Weibull analysis sits at the center of nearly every reliability program in Indonesia’s energy and manufacturing plants, used to justify replacement intervals, spare parts stocking, and turnaround scope. The trouble is that a Weibull curve built on the wrong dataset produces a confident-looking number that points maintenance planners in exactly the wrong direction, and most plants never find out until a component fails well ahead of its predicted interval. Getting the method right is less about statistical sophistication and more about discipline in how failure data gets recorded before the curve is ever fitted.

What the Shape Parameter Actually Tells You

A Weibull distribution describes time to failure using two numbers that matter to an engineer: the shape parameter, beta, and the scale parameter, eta.

Beta tells you which failure pattern you are looking at. A beta below 1 points to infant mortality, failures concentrated early in life, usually traceable to installation error, contamination, or a bad batch of parts. A beta near 1 means failures are random with respect to age, the classic signature of external events such as process upsets or electrical transients, where replacing the part early buys nothing. A beta above 1 signals wear-out, where failure probability rises with age and a time-based replacement strategy genuinely reduces risk.

Eta, the characteristic life, marks the age at which 63.2 percent of the population is expected to have failed. Together, beta and eta let an engineer calculate the probability of survival to any given age, which is what actually drives a defensible replacement interval or inspection frequency, rather than a round number pulled from an OEM manual written for a different climate and duty cycle.

The error most plants make is skipping straight to this calculation without checking whether the dataset represents one failure mechanism. A single curve fitted across every failure a pump has ever had is not a reliability model. It is an average of several different physical processes pretending to be one.

Where the Analysis Breaks Down in Practice

The most common mistake in industrial Weibull work is pooling failure data from a repairable asset, a pump, compressor, or motor, into one dataset without separating failure mode. Reliability analysts have pointed out that a single Weibull curve for a whole machine functions as noise disguised as insight, because it blends wear-out, random, and early-life events that have nothing in common mechanically. The resulting beta sits somewhere in the middle of all three patterns and describes none of them accurately.

Mixed failure modes. Impeller erosion, seal leakage, and bearing seizure on the same pump follow different physics and different timelines. Fitting them together produces a shape parameter that misrepresents all three, which is how a plant ends up extending a replacement interval for a wear-out mode while treating a lubrication-driven early-life failure as if it were simply bad luck.

Censored data handled incorrectly. Units still running at the time of analysis are suspensions, not failures, and leaving them out or treating them as failures at the analysis date biases the curve toward shorter life. Every working unit in the population needs to be accounted for as a right-censored data point.

Small sample sizes. Rotating equipment on a single platform rarely generates enough failures for a stable estimate. A beta calculated from four or five data points can shift dramatically with one additional failure, and confidence bounds on that estimate are usually wide enough to make the point estimate close to meaningless on its own.

Inconsistent failure coding. When technicians record a failure under a generic label like “component” instead of the specific part, defect, and cause, the dataset cannot be sorted by failure mode at all, which makes every downstream Weibull fit suspect regardless of how carefully the math is done afterward.

Reading the Numbers Without Fooling Yourself

Before trusting a beta value enough to change a maintenance interval, check the basics that most software packages will not flag automatically.

• Confirm the dataset covers one component, one defect type, and one root cause, not a mixed population of failure mechanisms.
• Check that suspended units, equipment still in service, are included as censored observations rather than dropped from the analysis.
• Look at the sample size relative to 🔒Contact us
• Plot the data on probability paper, or review the correlation coefficient of the fit, before accepting the parametric result. A poor fit usually means the population is still mixed.
• Cross-check the statistical result against the 🔒Contact us

Building a Weibull Program That Holds Up to Scrutiny

A Weibull result only carries weight with engineering management, and with auditors reviewing compliance against technical inspection regulations, if the underlying data architecture can defend it. That starts upstream of any statistics.

• Data architecture.
• Failure mode segregation.
• Review cadence.
• Linkage to regulatory assessments.

Making the Case to Management

None of this requires new software or a larger analytics team. It requires treating the failure report form as the most consequential document in the reliability program, because it is the raw material every later calculation depends on. A plant that cannot say with confidence whether a given failure was wear-out or random will keep setting replacement intervals by instinct and calling it data-driven. The fix starts with how technicians fill out the first line of a work order, not with how sophisticated the curve-fitting software is three steps later, and it is considerably cheaper to implement than any of the instrumentation upgrades competing for the same budget line.

Where to Get Help Building This

Cliste Rekayasa Indonesia helps reliability teams set up this kind of failure data structure and runs Weibull analysis training for maintenance staff. Reach out for a consultation or in-house training.

Contact us now!

Author: Jen Megah Bremanda S. (Reliability Engineer)

References

• Weibull Analysis Works Best When Failure Data Is Clean and Specific, Reliability Magazine, 2024
• Tingkatkan Pembinaan dan Pengawasan, Ditjen Migas Gelar Program Zero Accident dan Unplanned Shutdown, Ditjen Migas, 2020

Clarification Note: This article was developed with AI assistance for drafting, research support, and language refinement. The topic framework, technical direction, and key points presented were determined by the author.