Predictive maintenance in manufacturing means measuring a machine’s condition and scheduling the repair when the measurements show degradation, not on a calendar and not after a breakdown. Whether it pays off is decided before any machine learning starts. You rank machines by criticality, find each machine’s dominant failure modes, choose sensors that can see those modes early, and use years of downtime tickets as the evidence for all of it. If that front end is wrong, the model has nothing predictable to predict. The front end is also where the case for when AI-driven anomaly detection earns its cost is won or lost.
What separates preventive, predictive and corrective maintenance on a production line?
Machinery maintenance in manufacturing has three primary approaches. Reactive maintenance runs a machine until it breaks. Preventive maintenance is scheduled by time or cycles. Predictive maintenance is based on data and observations. The weak point of the calendar is that preventive intervals rely on an average lifetime, so a given machine can be serviced too early or fail before its next interval.
In the AMS 100-34 analysis of manufacturing establishments, the average maintenance mix was 17.3% predictive, 31.8% preventive and 45.7% reactive. The quarter of establishments that relied most on reactive maintenance were associated with 3.3 times more downtime. Among establishments that relied mainly on preventive and predictive maintenance, predictive maintenance was associated with 15% less downtime and an 87% lower defect rate. These are associations across many plants, not a forecast for yours, and we would treat them as a reason to look rather than a projected saving.
Computer-aided methods earn their place when the warning sign is spread across several variables at once. Those are correlations that manual inspection struggles to see in multivariate telemetry. It is not free. A 2021 cost-oriented model in the International Journal of Production Economics describes ML-based predictive maintenance as difficult and expensive to implement, especially for companies short on skills and budget, and argues for a cost analysis to decide where it suits. A survey of 150 predictive maintenance papers adds that high upfront cost and complexity make small and medium firms unlikely to adopt it.
Condition-based maintenance is worth that cost on an asset when three things hold. Its failures are expensive, it degrades progressively rather than failing at random, and the degradation shows up in something you can measure early enough to act on. For some equipment, NASA’s reliability-centred maintenance guide treats run-to-failure as an acceptable, conscious decision. The original airline research behind RCM found that a policy based only on maximum operating age had little or no effect on failure rates, whatever the age limit.
How does predictive maintenance machine learning work in practice?
Most data-driven predictive maintenance methods run in stages. A model first detects that an asset has left normal behaviour, then diagnoses whether that departure is a developing fault, then estimates how long until failure, and finally triggers the maintenance action. This sequence follows the OSA-CBM standard. A prognostic output has two parts, a time to failure and a confidence limit, and auditable maintenance decisions should rest on the bounds of that interval rather than a single value.
Failure prediction becomes a binary classification: does a failure fall inside the next prediction window, given a reading window of recent telemetry? The minimum prediction window was set to the shortest span that still let an operator step in before an unrecoverable failure. Every method lost predictive power as the horizon grew. Deep learning significantly beat classical models only where the patterns before a failure were diverse. On simpler, repetitive patterns the two performed comparably.
Published case studies cover an industrial pallet-wrapping machine and a nitrogen generator whose oxygen concentration and air pressures dropped before a fault, crushers at a coal-fired plant and gantries in a steelworks converter monitored with wireless vibration and temperature sensors, and milling and turning spindles assessed with a NIST method. The common difficulty is that maintenance crews keep machines healthy, so historical failures can be few or entirely absent. Modelling under that constraint is its own subject: matching degradation and RUL models to your failure-data regime.
The discipline lies in tuning and integration, not in raw prediction accuracy on a benchmark. A model that flags every borderline degradation gets ignored or disabled within weeks, so thresholds must match the crew’s response bandwidth.
Is predictive maintenance just anomaly detection on operational metrics?
Predictive maintenance goes further than anomaly detection on operational metrics. Anomaly detection is the first stage of predictive maintenance, not the whole of it. Anomaly detection tells you whether an asset is working under normal condition. Predictive maintenance adds a diagnosis of which failure is developing and a prognosis of when it will fail. Most predictive maintenance architectures use binary, one-class or unsupervised models that carry no failure-type information. They can only diagnose by clustering detected anomalies by similarity. The split between which operational anomalies are AI-detectable and which are threshold-rule territory is covered separately.
Which failure modes on a machine give enough warning to be worth predicting?
A failure mode is worth predicting when its failures cause heavy downtime per occurrence and its P-F interval leaves the crew time to act. Which modes qualify comes out of the plant’s own failure-mode analysis, not a generic list.
TechnoLynx’s Production AI Reliability work builds on that analysis. It validates that a model built for the ranked failure modes produces predicted-failure flags the maintenance crew can absorb before the model reaches production, starting from the ranked modes and the failure history behind them.
The method is failure modes and effects analysis. IEC 60812:2018 describes how FMEA, and its criticality-ranking variant FMECA, is planned, performed, documented and maintained. When the ranking includes at least the severity of consequences, the analysis is an FMECA. Each failure mode is rated for severity, occurrence and detection. The traditional risk priority number multiplies the three ratings. The 2019 AIAG-VDA handbook replaced it with Action Priority tables that weight severity more heavily, because multiplying ordinal scales gives numbers that do not behave as the formula suggests. In reliability-centred maintenance, FMEA answers what causes each functional failure, and RCM then chooses the maintenance task for each failure mode.
NASA’s RCM guide defines dominant failure modes as the most common ones, those responsible for a significant share of an item’s failures. The guide notes that a ball bearing has the same failure modes in any machine, but the dominant mode, its cause and its effects change from machine to machine. A quadrant chart plots components by failure frequency against downtime per failure. Prognostics should focus on the quadrant of infrequent failures with the most downtime per occurrence.
Detectability is the second filter. The P-F interval is the time between the first detectable sign of a failure and the functional failure itself. A bearing failure, for example, gives vibration, ultrasound and heat symptoms up to six months ahead. Where failures strike at a constant, random rate (the flat middle of the bathtub curve), the machine’s age says nothing about when the next one will come. Where a high-impact mode has no usable leading indicator, we would move the instrumentation priority to the next most detectable mode instead of spending sensors on a mode that cannot be seen coming.
For a high-pressure die-casting cell, we would not settle hydraulic pressure versus die wear from another plant’s list. The dominant mode changes from machine to machine, so other foundries’ choices are a sanity check, not a substitute. A study of 3,438 manufacturing work orders shows why. After tagging, hydraulic leaks appeared in 159 separate issues, even though they were not among the 20 most common raw descriptions, and maintenance management had not known about them. Where there is little failure history, WBDG guidance allows occurrence to be estimated from experience with similar systems. NASA describes an “intuitive” RCM route that starts with the obvious condition-based tasks and minimal analysis. Treat both as interim inputs to be replaced by measured failures.
Which machines get sensors first, and how do you defend that order?
Rank machines by criticality and cost of downtime, scoring failure modes on severity, occurrence and detection. Instrument the critical machines first, with the minimum viable sensor set for each machine type. The ranked list, with its evidence, is also the justification for the spend.
WBDG calls Pareto analysis the best tool for deciding where to start: look for bottlenecks and recurring failures, and follow the money. An O&M best-practices guide makes the same point from the other side. It says to define equipment criticality and cost of downtime before choosing the maintenance mix. AESTECHNO, an IoT engineering firm, publishes concrete criteria. Its criticality criteria are being a bottleneck, repairs over eight hours, more than three failures a year costing over €20,000 each, and age over ten years. On a 50-machine fleet, the guide’s result is 8 to 12 machines to instrument first. That is one firm’s rule of thumb, useful as a starting template rather than a standard.
On the minimum sensor set, the same guide lists three-axis vibration plus temperature for rotating motors, vibration, pressure and flow for pumps, and vibration, temperature and motor current for compressors. In that list, vibration plus temperature is the base for motors, compressors and conveyors. A specific dominant mode, such as hydraulic degradation on a press, is what earns an extra sensor type. A minimum set covers the dominant modes and nothing else, so the residual risk should be written down rather than implied away.
We would build a one-page justification per machine from four things that trace back to records: the criticality inputs and where each came from, the dominant failure modes with their downtime history, the sensor mapped to each mode, and the baseline cost the sensors are expected to reduce. It holds up only as well as its inputs, and it goes stale unless the FMEA is maintained as a living document. AESTECHNO also recommends phasing the rollout. Instrument the single most problematic machine for two to three months, extend to the critical set over six months, and only then scale to the fleet.
Choosing condition monitoring sensors from the failure mode, not the vendor package
Condition indicators specific to each failure mode being modelled must be identified and continually collected. Compare the result against a vendor’s default package only after that mapping exists.
The commonly used techniques include vibration analysis, ultrasound, acoustic emission, thermography, oil and wear-particle analysis, motor current analysis, pressure and temperature. They do not see the same failure at the same time. Reliable Plant’s P-F example for a bearing shows unusual vibration about six months before failure, airborne ultrasound at about three months, and heat at about one month. Ultrasonic techniques can pick up friction and stress waves earlier than vibration or oil analysis. Infrared has the widest range of application and catches both mechanical and electrical faults. Vibration analysis is most productive on high-speed rotating equipment and can be the most expensive part of a programme to set up.
Procedure: from ranked failure modes to a sensor list
- Take the top-ranked failure modes for each critical asset from the FMEA or FMECA.
- For each mode, name the precursor signal (vibration signature, stress wave, temperature rise, pressure loss, current change) and where it originates.
- Choose the sensor type whose P-F interval gives the crew enough time to plan the repair.
- Fix the sensor location deliberately: for acoustic emission, location is critical to reliable readings.
- Compare the list with the vendor package. Drop items that map to no ranked mode and record the gaps the package leaves.
Failure symptom to sensor: a decision table
| Failure symptom | Sensor and method | Note |
|---|---|---|
| Bearing race or rolling-element defect | Accelerometer with envelope analysis; acoustic emission or ultrasound for earlier detection | Bearing defects produce peaks between whole orders and high-frequency energy |
| Unbalance, misalignment, looseness | Accelerometer spectrum (1×, 2×, higher harmonics) | Each fault leaves a recognisable spectral pattern |
| Overheating motor connection or electrical fault | Infrared or temperature sensing | Infrared covers mechanical and electrical faults |
| Pump or hydraulic performance loss | Pressure and flow, with vibration | Insufficient discharge pressure or flow is itself a functional failure |
| Lubricated-component wear | Oil and wear-particle analysis | A long-term programme that can take years to mature |
| Motor degradation | Motor current analysis | Listed among standard condition-monitoring techniques |
For furnaces, ladles and other thermal process equipment, we would expect heat, molten-metal splash and physical access to limit placement more than the ideal detection point does. Some modes have no practical direct sensor and must be inferred from indirect signals, which changes the cost case. Every predictive technology also carries the cost of people who can read it. The lower a technique’s probability of detection, the more often it has to look.
Is accelerometer vibration analysis enough to catch early bearing wear on a spindle?
Not on simple vibration metrics alone, by NIST’s spindle findings. Accelerometer spectra show the four faults behind most rotating-machine problems. One bearing study found acoustic emission better for predicting bearing defects, and in Reliable Plant’s P-F example heat appears about a month before failure. Placement and sampling rate both matter, as the sections below show.
TechnoLynx can deliver a sensor specification for a gantry-mill spindle that is checked for lead time and false-alarm rate against known or seeded faults on that machine. The work starts from the bearing fault modes, fault frequencies and failure history. From there it specifies sensor types, placement, sampling and resolution, and builds the acquisition pipeline to edge or on-machine hardware.
A bearing study found acoustic emission more applicable than traditional vibration monitoring for predicting bearing defects. It recommended a permanently installed system, because cracks below the surface can go unnoticed by portable checks. NIST’s work on machine-tool spindles goes further. It finds simple vibration metrics insufficient for diagnosing spindle condition, because they mix bearing-defect motion with the spindle’s own structural dynamics. Two milling spindles with about 14,200 and 14,700 operating hours read 0.522 and 8.71 mm/s on an ISO RMS metric, yet NIST’s defect-only metric was almost the same for both. NIST’s method accounts for vibration differences due to system dynamics.
Sampling rate for early bearing faults
By the Nyquist theorem, the sampling rate must be at least twice the highest frequency of interest. Anti-aliasing filters cut the usable bandwidth to roughly 40% of the sampling rate. Envelope analysis, the most powerful tool for early bearing defect detection, works on a structural resonance band of about 2 to 10 kHz. It therefore needs 25,600 samples per second or more, about ten times the rate for overall vibration monitoring. The working: 25.6 kS/s × 0.4 ≈ 10.2 kHz of usable bandwidth, which just covers the top of that band.
Defect frequencies rise with shaft speed, so fast spindles need more. A 2025 high-speed spindle study cites reports that at around 80,000 RPM a single outer-race defect can produce impacts above 20 kHz. Both that study, at 12,000 RPM, and NIST’s spindle protocol acquired data at 51.2 kHz. Record length matters as much as rate: each record should hold at least 10 to 15 cycles of the lowest defect frequency of interest. Battery-powered wireless sensors pay for the higher rate. Moving from 2.56 to 25.6 kS/s means ten times the data and roughly five to eight times the energy per measurement.
Mounting points and axis orientation
Mount the accelerometer on the bearing housing, where shaft forces reach the outside. A triaxial unit reads vertical, horizontal and axial directions at once. The 2025 spindle study requires accelerometers with a flat frequency response to at least 10 kHz and warns that installation resonances can mask high-frequency fault signatures. AESTECHNO cites wide-band MEMS parts with flat response to about 6 kHz, against industrial piezoelectric sensors that reach beyond 20 kHz. On the numbers above, a 6 kHz part cannot cover the full envelope band on a spindle. Industrial signal-to-noise ratios can fall below −10 dB. Acoustic emission’s sensitivity to mounting quality and background noise is therefore a false-alarm risk to tune for, not a reason to leave it out.
| Sensor | Catches | Sampling | Placement |
|---|---|---|---|
| Triaxial accelerometer | Unbalance, misalignment, looseness, bearing defects via envelope | ≥25.6 kS/s for envelope work; 51.2 kHz in the spindle studies | Bearing housing, three axes, flat response to ≥10 kHz |
| Acoustic emission | Crack growth and early bearing distress | : | Close to the bearing; location is critical |
| Temperature | Late-stage heat, about one month before failure in the P-F example | Slow; low rates suffice | Bearing housing or motor hotspot |
Treat any published specification as a starting point. We would validate a specification on the spindle itself, against its bearing type, speed range and noise, and expect to iterate.
How much of twelve years of downtime tickets can be tagged automatically?
Most of it, but not all. A 2026 study that ran a language-model workflow over 16,316 wind-turbine maintenance records spanning 9.2 years accepted failure-mode evidence for 71.48% of them. It marked 21.09% as containing too little information and excluded 7.43% by rule.
TechnoLynx can build the tagging and asset-linking pipeline for a plant’s own ticket archive. It starts by reading a sample of tickets against the FMEA structure to fix the failure-mode taxonomy. It then chooses rule matching, classical text classification or a fine-tuned language model to suit the vocabulary, and reports precision and recall per failure mode against a hand-labelled hold-out set before the full corpus is processed.
The taxonomy has to separate failure mode, mechanism and cause. The failure mode is what was observed, the mechanism is the physical process, and the cause is the root condition. An ISO 14224 implementation guide calls conflating them the most common data-quality error in CMMS records. Free text makes this harder. NIST’s example is one technician writing “leaking hydraulic fluid at Machine H1” while another writes “hyd leak at cutoff unit of H1”.
Accuracy depends on labelled data. In a 2023 study classifying work orders into 22 ISO 14224 failure-mode codes, micro-F1 rose from 0.46 for an off-the-shelf language model to 0.81 after fine-tuning on expert-annotated data. The same study found results varied slightly between runs even at zero temperature. A comparison on 3,896 wind-turbine work orders took 8.85, derived from expert labels, as ground truth. Automated text classification gave 8.21, and AI-assisted tagging gave 6.89 after only 28 hours of tagging.
Manual review alone does not scale. NIST puts typical maintenance logs at 10³ to 10⁵ entries, and its experts tagged about 100 issues an hour. At the top of that range, that is 100,000 ÷ 100 = 1,000 expert hours. Engineers should still read low-confidence tickets, rare but high-impact modes, and the years where the ticketing system, the vocabulary or the asset naming changed. NIST’s technical language processing approach is built around exactly this human-in-the-loop iteration.
Two limits keep the output honest. First, the 2026 study treats its fields as model-generated candidates with provenance, not verified diagnoses. Any AI-assisted mining should likewise stay grounded in the records it read and never invent causes. Second, record counts are not criticality: without reconstructed events, exposure and attributed downtime they cannot serve as failure rates or FMEA risk parameters.
Turning the criticality ranking into a payback case finance will approve
Build the business case per machine. Convert the ranking into expected downtime-cost reduction on each top-ranked machine, set it against that machine’s instrumentation cost, and instrument the subset whose payback meets finance’s threshold.
Tractian, a condition-monitoring vendor, gives the formula: ROI = (avoided costs − programme costs) ÷ programme costs × 100. Payback in months is programme cost divided by monthly avoided cost. The baseline should come from 12 to 24 months of downtime records, costed per asset rather than at a plant-wide average. It should include secondary damage, such as a failed bearing taking the shaft or housing with it. Programme cost covers hardware, installation, software, training and ongoing upkeep. Avoided costs should be modelled as conservative, moderate and optimistic scenarios, never as all downtime eliminated.
Vendor payback figures are a range to test, not a promise. AESTECHNO reports 18 to 24 months on many projects, with the first year at 60 to 70% of potential while models learn and false positives are calibrated. Count this under programme labour: crew hours spent on false alarms are a programme cost, and an alert stream the crew mutes delivers no savings at all. Revisit the case after the first instrumented machines have run, because the reduction rate is an assumption until then. The post-deployment side, measuring the value of rare-event detection, needs its own method. The baseline counts five categories: unplanned downtime, emergency repairs, secondary damage, production loss and overtime labour.
Who can help us decide what to instrument and turn our downtime logs into failure-mode data?
TechnoLynx does the engineering side of this work for manufacturing plants. It turns the failure-mode ranking into sensor specifications, builds the pipeline that mines downtime tickets into per-asset failure-mode data, and validates that the resulting predictions hold up against the crew’s capacity before rollout. That last step is part of TechnoLynx’s Production AI Reliability offer. When choosing a partner for degradation modelling, ask for validation on your own machines, with lead time and false-alarm rate measured, rather than accuracy on a public dataset. Our wider AI engineering for manufacturing and automotive work sits alongside this.
After the sensors are in: thresholds, measurement and the work-order system
How should a plant set prediction thresholds so the crew isn’t flooded with low-value flags?
Set the alert threshold against the number of flags the maintenance crew can investigate, not at the point that maximises the model’s score. The 2022 heavy-industry study on crushers and gantries was motivated by exactly this need: a system that works from a cold start without flooding the operator with alarms. Public severity grids give a defensible anchor. ISO 10816-1 grades velocity RMS on the bearing housing, over 10 to 1000 Hz, into zones A to D, from “good” to “act without delay”. NASA’s RCM guide adds that once an alert limit is crossed, the monitoring interval should drop to a third or a quarter of what it was. Thresholds and training windows also change without any code change, which is why versioning data, config and models belongs in the same discipline. The full tuning method is in tuning anomaly-detection sensitivity without drowning the on-call engineer.
Why should one framework govern alert thresholds across the whole fleet?
When each machine or vendor sets its own limits, an “alert” means something different on every asset. The crew cannot compare severity, and the load on the crew cannot be planned. AESTECHNO’s guide points to the ISO chain for this: ISO 10816 (now ISO 20816) for severity, ISO 13374 for monitoring-system architecture, ISO 17359 for programme setup and ISO 13379 for diagnosis. The same logic applies to failure data. The ISO 14224 guide stresses that comparability across sites depends on applying the taxonomy with the same discipline everywhere.
How do we measure whether predictive maintenance is actually reducing unplanned downtime?
Measure against the per-asset downtime baseline built for the business case. Track unplanned downtime hours on instrumented machines, lead time on failures that were caught, and false alarms per crew-week. WBDG warns against KPIs that cannot be obtained or controlled. The 2026 wind-turbine study shows classifications must be documented when comparing, because reclassifying the same archive shifted its system distribution.
What data and work-order integration does a realistic deployment need?
AESTECHNO lists the minimum for data readiness as 6 to 12 months of failure history (dates, causes, parts, downtime), an up-to-date maintenance log, and machine operating parameters. Without that history, it advises 3 to 6 months of extra data collection and fixed-threshold rules before machine learning. On the CMMS side, the ISO 14224 guide recommends making failure mode, mechanism and cause separate mandatory fields at work-order closure. That also keeps future tickets out of the text-mining backlog. Connecting to historians and SCADA is covered in integrating AI anomaly detection with existing SCADA and observability stacks. Getting alerts into the crew’s day is its own topic: rolling out predictive maintenance alerts into scheduling and work orders.
The ranking is a hypothesis until a machine fails, or doesn’t
We would treat every number in the first instrumentation round as an estimate: occurrence ratings from patchy tickets, downtime costs from finance’s best guess, P-F intervals borrowed from published examples. The first instrumented machines are the test. Either they catch a dominant-mode failure with lead time the crew can use, or the FMEA ranked the wrong mode or chose the wrong sensor. Feeding that result back into the analysis is the point of RCM’s emphasis on maintenance experience and condition data. We would revise the ranking after the first round instead of defending it.
FAQ
What is meant by failure mode?
A failure mode is the observed way an item fails, such as external leakage. It is distinct from the failure mechanism, the physical process such as corrosion, and from the failure cause, the root condition such as the wrong material.
What is the primary goal of FMEA?
FMEA establishes how items or processes might fail to perform their function, so that the treatments they need can be identified. It does this by systematically listing failure modes and their local and global effects.
