Condition Monitoring Architecture for Rotating Machinery Health

A bearing rarely fails without warning. Weeks before a motor seizes, the outer race develops a microscopic spall, and every time a rolling element passes over it the machine emits a faint, repetitive impact — a defect frequency buried under shaft imbalance, gear mesh, and shop-floor noise. A well-built condition monitoring architecture exists to hear that impact early, quantify it against an accepted severity scale, and turn it into a work order before the failure cascades into a burned winding or a wrecked gearbox. This is not machine learning magic. It is deterministic signal processing, decades of standardized vibration science, and a disciplined edge-to-cloud data pipeline working together.

This post is the practitioner’s blueprint for that system: how vibration and acoustic energy become features, how ISO standards convert those features into severity zones and health states, and how the whole chain is wired from sensor to CMMS.

What this covers: the four sensing techniques and the P-F curve, a layered sensor-to-cloud reference architecture, the DSP chain (FFT, envelope demodulation, order tracking, bearing defect frequencies), the ISO standards that govern thresholds and data processing (ISO 20816, ISO 13374, ISO 17359), the trade-offs that produce false alarms, and how this deterministic layer feeds — rather than competes with — the ML predictive-maintenance stack.

Context and Background

Condition monitoring rests on four complementary sensing techniques, each sensitive to a different failure physics. Vibration analysis is the workhorse for rotating machinery: it detects imbalance, misalignment, looseness, gear defects, and bearing faults through the mechanical energy they radiate. Acoustic emission (AE) listens at ultrasonic frequencies (typically 100 kHz to 1 MHz) for the transient elastic waves released by crack propagation, incipient bearing damage, and lubrication breakdown — often the earliest detectable symptom of all. Infrared thermography captures the heat signatures of friction, electrical resistance, and cooling faults. Oil analysis and wear-debris monitoring examine the lubricant itself, quantifying particle count, ferrous density, and contamination to reveal what surfaces are shedding metal.

These techniques map onto the P-F curve, the central mental model of the discipline. “P” is the point where a potential failure first becomes detectable; “F” is functional failure. The P-F interval — the warning window — is longest for oil analysis and AE (early chemistry and micro-cracking), shorter for vibration (developed mechanical defects), and shortest for temperature and audible noise (late-stage). The architectural goal is to detect as far left on the curve as economically justified, because a longer interval means calmer, cheaper, planned intervention. A defect caught at “P” is a scheduled bearing swap; the same defect caught near “F” is an unplanned outage.

For rotating assets, vibration dominates because it offers the best balance of P-F interval, cost, and diagnostic specificity — a single spectrum can distinguish imbalance from a bearing spall from a gear tooth crack. That specificity is what the rest of this architecture is built to preserve and exploit. The most robust programs do not treat the four techniques as rivals but as corroborating evidence: a rising bearing-band vibration trend confirmed by an AE burst and a ferrous-particle spike in the oil is a far stronger signal than any one channel alone, and fusing them at the health-assessment layer sharply reduces false alarms while extending the effective warning window. For the machine-learning layer that consumes these signals, see our companion guide to predictive maintenance IoT and machine learning architecture. For the physics of the P-F curve, the Reliabilityweb reliability body of knowledge remains a useful open reference.

The Reference Architecture

A condition monitoring architecture is a layered pipeline that moves raw mechanical energy through progressively more compact, more meaningful representations: continuous vibration becomes a spectrum, the spectrum becomes a set of features, features become a severity zone, and the zone becomes an alarm and a diagnosis. Each layer discards data the next layer does not need, which is what makes continuous monitoring of thousands of measurement points affordable.

Figure 1: Reference architecture layers — the machine, sensing and signal conditioning, the edge (acquisition, feature extraction, ISO threshold logic), and the platform (transport, time-series storage, diagnosis/prognosis, dashboards and CMMS).

Long description: A vertical flow. The rotating asset feeds accelerometers and a tachometer through a signal conditioner into an edge gateway, which performs analog-to-digital conversion, FFT and envelope feature extraction, and ISO 20816 zone evaluation. Results move over MQTT or OPC UA to a time-series store, then to a diagnosis and prognosis layer, and finally to dashboards, CMMS, and work orders.

The sensing and conditioning layer

Everything downstream inherits the quality of this layer, so it deserves disproportionate care. The primary transducer for rotating machinery is the piezoelectric IEPE accelerometer, chosen for its wide usable bandwidth (a few Hz to 10–15 kHz for general-purpose models) and rugged linearity. Sensor selection is a frequency-domain decision: bearing defects and gear mesh live in the kilohertz range, so a sensor with a mounted resonance safely above the highest frequency of interest is mandatory. A tachometer or encoder providing a once-per-revolution pulse is nearly as important as the accelerometer itself, because shaft speed is the reference against which every diagnostic frequency is calculated.

Signal conditioning — IEPE current excitation, anti-alias filtering, and gain staging — happens as close to the transducer as possible to preserve signal integrity. For safety-critical machines governed by API 670, this layer also includes permanently installed proximity probes measuring relative shaft displacement, feeding a machinery protection system that can trip the machine independent of the analytics stack.

The edge processing layer

The edge gateway is where a condition monitoring architecture earns its economics. A single channel sampled at 25.6 kHz generates roughly 50 KB/s of raw waveform; multiply by dozens of channels and continuous streaming to the cloud becomes untenable. The edge collapses that firehose by extracting features locally: overall RMS velocity, band-limited RMS, crest factor, kurtosis, and — critically — the FFT spectrum and the envelope spectrum. Instead of streaming megabytes of waveform, the gateway emits a few kilobytes of spectral peaks and scalar indicators per interval, with full waveform capture triggered only on alarm. This is the single most important design decision for bandwidth, cost, and scalability.

The edge also hosts the first tier of decision logic: comparing overall levels against ISO 20816 zone boundaries and raising a preliminary alarm state. This keeps time-critical alerting fast and resilient to network outages. For a deeper treatment of pushing inference to the edge, see our predictive maintenance edge AI reference architecture.

The transport and platform layer

Features and events travel north over an industrial transport. MQTT suits lightweight, event-driven telemetry — small payloads, publish/subscribe fan-out, and graceful behavior over intermittent links. OPC UA suits richer, model-driven integration where the semantic structure of the asset matters and interoperability with control systems is required; the two are frequently combined via MQTT with OPC UA payload encoding, giving you a lightweight wire protocol with a self-describing information model on top. That information model is more than a convenience: when the transport carries the asset hierarchy, the measurement point, the unit, and the ISO machine class alongside the value, the diagnosis layer no longer has to guess what a number means.

On the platform, a purpose-built time-series database stores the scalar trends and spectral snapshots that make long-horizon degradation visible. Storage design here is a deliberate two-tier decision. Continuously ingested scalars — overall RMS velocity, crest factor, kurtosis, bearing-band energy, temperature — are cheap to retain at full resolution for years and are what reveal a slow, months-long climb toward Zone C. Spectral and raw-waveform snapshots are far larger and are retained selectively: on a rolling window, on schedule, and on every alarm. Downsampling and retention policies keep the cost bounded while preserving the fidelity analysts need to confirm a diagnosis after the fact. Above storage sit the diagnosis, alarm, and prognosis services, and finally the human-facing surfaces: dashboards, alarm consoles, and the CMMS integration that converts a health verdict into a scheduled work order. For the transport semantics, our OPC UA companion specification implementation tutorial covers the information-modeling details.

Deeper Walk-through

The intelligence of a condition monitoring architecture lives in its signal processing chain and in the standards that give the results meaning. This section walks both: first the DSP that turns a waveform into a defect diagnosis, then the ISO framework that turns numbers into decisions.

The DSP chain: from waveform to defect frequency

Figure 2: The vibration DSP chain — anti-alias filtering and sampling, windowing and FFT, band-pass around a bearing resonance, Hilbert envelope demodulation, and matching envelope peaks to computed bearing defect frequencies, with an order-tracking branch for variable speed.

Long description: A left-to-right flow. Raw vibration passes through an anti-alias filter, is sampled at 2.56 times the maximum frequency, windowed with a Hanning window, and transformed via FFT. The spectrum is band-pass filtered around a structural resonance, envelope-demodulated with a Hilbert transform, and re-transformed; peaks are matched against BPFO, BPFI, BSF, and FTF. A parallel order-tracking branch handles variable speed.

Everything begins with correct sampling. The Nyquist criterion requires sampling above twice the highest frequency of interest, but practical analyzers sample at 2.56 times the maximum analysis frequency (Fmax) to leave room for the anti-alias filter’s roll-off. That anti-alias filter is not optional: without it, energy above Nyquist folds back into the spectrum as false peaks that mimic real defects. Sample first through the filter, then digitize.

Frequency resolution is the next design tension. The spectral line spacing equals the sampling frequency divided by the number of FFT points, or equivalently the reciprocal of the acquisition time. Finer resolution — needed to separate a bearing defect frequency from a nearby harmonic — demands a longer time record. Because the FFT assumes the captured block is one period of a repeating signal, any discontinuity at the block edges smears energy across the spectrum (leakage). A Hanning window tapers the block ends to suppress leakage and is the default for continuous vibration; a flat-top window trades resolution for amplitude accuracy when precise peak magnitude matters.

For bearing diagnosis, the raw spectrum is often the wrong place to look. A developing bearing defect produces low-energy, repetitive impacts that excite a high-frequency structural resonance of the bearing housing. In the ordinary velocity spectrum these impacts are drowned out by high-amplitude, low-frequency components from imbalance and misalignment. Envelope analysis (also called demodulation or high-frequency resonance technique) solves this: band-pass filter the signal around the excited resonance to isolate the impact-ringing, apply a Hilbert transform to extract the amplitude envelope, then FFT that envelope. The result is a clean spectrum whose peaks sit exactly at the bearing’s characteristic defect frequencies — the single most powerful technique in the rotating-machinery toolkit.

Those characteristic frequencies are computed from bearing geometry and shaft speed. For a bearing with N rolling elements, ball diameter Bd, pitch diameter Pd, contact angle φ, and shaft rotation frequency fr:

BPFO (Ball Pass Frequency Outer race) = (N / 2) × fr × (1 − (Bd / Pd) × cos φ)
BPFI (Ball Pass Frequency Inner race) = (N / 2) × fr × (1 + (Bd / Pd) × cos φ)
BSF (Ball Spin Frequency) = (Pd / 2Bd) × fr × (1 − ((Bd / Pd) × cos φ)²)
FTF (Fundamental Train Frequency, the cage) = (fr / 2) × (1 − (Bd / Pd) × cos φ)

A peak in the envelope spectrum at BPFO, with harmonics, is a textbook outer-race defect; energy at BPFI modulated by shaft speed points to an inner-race fault. Because these frequencies are non-integer multiples of running speed, they are unmistakable once you know where to look — which is precisely why the tachometer reference matters.

Order tracking for variable speed

Fixed-frequency FFT assumes constant shaft speed. On a variable-speed drive, a wind turbine, or a machine during run-up, speed drifts, so a defect frequency smears across many spectral lines and disappears into the noise floor — an effect called spectral smearing. Order tracking fixes this by resampling the waveform against the tachometer pulse so that samples are equally spaced in shaft angle rather than in time. Analysis then happens in the order domain, where “1x” is running speed regardless of rpm and every defect frequency becomes a fixed order. This keeps diagnostic peaks sharp across the entire speed range and is mandatory for any asset that does not run at a locked speed.

The standards layer: ISO 20816, 13374, and 17359

Signal processing produces numbers; standards make them actionable. Three ISO standards form the backbone of a credible machinery health monitoring program.

ISO 20816 governs the measurement and evaluation of machine vibration, consolidating and superseding the older ISO 10816 (broadband severity) and ISO 7919 series. Its central, widely used contribution is the evaluation-zone scheme, which classifies a machine’s overall vibration into four zones:

Zone A — vibration of newly commissioned machines; healthy.
Zone B — acceptable for unrestricted long-term operation.
Zone C — unsatisfactory for long-term continuous operation; the machine may run for a limited period while remediation is planned.
Zone D — severe enough to cause damage; unacceptable.

The A/B and C/D boundaries anchor the alarm and trip thresholds in most programs. The specific velocity values (typically evaluated as RMS velocity in mm/s over a 10–1000 Hz band) depend on the relevant part of ISO 20816 for the machine class — Part 1 gives general guidelines, with dedicated parts for large machines, industrial machine sets, gas turbines, hydraulic machines, and reciprocating units. Crucially, the standard measures overall severity; it tells you how bad, not what is wrong. Diagnosis of the specific fault comes from the spectral and envelope analysis described above.

Figure 3: The ISO 13374 data-processing functional blocks — data acquisition, data manipulation, state detection, health assessment, prognostic assessment, and advisory generation — with ISO 20816 thresholds informing state detection and health assessment.

Long description: A left-to-right chain of six functional blocks. A sensor feeds DA (data acquisition), which feeds DM (data manipulation), then SD (state detection), HA (health assessment), PA (prognostic assessment), and AG (advisory generation), ending at the operator and CMMS. ISO 20816 thresholds branch into state detection and feed health assessment.

ISO 13374 defines the reference model for condition monitoring and diagnostics data processing, and it maps almost one-to-one onto the architecture layers already described. Its six functional blocks are:

DA — Data Acquisition: converting sensor output to digital parameters (the sensing/ADC layer).
DM — Data Manipulation: signal processing and feature extraction (the FFT/envelope chain).
SD — State Detection: comparing features to baselines and thresholds to flag abnormality (where ISO 20816 zones live).
HA — Health Assessment: diagnosing the fault and its severity.
PA — Prognostic Assessment: projecting future health and estimating remaining useful life.
AG — Advisory Generation: recommending maintenance actions and integrating with the CMMS.

The value of ISO 13374 is that it makes the pipeline vendor-neutral and modular: a vibration analyzer, a diagnostic engine, and a prognostic model from three different suppliers can interoperate because each speaks in terms of these blocks. It cleanly separates detection (SD) from diagnosis (HA) from prediction (PA) — the exact seam where deterministic signal analysis hands off to the machine-learning prognosis layer.

ISO 17359 is the umbrella “general guidelines” standard that ties the family together. It prescribes the overall condition monitoring procedure: identify critical machines through criticality analysis, select appropriate measurement parameters and locations, establish baselines and alarm criteria, and define the review-and-improve loop. If ISO 20816 sets the thresholds and ISO 13374 structures the data flow, ISO 17359 is the program-level playbook that says which machines to monitor, with what, and how to close the loop. Together the three answer the three questions every program must: what do we monitor (17359), how do we process it (13374), and when do we act (20816).

Trade-offs, Gotchas, and What Goes Wrong

The failure modes of a condition monitoring architecture are rarely in the physics; they are in the mounting, the sampling choices, and the compute split.

Figure 4: Alarm-and-diagnosis decision flow — a new measurement is checked against the ISO zone threshold, then envelope and spectral evidence route it to a bearing, unbalance/misalignment, or analyst-review verdict before a work order is raised.

Long description: A top-down decision tree. A new RMS and spectrum reading is tested against the Zone C threshold. Below threshold, it is logged as a trend. Above threshold, envelope evidence of bearing frequencies yields a bearing-defect diagnosis; otherwise dominant 1x running-speed energy yields an unbalance or misalignment diagnosis, and anything ambiguous is flagged for analyst review. Confirmed faults raise an alarm and create a work order.

Sensor mounting is the number-one source of bad data. The mounting method sets the usable frequency ceiling: a stud-mounted accelerometer preserves response to 10 kHz and beyond, a magnetic mount typically collapses to 2–3 kHz, and a handheld probe may be trustworthy only to a few hundred Hz. Mounting a magnet-base sensor and then trusting a 5 kHz bearing frequency is a classic self-inflicted false negative. Worse, a poorly chosen mounting location can place the sensor at a structural resonance, amplifying an irrelevant frequency into a phantom alarm. Mount rigidly, near the load zone, in the correct axis, and characterize the mounted resonance.

False alarms erode trust faster than missed detections. Overall-level alarms alone are noisy — a transient load change, a resonance, or a speed excursion can breach a threshold without any defect present. The remedy is corroboration: require agreement between overall severity (ISO 20816), spectral pattern (defect frequency present), and trend (sustained rise, not a single spike) before escalating, as Figure 4 illustrates. A single-reading trip with no spectral evidence should route to analyst review, not straight to a work order.

Sampling and resolution mistakes are permanent. If Fmax is set too low, the bearing resonance band is never captured and envelope analysis is impossible. If frequency resolution is too coarse, adjacent defect and harmonic peaks merge and diagnosis becomes guesswork. These are acquisition-time decisions; no downstream analytics can recover a frequency that was never sampled.

The edge-versus-cloud split is a genuine trade-off, not a slogan. Push too little to the edge and you drown the network in waveform data and lose alerting during outages; push too much and you lose the raw signal fidelity that analysts and ML models need for novel faults. The durable pattern is hybrid: extract features and evaluate thresholds at the edge for fast, resilient alerting, but retain triggered raw-waveform snapshots and periodic high-resolution captures in the cloud for deep diagnosis and model training. Data volume is the constraint that shapes everything — design the compute split around it deliberately.

Practical Recommendations

Treat the deterministic signal-and-standards stack as the foundation, and build the ML layer on top of it rather than instead of it. Start with a criticality assessment per ISO 17359 so sensing spend follows consequence, not convenience. Fix the physical layer before the analytics: correct transducers, rigid mounting, and a reliable tachometer reference are non-negotiable, because no algorithm recovers a signal the sensor never captured. Extract features at the edge to keep bandwidth and cost sane, but preserve triggered raw waveforms so analysts and models retain ground truth. Anchor alarms in ISO 20816 zones, but require spectral corroboration before dispatching a technician. And keep the ISO 13374 blocks modular so the diagnosis and prognosis layers can evolve independently.

A practical starting checklist:

Rank assets by criticality (ISO 17359) and match monitoring intensity to consequence of failure.
Specify sensors by bandwidth: confirm mounted resonance sits well above your highest bearing/gear frequency of interest.
Install a speed reference (tacho or encoder) on every variable-speed asset; enable order tracking.
Set Fmax and resolution deliberately so bearing resonance bands and closely spaced peaks are both captured.
Extract features at the edge (RMS, crest, kurtosis, FFT, envelope) and stream those, not raw waveform, by default.
Calibrate alarm zones to ISO 20816 for the correct machine class, and require spectral + trend corroboration.
Capture triggered raw waveforms on alarm for diagnosis and to build labeled datasets for the ML layer.

Frequently Asked Questions

Is condition monitoring the same as predictive maintenance?

No — they are complementary layers. Condition monitoring is the deterministic sensing-and-standards architecture that measures machine health in real time using physics-based signal analysis and ISO thresholds. Predictive maintenance is the broader program that uses that health data — often with machine-learning models — to forecast remaining useful life and optimize intervention timing. Condition monitoring produces the trustworthy, labeled features; the ML layer consumes them for prognosis. Build the deterministic foundation first, because ML models trained on poorly acquired signals inherit every mounting and sampling error underneath them.

Why use envelope analysis instead of the normal vibration spectrum?

Because early bearing defects are low-energy, repetitive impacts that excite high-frequency housing resonances, and in the ordinary velocity spectrum they are buried under high-amplitude, low-frequency components from imbalance and misalignment. Envelope analysis band-pass filters around the excited resonance, demodulates the impact-ringing with a Hilbert transform, and transforms the result, producing a clean spectrum whose peaks land exactly on the bearing defect frequencies (BPFO, BPFI, BSF, FTF). It routinely detects bearing faults weeks earlier than broadband analysis.

What does ISO 20816 actually tell me?

ISO 20816 gives you severity, not diagnosis. It classifies overall machine vibration into four evaluation zones — A (new/healthy), B (acceptable long-term), C (unsatisfactory for continuous operation), and D (damaging) — with velocity thresholds that depend on the machine class and the relevant part of the standard. The A/B and C/D boundaries typically become your alarm and trip levels. It tells you how bad the vibration is; the spectral and envelope analysis tells you what is wrong. ISO 20816 supersedes the older ISO 10816 series.

How does ISO 13374 relate to my data pipeline?

ISO 13374 is the reference model your pipeline should mirror. Its six functional blocks — data acquisition, data manipulation, state detection, health assessment, prognostic assessment, and advisory generation — map directly onto sensing, feature extraction, thresholding, diagnosis, RUL prediction, and CMMS integration. Adopting it makes the architecture modular and vendor-neutral: components from different suppliers interoperate because each speaks in the same functional terms. It also marks the clean seam where deterministic analysis (state detection, health assessment) hands off to predictive modeling (prognostic assessment).

How much data should I send to the cloud?

By default, send features — not raw waveforms. A single channel of continuous raw vibration is roughly 50 KB/s; extracting RMS, crest factor, kurtosis, and spectral peaks at the edge collapses that to a few kilobytes per interval. Stream those features and ISO zone states continuously for trending and alerting, and capture full-resolution raw waveforms only when an alarm triggers or on a periodic schedule. This hybrid split keeps bandwidth and storage affordable while preserving the raw fidelity that deep diagnosis and ML model training require.

Do I need order tracking if my machine runs at constant speed?

Not necessarily. On a machine locked to a fixed speed, standard fixed-frequency FFT works well because defect frequencies stay put. Order tracking becomes essential when speed varies — variable-speed drives, wind turbines, or any analysis during run-up and coast-down — because speed drift smears defect frequencies across many spectral lines and hides them. Order tracking resamples the signal against the tachometer so analysis happens in the shaft-angle (order) domain, keeping every defect frequency sharp regardless of rpm.

Condition Monitoring and Machinery Health Architecture (2026)