3. Exploratory data analysis for reliability

Before fitting any life distribution or building a failure-prediction model, you must look at the data. Exploratory data analysis (EDA) for reliability is the disciplined practice of summarizing time-to-failure or degradation data visually and numerically so that distributional shape, time trends, and anomalous observations become visible. Skipping this step is the most common reason a Weibull fit looks "wrong": the analyst has forced a model onto data that is bimodal, contaminated by a mixed failure mode, or drifting over time. This section covers the four core EDA instruments — histograms, probability plots, trend tests, and multivariate visualization — and shows how to read them together.

1. Histograms and Empirical Density: Reading Shape First

A histogram bins the observed failure times (or cycles, or kilometers) and plots counts per bin. Its purpose is not precision but shape recognition. For reliability data, ask three questions of any histogram:

• Unimodal or multimodal? A second hump usually signals a second failure mechanism (e.g., infant mortality plus wear-out) or a mixed population.
• Skew direction? Time-to-failure data are almost always right-skewed; strong left skew suggests a data-entry or censoring artifact.
• Where is the mode relative to the warranty or maintenance interval? A mode inside the warranty window is a commercial risk, not just a statistical curiosity.

Bin width matters. With n = 40 failures, use roughly 6–8 bins; the Freedman–Diaconis rule, bin width ≈ 2·IQR·n^−1/3, is a defensible default. Always overlay the empirical CDF or a kernel density estimate, because a histogram alone hides the tail — and in reliability the tail is where the money is.

2. Probability Plots: Turning a Curve into a Line

A probability plot transforms the vertical axis so that a candidate distribution plots as a straight line. The Weibull plot uses y = ln(−ln(1 − F(t))) against x = ln(t); the lognormal plot uses the normal quantile Φ⁻¹(F(t)) against ln(t). The slope of the fitted line gives the shape parameter β (Weibull) or σ (lognormal), and the intercept gives the scale parameter η or the median.

Reading a probability plot is a skill built on three diagnostics:

1. Straightness. Systematic curvature means the distribution is wrong. Concave-up curvature on a Weibull plot often indicates a decreasing hazard (β < 1) mixed with a wear-out population.
2. Points at the extremes. The first and last few ordered points carry the most leverage. A single point far off the line at the low end is usually an early-life outlier, not evidence of a new distribution.
3. Parallelism across groups. If you plot several production lots or test cells on the same axes, parallel lines imply a common shape with different scales — a strong hint that one physical mechanism dominates.

Use median rank plotting positions, F_i ≈ (i − 0.3)/(n + 0.4), rather than i/n, especially for small samples. For right-censored data, use the Kaplan–Meier estimate and plot the censored points as separate markers so you can see where information ends.

3. Trend Tests: Is the Failure Rate Changing?

Reliability data are often collected in time order — successive failures of a repairable system, or units entering a test at different times. A trend test asks whether the inter-arrival times are homogeneous or whether the intensity is increasing (deterioration) or decreasing (reliability growth).

The two workhorses are the Laplace test (also called the centroid test) and the Mann–Whitney / military handbook trend test. The Laplace statistic for a renewal process observed to time T with n failures at times t₁…t_n is:

U = ( Σ t_i / n − T/2 ) / ( T · sqrt(1 / (12 n)) )

Interpretation:
  U ≈ 0        → no trend (homogeneous Poisson)
  U > +1.96    → increasing intensity (wear-out / deterioration)
  U < −1.96    → decreasing intensity (reliability growth)

For a small example, suppose a repairable pump accumulates 8 failures over T = 4000 h at times 210, 640, 1180, 1620, 2210, 2870, 3310, 3760 h. The mean inter-arrival time is 470 h, and Σt_i/n = 1975 h, which is less than T/2 = 2000 h, so U is slightly negative — no significant trend at the 5% level. If the later intervals had shortened (e.g., failures at 210, 640, 1180, 1620, 1900, 2100, 2250, 2380 h), Σt_i/n would drop well below T/2 and U would exceed +1.96, flagging deterioration. This is a worked illustration with assumed values, not a measured result.

4. Visualization Techniques for Patterns and Outliers

Beyond the histogram and probability plot, four plots earn their place in a reliability EDA workflow:

Outlier detection should be mechanism-driven, not purely statistical. A point that fails the 1.5·IQR rule and corresponds to a known field incident, a test-rig anomaly, or a data-entry error is a genuine outlier to investigate. A point that is statistically extreme but physically plausible is often the most informative observation in the dataset — the earliest sign of a new failure mode. Never delete it silently; annotate it and carry it through the analysis.

5. Putting the Instruments Together

The four tools are complementary, and the order matters. Start with a time-ordered scatter or trend test to establish whether the process is stationary. If a trend exists, model the intensity (e.g., a non-homogeneous Poisson process) rather than a lifetime distribution. If no trend exists, pool the data and examine the histogram for modality. Then use the probability plot to select and validate a distribution, and finally use box plots or Q–Q plots to check group differences and tail fit. Outliers identified along the way should be annotated and, where possible, traced back to a physical cause before any of them are excluded.

A compact, reproducible EDA sequence in Python or R looks like this:

# R sketch — reliability EDA
library(ggplot2); library(fitdistrplus); library/reliaR

# 1. Trend test on ordered failure times
laplace.test(times)          # returns U statistic and p-value

# 2. Histogram on log scale
ggplot(df, aes(x = log(time))) + geom_histogram(bins = 8)

# 3. Weibull probability plot with censoring
fit <- fitdistcens(cens_data, "weibull")
plot(fit, demp = TRUE)

# 4. Group comparison
ggplot(df, aes(x = lot, y = time)) + geom_boxplot()

Treat the output of each step as a hypothesis, not a conclusion. The histogram suggests modality; the probability plot tests distributional fit; the trend test decides whether pooling is legitimate; the box plot exposes group effects. Only when these four views agree should you commit to a distribution and move on to parameter estimation and prediction. Documenting each plot and the decision it drove is what makes a reliability analysis auditable — and auditable analysis is what survives a design review.