).

A reed design is stable only when the closed-state geometry margin is large enough that vibration produces few or no threshold crossings. Debounce should be a measured parameter (burst duration/edge count), not a fixed default.

Contact Sensing Path B: Hall Sensor

Hall sensing turns “magnet geometry” into a measurable signal with tunable thresholds. Compared with a reed switch, Hall enables programmable hysteresis and installation-aware calibration, which directly reduces false opens caused by door vibration, magnet misalignment, and metal door-frame distortion. The design objective is not higher sensitivity, but stable margin: the closed-state reading must remain far from the switching region across temperature and mechanical drift.

• Distance curve: raw Hall code vs distance/offset; confirm the “safe gap” between closed and open distributions.
• Temperature curve: raw code drift vs temperature; verify thresholds remain outside drift envelope.
• Replay buffer: on any false event, store the last N seconds of raw codes + threshold/hysteresis values for replay.

A Hall design is stable when the closed-state distribution (across temperature and alignment) remains well separated from the switching band. Calibration should be bounded and replayable so that false events can be proven with raw data.

Glass-Break AFE: Mic → AFE → ADC/Comparator

Glass-break detection succeeds when the analog front-end preserves the signature without distortion. The key engineering variables are bias stability, gain and dynamic range, band shaping, and clipping avoidance. A practical endpoint can separate noise from true events using two-band energy (LF impact + HF shatter burst) and a time gate, without relying on cloud inference.

• Time-domain snippet: confirm no clipping and that the burst exists (impact + shatter segment).
• Two-band envelopes: LF rises first, HF shows a short burst above baseline, within a defined window.
• Baseline capture: noise floor measured in the installation environment to prevent fan/voice false triggers.

Priority order for stable detection: prevent clipping first, then set ratio-based thresholds to the local baseline, and finally enforce LF→HF timing. This keeps the decision explainable and field-debuggable without complex models.

Local Classification & False-Alarm Control (Explainable)

False-alarm control must be explainable on the endpoint. A robust local decision uses two-band energy (LF impact + HF shatter burst), duration limits, and a time gate (LF → HF order) with thresholds referenced to an adaptive noise floor. Each report should carry a compact “why” record: Rule ID, thresholds, baseline, and feature snapshots.

• Rule ID hit list: which rules fired (e.g., R1+R3 but missing R2).
• Baseline snapshot: Baseline_LF/HF at trigger time (night/AC/rain drift visibility).
• Feature snapshot: LF_peak, HF_peak, HF_dur, Δt, ratio — logged for replay.

A stable endpoint report is one that can be replayed and explained: which rules fired, what baseline was assumed, and what feature values crossed thresholds.

Long-Life Power Architecture (Current Waveform Managed)

Long battery life is achieved by managing the current waveform, not by slogans. The lifetime loop closes only when sleep current, wake sampling window, TX burst peak, and battery internal resistance (especially at low temperature) are measured together. The critical failure mode is VBAT droop during peaks crossing BOR/UVLO, causing resets and missed events.

• Sleep current (S0): measure nA/µA-class current after stabilization.
• TX peak & VBAT droop (S3): scope VBAT and tag the TX moment; verify droop margin vs BOR/UVLO.
• Low-temp reproduction: repeat bursts at low temperature to reveal increased internal resistance and deeper droop.

A credible lifetime claim requires replacing all placeholders with measured currents, measured durations, measured event rates, and scoped VBAT droop under worst temperature.

Thread/Zigbee Radio Coexistence (Power + Antenna Evidence)

Endpoint radio stability is determined by two measurable chains: TX burst → VBAT droop (and possible BOR) and antenna efficiency under metal/battery shielding. Coexistence here focuses on hardware evidence only: align the TX moment to VBAT waveform and correlate packet performance statistics with distance and installation geometry.

• Keepout: reserve a no-metal/no-battery region around the antenna area; avoid screws and large copper islands in the near-field.
• Ground reference: provide a controlled ground reference; avoid “floating” grounds that make tuning unstable across units.
• Metal proximity risk: door frame and battery can detune and block radiation; validate by rotating/offsetting a few centimeters and re-logging RSSI/LQI.

• VBAT waveform aligned to TX marker: droop depth + recovery time + BOR flag correlation.
• Packet performance vs distance: simplified PER curve and RSSI/LQI distribution; compare “near metal” vs “free-space.”

A “range drop” is not a protocol mystery until VBAT droop and antenna shielding have been ruled out with aligned evidence.

Ruggedness: ESD/EMC/Tamper (Minimum Set + Local Flags)

Endpoint sensors most often fail from the real environment: ESD from touch and door frames, EFT-like fast transients from long leads, and repeated coupling into sensitive inputs. Ruggedness is achieved by a minimum protection set (TVS + series-R + RC where needed) and by verifying return paths. Tamper mechanisms must remain local and evidence-based: case-open and magnet interference should set local flags and freeze adaptive learning.

• Sleep current drift: rising sleep current indicates leakage damage; compare pre/post stress measurements.
• GPIO behavior: input threshold/leakage changes show up as unstable levels or abnormal edge counts.
• AFE baseline drift: mic/AFE bias shift raises noise floor and increases false triggers; log baseline before/after.

• Separate power vs input damage: if sleep current rises immediately, suspect leakage; if only triggers rise, suspect input baseline shift.
• Check return-path effectiveness: repeat stress while observing baseline and droop; improvements should be visible in waveforms/logs.
• Freeze learning on instability: prevent adaptive baselines from masking damage and making faults non-replayable.

Many “passing” units still degrade: higher sleep current, unstable GPIOs, or elevated AFE noise floors. Logging these before/after stress is the fastest proof of damage.

Validation & Production Test Plan (Manufacturable, Reproducible)

A sensor is manufacturable only when failure modes are caught by a minimum, repeatable test set. This chapter defines a 3-layer closure loop: EOL (factory) → Sampling (QA) → Field repro. Every test has: pass criteria + failure code + next evidence to isolate the root cause.

• EOL (End-of-Line): fast screens for assembly, thresholds, and gross defects (minutes, not hours).
• Sampling (QA): stress the corner conditions (distance/angle/temp, noise scenes, IR simulation) to catch systematic drift.
• Field repro: minimum evidence capture with aligned markers: VBAT_min@TX + rule_id + PER/RSSI stats.

The goal is not to test “everything”, but to lock a repeatable tolerance window for installation. The same matrix structure works for reed and Hall; only the logged fields differ.

Glass-break validation must stay locally explainable: two-band energy + duration + time order + adaptive noise floor. The scenario set focuses on “looks similar but should not trigger”.

“Range drop” and “random reset” often collapse to one measurable chain: TX burst → VBAT droop → BOR. IR simulation uses a series resistor to emulate battery internal resistance rise at low temperature or aging.

A test is only “production-ready” when a failure automatically produces: (1) a code, (2) a minimum evidence snapshot, and (3) a deterministic next step.