Three independent layers of AI monitor your unit 24/7 — from the accelerometer on the compressor housing to plain-English diagnosis on your phone. Detects eight specific fault types weeks before they become failures.
See how it works Order yoursNo subscription · Ships ready to mount · Installs in minutes
Heat pumps and central AC units give almost no warning before failure. Bearings degrade over months. Fan blades slowly go out of balance. Coils ice up gradually. By the time you hear something wrong, the damage is usually done — and the timing is always the worst possible: the hottest day of summer or the coldest night of winter.
Professional vibration monitoring systems exist, but they cost hundreds to thousands of dollars per year in subscription fees — designed for commercial facilities, not homes. HVAC Vibration Guard brings the same physics-based fault detection to any homeowner for the price of a service call.
A single prevented compressor failure covers the cost of protection many times over.
Each layer catches what the others might miss. No single point of failure in the detection chain.
Signal processing at the source
400 Hz sampling — fast enough to capture all five diagnostic frequency bins up to 120 Hz (Nyquist 200 Hz)
Real FFT — 256-sample Hamming-windowed DFT resolves 33, 60, 66, 99, and 120 Hz with 1.5 Hz bin resolution
Per-axis RMS — X, Y, Z measured independently; asymmetry between X and Y is a fan imbalance signature
State machine — tracks OFF / STARTING / RUNNING / STOPPING so every measurement carries operating context
MQTT over TLS — payload published every 500 ms to HiveMQ Cloud; Android receives it in real time
Classical algorithms on your phone
Baseline learning — first 60 seconds of each run cycle, the system learns what normal looks like for your specific unit
EMA adaptive reference — the baseline drifts slowly with the unit's natural aging so hot summer days don't cause false alarms
CUSUM drift detection — accumulates tiny deviations over time; catches bearing wear weeks before any threshold is crossed
Mahalanobis distance — treats all five FFT bins as a correlated vector; detects fingerprint shifts no single-bin ratio can see
8 fault classifiers — each scored and blended; the dominant fault wins with a confidence percentage
Plain English for the homeowner
Fault-triggered — Gemini is called only when the active fault type changes, not on every sample; no noise, no spam
Full context — receives fault type, severity, confidence, frequency readings, rate of change, and operating history before responding
Plain English output — 2–3 sentence diagnosis in homeowner language: what it is, how serious, what to do next
Chat interface — ask follow-up questions. "Should I call someone today?" "Is this covered by warranty?" — answered in context
Three classical algorithms, each solving a different detection problem. No black box — every alert has a traceable mathematical cause.
A fixed threshold set in June will false-alarm every August — hotter weather legitimately increases vibration. Exponential Moving Average continuously adjusts the reference point with a time constant of ~8 minutes, so seasonal drift is absorbed automatically.
Bearing wear happens over months — each reading is only fractionally above normal. A 30-second rolling average never sees it. CUSUM accumulates tiny consistent deviations like a savings account; when the balance crosses a threshold set at 5 standard deviations, the alarm fires. Single spikes don't trigger it — only sustained drift does.
Fan imbalance simultaneously elevates 33 Hz and 99 Hz while leaving 60 Hz unchanged — a correlated five-bin fingerprint that no single ratio check can see. Mahalanobis distance measures how far the current FFT vector is from the baseline distribution, accounting for the correlations between all five bins at once.
Every alert identifies the specific mechanical cause and the algorithm that caught it — not a generic "vibration elevated" warning.
| Fault | What's happening | Detection signal | Algorithm | |
|---|---|---|---|---|
| ⚙ | Fan Imbalance |
Dirty blades, bent fin, or assembly imbalance causing uneven rotation | 33 Hz BPF elevation + X/Y axis asymmetry | EMAMahal |
| ⟳ | Bearing Wear |
Motor or fan bearing degradation — the earliest detectable mechanical failure | 99 Hz + 120 Hz harmonic rise, slow RMS drift | CUSUMMahal |
| ⚠ | Debris / Obstruction |
Leaves, ice, or foreign material contacting a rotating component | Crest factor spikes (peak/RMS > 3.5×) in >10% of samples | Crest |
| 〰 | Panel Rattle |
Loose sheet metal cover or housing panel vibrating at grid harmonics | 99 Hz / 120 Hz bin dominant relative to total RMS | Ratio |
| ❄ | Coil Icing |
Ice formation on evaporator coil restricting airflow and straining compressor | Compressor 60 Hz rise + sustained RMS uptrend combined | EMATrend |
| ⚡ | Startup Stress |
Abnormal compressor startup transient — signals capacitor or motor wear | Startup amplitude profile deviation from baseline timing | State |
| ◈ | Spectral Anomaly |
FFT fingerprint changed in a way that doesn't match any named fault pattern | Mahalanobis D² > χ²(5, p=0.99) = 15.09 threshold | Mahal |
| ↑ | Elevated Vibration |
Sustained RMS above adaptive baseline — general mechanical stress indicator | Rolling average >20% above EMA baseline for >30 seconds | EMATrend |
HVAC Vibration Guard ships fully assembled and pre-configured. No firmware to flash, no wiring, no broker to set up.
Ships assembled in a weatherproof enclosure with magnet mount, power adapter, and everything needed for installation.
The neodymium magnet sticks directly to your compressor housing — no drilling, no brackets. Takes about 60 seconds.
Scan a QR code to configure your network on first boot. The device connects to the Sensee cloud over TLS — no port forwarding or router changes needed.
Scan a second QR code to pair the app to your device. Live data appears immediately.
Within the first run cycle, the system learns your unit's normal vibration fingerprint. EMA, CUSUM, and Mahalanobis models go live automatically. You'll hear from it only when something changes.
Typical install time: under 5 minutes. If you run into anything, we respond personally to every support message — usually same day.
Early access, partnership, installation help, or just a question — we read every message and respond personally.