Engineering Statistics and Reliability Data Formula Sheet

This formula sheet collects engineering statistics and reliability data formulas for estimation, test planning, experimental design, model validation, life data, censored observations, and decision evidence. Use it to make calculations traceable, not to hide weak assumptions behind statistical notation.

Every formula needs a decision boundary: what population is represented, what measurement is used, what error is costly, and what action follows from the result.

Symbols and Conventions

Common symbols:

Symbol	Meaning
n	sample size
x_i	measured observation
\bar{x}	sample mean
s	sample standard deviation
\mu	population mean
\sigma	population standard deviation
SE	standard error
z	standard normal multiplier or z-score
t	Student t multiplier
p	probability or proportion
\lambda	failure rate for an exponential model
T	total exposure time, cycles, distance, or transactions
R(t)	reliability over mission time t
\beta	Weibull shape parameter
\eta	Weibull scale parameter

Keep units attached to the engineering variable. Statistical multipliers are dimensionless; the interval or estimate has the same unit as the measured quantity.

Mean, Variance, and Standard Deviation

Sample mean:

Sample variance:

Sample standard deviation:

Use n-1 for the usual unbiased sample variance estimate. Before interpreting s, check whether observations are independent, representative, and measured with the same method.

Weighted Mean

For observations with weights w_i:

Weighted means are useful when measurements have different exposure, duration, uncertainty, or population weight. They are misleading if the weights are chosen after seeing the result.

Z-Score

For an observation x:

For sample-based screening:

Use z-scores only when the distribution and spread are meaningful for the decision. A large z-score can indicate a real shift, a special-cause event, a measurement error, or a model mismatch.

Standard Error of a Mean

If observations are independent:

Standard error measures uncertainty in the mean estimate, not the spread of individual observations. A small standard error does not imply that every part, patient, packet, sensor, or test article is close to the mean.

Confidence Interval for a Mean

When s is estimated from data:

For large samples or known \sigma:

Worked Check

A calibration offset has:

Use:

Half-width:

Interval:

Engineering Comment

If the allowable mean offset is 0.50\ \text{mm}, this interval crosses the limit. The result is not a clean release; it calls for correction, more representative data, or a guarded decision rule.

Confidence Interval for a Proportion

For a simple large-sample screen:

Approximate interval:

This approximation is weak for small samples, rare failures, or proportions near 0 or 1. Use an exact or adjusted method when release risk is high.

Sample Size for Mean Precision

To estimate a mean with target half-width h:

Round up.

Worked Check

If:

target half-width is:

and:

then:

Round up:

Engineering Comment

This assumes stable variance and representative sampling. If results differ by operator, lot, temperature, site, or duty cycle, the design should be blocked or stratified rather than simply enlarged.

Sample Size for Proportion Precision

For a target half-width h:

If p is unknown, use p=0.5 for a conservative maximum variance screen.

Difference Between Two Means

For independent samples:

Standard error:

Signal-to-noise screen:

This is not a full test unless the assumptions and degrees of freedom are handled. It is useful for quick engineering review of whether a claimed improvement is large relative to noise.

Guarded Acceptance with Measurement Uncertainty

For a symmetric limit L on absolute error and expanded uncertainty U:

Accept only when:

Worked Check

If:

then:

A measured error of:

fails the guarded rule even though it is below the nominal 3.0\% limit.

Engineering Comment

Guard bands reduce false acceptance risk but can increase false rejection. The project should define the rule before testing.

Process Capability

For two-sided specification limits:

Upper capability:

Lower capability:

Overall centered capability:

Capability values require a stable process, credible distribution, adequate measurement system, and specification limits that actually represent engineering requirements.

Factorial Effect Estimate

For a two-level factor A coded low/high, the main effect can be screened as:

For an interaction between factors A and B:

where AB+ means runs where the product of coded factor signs is positive.

Randomization, replication, blocking, and pre-defined response variables matter more than the arithmetic. A factorial effect from a biased run order can be a temperature drift, tooling drift, operator sequence, or battery state rather than a real factor effect.

Linear Regression

Simple linear model:

Least-squares slope:

Intercept:

Predicted value:

Regression should be checked with residuals, leverage, operating range, measurement uncertainty, and physical plausibility. A good fit inside a test range is not proof outside that range.

Bias and RMSE

Prediction residual:

Mean bias:

Root mean square error:

Bias shows systematic offset. RMSE combines bias and scatter. For validation, compare both with the engineering margin.

Failure Rate and MTBF from Exposure

For N_f failures over total exposure T:

For a constant-rate model:

Use consistent exposure units: operating hours, cycles, starts, transactions, distance, or mission count. Do not mix calendar time with operating exposure without a defensible conversion.

Exponential Reliability

For constant failure rate:

Probability of failure by time t:

Mean time to failure:

This model is appropriate only when the failure rate is approximately constant over the use interval. It does not represent early-life screening failures or wear-out unless the interval is limited and justified.

Zero-Failure Demonstration Screen

For zero observed failures over exposure T, a one-sided confidence lower bound on MTBF under an exponential model can be screened as:

where C is confidence level.

Worked Check

Ten units run:

with zero failures. Total exposure:

For:

Therefore:

Engineering Comment

This is not a field reliability guarantee. It depends on the exponential assumption, failure definition, unit representativeness, test stress, censoring, and whether the released configuration matches the tested configuration.

Weibull Life Data

Weibull cumulative failure probability:

Reliability:

Hazard rate:

Interpretation:

• \beta<1 suggests decreasing hazard, often early-life or screening-dominated failures;
• \beta=1 gives constant hazard, equivalent to exponential life;
• \beta>1 suggests increasing hazard, often wear-out or degradation.

The interpretation is engineering evidence, not automatic truth. Check failure modes, censoring, operating conditions, and confidence bounds.

Censored Data Exposure

Right-censored units contribute exposure even when they have not failed. A simple exponential failure-rate estimate with complete exposure is:

where failed and censored units both contribute observed time t_i.

Worked Check

Four devices are observed:

Unit	Observation
failed	120\ \text{h}
failed	180\ \text{h}
still running	240\ \text{h}
still running	240\ \text{h}

Failures:

Exposure:

Failure rate estimate:

MTBF estimate:

Engineering Comment

Discarding censored units would underuse valid exposure. Treating censored units as failures would be overly pessimistic. The right method depends on the censoring mechanism and life model.

Monte Carlo Propagation

For a model:

Monte Carlo simulation draws random samples from input distributions and evaluates:

Estimated mean:

Estimated probability of exceeding a limit L:

Monte Carlo results are only as valid as the input distributions, correlations, model form, sample count, and validation evidence.

Queueing and Waiting-Time Screen

For a stable single-server M/M/1 screen:

Average number in system:

Average time in system:

This model assumes Poisson arrivals, exponential service times, one server, first-in-first-out behavior, and steady state. Use it as an order-of-magnitude check for test labs, service queues, packet systems, inspection stations, or maintenance response.

Validation Acceptance Metrics

Normalized error:

Coverage check for uncertainty:

where U_{combined} should include model uncertainty, measurement uncertainty, input uncertainty, and numerical uncertainty when relevant.

Validation should specify:

1. operating range covered;
2. independent data source;
3. acceptance threshold;
4. residual pattern review;
5. uncertainty basis;
6. action if the model passes only part of the range.

Common Mistakes

Common engineering statistics mistakes include:

• reporting a mean without spread, confidence, or measurement uncertainty;
• using more samples to compensate for biased sampling;
• treating repeated measurements on one unit as if they represented many units;
• using a normal model for bounded, censored, rare-event, or strongly skewed data without checking;
• using MTBF without failure definition, exposure basis, confidence, or failure-mode context;
• ignoring censored units in reliability data;
• accepting a regression model outside the tested range;
• using Monte Carlo output without validating input distributions and correlations.

Formula Selection Checklist

Before using a statistical result for release or design change, confirm:

1. the engineering decision is stated;
2. the measured variable and units are defined;
3. the population and sampling method match the decision;
4. the measurement system is adequate;
5. the assumptions behind the formula are credible;
6. uncertainty or confidence is compared with engineering margin;
7. the result changes an action: release, hold, retest, redesign, monitor, or escalate.