Home / Courses / Probabilistic Design and Reliability / Probability Models for Engineering Variables

Probabilistic Design and Reliability | Module 2 of 10

Probability Models for Engineering Variables

A reliability result is only as defensible as the probability models behind it. Distribution choice is an engineering claim, not a curve-fitting decoration.

Readiness check

Tick only what you can do closed-notes before using this module for design decisions.

Compute a mean and standard deviation.
Read a cumulative probability from a distribution.
Explain why dimensions, loads, and strengths can vary.
Use a histogram without overinterpreting every bin.
Recognize that correlation means variables move together.

0 or 1 weak itemsContinue with this module.

2 weak itemsReview probability and statistics in Mathematics, Module 15.

3 or more weak itemsReview measurement scatter in Measurements, Module 4 before continuing.

The core idea

A probability model maps real engineering variation into a mathematical random variable. The model must fit the physics, data source, bounds, units, and dependence structure of the variable.

F_X(x) = P[X <= x]COV = sigma / murho_XY = cov(X,Y) / (sigma_X sigma_Y)

Loads, strengths, dimensions, friction coefficients, and lifetimes are not interchangeable random numbers. A normal model can be reasonable for tightly controlled dimensions, but it can assign impossible negative values if scatter is large. A lognormal model keeps positive variables positive and is often useful for multiplicative variation. A Weibull model can describe lifetime or strength scatter when weakest-link or aging behavior is plausible, but a good visual fit does not prove a mechanism. Uniform variables are useful for bounded tolerances when no location inside the range is favored. Exponential lifetime models imply a constant hazard rate, which is a strong assumption. Dependence matters as much as the marginal distributions: if high load and high temperature occur together, assuming independence can understate risk.

The skill works when: the failure event, units, assumptions, distributions, and checks are visible.

The skill breaks down when: a probability is reported without the mechanical model and evidence boundary.

The concept. A probability model maps real engineering variation into a mathematical random variable. The model must fit the physics, data source, bounds, units, and dependence structure of the variable.

The skills, taught in order

Choose variables before distributions

Name the physical quantity, unit, data source, and use condition before choosing a mathematical family.

Use CDFs for failure probability

Failure often asks for P[X <= limit] or P[X > limit], so the cumulative distribution is the working tool.

Estimate parameters cautiously

Small samples support rough parameter estimates, not high-confidence rare-event tails.

Check physical bounds

A distribution that predicts impossible negative strength or dimension may be unacceptable even if the mean and standard deviation match.

State dependence

Independence is an assumption that needs evidence. If evidence is unavailable, test sensitivity to plausible correlation.

Engineering connection: use these skills to turn variability into a design decision, not just a plot.

Worked example 1: selecting a model for yield strength

Ten coupon tests of yield strength in MPa are 352, 358, 349, 355, 362, 351, 357, 354, 360, and 353. Estimate the sample mean and standard deviation, then state a cautious probability model for preliminary design.

ProblemTen coupon tests of yield strength in MPa are 352, 358, 349, 355, 362, 351, 357, 354, 360, and 353. Estimate the sample mean and standard deviation, then state a cautious probability model for preliminary design.
AssumptionsCoupons come from one material batch and test method. The sample is too small for rare-tail claims.
Modelx_bar = sum(x_i)/ns = sqrt(sum((x_i - x_bar)^2)/(n - 1))preliminary model: S_y ~ Normal(x_bar, s)
SolveThe mean is 355.1 MPa. The sample standard deviation is about 4.1 MPa. A preliminary normal model can be used near the center of the data because values are tightly clustered and far from zero.
CheckDo not claim the 0.001 lower-tail strength from ten observations. The sample supports central behavior, not extreme reliability.
ConclusionUse S_y ~ Normal(355.1 MPa, 4.1 MPa) only as a preliminary model, with a warning about limited tail evidence.

Result. Use S_y ~ Normal(355.1 MPa, 4.1 MPa) only as a preliminary model, with a warning about limited tail evidence.

Worked example 2: dependence changes load combinations

A bracket sees a vertical load V and a lateral load H. Each has mean 1.0 kN and standard deviation 0.2 kN. Estimate the standard deviation of total load T = V + H for independent loads and for correlation rho = 0.7.

ProblemA bracket sees a vertical load V and a lateral load H. Each has mean 1.0 kN and standard deviation 0.2 kN. Estimate the standard deviation of total load T = V + H for independent loads and for correlation rho = 0.7.
AssumptionsLoads are approximately normal near the design condition. Correlation is linear.
Modelvar(T) = var(V) + var(H) + 2 rho sigma_V sigma_H
SolveIf independent, sigma_T = sqrt(0.04 + 0.04) = 0.283 kN. If rho = 0.7, sigma_T = sqrt(0.04 + 0.04 + 2(0.7)(0.2)(0.2)) = sqrt(0.136) = 0.369 kN.
CheckPositive correlation increases the upper tail of combined load because high V tends to arrive with high H.
ConclusionDependence raises combined-load scatter by about 30 percent in this case.

Result. Dependence raises combined-load scatter by about 30 percent in this case.

Python activity

Engineering question: Fit simple descriptive models to synthetic strength data and compare lower-tail estimates.

Assumptions and units: Use the units shown in the code comments. Keep the random seed fixed while learning so the result is reproducible.

from statistics import mean, stdev

def percentile(values, percent):
    ordered = sorted(values)
    position = (len(ordered) - 1) * percent / 100.0
    lower = int(position)
    upper = min(lower + 1, len(ordered) - 1)
    weight = position - lower
    return ordered[lower] * (1.0 - weight) + ordered[upper] * weight

strength = [352, 358, 349, 355, 362, 351, 357, 354, 360, 353]
mean_strength = mean(strength)
sd = stdev(strength)
lower_5 = percentile(strength, 5)

print("mean MPa", round(mean_strength, 2))
print("sample sd MPa", round(sd, 2))
print("empirical 5th percentile MPa", round(lower_5, 2))
print("engineering note: do not infer rare tails from this small sample")

Numerical checks: The computed mean and standard deviation should match the worked example. The empirical 5th percentile is descriptive, not a certified design allowables calculation.

Limitations: Synthetic or small classroom data are for learning the workflow. Real design distribution choices need traceable test conditions and expert review.

Misconceptions and diagnostics

Mistake	Diagnostic question	Correction
The mean value is the most likely value in every distribution.	Ask: is the distribution symmetric, skewed, multimodal, or bounded?	Correct the assumption, then rerun the calculation or rewrite the report.
All engineering variables are normally distributed.	Ask: can the variable go negative, and what physical process creates the scatter?	Correct the assumption, then rerun the calculation or rewrite the report.
A fitted Weibull distribution proves the physical failure mechanism.	Ask: what mechanism evidence supports the distribution beyond a plot?	Correct the assumption, then rerun the calculation or rewrite the report.
Independent variables can be assumed whenever correlation data are unavailable.	Ask: what common cause might link the variables?	Correct the assumption, then rerun the calculation or rewrite the report.

Practice ladder

Recognize

Name one variable that must remain positive and may not be well represented by a broad normal distribution.

Solution guidance

Write the governing equation, state units and assumptions, then compare the answer with the relevant limit state or design decision.

Calculate

A load has mean 200 N and standard deviation 20 N. Compute its coefficient of variation.

Solution guidance

Write the governing equation, state units and assumptions, then compare the answer with the relevant limit state or design decision.

Diagnose

A fitted normal model predicts negative bolt diameter with non-negligible probability. Identify the problem.

Solution guidance

Write the governing equation, state units and assumptions, then compare the answer with the relevant limit state or design decision.

Design

Specify a data plan for choosing a distribution for fatigue life of a small spring.

Solution guidance

Write the governing equation, state units and assumptions, then compare the answer with the relevant limit state or design decision.

Working with AI, and proving it yourself

Useful AI support

Ask for alternative explanations of terminology, candidate code tests, or a checklist of assumptions to verify.

Ask it to challenge a derivation, then prove the result with units, limiting cases, and an independent calculation.

Do not let AI invent material distributions, fabricate failure data, claim standards compliance, or make final safety decisions.

Portfolio evidence

Create one short reliability note for this module. It must include the engineering question, assumptions, input definitions, equations, calculation, checks, interpretation, and limitations.

Verification required: dimensional analysis, a benchmark case, convergence or sample-size check, and a plain-language residual-risk statement.

Retrieval and spaced review

Closed notes. Answer out loud, then reveal.

What does a CDF give?

P[X <= x], the probability a random variable is at or below a value.

Why is distribution choice an engineering decision?

It encodes physical bounds, mechanisms, data source, and intended use.

What does COV measure?

Relative scatter, sigma divided by mean.

What does positive correlation do to a sum of loads?

It increases the variance of the sum.

Why are rare tails hard to estimate?

They require much more relevant data than central trends.

TodayFinish the review and first two practice levels.

+1 dayRebuild one worked example from a blank page.

+3 daysChange one assumption and predict the effect.

+7 daysConnect this module to the capstone bracket.

+30 daysReuse the method in a new mechanical component.

References and source discipline

This course synthesizes original MechCompass explanations from established reliability engineering sources. Confirm current editions and standards before using them for formal design approval.

Topic in this module	Reference direction
Distribution fitting and probability plots	NIST/SEMATECH Engineering Statistics Handbook
Reliability distribution models	Elsayed, Reliability Engineering, 3rd edition
Model uncertainty and reliability inputs	Haldar and Mahadevan, Probability, Reliability, and Statistical Methods in Engineering Design