One eight-microphone front end, seven kinds of site. Each case below answers the same eight questions — scope, failure mode, interference, array and sample rate, mounting, acceptance, output, and what is actually known versus inferred. The engineering derivation behind the numbers lives in the companion Bestom technical note, linked at the end.
These are engineering scenarios, not delivered projects. We have not instrumented any of these sites and no customer is named or implied. Every number here is fixed by four derivation inputs — the site's band, the required sample rate, the judgement timescale, and where the decision lands — and each item is tagged by source: [chip] component capability, [physics] derived from the speed of sound and array geometry, [industry] ranges in common use, [engineering] our inference. Where a number would only be meaningful after a site survey, we say so instead of inventing one.
The same eight-capsule front end appears in every case below. What changes is everything around it.
A bearing defect and a tower mode do not live in the same part of the spectrum. Aperture is chosen against the failure mode you are being paid to catch — which is why we ask what the maintenance team already suspects before we quote an array.
Frequency resolution is Δf = 1/T, so resolving a feature at frequency f requires observing for at least 1/f. A rockfall and a transformer trend therefore cannot share one analysis window. This is the single most common reason a monitoring pilot quietly stops working.
Heat, dust, washdown, vibration and the dominant interferer decide the enclosure, the mount and the cable run. Most pilots do not fail at the algorithm; they fail at month three, when the bracket has moved.
Each row is derived from the scenarios inside it — no row is specified independently.
| Case | Site | Band union | Peak sample rate | Judgement timescale | Longest analysis window |
|---|---|---|---|---|---|
| Case 1 · Industrial site | Rotating & Reciprocating Machinery | 10 Hz – 20 kHz | 48 kHz | 1 ms – 3600 s | 100 ms |
| Case 2 · Energy site | Wind Turbines, Towers & Transformers | 0.1 Hz – 10 kHz | 24 kHz | 60 s – 3600 s | 10 s |
| Case 3 · Livestock site | Livestock Houses: Fans, Animals & Feed Lines | 20 Hz – 10 kHz | 24 kHz | 1 s – 60 s | 50 ms |
| Case 4 · Geohazard site | Rivers & Valleys: Debris Flow and Rockfall | 10 Hz – 20 kHz | 48 kHz | 1 ms – 60 s | 100 ms |
| Case 5 · Infrastructure | Bridges, Roads, Rail, Cables & Buildings | 0.5 Hz – 20 kHz | 48 kHz | 1 ms – 60 s | 2 s |
| Case 6 · Indoor & home | Indoors: Glass, Falls, Dry-Boil & Pipes | 20 Hz – 50 kHz | 100 kHz | 1 ms – 60 s | 50 ms |
| Case 7 · Perimeter security | Airspace: Drone Intrusion | 100 Hz – 8 kHz | 24 kHz | 1 s – 1 s | 10 ms |
“Longest analysis window” is set by the lowest frequency in the group: at 0.1 Hz it reaches 10 s. That is 3,600,000× the shortest event on this page (1 ms), spanning 6.6 orders of magnitude.
Read down the sample-rate column and the window column together. They do not move in the same direction — and that is the whole reason this page is organised by site rather than by product.
Twenty-two scenarios in total, grouped by where they actually occur rather than by the algorithm they use.
Pumps, fans, motors, gearboxes and compressors on a running line. The array sits one to three metres from the asset, and the plant will not stop for you.
Pumps, fans, motors, gearboxes and compressors — continuously running process machinery. One monitoring point per asset, array one to three metres from the housing.
Bearing pitting and wear, gear-mesh sidebands, cavitation, and valve-surge pressure pulsation. All of them periodic, which is exactly what makes them catchable by a fixed array.
Neighbouring machines on the same frame, VFD switching noise, and plant reverberation. A T60 above one second smears impulsive content badly, and most machine halls are worse than that.
Eight capsules, peak sample rate 48 kHz, band union 10 Hz – 20 kHz. Aperture follows the defect frequency, not the size of the enclosure we happen to have.
Magnetic base or stud mount on a machined pad, inside a protective cage, out of the direct cooling-airflow path. No reflecting surface larger than 200 mm facing the array within 1.5 m.
Background-noise spectrum at the same point within ±2 dB across three cold starts, and the defect repetition rate reproducible run to run. If the baseline moves between starts, the installation is moving.
Repetition rate and trend leave the board over Modbus or MQTT into the SCADA the plant already runs. No new acquisition cabinet and no second historian.
Bands and sample rates are derived from rotation speed and defect frequencies [engineering]; repetition-rate magnitudes follow vibration-monitoring practice [industry]. No measured field data on this page.
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| Motor / pump — unbalance, misalignment, bearing | 10 Hz – 10 kHz | ≥ 24 kHz | hours | On device |
| Gearbox / reducer — mesh, tooth breakage | 100 Hz – 20 kHz | ≥ 48 kHz | hours | On device |
| Press / machine tool — impact, tool wear | 200 Hz – 20 kHz | ≥ 48 kHz | milliseconds | On device |
The slowest content on this list lives here. A tower mode near 0.1 Hz needs a 10 s window — which is why the same array cannot also catch a 20 kHz discharge burst.
Nacelle, tower, foundation and oil-immersed transformers — remote assets, frequently with nobody on site for weeks at a time.
Tower and blade modes below 1 Hz, bolt and weld looseness, transformer core and winding hum drift, and partial discharge in the ultrasonic band. Two very different physics in one case.
Wind-induced broadband noise that tracks gust speed, rain on the nacelle, and the yaw system's own mechanics. The interferer is not stationary, so a fixed threshold will not hold.
Eight capsules, peak sample rate 24 kHz, band union 0.1 Hz – 10 kHz. This case needs two parallel paths — a fast channel for ultrasonic discharge and a slow channel for sub-hertz modes. Of 3 scenarios, 2 are judged in the cloud.
IP-rated enclosure on the tower wall or in the nacelle, electrically isolated from the structure to avoid ground loops. Cabling must not bridge to a vibrating panel — that turns the cable into a microphone.
A gust-to-gust reference: the mode-frequency estimate must stay stable across wind conditions. If it tracks wind speed, you are measuring the wind rather than the tower.
Mode frequency and damping trend, plus looseness events, pushed to the fleet platform. The cloud does the cross-turbine comparison; the board only decides whether a sample is worth keeping.
The sub-hertz requirement comes directly from Δf = 1/T [physics]; discharge bands follow condition-monitoring practice [industry]. These are design targets, not field measurements.
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| Wind turbine — main bearing, blade, yaw | 0.1 Hz – 10 kHz | ≥ 24 kHz + slow channel | hours | Trigger + review |
| Tower / blade — modal and low-frequency response | 0.1 – 20 Hz | ≥ 100 Hz | minutes | Trigger + review |
| Transformer / reactor — abnormal hum | 50 Hz – 5 kHz | ≥ 12 kHz | hours | On device |
A livestock house is hot, humid and chemically aggressive, and the loudest broadband source in the building is the thing you are trying to look after.
Tunnel-ventilated houses: wall fans, feed lines, drinkers, and the animals themselves. Array on a column or truss, peak sample rate 24 kHz.
Fan bearing and belt degradation, an auger jamming or running dry, water-line hammer, and abnormal vocalisation as a welfare signal.
The animals are the dominant broadband source; humidity and ammonia attack every connector; and fan noise masks a failing bearing until it is quite far advanced.
Eight capsules, peak sample rate 24 kHz, band union 20 Hz – 10 kHz. Timescales run 1 s to 60 s — a distress call is an event, a fan trend is a slow drift.
IP65 minimum with a hydrophobic membrane, cable glands facing down, out of the washdown jet path. The enclosure must not become a perch or a rubbing point.
The system has to separate a failing fan from a merely dirty fan at the same duty point. That is a per-site threshold, so the acceptance test is a deliberate dirtying, not a datasheet number.
Alarm plus a one-line reason to the house controller; vocalisation statistics to the farm management system as a daily index. No audio leaves the site by default.
Fan signatures follow rotating-machinery practice [industry]. The vocalisation use is our inference from published acoustic livestock work and is the least settled item on this page [engineering].
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| Ventilation fan — bearing, belt, rotor lock | 20 Hz – 10 kHz | ≥ 24 kHz | minutes | On device |
| Animal vocalisation — cough, alarm, farrowing | 50 Hz – 8 kHz | ≥ 24 kHz | seconds | On device |
| Feed line / scraper — jam and abnormal impact | 20 Hz – 8 kHz | ≥ 24 kHz | minutes | On device |
This is the case where trigger and diagnosis genuinely cannot share one window: a rock impact lasts about 1 ms, but you need 10 ms of signal at 100 Hz to tell a rock from a passing truck.
Channel beds, gullies, and slope toes above roads, railways or villages. Unattended, solar powered, frequently no backhaul at all.
Debris-flow surge arrival, individual rockfall impacts, and the change in bedload noise that precedes a surge. The useful part is the onset — by definition the first few hundred milliseconds.
Rain, wind in vegetation, wildlife, and road traffic in the valley below. The river's own continuous flow noise sets a broadband floor that rises with discharge.
Eight capsules, peak sample rate 48 kHz, band union 10 Hz – 20 kHz. Rockfall is one of 4 scenarios here where the required analysis window is longer than the event itself, so the trigger and the diagnosis must be separate windows.
Rock-anchored mast or boulder-mounted plate, set above the expected surge height. The cable run is the weak point on this kind of site — bury it or armour it.
A deliberate drop test at a known distance, plus a recorded passage of a known vehicle, must separate under one threshold. Without both halves, a false-positive rate is unknowable and any quoted accuracy is meaningless.
An event record with onset timestamp and a one-line classification, pushed over a low-bandwidth link. Raw audio stays on site by default because the link cannot carry it and the privacy case does not need it.
The two-window requirement follows from Δf = 1/T [physics]; debris-flow acoustic bands follow published geohazard monitoring practice [industry]. The site has not been instrumented.
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| River acoustics — discharge and turbulence shift | 20 Hz – 5 kHz | ≥ 12 kHz | minutes | On device |
| Debris flow / flash flood — low rumble plus sustained scour | 10 Hz – 2 kHz | ≥ 6 kHz | minutes | Device + cloud |
| Rockfall / slope collapse — impact transient and rolling rhythm | 100 Hz – 20 kHz | ≥ 48 kHz | milliseconds | On device |
Five scenarios in one case. The common thread is that the source moves past a stationary array, so the array must answer what and how severe without ever knowing exactly where.
Bridge decks and bearings, road pavements, rail track and joints, stay cables and hangers, building facades and glazing. Fixed array beside the structure.
Bearing and expansion-joint degradation, delamination and pothole growth, rail joint and wheel-flat impact, cable damping loss, and facade panel or glass looseness.
Traffic is both the excitation and the interferer. Wind on cables, rain on the deck and temperature-driven geometry changes all move the baseline over a day.
Eight capsules, peak sample rate 48 kHz, band union 0.5 Hz – 20 kHz, timescales 1 ms to 60 s. 5 scenarios, of which 3 are judged entirely on the board.
Permanent bracket or adhesive pad on a non-structural face with a strain-relief loop. On stay cables the mount must not add damping — adding damping changes the quantity you are trying to measure.
Repeatability across a known vehicle class. One passage at a known axle load has to produce the same features within a stated tolerance; if it does not, the mount is drifting or the reference is wrong.
Event records joined to axle counts or train timetables. The board classifies; the cloud correlates with traffic, weather and temperature.
Bands and windows follow structural-health-monitoring practice and Δf = 1/T [industry][physics]. No instrumented structure is referenced or implied.
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| Bridge — vehicle passage response, expansion-joint impact | 1 Hz – 2 kHz | ≥ 6 kHz | seconds | Device + cloud |
| Pavement — tyre noise as a proxy for damage and ponding | 200 Hz – 5 kHz | ≥ 12 kHz | seconds | On device |
| Rail track — wheel/rail noise, joints, corrugation, switches | 100 Hz – 20 kHz | ≥ 48 kHz | milliseconds | On device |
| Rope and stay cable — tension, wire break, rain-wind vibration | 0.5 Hz – 5 kHz | ≥ 12 kHz | minutes | Trigger + review |
| Building — curtain wall, structural noise, lifts, plumbing | 50 Hz – 10 kHz | ≥ 24 kHz | seconds | On device |
The only case here that reaches 50 kHz, and the only one where the people being monitored never asked for a microphone. Both facts shape the delivery.
Hotel rooms, apartments, care facilities and plant rooms. Glass breakage, human falls, unattended cooking, and concealed pipe leaks.
Glass fracture — a broadband burst extending above 20 kHz — human fall impact, dry-boil and overheated-cookware signatures, and pressurised pipe leak noise.
Speech, television, doors and HVAC occupy the same band as the events. This is the hardest privacy case on the page, so the design assumes the audio never leaves the room — and should not.
Eight capsules, peak sample rate 100 kHz, band union 20 Hz – 50 kHz — the widest band on this page. Timescales 1 ms to 60 s, all 4 scenarios judged on the board.
Ceiling or high wall corner for falls and glass, near the pipe run for leaks, clear of air vents. In care settings the unit has to be visually neutral rather than camera-shaped.
False alarms under ordinary daily activity must be demonstrated rather than asserted. The acceptance test is a multi-day soak with an activity log kept alongside.
A local alarm and a coarse event class only: band-limited features plus a label. Raw audio is not an output of this product.
Glass and fall bands follow established detection practice [industry]. Pipe-leak and dry-boil signatures are our engineering inferences and need per-site thresholding [engineering].
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| Glass breakage | 2 – 8 kHz | ≥ 24 kHz | milliseconds | On device |
| Fall — impact plus vibration | 20 Hz – 2 kHz | ≥ 8 kHz | milliseconds | On device |
| Hob left dry-burning — whistle, vaporisation, empty pan | 1 – 16 kHz | ≥ 40 kHz | seconds | On device |
| Pipe — leak, cavitation, blockage | 100 Hz – 50 kHz | ≥ 100 kHz | minutes | On device |
One scenario, and the narrowest specification on the page — which is precisely why it is the easiest one to ruin by over-specifying it.
Perimeter and rooftop monitoring against small multirotor intrusion. Fixed mast position with an unobstructed sky sector.
Approach, hover and departure of a small multirotor, identified by its harmonic blade-passage signature rather than by level. The signature is what separates it from a distant aircraft or a leaf blower.
Wind in foliage and structures, rooftop HVAC and road traffic — all broadband and all in the same band. Bearing estimation rather than loudness is what carries the decision.
Eight capsules, peak sample rate 24 kHz, band union 100 Hz – 8 kHz, judgement timescale 1 s: the narrowest specification on this page.
A rigid mast rather than a wall bracket, and the mast must not sway in the same band as the blade passage. Wind noise across the capsule faces is the limiting factor, so the windscreen has to be rated for the site.
A controlled approach from each sector at a stated altitude, with wind and traffic false alarms measured at the same time. Any detection distance we quote has to be quoted with the wind condition it was measured in.
Bearing plus a confidence value to the security system. Recording is kept only on a positive class decision, and retention is a policy the customer sets — not a product default.
Blade-passage bands follow published counter-UAS acoustic work [industry]; the geometry is our own derivation [engineering]. Detection range is strongly site-dependent and is not claimed here.
| Scenario | Band | Sample rate | Judgement timescale | Where the decision lands |
|---|---|---|---|---|
| Drone intrusion — blade-pass frequency and harmonics | 100 Hz – 8 kHz | ≥ 24 kHz | seconds | On device |
In 4 of 22 scenarios the window needed to resolve the frequency is longer than the event being detected. Triggering and diagnosing therefore have to be two different operations.
| Scenario | Event scale | Required window | What we deliver |
|---|---|---|---|
| Press / machine tool — impact, tool wear | 1 ms | 5 ms | trigger window fires, analysis window classifies the impact |
| Rockfall / slope collapse — impact transient and rolling rhythm | 1 ms | 10 ms | trigger window fires, analysis window examines the onset spectrum |
| Rail track — wheel/rail noise, joints, corrugation, switches | 1 ms | 10 ms | trigger window fires, analysis window separates joint from wheel defect |
| Fall — impact plus vibration | 1 ms | 50 ms | trigger window fires, analysis window removes the speech band |
The delivered firmware therefore runs two windows per site. The pre-trigger buffer is not an optimisation to save memory — without it the onset, which is the diagnostic part, is already gone by the time the threshold is crossed.
Largest gap between the two: 50×. That is not a tuning problem; it is a consequence of Δf = 1/T.
A quotation that names only the array is not a specification. Each of these is delivered per site.
| # | Item | Delivered per site |
|---|---|---|
| 1 | Eight-capsule front end | Array, enclosure rated for the site, and the mounting hardware for the surface actually present. |
| 2 | Edge firmware | Calibration, beamforming, feature extraction, per-user decision logic and event recording on the board. |
| 3 | Per-site parameter sheet | Band, sample rate, window lengths, thresholds and where each decision lands — written down, not left in the firmware. |
| 4 | Integration interface | Modbus, MQTT or HTTP against the system the site already runs. No parallel dashboard required. |
| 5 | Acceptance test plan | Background-noise repeatability, deliberate triggers, and a soak period with an activity log. |
| 6 | Data contract and retention | What leaves the device, what never does, and for how long either is kept. |
These are the tests we are willing to be held to. If they cannot be run, the site is not ready for a pilot.
| # | Test | Pass condition |
|---|---|---|
| 1 | Background repeatability | Same point, three cold starts, spectrum within ±2 dB. A drifting baseline invalidates every later comparison. |
| 2 | Deliberate triggers separate | A known event and the known dominant interferer must fall on opposite sides of one threshold. |
| 3 | No slow drift | Over the soak period the baseline and threshold must not drift monotonically. Drift has to be attributable to environment, not to the mount. |
| 4 | False alarms accounted for | Every false alarm in the soak window has a matching entry in the activity log. |
| 5 | Data contract holds | Traffic upstream matches the agreed contract, and raw audio demonstrably has not left the device unless the customer turned it on. |
Stated plainly, because a solution page that lists only capabilities is not usable for a decision.
We do not quote detection rate or detection distance without a site measurement. Where a figure depends on wind, ground or background, we quote it with the condition attached or not at all.
We do not claim one array covers unrelated site types. Bands can coexist on one board; observation windows cannot, and the window is set by the event.
We do not stream raw audio upstream to buy accuracy. It costs bandwidth twice — once on the link, once in the privacy review.
We do not treat “it is installed” as “it still works”. Threshold stability and mount stability are separate acceptance items, not assumptions.
No. They are engineering scenarios, and we say so at the top of the page. We have not instrumented these sites, no customer is named, and the numbers are design targets rather than measurements.
One array can share a band across sites — that is what the aperture table in the companion note shows. It cannot share an observation window, because the window follows the event, and events here span 6.6 orders of magnitude.
No. Eight is where the band and the spatial filtering both become useful for these cases. Fewer elements means either a narrower band or no ability to place a null on the dominant interferer.
Because a detection rate without a measured background is not a property of the product. It is a property of one installation on one afternoon. We would rather quote the method than a number we cannot stand behind.
Not by default. The board emits features and event records. Raw audio stays on site; where a cloud component is needed — fleet comparison for towers, traffic correlation for bridges — it consumes features.
Same data source, different lens. The Bestom note derives the numbers and lists all 22 scenarios as a specification grid. This page asks what a site actually looks like, what gets delivered, and how we would prove it works.
The companion Bestom note: how aperture, sampling rate and observation window fix the specification, and where each of the 22 scenarios sits.
ENThe Bestom case set: the same seven sites described on site terms — what is there, what to listen for, and what interferes.
ENThe delivered eight-microphone acoustic diagnostics front end, with the aperture and data-rate figures for each configuration.
The specification starts from what your maintenance team already suspects, not from our catalogue. Send the asset type, the fault you are trying to catch earlier, and the reaction time you need.
Discuss a siteWe will tell you which of the seven cases it resembles, what the band and window requirement implies, and where a site survey is unavoidable.
Component capability [chip] — Rockchip RK3308 product page and datasheet: 8-channel PDM, up to 192 kHz per channel at 16–24 bit over a 5-wire interface; 2 × 8-channel I2S/TDM at 48 kHz; embedded CODEC with 8 × 24 bit ADC; hardware VAD with pre- and post-detection buffering. Derived [physics] — aperture from c = 343 m/s, and observation windows from Δf = 1/T, both computed locally in this page's build. Industry ranges [industry] — vibration monitoring, acoustic emission, structural health monitoring and counter-UAS, used as engineering guidance rather than as measured values. Inferences [engineering] — everything stated as a design target, a mounting rule or an acceptance criterion. No site on this page has been instrumented and no measured deployment data is included.