2. Reliability data types and sources

Reliability engineering is only as good as the data feeding it. Before any Weibull fit, acceleration model, or remaining-useful-life (RUL) prediction can be trusted, you must know what kind of data you have, where it came from, and what it can and cannot tell you. This chapter maps the six dominant data types encountered in real reliability programs — time-to-failure, censored data, degradation signals, field returns, warranty claims, and sensor telemetry — and shows how each is structured, collected, and interpreted.

Time-to-Failure Data: The Classical Backbone

Time-to-failure (TTF) data records the elapsed time until a unit fails, measured in cycles, hours, kilometers, or any meaningful usage metric. It is the canonical input to life distributions such as Weibull, lognormal, and exponential. A TTF record minimally contains a unit identifier, a start time, a failure time, and a failure mode.

The critical discipline with TTF data is defining the time origin and the usage metric consistently. A bearing that fails after 8,000 operating hours at 1,500 rpm is not comparable to one that fails after 8,000 hours at 3,000 rpm unless you convert to revolutions. Always record the raw usage counter alongside calendar time.

unit_id,start_ts,fail_ts,usage_metric,usage_value,failure_mode
BRG-0142,2024-01-03T08:00,2024-09-11T14:22,revolutions,7.2e8,outer_race_spall
BRG-0143,2024-01-03T08:00,,revolutions,6.1e8,censored_running

Censored Data: When Failure Hasn't Happened Yet

In practice, most units have not failed by the time you analyze the data. Ignoring them biases life estimates downward severely. Censoring is the formal mechanism for including partial information.

• Right-censored: the unit survived past its last observation time. This is the most common case — a test ends, a fleet is still running, or a customer simply hasn't reported a failure.
• Left-censored: failure occurred before the first inspection, so the exact time is unknown but bounded above.
• Interval-censored: failure occurred somewhere between two inspection points, common with periodic maintenance.

Every censored record still contributes information: it tells the likelihood function that the unit survived at least that long. In a Weibull maximum-likelihood fit, right-censored units enter the likelihood as the survival function S(t) = exp(-(t/η)^β) rather than the density f(t). Dropping them is one of the most frequent and damaging errors in reliability analysis.

Degradation Signals: Measuring the Path to Failure

Many failures are preceded by a measurable drift: battery capacity fades, bearing vibration amplitude rises, insulation resistance drops, LED lumen output decays. Degradation data captures this continuous or periodic signal, enabling prognostics rather than just life statistics.

A degradation dataset is a time series per unit, with a defined failure threshold. For example, a lithium-ion cell might be considered failed when capacity falls below 80% of nominal. The time at which the fitted degradation path crosses that threshold becomes a pseudo-failure time, which can then feed TTF models.

unit_id,cycle,capacity_Ah,internal_resistance_mOhm
CELL-07,0,3.00,22.1
CELL-07,200,2.81,26.4
CELL-07,400,2.58,31.9
CELL-07,600,2.31,39.7

Degradation modeling typically fits a parametric path (linear, exponential, or power-law) per unit, then treats the path parameters as random variables across the population. This yields RUL distributions conditioned on the current measurement — the foundation of predictive maintenance.

Field Returns and Warranty Claims

Field returns are physical units shipped back from customers, often accompanied by a failure report. They are invaluable because they allow failure analysis (FA) to confirm the actual root cause, distinguishing true failures from no-fault-found (NFF) cases. A high NFF rate (often 30–60% in electronics) signals that your field failure definition is too loose or that diagnostics are inadequate.

Warranty claims are the transactional record: claim date, repair action, part replaced, mileage or usage, and sometimes a symptom code. They are large-volume and cheap to obtain, but noisy — a claim may reflect a customer complaint rather than a confirmed failure. Warranty data is typically left-truncated (you only see units that failed within the warranty window) and subject to reporting delays.

Sensor Telemetry: The Modern Data Firehose

Connected products stream telemetry — temperature, vibration, current, pressure — at high frequency. This enables condition monitoring and early fault detection, but introduces challenges: missing data, clock drift, sensor calibration drift, and sheer volume. Telemetry is best used to derive features (RMS vibration, kurtosis, spectral peaks) that correlate with degradation, rather than raw signals fed directly into life models.

Choosing and Combining Sources

The right data strategy depends on the question. For population life estimation, TTF plus censored data suffices. For prognostics, you need degradation or telemetry. For root-cause and corrective action, field returns with FA are essential. Mature programs triangulate: telemetry detects anomalies, warranty claims flag population trends, and field returns confirm mechanisms.

Practical Data-Quality Checks

Before analysis, run a small set of sanity checks on any reliability dataset:

1. Time-origin consistency: confirm every unit's start time uses the same reference (manufacture date vs. installation date).
2. Censoring flags: verify that every record has an explicit event indicator (1 = failed, 0 = censored).
3. Usage vs. calendar time: check correlation; if usage varies widely, model in usage units.
4. Duplicate and orphan records: warranty claims sometimes duplicate; telemetry sometimes loses device IDs.
5. Threshold definition: for degradation data, confirm the failure threshold is documented and applied uniformly.

These checks take minutes but prevent the most common analytical failures. A dataset that passes them is ready for distribution fitting, degradation modeling, or prognostics — and the conclusions drawn from it will be defensible.