Weibull analysis is a powerful tool for modeling reliability data. It helps engineers estimate life characteristics, identify failure modes, and make data-driven decisions. Understanding how systems fail in real-world conditions is key to improving reliability and planning maintenance. In this post, we explore practical techniques for analyzing in-service data using Weibull distribution.
This article starts with a technical overview of the Weibull analysis. It highlights its core functions and explains how its parameters affect failure behavior. It then addresses the parameter estimation from in-service data, including the treatment of both complete and censored datasets. Two estimation techniques – Median Rank Regression (MRR) and Maximum Likelihood Estimation (MLE) – are outlined and compared.
The final section focuses on the analysis of systems with multiple failure modes, introducing approaches such as data clustering, Monte Carlo simulation, and logical combinations to characterize overall system reliability.
Understanding the Weibull Distribution
Background and Core Functions
Waloddi Weibull created the Weibull distribution in 1939 and published it twelve years later. This distribution has three parameters: the shape parameter, \( \beta \), the scale parameter, \( \eta \), and the location parameter, \( \gamma \). For simplicity’s sake, we’ll assume the location parameter equals zero throughout this article.
This statistical distribution can be described by the Cumulative Distribution Function (CDF), \( F(t) \), and the Probability Density Function (PDF), \( f(t) \).
The CDF represents what it is commonly known as the “probability of failure”:
\(\)\( F(t) = 1 – e^{- \left( \frac{t – \gamma}{\eta} \right)^\beta } \)
You can linearize this function, i.e., be represented as \( y=mx+n \), which is useful for parameter estimation. The linearized CDF, considering \( \gamma=0 \), is represented by the following expression:
\( \ln(-\ln(1 – F(t))) = \beta \ln t – \beta \ln \eta \)
Many interpret the PDF as the probability for a specific value of the random value, in this case the time \( t \), but that’s incorrect since it is always zero. Instead, the PDF calculates the probability over an interval of values. When that interval starts in zero to a certain value T, that is equivalent to the CDF.
\( f(t) = \frac{\beta}{\eta} \left( \frac{t – \gamma}{\eta} \right)^{\beta – 1} e^{- \left( \frac{t – \gamma}{\eta} \right)^\beta } \)
An important function in RAMS Engineering is the failure rate function. In contrast to the Exponential Distribution, the value of the failure rate in the Weibull Distribution changes over time instead of remaining constant. It can be represented as a combination of the PDF and the CDF as stated in the following equation:
\( \lambda(t) = \frac{f(t)}{1 – F(t)} = \frac{\beta}{\eta} \left( \frac{t – \gamma}{\eta} \right)^{\beta – 1} \)
Parameter effects on Reliability Behavior
Shape parameter effect
Figure 1 explores how varying the shape parameter \( \beta \) influences its key functions. When \( \beta \) is less than 1, the PDF exhibits a monotonically decreasing behavior, indicative of early-life failures. As \( \beta \) increases beyond 1, the PDF transitions to a bell-shaped curve with a peak that shifts closer to the scale parameter \( \eta \), reflecting a wear-out failure pattern.
This shift also affects the CDF, which evolves from a concave shape to an S-curve. In the latter case, even small increases in time lead to a sharp rise in the probability of failure — a characteristic visible in the linearized plot as the curve approaches a vertical line.
The Failure Rate function follows a pattern familiar to reliability engineers — the bathtub curve — showing decreasing, constant, or increasing failure rates depending on the value of \( \beta \), which will be described further below.
Scale parameter effect
On the other hand, Figure 2 examines the effect of varying the scale parameter \( \eta \). When \( \eta \) is small, the PDF peaks near the origin. This means failures happen earlier, and the system has a shorter expected life. As \( \eta \) increases, the PDF shifts right and flattens. Failures occur later and spread over a longer time.
In other words, a larger \( \eta \) extends the expected lifetime without altering the inherent failure behavior. Because of this time-stretching effect, \( \eta \) is referred to as the scale parameter — it scales the distribution along the time axis without changing its shape, which is governed by \( \beta \).
The CDF also reflects this behavior. As \( \eta \) increases, the CDF curve moves to the right, taking longer to approach 1. Considering a fixed probability of failure, this means it takes more time to reach that value.
Failure Rate variation with time – Bathtub curve
As stated previously, from Figure 3 the bathtub curve can be inferred. It’s a simple yet powerful way to describe how failure rates change over the life of a product and the Weibull distribution is one of the most flexible tools to model it.
Figure 3 explores how the Weibull distribution’s two key parameters, \( \beta \) and \( \eta \), shape the Failure Rate function, denoted as \( \lambda(t) \).
The shape parameter, \( \beta \), controls how the failure rate evolves over time:
- When \( \beta<1 \), the failure rate decreases with time. This represents infant mortality, where early-life failures are more common, often due to manufacturing defects or initial weaknesses.
- When \( \beta=1 \), the failure rate remains constant. This scenario shows random failures, like the Exponential Distribution. The system doesn’t age, and the failure risk stays constant.
- As \( \beta \) increases above 1, the failure rate begins to rise. This marks the wear-out phase, where components degrade with time. The larger the \( \beta \), the steeper the increase. This is especially noticeable when \( \beta>4 \), which is typical for end-of-life degradation.
While \( \beta \) tells us how failure progresses, \( \eta \) determines when that progression happens: increasing \( \eta \) shifts the entire curve to the right, meaning failures occur later in time. So, for a fixed value of \( \beta \), a higher \( \eta \) means a lower failure rate at any given point — i.e., the component lasts longer before entering its critical failure phase.
Together, \( \beta \) and \( \eta \) allow the reliability engineer to model a wide range of real-world scenarios, from early failures in consumer electronics to long-term degradation in mechanical systems. This flexibility is a key strength of the Weibull Distribution — it adapts to match the behavior of any failure data.
How to Estimate the Weibull Analysis Parameters from Reliability Data
One of the mantras for Reliability Engineering predictions is the following:
“The accuracy of any prediction is directly proportional to the quality, accuracy and completeness of the failure data”
When estimating in-service reliability, engineers must first assess the type of life data collected from the field. Not all data comes neatly packaged — in fact, most of it doesn’t. That’s why understanding the different types of life data is essential for an accurate analysis.
Broadly, field data can be classified into two main categories: complete data and censored data.
Complete Reliability Data
This is the ideal case where a clear record of when a unit has failed is obtained. That unit has reached their end of life, and the exact time of failure is known. If the records of each unit meet these criteria, a straightforward modelling can be made. In Figure 4, this is shown in the top-left quadrant. Each unit has a red “X” at the point in time when it failed.
Censored Reliability Data
This type of data is much more common in real-world reliability studies. It occurs when some aspect of the failure timing is unknown. The data is still usable, but it requires specialized techniques for accurate analysis. There are three main types:
- Right-Censored Data (Suspensions): This occurs when a unit still works at the end of observation. We know it hasn’t failed, but we don’t know when it eventually will. It can be seen as the failure time being somewhere to the right of the current point on the timeline. In the top-right quadrant of Figure 4, green arrows mark the units still running when observing them.
- Interval-Censored Data: Sometimes, it is not known the exact failure time, but it is known that it happened between two known points – e.g., during scheduled maintenance interval. If a unit was working during the last check but failed by the next one, its failure occurred somewhere in that interval. This is visualized in the bottom-left quadrant of Figure 4, where each failure is bracketed by two points.
- Left-Censored Data: Left-censoring occurs when it is known the failure happened before a certain time, but not exactly when. This often arises when systems are inspected for the first time after they have already failed, and the failure time is not recoverable. It can be considered a special case of interval censoring.
You must classify the data type correctly to choose the right statistical methods.
Weibull Analysis Parameters Estimation Methods
Median Rank Regression
Median Rank Regression (MRR) is a widely used method in Weibull analysis that fits a straight line to failure data using least squares. It works by plotting failure times (X-axis) against estimated failure probabilities (Y-axis), which are typically calculated using median ranks — a statistical estimate of the cumulative probability of failure:
\( P = 0.5 = \sum_{k=i}^{N} \frac{N!}{k! (N – k)!} MR^k (1 – MR)^{N – k} \)
The line is fitted to minimize either vertical or horizontal deviations, depending on whether the regression is being performed on Y (top of Figure 5) or X (bottom of Figure 5). This method works best with complete data, but you can adjust it to handle censored data. MRR is popular because it is simple, visual, and provides reasonably accurate estimates of the Weibull parameters (\( \beta \) and \( \eta \)).
Maximum Likelihood Estimation
Maximum Likelihood Estimation (MLE) is a powerful statistical method in Weibull analysis. It estimates parameters by maximizing the likelihood of observing the given data. It works by constructing a likelihood function based on the time-to-failure data — whether complete or censored — and finding the parameter values (shape and scale) that make the observed data most probable. We construct the likelihood function as shown below:
\( L = L(\beta, \eta, \gamma \mid t_1, \ldots, t_R, s_1, \ldots, s_M, I_{r_1}, I_{l_1}, \ldots, I_{r_P}, I_{l_P}) \)
Then,
\( L = \prod_{i=1}^{R} f(t_i; \beta, \eta, \gamma)\cdot \prod_{j=1}^{M} \left[ 1 – F(s_j; \beta, \eta, \gamma) \right]\cdot \prod_{m=1}^{P} \left[ F(I_{r_m}; \beta, \eta, \gamma) – F(I_{l_m}; \beta, \eta, \gamma) \right] \)
And finally,
\( L = \prod_{i=1}^{R} \left( \frac{\beta}{\eta} \left( \frac{t_i – \gamma}{\eta} \right)^{\beta – 1} e^{ – \left( \frac{t_i – \gamma}{\eta} \right)^\beta } \right)\cdot \prod_{j=1}^{M} e^{ – \left( \frac{t_j – \gamma}{\eta} \right)^\beta }\cdot \prod_{m=1}^{P} \left[ e^{ – \left( \frac{t_{l_m} – \gamma}{\eta} \right)^\beta } – e^{ – \left( \frac{t_{r_m} – \gamma}{\eta} \right)^\beta } \right] \)
We do this by taking the partial derivatives of the log-likelihood function \( \Lambda=\ln L \) with respect to the parameters, setting them to zero, and solving the equations numerically.
\( \frac{\partial \Lambda}{\partial \beta} = 0 \Rightarrow \hat{\beta}, \quad
\frac{\partial \Lambda}{\partial \eta} = 0 \Rightarrow \hat{\eta}, \quad
\frac{\partial \Lambda}{\partial \gamma} = 0 \Rightarrow \hat{\gamma} \)
MLE is especially valuable for handling censored data and provides efficient, consistent estimates even with small sample sizes. Note that the estimation may be biased, so apply corrections like those suggested by Cox and Snell (1968) to obtain unbiased parameter estimators.
MRR vs MLE
Figure 6 shows Weibull parameter estimates from MRR and MLE. The data varies in size but comes from a known reference distribution (red dotted line).
When the sample size is small, MLE tends to produce more accurate parameter estimates compared to MRRs, especially in capturing the curve’s shape. As the number of failures increases, the differences between methods diminish, and all approaches converge toward the true distribution. This example shows that MLE usually works better with small datasets, while all methods work well with larger samples.
Is Weibull Analysis useful with multiple failure modes?
The previous case assumed a single failure mode, but real-world data often includes a mix of failure modes, making parameter estimation much more complex. Lumping different modes together can distort the Weibull fit. This issue appears clearly in the first plot of Figure 7. This contrasts with the middle plot, where the failure modes were segregated manually during the simulation of the failures. In real-world scenarios, engineers often can’t manually segregate failure modes.
To address this problem, in the right plot of Figure 7 a simple K-means clustering algorithm is used to separate the failure data, which significantly improves the goodness-of-fit compared to the first approach. While K-means is just a basic approach, it demonstrates how automated clustering can be a useful first step before applying more advanced classification methods in practical reliability analysis.
In case of dealing with multiple failure modes or subsystem Weibull distributions, where the overall system behavior is needed, there are two strategies:
One is to run a Monte Carlo simulation, combining the distributions to directly estimate the system’s equivalent Weibull analysis parameters — as shown in the left plot of Figure 8.
The other is a logical “OR” combination, similar to Fault Tree Analysis, where you estimate the system’s unreliability by combining the individual curves — this gives a more conservative system behavior (right plot of Figure 8), but not explicit Weibull parameters.
You can use each method based on the specific needs of your analysis.
Conclusion: Unlocking Real-World Reliability with Weibull Analysis
Engineers use the Weibull analysis to model how systems fail over time under real-world conditions. It accommodates a wide range of failure behaviors, from early-life issues to wear-out mechanisms. By applying estimation techniques like MRR and MLE, RAMS engineers can accurately model component lifetimes — even when data is incomplete or censored.
Moreover, the approach scales to more complex situations, such as systems with multiple failure modes, through techniques like clustering or system-level modeling.
Connect with Anzen to discover how Weibull analysis can enhance your system reliability and support more informed design and maintenance strategies.
About the author
Samuel García is an aeronautical engineer with experience in Safety and Reliability engineering, focusing in recent years on the model-based safety analysis.
At Anzen, Samuel is divided into two main roles: one is to lead or support industrial projects as a consultant, and the other is to participate in the activities performed in the digital engineering department, leveraging those activities and ATICA development by feeding them with actual industrial needs.
