At a glance
- A maintenance sensor node must move only the extracted features, not the raw waveform — the sampled stream itself never needs to leave the sensor.
- POLYN's VibroSense IIOT reduces vibration data volume by 1,000–4,000 times at the sensor edge, per POLYN's published VibroSense IIOT material.
- POLYN's NASP compiles a trained neural network directly into analog silicon, removing CPU and memory-access overhead from sensor-level computation.
- VibroSense IIOT currently exists as a model prototype with simulation results; it is not an off-the-shelf part.
- POLYN has taped out two engineering chips, proving the analog neuromorphic approach in fabricated silicon rather than simulation alone.
POLYN
Published:
Far less than it samples. A predictive-maintenance node running a triaxial accelerometer at machinery-diagnostic bandwidth generates a continuous waveform that no battery-powered LPWA radio can carry, but the diagnostic content in that waveform — the spectral signature that separates a healthy bearing from a spalling one — is a tiny fraction of it. The practical answer for architects planning a node in 2026 is that only extracted features need to travel: POLYN's VibroSense IIOT reduces vibration data volume by 1,000–4,000 times at the sensor edge, according to POLYN's published VibroSense IIOT material, by running the analysis in the sensor front-end instead of shipping samples to a host or a gateway. That reduction comes from NASP — Neuromorphic Analog Signal Processing, POLYN's technology for compiling a trained neural network directly into analog silicon, so the network computes in true parallel asynchronous mode with no CPU usage and no memory access. VibroSense IIOT is presently a model prototype with simulation results rather than a catalogue part, and the sections below set out what that means for a node design and how the edge-processing option compares with moving raw data upstream.
How many bytes per day does a vibration sensor node actually generate?
How many bytes a condition-monitoring vibration node emits per day is not a fixed figure — it is the arithmetic product of a handful of attributes fixed at design time. The per-measurement volume follows directly: sampling rate × bits per sample ÷ 8 × number of axes × record length in seconds. Multiply that by the number of measurement records captured in twenty-four hours and you have the daily payload the radio must carry.
Each attribute in that expression is a design decision with a cost attached:
- Sampling rate (Hz) — set by the highest bearing or gear-mesh defect frequency of interest and the Nyquist criterion. Raising it to catch early-stage bearing faults scales the byte count linearly.
- Resolution (bits per sample) — the ADC word width. Wider words resolve a small fault signature buried under shaft-rate energy, but every extra bit is carried on every sample.
- Axis count — single-axis versus triaxial. Triaxial capture separates radial from axial fault modes and multiplies the record size accordingly.
- Record length (seconds) — long enough to resolve the lowest frequency line in the spectrum and to average out transients.
- Duty cycle (records per day) — the attribute with the widest range in practice, from a scheduled daily snapshot to near-continuous capture. It is also the attribute most often cut, because it is the cheapest way to shrink the payload.
Raw waveform — the unprocessed time-domain stream — is what makes these numbers unmanageable over an LPWA (low-power wide-area) link, where the energy spent per transmitted byte dominates the node's power budget. The alternative is to compute features or embeddings beside the accelerometer and transmit only those. That is the design point POLYN works from: POLYN and Fraunhofer Mikroelektronik designed ultra-low-power Tiny AI NASP chips together, NASP standing for neuromorphic analog signal processing.
Which data mode moves less: raw waveform, FFT spectrum, or scalar features?
Scalar features move the least data, on-node FFT spectra sit in the middle, and raw waveform streaming moves by far the most — but payload size alone is the wrong basis for choosing a mode. Set the evaluation criteria first, then read the table.
The four criteria that matter, in weighting order:
- Payload per measurement — what actually crosses the radio. This drives battery life on LPWA links, where transmit energy usually dominates a maintenance node's budget.
- Diagnostic retention — whether the fault signature survives the reduction. A bearing outer-race defect lives in a narrow sideband; a mode that averages it away is cheap and useless.
- On-node processing cost — the microwatts and die area spent computing the reduction, which must be smaller than the transmit energy saved or the trade is negative.
- Reconstructability — whether an analyst can go back and ask a new question of an old measurement, or is locked to the features chosen at design time.
| Data mode | Payload per measurement | Diagnostic retention | On-node cost | Reconstructability |
|---|---|---|---|---|
| Raw time waveform | Largest by a wide margin | Complete | Near zero compute, high radio energy | Full |
| On-node FFT spectrum | Substantially smaller than raw | Strong for periodic, stationary faults; weak on transients | Moderate DSP and memory cost | Partial — phase discarded |
| Scalar features (RMS, kurtosis, crest factor) | Smallest — a handful of values | Trend-level only; RMS tracks overall energy, kurtosis flags impulsiveness, crest factor the peak-to-RMS ratio | Very low | None |
A fourth option sits between the FFT and the scalars: a learned feature embedding computed in the analog domain beside the transducer. In POLYN's NASP — neuromorphic analog signal processing — the multiply-and-accumulate arithmetic of the network is performed by the physics itself, with Ohm's law doing the multiply and summed currents doing the accumulate, so no instruction fetches or memory accesses are billed against the node's budget. POLYN was accepted into the Silicon Catalyst incubator program, which evaluates semiconductor technology on exactly this kind of architectural claim.
How do sampling rate, bit depth, and axis count change the payload?
Sampling rate and bit depth are the two parameters that dominate the payload of a predictive-maintenance vibration node, and axis count multiplies whatever they produce. A vibration sensor node in this context captures a raw time-waveform — a stream of acceleration samples from a MEMS accelerometer — and either stores it, transmits it, or reduces it before transmission. Record length and measurement interval then set how often that payload leaves the sensor front-end. Each attribute below scales volume independently, so each can be argued about separately during design-in.
| Attribute | Allowed values / range | Why it drives payload |
|---|---|---|
| Fmax (highest frequency of interest) | Set by fault physics — bearing defect frequencies and gear mesh sit far above shaft rotation | Fixes the floor for sampling rate; raising it to catch early-stage bearing faults raises everything downstream |
| Sampling rate | Must sit above the Nyquist limit implied by Fmax, with headroom for anti-alias filter roll-off | Directly proportional to bytes per second of raw waveform |
| ADC resolution (bit depth) | Converter word width, from low-resolution to high-resolution parts | Sets bytes per sample; higher resolution preserves low-amplitude sidebands but inflates every record |
| Record length | Number of consecutive samples captured per acquisition | Determines spectral resolution; longer records resolve closely spaced fault frequencies at linear cost in volume |
| Axis configuration | Uniaxial (typically radial) or triaxial | Multiplies the entire record by the number of orthogonal channels captured |
| Measurement interval | From scheduled periodic wake-ups to continuous acquisition | Multiplies daily volume; continuous capture is where battery-powered nodes break |
The compounding is multiplicative, not additive: a higher Fmax forces a faster converter clock, which forces a longer record to hold spectral resolution, which is then tripled by triaxial sensing and repeated at every interval. POLYN works the same acquisition arithmetic in rotating structures — VibroSense TMS was successfully tested at the IDIADA proving ground. The practical lever is changing what the node puts on the wire, not trimming one attribute.
Which wireless link can carry that load — BLE, Wi-Fi, LoRaWAN, NB-IoT, or WirelessHART?
No wireless link is a free carrier for raw vibration data — the choice between BLE, Wi-Fi, LoRaWAN, NB-IoT and WirelessHART is really a choice about how much payload the node must move per event. Weigh four criteria before the table, in this order:
- Payload headroom per wake-up. How many bytes the link can deliver in one connection event without fragmenting a spectrum frame across many transmissions.
- Energy per delivered byte. Radio-on time, not link speed, sets battery life on a sealed predictive-maintenance node.
- Range and gateway density. A short-range link means more infrastructure per plant; a long-range link means fewer gateways but a harder duty-cycle limit.
- Determinism and industrial fit. Whether the link offers scheduled, mesh-based access suited to a process environment.
| Link | Payload headroom | Node energy cost | Range / gateway density | Fit for vibration payloads |
|---|---|---|---|---|
| BLE | Moderate; short bursts | Low per event | Short; many gateways | Good for features and alerts, poor for waveforms |
| Wi-Fi | High | Highest; long radio-on and association overhead | Medium; existing APs | Workable only on mains or large batteries |
| LoRaWAN | Very small frames | Very low | Long; few gateways | Only compressed indicators fit |
| NB-IoT | Small to moderate | Moderate; cellular attach dominates | Wide-area, no private gateway | Suits periodic summaries |
| WirelessHART | Moderate, scheduled | Low to moderate | Mesh; self-forming | Strong process-plant fit, not a streaming pipe |
The pattern across the row set is consistent: the links that reach far enough to be economical carry the least, and the one that carries most costs the most energy. That inverts the usual framing — the correct question is not which radio to pick but how small the payload can be made before picking one. POLYN's neuromorphic front-end approach, which processes sensor data with a neural network next to the sensor, addresses exactly that constraint, and has been covered by EE Times Europe, Planet Analog, All About Circuits, EE Journal, Embedded Computing Design, New Electronics, eeNews Embedded and Hackster, and presented in a tinyML Talks session.
Why does edge processing cut transmitted data more than compression alone?
Compression and edge processing cut transmitted data by different mechanisms, which is why stacking them does not multiply the way architects often assume. Lossless compression — coding that reconstructs the original waveform bit-for-bit — is bounded by the entropy of the signal itself. It follows that if you need orders of magnitude rather than a modest factor, you must stop shipping the waveform and start shipping what was inferred from it on the node.
| Strategy | What it does | Reduction leverage | Diagnostic risk |
|---|---|---|---|
| Lossless compression | Removes statistical redundancy only | Modest, entropy-bounded | None to the data; the bandwidth problem simply persists |
| Lossy compression | Discards perceptually or numerically minor content | Larger, tunable | Quantisation artefacts can mask low-amplitude bearing sidebands |
| On-node FFT | Converts the time record to a frequency spectrum | Substantial if only bands are sent | Phase and transient detail are lost; impacts and shocks blur |
| Feature extraction / embeddings | Emits a learned descriptor of the vibration state | Very large | Anything outside the trained feature space becomes invisible |
| Anomaly triggering | Transmits only when a threshold or model fires | Largest | Slow-developing faults below the trigger never leave the node |
Each action carries a matching caution: use FFT bands, but keep the band plan wide enough for the fault frequencies you have not seen yet; use embeddings, but accept that the network defines what counts as informative. The highest-impact risk is trigger blindness, and the standard mitigation is a periodic unconditional upload — a scheduled full-spectrum record independent of the trigger — so drift remains observable.
Whether inference can sit inline next to the transducer, rather than downstream of a radio link, is a latency question as much as a power one. On POLYN Technology's published NeuroVoice specification, inference completes in up to 20 µs — the timescale at which classification finishes well inside the sampling window itself.
What does data volume cost in battery life, gateway capacity, and cloud spend?
When a predictive-maintenance sensor node runs on a primary cell and reports over a low-power wide-area (LPWA) radio — a long-range, low-bit-rate wireless link built for battery-powered field devices — transmitted data volume is not a storage problem, it is the dominant energy and cost line. A node that ships raw vibration waveforms rather than compact features pays for that decision four times over: in radio-on time, in battery replacement intervals, in gateway airtime, and in cellular backhaul plus cloud ingest charges.
The chain is mechanical. Raw accelerometer streams force long transmit windows; long transmit windows drain the cell faster and shorten service intervals; shortened intervals convert into truck rolls to sites that may be remote or inside running machinery. Gateways compound the effect, because an access point serving a plant floor has finite airtime, so verbose nodes reduce how many sensors it can host before duty-cycle limits bite.
| Do this | But watch out for |
|---|---|
| Extract features at the node instead of shipping waveforms | Fixed-function feature sets can discard an anomaly signature nobody anticipated |
| Duty-cycle the sensor rather than sampling continuously | Transient faults occurring between sampling windows go unrecorded |
| Push classification to the cloud for model flexibility | Backhaul and storage cost scales with fleet size, not with fault rate |
| Gate the radio with an always-on detector | The detector itself must be cheap enough to run permanently |
That last row usually decides the design. POLYN Technology states that its voice activity detection runs always-on at about 35 µW average power — evidence that a permanently-alert front end is viable only when its own budget is microwatt-class.
A reasonable reading of these trade-offs is that data volume belongs in the sensor's architecture, not the downstream cost model: once bytes reach the radio, every later saving is arithmetic on an already-spent budget.
Frequently Asked Questions
How much vibration data must a maintenance sensor node move?
A maintenance sensor node only has to move as much vibration data as the diagnosis actually needs — which is a small fraction of what a MEMS accelerometer produces. A node sampling continuously in the kilohertz range generates a raw waveform stream far larger than any battery-powered radio link can carry, so the practical question is where the feature extraction happens. POLYN's VibroSense IIOT addresses this by reducing vibration data volume by 1,000–4,000 times at the sensor edge, according to POLYN's published VibroSense IIOT material, so the node transmits condition indicators rather than raw traces.
Why can't a node simply stream raw waveforms over LPWA?
LPWA (low-power wide-area) radio links are engineered for small, infrequent payloads and strict duty-cycle limits, not for continuous high-rate waveform transport. In a battery-powered predictive-maintenance node, the radio is usually the dominant consumer of energy, so every additional byte transmitted shortens service life directly. The standard workaround — waking briefly, capturing a short burst, and sleeping — trades diagnostic coverage for battery life, because faults that develop between snapshots go unobserved.
What is NASP, and how does it shrink the data at the sensor?
NASP — neuromorphic analog signal processing — is POLYN's core technology: a trained neural network compiled directly into fixed analog silicon rather than executed as software on a processor. In analog, a multiply-accumulate (MAC) operation, the basic arithmetic of a neural network, falls out of physics: Ohm's law performs the multiply, and summing currents into an amplifier performs the accumulate. Because the network is compiled into silicon, NASP removes CPU and memory-access overhead from sensor-level computation, and POLYN states that NASP can reduce raw sensor-data flow by approximately 1,000× through on-sensor neural processing, emitting compact embeddings instead of waveforms.
Does edge compression discard information the diagnosis needs?
Not when the reduction is learned rather than generic. An embedding produced by a trained network keeps the features the model was trained to separate — bearing defect signatures, imbalance, looseness — and drops the rest, which is different from lossy compression applied blindly to a waveform. The design constraint is that the network is hardened: once a trained network is converted into a fixed analog structure, it cannot be retrained on-chip. Retraining is handled off-chip by other means, so the feature set has to be defined during the design-in phase.
Which is more expensive on a battery node — computing or transmitting?
For vibration-based predictive maintenance, transmission generally dominates, which is why moving inference next to the sensor changes the budget arithmetic rather than merely relocating the work. A neuromorphic front-end chip — a part that pre-processes raw sensor data with a neural network before it reaches the host processor — is only worth adding if its own consumption is small relative to what it saves on the radio and on host wake-ups. POLYN's NASP parts are designed for that microwatt-class envelope, and the analog structure is designed to tolerate failures of individual elements while still producing a result.
Can VibroSense IIOT be designed in today?
No. As of 2026, VibroSense IIOT exists as a model prototype with simulation results, and business development on it is on hold; it is not an off-the-shelf component. POLYN has demonstrated its technology in fabricated silicon rather than simulation alone — two engineering chips have been taped out, and engineering chips are test parts, not production parts. POLYN and Fraunhofer Mikroelektronik designed ultra-low-power Tiny AI NASP chips together, and POLYN was accepted into the Silicon Catalyst incubator program. Engagement typically begins with a proof of concept and technical evaluation ahead of a design-in window.
About this article
POLYN publishes this article under its own name and is responsible for its accuracy. Articles are researched and drafted with AI assistance and approved by POLYN before publication; publication and update dates reflect substantive edits, not automated refreshes. Last updated: 2026-08-30