ML-based Condition Monitoring at the OT Edge

AI TECHNICAL BRIEF: We explore the realities of deploying ML at the OT Edge

Condition monitoring of rotating machinery has become the flagship use case for machine learning at the OT edge, and for a structural reason: sound and vibration are dense signals. A single sensor emits hundreds of kilobytes per second, and no plant can round-trip that volume to a cloud and expect real-time answers. Interpretation has to happen where the data is born.

Add the environments that define our practice — disconnected sites, DMZ-segmented networks, assets that must protect themselves when the uplink is down — and on-device inference stops being an optimization and becomes an architectural requirement.

The market has responded up and down the stack, from sub-dollar microcontrollers with neural accelerators to safety-rated gateway platforms to packaged smart sensors. This brief maps that stack in four layers. It also names the problem with the label itself: “edge AI” in vendor marketing spans at least three technically distinct things, and buyers who conflate them will procure the wrong one.

Three Things “Edge AI” Actually Means

Before the matrix, a taxonomy. When an OT vendor says its product does AI at the edge, it means one of the following:

•    Category A — Embedded signal processing plus rules. Deterministic feature extraction (FFT, envelope analysis, proprietary methods like Emerson's PeakVue) with threshold and rule logic on top. Mature, explainable, and — not incidentally — the only category that fits comfortably inside existing functional-safety frameworks. Frequently marketed as AI; it is not learned inference.

•    Category B — Genuine on-device model inference. A trained model (anomaly detector, classifier, small neural network) executing on the device itself: OtoSense at the edge, MAX78000-class sensor designs, the STM32N6 demo class. This is the category the silicon layer was built for, and the one that delivers true disconnected autonomy.

•    Category C — Cloud ML behind an edge sensor. The device collects and transmits; the learning and diagnosis happen in a remote service (SKF's diagnostic centres, Augury). Often the right commercial answer — cross-fleet training data is a real advantage — but it is not edge intelligence, and it fails exactly when the disconnected environment needs it most.

More often than not buyers who cannot distinguish these three categories will pay Category B prices for Category A capability or accept Category C dependencies in a Category B environment.

Layer 1 — Silicon: The NPU Becomes Standard Equipment

The defining trend of 2025–2026 is vertical integration: silicon vendors are building neural accelerators into mainstream microcontrollers and pairing them with proprietary toolchains that carry a model from training to deployment. Arm's Ethos-U cores now appear across ST, NXP, Renesas, and Infineon parts, while TI ships its own TinyEngine NPU. Disconnected inference at the sensor node is now a sub-dollar bill-of-materials decision, not an exotic capability.

Vendor	Offering	What it signals
STMicroelectronics	STM32N6 (Neural-ART NPU, 600 GOPS, Cortex-M55); STM32Cube.AI; NanoEdge AI Studio	First STM32 with an integrated NPU. Demonstrated vibration anomaly detection, vision, and audio classification running entirely locally, with no cloud connection. NanoEdge targets anomaly detection specifically.
Texas Instruments	MSPM0G5187 and AM13Ex MCUs with TinyEngine NPU	Edge inference at under $1 in volume; TI claims up to 90x lower latency and 120x lower energy per inference versus unaccelerated MCUs. Deep TUV-certified functional-safety pedigree elsewhere in the portfolio — but not for the ML layer.
NXP	i.MX RT crossover MCUs; eIQ ML development environment; EdgeLock security	eIQ spans the processor range and integrates with NXP SDKs; hardware root of trust via EdgeLock. Crossover class blurs the MCU/MPU boundary.
Renesas	RA8P1 (Cortex-M85 + Ethos-U55, 256 GOPS); RZ/V series (DRP-AI); Reality AI Tools	Reality AI is notably strong on vibration and acoustic signature analysis — the closest tooling fit for rotating-machinery work. RZ/V targets deterministic, low-power industrial vision.
Infineon	PSOC Edge (Cortex-M55 + Ethos-U55 microNPU); NVIDIA TAO integration	Vision-model fine-tuning workflow (TAO) brought down to MCU class — a cloud-trained, edge-deployed pattern made explicit.
Analog Devices / Maxim	MAX78000 AI microcontroller; Voyager4 wireless vibration reference design	The architectural pattern in miniature: the MCU wakes, runs inference locally, and fires the radio only when an anomaly is detected. ML as a power and bandwidth strategy, not just an analytics one.

Layer 2 — Safety-Adjacent Compute

Between the sensor node and the plant network sits a gateway class of compute, and it is here that the first serious functional-safety engineering appears.

Vendor	Offering	What it signals
NVIDIA	IGX Orin / IGX Thor industrial edge platforms	The outlier: hardware-based functional safety with ISO 26262 and IEC 61508 compliance, a dedicated safety MCU, real-time Linux, and extended lifecycle support. The safety case covers the platform — not the model you run on it. Aimed at robotics and human-machine collaboration, not process safety.
AMD / Xilinx	Kria K26 SoM; Zynq UltraScale+ MPSoC	FPGA fabric for deterministic inference pipelines — relevant wherever timing guarantees matter more than framework convenience.
Industrial PC vendors (MiTAC et al.)	Fanless, wide-temperature edge systems	Continuous FFT plus local inference mounted next to the machine — the gateway-class deployment pattern for brownfield sites.

Layer 3 — Tooling and Model Pipelines

Edge Impulse — acquired by Qualcomm in 2025 — remains the dominant cross-vendor TinyML pipeline: data collection, training, optimization, and deployment across heterogeneous targets. Against it stand the silicon vendors' own stacks (ST NanoEdge and Cube.AI, NXP eIQ, Renesas Reality AI, TI's model libraries), which compete by being free and vertically integrated with their hardware.

A model optimized for one vendor's NPU through one vendor's compiler does not port cleanly, and the more the pipeline automates, the less visible that coupling becomes until a redesign forces the question.

Layer 4 — Packaged Condition-Monitoring Products

At the product tier, the taxonomy of Section 2 does real work. The offerings below are routinely grouped under one “AI-powered monitoring” label; where the inference actually lives differs decisively.

Vendor	Offering	Where the inference lives
Analog Devices	OtoSense Smart Motor Sensor; OtoSense Insights	Edge-first by explicit design thesis: sound and vibration data are too dense to ship to a cloud for real-time analytics, so interpretation happens continuously at the edge. The clearest disconnected-operation story in the product tier.
Emerson	AMS Asset Monitor; AMS Wireless Vibration Monitor; PeakVue	Embedded auto-analytics at the edge device (CHARMs-based) flag common faults on pumps, fans, and gearboxes; the wireless monitor carries embedded prescriptive analytics. PeakVue is patented signal processing plus rules — Category A in the taxonomy — marketed under the AI banner.
SKF	Enlight Collect IMx-1	The sensor collects; AI/ML analysis happens in SKF's remote diagnostic centres, sold as a service wrapper around the bearing business. Cloud/service-centric — the instructive contrast to OtoSense.
Siemens	Sitrans MS200 / Sitrans Connexa; Industrial Edge	Platform play: edge runtime plus app ecosystem rather than a single smart sensor. Also an early NVIDIA IGX partner on the autonomous-factory vision.
Augury	Halo sensors + cloud diagnostics	Cloud-centric contrast case: the value proposition is the diagnostic service and its cross-fleet training data, not on-device autonomy.

The Missing Phase Reference

A structural omission separates the new edge-AI tier from classical machinery diagnostics: the phase reference. Order tracking, angle-domain resampling, synchronous averaging, and phase-angle discrimination — imbalance versus misalignment, for instance — all require a tachometer or keyphasor channel acquired coherently with the vibration channel: same timebase, matched group delay, deterministic latency. The tolerance is unforgiving. At 3,600 RPM one shaft revolution is 16.7 ms, so a single degree of phase corresponds to roughly 46 microseconds of timing error, and the budget tightens proportionally at higher orders.

The silicon is not the obstacle. Timer input capture and quadrature encoder peripherals are standard equipment on these MCUs, and a hardware-triggered ADC clocked from the same timer that captures tach edges can acquire a wideband analog channel coherently — TI's IEPE reference design, with four synchronous wideband channels read by Cortex-R5F MCUs through programmable real-time cores, demonstrates exactly that acquisition class. The obstacle is the architecture the TinyML tier has converged on: digital-output MEMS accelerometers running on internal sample-rate oscillators with percent-level tolerance, multi-MCU sensor nodes split across sleep/wake power domains, and a radio link between the measurement and any external reference. None of it preserves a coherent timebase between a tach pulse and an accelerometer sample.

The product tier reflects this. No TinyML-class wireless sensor surveyed here offers a phase reference input; speed, where addressed at all, is inferred — OtoSense estimates motor speed from magnetic sensing rather than a wired tach. That suffices for anomaly detection on fixed-speed machines, and it degrades precisely where order analysis matters most: variable-speed, VFD-driven equipment, where spectra smear without angle-domain resampling. The wired Category A tier — SKF IMx racks, Bently Nevada, Emerson AMS 6500 — retains keyphasor channels because API 670-class machinery protection demands them.

The edge-AI tier did not merely relocate the analysis. It quietly dropped the phase reference, and it is compensating with statistics.

For buyers, this is a capability regression masked by the category confusion of Section 2: replacing a wired or route-based program with smart wireless sensors can trade away order-domain and phase diagnostics without anyone deciding to. For vendors, it is an open design gap — a sensor node that hardware-timestamps a tach edge against a coherently clocked analog channel is feasible with today's parts, and no one is shipping it at this tier.

The Certification Path for Protective Action

At this writing, no vendor sells machine learning inside a protection function. Every product above is advisory — detect, alert, schedule. The moment the requirement becomes “take protective action,” the ML layer stops at the boundary of the SIL-rated logic and hands off to deterministic trip circuitry.

The reason is structural, not commercial timidity. IEC 61508 and its descendants were built for deterministic systems with enumerable failure modes; a statistical inference engine has neither property. The standards world is moving — ISO/IEC TR 5469 (2024) frames AI functional safety generically, ISO PAS 8800 (2024) does so for road vehicles, and proposals from groups like CARIAD add ML-specific lifecycle phases (prepare data, train model, deploy model) to close the gaps — but a technical report is not a certification path, and today there is none for an ML model executing inside a safety function. NVIDIA's IGX line is the closest scaffolding on the market, and its safety case deliberately covers the platform and its safety island, not the model.

This is the TCL3 asymmetry relocated from the test bench to the runtime: the qualified boundary encloses the deterministic execution layer, while the learned layer — the part generating the intelligence — sits outside it. And it compounds with the operational question this series raised in “The Patching Divide”: a model that drifts is a component that needs patching, except no CVE announces it, no vendor bulletin schedules it, and in a disconnected environment no telemetry may even reveal it. Who re-validates a fielded model, against what baseline, on whose liability? That question is, not the silicon — is where the next installment lives.

The story: everyone sells ML that advises; no one sells ML that trips. It is important to note that an ML enabled condition monitoring system can be architected that is integrated with a SIL certified device to perform analysis heavy lifting. What matters is that the ML inferences cannot be used to trigger safety faults - This needs to be the domain of the SIL certified safety function.