Failure Mode and Effect Analysis with Data [/ˈfeɪljər ˈmoʊd ənd ˈifɛkt æˈnælɪsɪs wɪθ ˈdætə/] n - Failure mode and effect analysis is traditionally a qualitative exercise: a team assembles, lists the ways each component can fail, rates the severity and likelihood of each mode, and assigns a risk priority number. The method is only as good as the team’s memory and judgment.

A data-driven FMEA replaces subjective ratings with statistics drawn from the actual incident record; the prioritization then reflects what has happened, not what engineers fancy might happen.

Of Frequency, Read From the Incident Logs

A well-kept ITSM system records every incident with a category, equipment identifier, root cause, and duration. From this the frequency of each failure mode may be estimated:

$${\hat{\lambda }}_k=\frac{n_k}{T_{\text{obs}}}$$

where $n_k$ is the number of observed failures of type $k$ and $T_{\text{obs}}$ is the total observation time in equipment-years. This is the maximum likelihood estimate of the failure rate under a Poisson process assumption.

When the counts are small, the estimate is noisy. A Bayesian approach lays a gamma prior upon the failure rate and updates it with the observed count:

$$p({\lambda }_k\mid n_k)\propto {\lambda }_k^{n_k}{e}^{-{\lambda }_k{T}_{\text{obs}}}\cdot {\lambda }_k^{a-1}{e}^{-b{\lambda }_k}$$

The posterior mean is $(n_k+a)/(T_{\text{obs}}+b)$. For rare failures the prior wards the estimate against being driven by one or two events. This is Bayesian Updating for Beliefs applied to reliability data.

Of Impact, Reckoned

The severity of a failure mode joins the probability that it causes an outage to the cost of that outage:

$${\text{Risk}}_k={\lambda }_k\cdot P(\text{outage}\mid {\text{failure}}_k)\cdot C_{\text{outage}}$$

The outage probability given a failure demands an understanding of the redundancy topology. In an N+1 configuration, the first failure causes no outage; the second within the recovery time does. The probability of a second failure within time $\tau$ is:

$$P(\text{second failure within }\tau )=1-e^{-{\lambda }_k\tau }$$

This compound risk is the one that matters. It depends on the failure rate and the mean time to repair the first failure. Swift repair narrows the window of exposure; slow repair (for want of parts, by reason of access restrictions, or from crew certification requirements) extends it past safety.

Where the Monitoring Money Should Go

The data-driven FMEA ranks failure modes by expected annual cost, and the top modes are the candidates for predictive monitoring investment. A vibration sensor on a bearing contributing $500,000peryearinexpectedriskisabetterpurchasethanasensoronacomponentcontributing$5,000.

The incremental value of a predictive monitor for mode $k$ is the risk reduction it buys:

$$\Delta V_k={\lambda }_k\cdot P(\text{detectable before outage})\cdot P(\text{outage}\mid {\text{failure}}_k)\cdot C_{\text{outage}}-C_{\text{sensor}}$$

Modes of high frequency, high impact, and high detectability return the most on sensor investment. Modes that occur rarely, cost little, or fail too fast to detect are poor candidates, however fearsome they appear upon a whiteboard.

Keeping the Ledger Current

The data-driven FMEA is no one-time deliverable. As the incident record grows, the failure rate estimates improve; as the facility ages, the rates may shift; as equipment is replaced or upgraded, some modes vanish and new ones emerge.

A rolling FMEA refreshes the posterior estimates quarterly on the most recent incident data. A sudden increase in a once-quiet failure mode (flagged by a Bayesian change point detector or an EWMA control chart on the event count) triggers a review of the affected equipment. This closes the loop between Anomaly Detection for Infrastructure Sensors, the maintenance schedule, and the risk prioritization.

See Remaining Useful Life Estimation for how the equipment-level health models join the system-level risk assessment, and IOC operations for how the prioritization is carried into the day-to-day work of the operations center.