Without failure-rate data, failure modes can still be ranked qualitatively. First rank every machine by consequence of failure multiplied by likelihood of failure. Then run failure-mode-level FMEA or FMECA only on the shortlist that this asset-level pass produces. The result holds up in front of a capital committee when the scales, weights and reconciliation rules are fixed before anyone scores, and when the finished list has been checked against what technicians actually see breaking.
Is there a standard FMECA framework, or do we have to invent our own?
IEC 60812 (Edition 3.0, 2018) defines FMEA as a systematic method for identifying the ways an item or process might fail, and the effects of each failure mode on performance, the surrounding environment and personnel. Under IEC 60812, the analysis becomes FMECA when the ranking of criticality involves at least the severity of consequences, and often other measures of importance. The 2018 edition added alternative ways of calculating risk priority numbers and a method based on a criticality matrix.
FMECA is formalised in MIL-STD-1629A, the US military procedure. It defines criticality analysis as ranking each potential failure mode by the combined influence of severity and probability of occurrence. A 2021 paper on FMEA in engineering design traces the method back to MIL-P-1629 in 1949, when the process was already called FMECA. In automotive supply chains, the AIAG and VDA harmonised handbook published in 2019 replaced earlier practice with a seven-step approach, and IATF 16949 treats that handbook as the reference methodology.
A 2025 FMECA study of laser-cutting defects notes that the standard does not impose a fixed numerical range and allows the scales to be adapted to the analysis. So pick the standard, adapt the scales to your plant, and write the adaptation down.
Asset criticality analysis and FMECA answer different questions
| Asset criticality analysis | FMEA / FMECA | |
|---|---|---|
| Level of analysis | Whole assets: which machines matter most | Individual failure modes: how each component fails, its effects, how to detect or prevent it |
| Speed | Fast enough to cover a whole plant | Detailed and slow, kept for the most important systems |
| Output | A criticality score per asset and a tier (Critical, Important, Non-Critical) | Ranked failure modes with actions, owners and dates |
| Where it is applied | The full asset base | The 10 to 30 percent of assets where deeper analysis is justified, per Reliamag’s guide |
Failure-mode analysis belongs on the top tier that the asset-level pass produces.
Two machines broke down: a one-off, or a failure mode the other ten share?
To find out whether two breakdowns in a fleet of twelve high-pressure die-casting machines are isolated or fleet-wide, trace each breakdown to its failure mode and cause. Then check whether the other ten machines share the condition behind it: age, duty cycle, maintenance history, component batch.
TechnoLynx can take that comparison beyond the spreadsheet through its AI engineering for manufacturing work. It defines which maintenance records, sensor logs and failure timelines are needed to establish root cause, builds the data pipeline that compares failure signatures across all twelve machines, checks the root-cause hypothesis against the ten machines that have not failed, and sets up condition monitoring if the shared mode is confirmed. The first step is defining which maintenance records, sensor logs and failure timelines are needed to establish root cause.
FMEA’s discipline matters here because it keeps three things apart. A failure mode should be a simple description of the failure, stated without its cause or its effect. Causes are the reasons that lead to the failure mode, and effects are the negative impacts if it materialises. Once the two breakdowns are written that way, the generalisation question becomes testable. If the cause is a design, age or duty-cycle factor that all twelve machines share, treat it as fleet-wide risk and schedule inspection or replacement across the other ten. If the cause belongs to those two machines alone (an unusual load profile, a single defective part), monitor the rest rather than intervening fleet-wide.
The published failure analyses give a starting list of candidates, not a verdict. A 2021 study in Engineering Failure Analysis of an H13 steel die used for high-pressure die casting of aluminium describes the die working under high pressure, high temperature and corrosion, all of which limit its service life. The literature on these dies places thermal fatigue among the main factors limiting hot-work die steel, with fatal cracking around fillet radii and holes or inserts where local stresses are high. On the machine itself, an analysis of a 150-tonne plastic injection-moulding machine found a tie bar that failed by fatigue at the thread root. The cause was uneven tension across the four tie bars, from improper moulding parameters, combined with material defects. That machine had no tie-bar tension measurement. The case comes from injection moulding, not die casting. The tie-bar mechanism is a candidate to check on the die-casting machines: measuring tie-bar tension on all twelve would test it.
Sensor history is not a precondition for the ranking. MIL-STD-1629A provides for exactly this case: when failure rate data are not available, failure modes are assessed in qualitative probability levels, running from Level A (frequent) to Level E (extremely unlikely). A qualitative ranking built from maintenance logs and the judgement of the people in the room should be re-scored once condition data starts arriving.
Turning downtime logs into an asset criticality ranking a capital committee will trust
The standard methodology for a ranked list is semi-quantitative weighted scoring. Score each asset’s consequence of failure across defined categories, multiply by likelihood, and sort the result into tiers.
Settle the frame first. Assetivity’s streamlined method has three steps: agree the risk matrix, assemble the equipment hierarchy, then assess the failure risk of each item. Assetivity advises lowering a corporate risk matrix’s trigger levels for equipment-level use, to spread the results. Operations and maintenance must agree what counts as a maintainable asset before scoring starts, because boundary decisions made mid-analysis cannot be reconciled later. Weightings across consequence categories have to be documented, approved by leadership and locked, since changing them part-way invalidates the earlier scores. Discrete manufacturing typically weights production and quality more heavily.
Then get the right people to score. Reliamag’s guide requires a cross-functional team including operations, maintenance, safety, engineering and quality. It sets one rule that drives consistency: score consequence against the worst credible failure mode, not the most likely one, and document which mode was considered. The same asset scored separately by operations, maintenance and safety produces three different scores, which is why single-person scoring cannot be defended in an audit. Who is in the room also decides coverage: whole categories of failure mode get missed because they live in the heads of people who were not invited. The 5xWhys facilitation guide goes further and recommends postponing the workshop if an essential role cannot attend.
Two of Assetivity’s tips keep the asset pass streamlined but robust. Score one event per asset, the maximum reasonable outcome, and agree it in a workshop with operations and maintenance. Assess only the risk dimension that dominates that event, rather than averaging every dimension. Guard against letting asset value drive the score. An expensive machine with a cheap standby is not critical, because consequence to the operation counts, not replacement cost.
Which failure modes count as line-stopping?
Reserve the top production-severity band for failure modes that halt the whole line, not a single station. A torque wrench fault that a spare tool covers within the station’s buffer is a station-level event. A furnace failure that starves the casting cell beyond its buffer stops the line. In Gopalakrishnan and colleagues’ Chalmers case studies, criticality classification worked better in tightly coupled assembly and foundry layouts, where the effect of one machine failing on the rest of the system is easy to capture. The classification depends on today’s layout and buffer sizes, so any change to either is a trigger to re-score.
Can a proper FMEA fit into six weeks? Yes, if it is scoped. Reliamag’s guide puts a simple PFMEA for a single process step at 4 to 8 hours of team time, and a system-level FMEA of a complex assembly at 200 hours or more. Limiting the failure-mode analysis to the top tier is what keeps a six-week schedule realistic. The cost of that scoping is that low-frequency, high-severity modes absent from the logs can be missed. Scoring the worst credible event at asset level is the main guard against that.
A severity, occurrence and detection scale that stays comparable across machine types
Comparable RPNs need anchored scales, not a bare 1 to 10. The 5xWhys guide recommends the AIAG-VDA scales for automotive and aerospace work, because generic scales are too abstract to give a team concrete reference points. The AIAG-VDA process FMEA tables, as summarised by QHubio, read as follows.
| Rating | Severity (effect) | Occurrence (indicative ppm) | Detection (control capability) |
|---|---|---|---|
| 10 | Hazardous without warning, safety or regulatory failure | Failure inevitable, > 100,000 | No control, cannot detect |
| 9 | Hazardous with warning | ~ 50,000 | Very remote, random visual |
| 8 | Loss of primary function | ~ 20,000 | Remote, manual visual at the workstation |
| 7 | Degraded primary function | ~ 10,000 | Very low, sample visual or dimensional check |
| 6 | Loss of secondary function | ~ 2,000 | Low, sample SPC or variable gauge |
| 5 | Degraded secondary function | ~ 500 | Moderate, frequent variable measurement |
| 4 | Appearance or noise, most customers notice | ~ 100 | Moderately high, 100% gauging without feedback |
| 3 | Appearance or noise, some customers notice | ~ 10 | High, 100% automated inspection with feedback |
| 2 | Minor inconvenience | ~ 1 | Very high, poka-yoke detects the defect |
| 1 | No discernible effect | < 0.01 | Almost certain, error-proofing prevents the defect |
The ppm values are indicative process-FMEA anchors, to be calibrated to your own process and history. For equipment, re-anchor occurrence as failure frequency per operating period, taken from your own logs. The calibration practice carries over directly: pick two or three anchor examples per rating from your own history, publish the anchored tables as a controlled document, and re-anchor occurrence whenever the data changes meaningfully. Another option is the five-point scale in the PQRI FMEA training guide. It favours non-consecutive values such as 1, 3, 5, 7, 10 for more distinction and less debate, and advises piloting the definitions on a few examples to check their clarity. Calibration works when severity 7 means roughly the same thing on row 12 as on row 234.
The 5xWhys rule is that if the team cannot name the specific control (which inspection, where, by whom), detection is 10 by definition. On machines without condition monitoring, many modes will score that way honestly, and those high-severity, undetectable modes are where the first sensors should go.
Why equal RPNs can hide very different risks
The risk priority number is the product of the three ratings: RPN = S × O × D. Multiplying ordinal ratings is statistically inappropriate. Reliamag’s guide shows S=10, O=2, D=5 (catastrophic but rare), S=2, O=5, D=10 (minor but undetectable) and S=4, O=5, D=5 all producing an RPN of 100, despite very different risk profiles. The 5xWhys guide offers three practical counters. Any score change after initial scoring needs evidence cited aloud. Severity 9 and 10 items always trigger action regardless of RPN. And the AIAG-VDA Action Priority method can replace an RPN threshold. Action Priority is read from a fixed rule table, and the three ratings are not multiplied.
When the workshop’s scores disagree with the hard downtime numbers, which should win?
Neither should overrule the other without a record of why. Agree before scoring which input sets which rating, and send any failure mode where the two diverge sharply to an explicit review, with the decision and its reason written down.
The case for leaning on the logs is real. Reliamag’s guide calls occurrence the most subjective of the three ratings and says teams with access to failure history produce more accurate occurrence ratings than teams scoring from intuition. Scoring likelihood on gut feel instead of records turns the analysis into an opinion poll. The logs have weaknesses of their own, though. A PHM Society study of maintenance work orders found that the datasets commonly have poor data quality, with fields left empty during manual entry. The authors trace this to personnel entering data under time pressure, and they argue that the errors are likely to be non-random, so they distort the resulting indicators in unexpected ways. Where incidents are scarce, particularly for high-risk failure modes, the DeepFMEA authors set assumptions by agreement between several domain experts and set failure-mode thresholds by expert consensus.
A defensible combination is a documented choice, not a universal formula. One rule set looks like this:
| Situation | Which input leads |
|---|---|
| Severity of any failure mode | Workshop consensus against the documented effect; frequency never changes severity |
| Frequent mode, log covers several years and has passed a data-quality check | Log frequency sets occurrence; the workshop can move it only with cited evidence |
| Rare or high-severity mode with few recorded incidents | Expert consensus sets occurrence, with the reasoning recorded |
| Workshop and log disagree sharply | Flagged for review; the record states which input dominated and why |
The guidance on the AIAG-VDA tables points the same way: keep severity essentially fixed across reviews and recalibrate occurrence when the data changes. As the log grows and its quality is checked, shift weight towards it deliberately, and record that change in the rule.
Keeping the FMEA workshop to about one day per line
One day per line is tighter than the two days a focused PFMEA on one line might take. The 5xWhys guide describes FMEA workshops of 6 to 12 people working over one to four days. Carl Carlson’s position is that FMEAs can be sped up by prioritising, but essential preparation cannot be skipped. Under a hard deadline, his approach is to crash-train the team, complete the preparation in priority order and blitz the analysis.
A calibrated team scores 20 to 30 failure modes per hour, the first hour after a break is the slowest, and the last hour of a long day is the most error-prone. Anything larger should be pre-filtered by the asset-level pass. Two facilitation habits protect quality across the day. Re-score the last two or three items after every break and refresh the scale anchors if scores drift by more than one point. And keep severity and occurrence separate out loud: severity is how bad the effect is, ignoring probability, and occurrence is how often it happens, ignoring badness.
Comparable RPNs across die-casting cells, furnaces and assembly stations need one anchored scale, published as a controlled document and referenced in every FMEA. Settle operator–technician disagreements with the 5xWhys rule: every score change needs evidence cited aloud. If the evidence does not settle it, the facilitator records both views and calls the score, and a one-day timebox should not be allowed to force consensus on a complex mode that deserves a follow-up session.
What matters is that the facilitator is independent and trained, wherever they come from. The 5xWhys guide says the facilitator should be independent of the technical lead, because combining the roles lets scoring drift towards the lead’s opinions. A quality engineer trained in FMEA, a Lean specialist or a methodology coach can fill the role. A template is enough when someone in-house already has that training and the time. Otherwise an outside facilitator is the safer way to hold a six-week schedule.
What is the minimum viable FMEA scope we can defend in six weeks?
The defensible minimum for six weeks is an asset-level criticality ranking across every machine in scope, plus failure-mode FMEA on the top tier only. Present it as a scoped, phased result, not a complete analysis.
TechnoLynx works alongside internal engineers to deliver a six-week result the committee can defend and a documented process the team can continue after week six, starting by framing the scope with the criticality ranking.
Carlson’s advice on duration is to keep the scope defined by the block diagram or process flow diagram rather than attempting everything at once. He contrasts a blitz, pushing through over consecutive days, with a few hours a week spread alongside other work. A six-week plan can combine the two:
- Week 1: ground rules and data. Write down the ground rules, analysis assumptions and failure definitions before analysis starts, as MIL-STD-1629A requires. Fix hierarchy anomalies such as items with no parent, because a clean functional hierarchy speeds the analysis.
- Week 2: asset-level ranking. One cross-functional session scores every asset’s maximum reasonable outcome and produces the tiers.
- Weeks 3 and 4: failure-mode workshops. One time-boxed workshop per line on top-tier assets, with technician and operator interviews running in parallel.
- Week 5: reconciliation and validation. Apply the pre-agreed rule to divergences between workshop scores and logs, then review the ranking with the shop floor.
- Week 6: committee pack. Ranked list, top-tier failure modes with actions, the investment case and an explicit list of what is not yet covered.
If the full analysis is not finished by week six, a credible interim deliverable is the prioritised critical-asset list plus the completed part of the failure-mode analysis, clearly labelled as incomplete. There is precedent in the standards: MIL-STD-1629A expects interim reports at each design review that highlight Category I and II failure modes and single failure points. Agree the phased deliverable with the committee at the start. Asking for that agreement at week six reads as a missed deadline.
How do we sanity-check the ranked list against what technicians actually see breaking?
Before the list goes to committee, hold structured interviews or review sessions with maintenance technicians and line operators in every cell. If their input reorders the ranking, revise the underlying scores and document why, rather than editing the order by hand.
Gopalakrishnan and Skoogh (2018) found that manufacturing companies did not base maintenance prioritisation on machine criticality, because their criticality assessments were static, non-factual and lacked a systems view. A follow-up study by Gopalakrishnan and colleagues found that many machines ended up classified as highly critical, so the classification no longer separated them. It also found that different people running the assessment arrived at different critical machines. Check for both problems in the shop-floor review: an over-inflated top tier, and a ranking that reflects whoever happened to score it.
Should you interview line operators as well as technicians? Yes. Operations brings a failure’s production impact, and maintenance brings its repair complexity and failure history. A ranking missing either view is the cross-functional gap that undermines scoring in the first place. Moss and Woodhouse, as cited by Gopalakrishnan and colleagues, hold that human psychology matters as much as statistics in criticality assessment, so the whole relevant workforce should help set priorities. Use semi-structured interviews; the 2018 Chalmers study used that format to explain its quantitative findings. Useful prompts include: which machine in the top five surprises you, which machine is missing, and when did each top-tier asset last stop the line.
Gopalakrishnan and colleagues’ results also help explain operator-influenced, experience-driven decision making in place of fact-based decisions. When an interview contradicts the ranking, go back to the records before changing a score. Running the interviews in parallel with the workshops keeps them from eating into the six weeks.
One failure-mode and criticality definition for every engineer, and one owner
Run one FMECA framework across the fleet: a shared failure-mode taxonomy, one scoring rubric kept as a controlled document, a named owner, and a single threshold policy that replaces per-engineer cut-offs.
TechnoLynx builds that shared layer. It catalogues the failure-mode taxonomies and RPN practices each asset class currently uses to find where they diverge, defines one FMECA schema and scoring rubric, implements it as a structured data model and a versioned threshold policy with assigned ownership, and validates the thresholds against historical work-order outcomes. The work starts with the inventory of existing practice.
Shared definitions matter because FMEA ratings vary between teams. Reliamag’s guide cites inter-rater reliability studies that document significant variation in RPN values across teams analysing identical failure modes, and notes that calibration reduces this subjectivity without eliminating it. QHubio’s AIAG-VDA calibration guidance supplies the governance that keeps a scale honest. The anchored table carries a version, an owner and a revision date, and cross-team consistency is audited twice a year by re-rating five random rows from different FMEAs as a group.
A fleet-wide threshold policy can follow the trigger the 5xWhys guide describes: act when the RPN reaches the threshold, or severity is 9 or above, or Action Priority is high. Reliamag’s guide adds a tiering safeguard: a top score in safety, environment or production usually warrants Tier 1 regardless of the aggregate. Because RPN multiplies ordinal ratings, a single numeric cut-off treats very different risk profiles alike. Where failure physics genuinely differ, keep class-specific sub-scales but one policy document and one owner.
Ownership should sit with a reliability lead who has authority across every asset class in scope. Someone without that authority becomes a bottleneck at escalation. Reliamag’s guide puts the review cycle at quarterly for critical assets and annually for important ones, documented in the asset management plan as ISO 55000 requires.
Turning the criticality ranking into an investment case with expected downtime savings
An investment case built on the plant’s own failure history and conservative benefits is the more credible one. Benchmark returns from elsewhere carry far less weight. Start with a baseline: Tractian’s guidance is to document, for the assets in scope, how many unplanned failures occurred in the past 12 months, what each one cost in parts, labour and lost production, and the total annual maintenance spend. Vista Projects, an engineering vendor, estimates that gathering 12 to 24 months of this data takes 40 to 80 analyst hours for a mid-sized facility, and more the first time because the records need reconciling.
The arithmetic is simple. Return on investment is total benefits minus total costs, divided by total costs, times 100. For a criticality-ranked asset, expected downtime savings follow from the definitions: failures expected to be avoided per year multiplied by the average documented cost of one failure on that asset, plus any repair-time reduction priced at production value per hour. MTBF drives the avoided-failure term, and MTTR drives production loss per failure. Use the lowest defensible estimate for each benefit, and value failure avoidance at the average historical failure cost, not the worst event. Industry averages are fine for raising awareness, but a serious capital proposal needs your own failure history, labour rates and production value per hour. Safety and compliance benefits belong in the case as qualitative notes, not in the core calculation.
The ranking shows which assets to instrument first. Tractian argues that instrumenting the assets whose failure carries the most severe consequences first concentrates spending where the return is highest. The same vendor claims that criticality analysis typically identifies 10 to 20 percent of equipment driving 60 to 80 percent of downtime and maintenance spend. Treat that as a vendor’s claim to check against your own Pareto chart, not as an expectation. Vista Projects sets a threshold: consider predictive maintenance where annual downtime cost exceeds three times annual monitoring cost.
One published result is useful in the committee room because it involves no new investment at all. In a simulation of an 11-machine serial line, Gopalakrishnan and Skoogh found that prioritising reactive maintenance on throughput-critical machines raised throughput by about 5 percent compared with first-come-first-served, without changing anything else. Part of the return from a good ranking arrives before the first sensor is bought, and the case is stronger for saying so. After that, record every failure the new monitoring catches and the cost avoided, because that tracking is the evidence the next investment case will need. The first instrumented tranche also puts the workshop’s occurrence scores to the test. In the next review cycle, the qualitative ranking should give way to measured failure data on the machines where it now exists.
FAQ
When should FMEA be performed?
MIL-STD-1629A says FMECA should start as soon as preliminary design information is available, and be extended to lower levels as more information arrives. IEC 60812 notes that FMEA can be carried out several times during the lifetime of the same item or process.
What are the different types of FMEA?
The variants are design FMEA, process FMEA, system FMEA and service FMEA, with FMECA, which adds criticality ranking, used heavily in aerospace and military work.
