Most acoustic monitoring pilots fail after the demo. The algorithm was fine; the installation quietly changed the signal. We specify and deliver the whole chain against the asset's failure mode and your reaction time.
None of them is the classifier.
A wind-turbine main bearing and a breaking window occupy different decades of the spectrum. An 8-element array gives a 1.8-octave window wherever you put it, so the array is specified against the asset, not picked from a catalogue.
Events are classified and graded locally. What leaves the site is a feature vector and an event record — never the raw audio, and never a continuous stream.
Calibration, mounting resonance, wind and rain, clock discipline between nodes. These are the reason pilots stall, so they are part of the deliverable rather than an afterthought.
The delivered chain. Raw audio never leaves the site; a feature vector and an event record do.
Every monitoring point is fixed by the same four numbers: the band you must be sensitive to, the sampling rate that follows, how fast the judgement is needed, and where it can live.
| Group | Scenario | Band | Sampling | Response | Decision |
|---|---|---|---|---|---|
| Rotating and reciprocating machinery | Motor / pump — unbalance, misalignment, bearing | 10 Hz – 10 kHz | ≥ 24 kHz | hours | On device |
| Rotating and reciprocating machinery | Gearbox / reducer — mesh, tooth breakage | 100 Hz – 20 kHz | ≥ 48 kHz | hours | On device |
| Rotating and reciprocating machinery | Press / machine tool — impact, tool wear | 200 Hz – 20 kHz | ≥ 48 kHz | milliseconds | On device |
| Energy assets | Wind turbine — main bearing, blade, yaw | 0.1 Hz – 10 kHz | ≥ 24 kHz + slow channel | hours | Trigger + review |
| Energy assets | Tower / blade — modal and low-frequency response | 0.1 – 20 Hz | ≥ 100 Hz | minutes | Trigger + review |
| Energy assets | Transformer / reactor — abnormal hum | 50 Hz – 5 kHz | ≥ 12 kHz | hours | On device |
| Livestock houses | Ventilation fan — bearing, belt, rotor lock | 20 Hz – 10 kHz | ≥ 24 kHz | minutes | On device |
| Livestock houses | Animal vocalisation — cough, alarm, farrowing | 50 Hz – 8 kHz | ≥ 24 kHz | seconds | On device |
| Livestock houses | Feed line / scraper — jam and abnormal impact | 20 Hz – 8 kHz | ≥ 24 kHz | minutes | On device |
| Hydrology and slope hazards | River acoustics — discharge and turbulence shift | 20 Hz – 5 kHz | ≥ 12 kHz | minutes | On device |
| Hydrology and slope hazards | Debris flow / flash flood — low rumble plus sustained scour | 10 Hz – 2 kHz | ≥ 6 kHz | minutes | Device + cloud |
| Hydrology and slope hazards | Rockfall / slope collapse — impact transient and rolling rhythm | 100 Hz – 20 kHz | ≥ 48 kHz | milliseconds | On device |
| Transport and structures | Bridge — vehicle passage response, expansion-joint impact | 1 Hz – 2 kHz | ≥ 6 kHz | seconds | Device + cloud |
| Transport and structures | Pavement — tyre noise as a proxy for damage and ponding | 200 Hz – 5 kHz | ≥ 12 kHz | seconds | On device |
| Transport and structures | Rail track — wheel/rail noise, joints, corrugation, switches | 100 Hz – 20 kHz | ≥ 48 kHz | milliseconds | On device |
| Transport and structures | Rope and stay cable — tension, wire break, rain-wind vibration | 0.5 Hz – 5 kHz | ≥ 12 kHz | minutes | Trigger + review |
| Transport and structures | Building — curtain wall, structural noise, lifts, plumbing | 50 Hz – 10 kHz | ≥ 24 kHz | seconds | On device |
| Indoor and home | Glass breakage | 2 – 8 kHz | ≥ 24 kHz | milliseconds | On device |
| Indoor and home | Fall — impact plus vibration | 20 Hz – 2 kHz | ≥ 8 kHz | milliseconds | On device |
| Indoor and home | Hob left dry-burning — whistle, vaporisation, empty pan | 1 – 16 kHz | ≥ 40 kHz | seconds | On device |
| Indoor and home | Pipe — leak, cavitation, blockage | 100 Hz – 50 kHz | ≥ 100 kHz | minutes | On device |
| Security | Drone intrusion — blade-pass frequency and harmonics | 100 Hz – 8 kHz | ≥ 24 kHz | seconds | On device |
Bands are industry practice across vibration monitoring, acoustic emission and structural health monitoring. Sampling rates are our engineering figures: the band doubled with margin, then mapped to a rate the RK3308 audio paths actually produce.
Below c/D an array has no directional gain; above c/2d it produces grating lobes. The window between them is 1.8 octaves for 8 elements, and spacing only slides it.
| Element spacing d | Array length (N−1)d | Usable from c/D | Aliasing above c/2d | Where it fits |
|---|---|---|---|---|
| 20 mm | 140 mm | 2.4 kHz | 8.6 kHz | In-module array. High-frequency transients only — glass break, impact. |
| 40 mm | 280 mm | 1.2 kHz | 4.3 kHz | Panel-mounted array. Hand tools, small gearboxes, hob and pipe noise. |
| 100 mm | 700 mm | 490 Hz | 1.7 kHz | Cabinet or room array. Mid band — ventilation fans, pumps, rail. |
| 300 mm | 2100 mm | 163 Hz | 572 Hz | Machine- or mast-mounted. Low-frequency machinery and water noise. |
| 1000 mm | 7000 mm | 49 Hz | 172 Hz | Outdoor structure or tower. Structural modes, sway, cable force. |
Most projects need two apertures rather than one: a dense array for transients and a sparse, long array for structural low frequencies, sharing one board. Widening a single aperture is not possible; adding elements is — 16 elements would give 2.9 octaves.
RK3308 reaches microphones three ways and has no NPU. Both facts shape the delivery.
| Item | Capability | What we use it for |
|---|---|---|
| PDM | 8 channels, up to 192 kHz, 16–24 bit on a 5-wire interface | Impact, leak and other rows whose answer lives above 20 kHz |
| I2S / TDM | 2 × 8 channels at 48 kHz | Standard digital microphones, mid-band work, slow structural channels |
| On-chip CODEC | 8 × 24-bit ADC, differential or single-ended, programmable gain | Analog measurement microphones wired straight to the board |
| Hardware VAD | Analog, I2S and PDM microphones; array pre / post buffering | Keeping the onset in the recording — the part a threshold-triggered recorder loses |
| Raw data rate | 93 GB per node per day at 48 kHz (371 GB at 192 kHz) | Why capture is event-driven and features are continuous |
| Compute | 4×Cortex-A35 @ 1.3 GHz, no NPU | Classical DSP plus small classifiers on the board; heavier models and fleet reasoning in the cloud |
We would rather state the ceiling than discover it in your pilot. If your requirement needs a large model on the same box, the right answer is a platform with an NPU behind an RK3308 acquisition front end — and we will say so.
Four work packages. The last one is the one that usually gets skipped.
Aperture, element count, spacing and enclosure specified against the asset. Brackets and windscreens included, because mounting resonance ends up in the measured spectrum.
RK3308 bring-up on the PDM, I2S/TDM and analog paths; per-channel gain and delay calibration; hardware-VAD pre/post buffering with timestamps taken at the capture point.
Feature extraction, a commissioned baseline for that specific asset, novelty scoring, event classification and a severity grade. All of it runs on the board, with no round trip.
Event records and feature vectors into your platform. Optional cloud layer for explanation, work orders and fleet comparison. Raw audio remains on site.
A pilot that classifies events is not the deliverable. The deliverable is a monitoring point that still works in month six, with a baseline that belongs to that asset and a false-alarm rate you can live with.
| Step | What happens | What we need from you |
|---|---|---|
| Failure mode and reaction time | Fix the band, the sampling rate and the latency budget from the failure you care about | The equipment, the failure mode, and how fast you must know |
| Aperture and placement | Array geometry, element count and mounting points; simulation where the structure is complex | Drawings, access, and the noise sources around the asset |
| Capture and calibration | Firmware bring-up, channel calibration against a known source, stored gain and delay offsets | A quiet window for commissioning |
| Baseline and rules | A commissioned fingerprint for that asset, novelty thresholds, event classes and grades | Two to four weeks of normal operation to build the baseline |
| Integration and acceptance | Events into your platform, false-alarm review, threshold tuning, handover with the baseline file | The platform endpoint, and the people who will act on the alerts |
Step four is the one that cannot be compressed. A baseline measured over two days is a baseline of two days' weather, load and traffic — not of the asset.
No. An 8-element array gives a 1.8-octave window wherever you place it. Assets whose failure bands sit far apart need different apertures, or several monitoring points.
No, and we would argue against it. 93 GB per node per day of raw 8-channel audio at 48 kHz is not a cloud workload. What leaves the site is features and events.
Contact accelerometers are better above a few kilohertz and on surfaces you can reach. Microphones are non-contact, faster to deploy and cover a wide area from one point. The two are complementary, and we deploy both on most sites.
That depends on installation discipline, not on the model. We commission a baseline per asset and tune thresholds against your tolerance before handover. Anyone quoting a figure before the installation has been surveyed is guessing.
For classical DSP and small classifiers, yes — the audio front end is the hard part, and eight-channel capture with hardware VAD is what the board is built for. If you need larger models on the same box, we move the compute to a platform with an NPU and keep RK3308 as the acquisition front end.
One monitoring point on one asset with a known failure mode, run long enough to build a real baseline. Decide from that evidence whether to scale, rather than from a demonstration.
When the on-board classifier is not enough and the compute has to stay on site.
The same edge-versus-cloud split applied to language: what stays local and what is worth sending.
The household end of the same problem — glass, falls, dry-burning and pipe noise.
Chip capability [chip]: Rockchip RK3308 product page and datasheet — 8-channel PDM up to 192 kHz, 16–24 bit, 5-wire interface; 2 × 8-channel I2S/TDM at 48 kHz; on-chip CODEC with 8 × 24-bit ADC; hardware VAD accepting analog, I2S and PDM microphones with multi-microphone pre/post buffering. Aperture figures [physics]: c = 343 m/s, 8 elements, total length (N−1)d. Data rates [physics]: 8 channels × 24 bit at the stated rate, in binary gigabytes. Scenario bands [industry]: ranges in common use across vibration monitoring, acoustic emission and structural health monitoring. Figures describe platform capability and engineering method; they are not field measurements from a deployed array.
Send us the asset, the failure you care about and the reaction time. We will come back with the aperture, the sampling rate and the decision split.
Inquire