Diagnostics · Prognostics · Reference
Every piece of jargon used across the diagnosis series — fault-finding, statistics, reliability, prognostics, and the systems that run them — in plain English.
Around two hundred terms, grouped by theme and defined in one plain sentence each — enough to follow any post in the series without a textbook beside you. Where a word has a precise technical meaning, the gloss gives the working idea, not the formula. Jump to a section, or read it as a tour of the field.
The basic vocabulary of things breaking.
Fault — an abnormal condition in a component or system; the thing that is wrong (not yet necessarily a loss of function).
Failure — the moment a fault stops the system doing its job; the loss of intended function.
Error — a deviation between actual and correct behaviour; a fault that has started to affect the output.
Defect — a flaw introduced in design or manufacture that can later become a fault.
Symptom — an observable effect of a fault (a code, a reading, a complaint); evidence, not the cause.
Anomaly — any reading that departs from normal/expected behaviour, whether or not a fault is behind it.
Root cause — the deepest controllable reason the failure happened; fixing it prevents recurrence.
Proximate cause — the immediate trigger nearest the failure (often a symptom of the root cause).
Contributing cause — a factor that made the failure more likely or worse, but isn’t the root.
Mechanism — the physical chain of events by which a cause produces the failure.
Common-cause / common-mode failure — one shared cause that takes down several parts (or redundant copies) at once.
Cascading failure / fault propagation — one fault that triggers others down the line.
Latent (dormant) fault — a fault already present but not yet showing, waiting for the right condition.
Incipient fault — an early-stage fault still growing toward failure (the target of prognostics).
Intermittent fault — a fault that appears and disappears, often hiding between measurements.
Hard vs soft fault — an abrupt, total break vs a gradual drift or partial loss.
Phantom fault — a code or alarm with no real underlying fault (often a bad detector or sensor).
The disciplined reasoning the first posts build.
Provenance grading — judging what each piece of evidence can and can’t establish, and how far to trust it, before using it.
Trust tier — a high/medium/low reliability label on a source.
Abduction (inference to the best explanation) — reasoning from a surprising symptom to the hypothesis that best explains it.
Hypothesis — a candidate explanation, ideally written as a testable causal chain.
Causal chain — a hypothesis stated as root cause → mechanism → observable signature.
Signature — the specific data pattern a given cause would produce; what you go and check.
Differential diagnosis — the short list of competing hypotheses you are deciding among.
Impostor — a correlation that mimics the cause but is really a co-effect of the true cause.
Referee — an independent confirmation: physics/a model regenerating the signature, or tracing the code path.
Capability vs occurrence — code or physics shows what can happen; you must still prove it did under the recorded conditions.
Fusion (weight, not vote) — combining evidence by trust × strength, counting independent items only.
Independence / double-counting — many reports of one cause are one fact, not many; only independent evidence multiplies.
Convergence / triangulation — accepting a cause only when several independent lines all point to the same mechanism.
Sufficiency gate — the check for whether you actually have enough evidence to conclude.
Discrimination / observational equivalence — whether your data can tell two hypotheses apart; “equivalent” means it can’t.
Discriminating measurement — the one observation that reads differently under the rival hypotheses.
Coverage / orphan anomaly — whether the explanation accounts for everything; an unexplained leftover is an “orphan” (often a second fault).
Robustness (leave-one-out) — does the verdict survive deleting its single strongest clue?
Three exit states — Confirmed (one cause clears every check), Inconclusive (collect a specific next measurement), Incoherent (re-audit a source or add a hypothesis).
Kill-switch (disconfirming test) — the single observation that, if it came out wrong, would overturn the conclusion.
Audit trail — the evidence, tool outputs, and approvals attached to a verdict so it can be re-litigated.
The control-engineering machinery beneath on-board and offline detection.
Fault detection — deciding that something is wrong.
Fault isolation — deciding which component is wrong.
Fault identification — characterising the size/type of the fault.
Detectability / isolability / diagnosability — whether your sensors can, in principle, detect / locate / distinguish a given fault.
Residual — the gap between a measured signal and what a model says it should be; near zero when healthy, nonzero under a fault.
Observer (e.g. Luenberger, Kalman) — a model that estimates a system’s internal state from its inputs and outputs, used to form residuals.
Parity relation / parity space — a redundancy equation among signals that should hold when healthy; its violation flags a fault.
Analytical redundancy — using a model (instead of a duplicate sensor) to cross-check a measurement.
Hardware redundancy / voting / TMR — duplicate sensors and a majority vote; Triple Modular Redundancy = three units, vote 2-of-3.
Structural analysis — using the system’s equation structure to find which faults are detectable and isolable.
Fault-symptom (signature) matrix — a table of which symptoms each fault produces; used to isolate the culprit.
Consistency-based diagnosis — find the set of faulty parts that makes the observations consistent again.
Conflict set — a group of components that cannot all be healthy given what was observed.
Minimal diagnosis / candidate / hitting set — the smallest set of faulty parts that explains every conflict.
Plausibility / range / rate check — cheap on-board tests that reject impossible or impossibly-fast readings (a likely sensor fault).
Threshold / hysteresis / debounce — the trip level / a gap to avoid chattering / waiting before acting on a crossing.
False positive / false alarm — flagging a fault that isn’t there.
False negative / missed detection — failing to flag a fault that is there.
Sensitivity / specificity — fraction of real faults caught / fraction of healthy cases correctly cleared.
ROC curve — the trade-off curve between catching faults and raising false alarms as you move the threshold.
The tools that decide whether a pattern is real and whether you know enough.
Prior / posterior — your belief before / after seeing a piece of evidence.
Likelihood / likelihood ratio — how probable the evidence is under a hypothesis / how much more probable under one than another; the engine of updating.
Bayesian vs frequentist — reasoning about the probability of hypotheses given data vs about long-run frequencies of data.
Bayes factor / marginal likelihood (evidence) — the weight of the data for one model over another / the data’s overall probability under a model.
p-value — the chance of seeing a result this extreme if nothing were really going on; small = surprising under “no effect.”
Significance level α / power (1−β) — the false-alarm rate you accept / the chance of catching a real effect.
Type I / Type II error — a false alarm / a missed real effect.
Confidence interval (CI) — the plausible range around an estimate; tight = lots of data.
Chi-square test — a test for whether a difference between groups in a table is more than chance.
Calibration / Brier score — whether stated confidence matches reality / a score for how well-calibrated probabilities are (lower is better).
Entropy — a number for how uncertain you are; maximal when all options are equally likely.
Mutual information / information gain / KL divergence — how much one variable tells you about another / how much a test cuts uncertainty / a measure of how far two distributions differ.
Value of information (VoI) — how much a candidate measurement would shrink uncertainty; collect the highest.
SPRT (sequential probability ratio test) — accumulate evidence and stop the instant it crosses a decision threshold.
CUSUM / EWMA / Shewhart chart — running monitors that accumulate / smooth / spot-check deviations to catch a shift early.
Change-point detection — finding the moment a series’ behaviour changed.
Multiple comparisons — testing many things inflates false hits; the “look-elsewhere” problem.
Bonferroni / FDR (Benjamini–Hochberg) — corrections that raise the bar when you ran many tests; FDR controls the false-discovery rate.
Overfitting / cross-validation — a model that memorises noise / testing on held-out data to catch that.
The terms that separate a real cause from a convincing coincidence.
Correlation ≠ causation — two things moving together needn’t mean one causes the other.
Confounder (confounding variable) — a hidden common cause of both things you’re comparing, making them look related; the classic impostor-maker.
Driver — a variable that actually drives an outcome or rate (a causal/risk/cost driver); the thing worth acting on.
Mediator — a variable on the causal path between cause and effect (cause → mediator → effect).
Moderator / effect modification / interaction — a factor that changes how strong a cause’s effect is; an effect that only appears when two factors combine.
Collider — a variable caused by two others; controlling for it creates a false association (the collider trap).
Covariate — any background variable you measure and may adjust for.
Adjustment / controlling for / stratification — holding a confounder fixed (statistically, or by splitting into subgroups) so a comparison is fair.
Partial correlation — the correlation that remains after removing a third variable’s influence; the test that unmasks an impostor.
Propensity score / matching — comparing treated and untreated units with similar profiles to mimic a fair experiment.
Spurious correlation — an association with no causal link (chance, or a shared cause).
Regression to the mean — extreme readings tend to be followed by more normal ones, faking an “improvement.”
Base rate / base-rate fallacy — how common something is overall / ignoring it and over-reading a test result.
Simpson’s paradox — a trend that reverses once you look inside the subgroups.
Berkson’s paradox — a false negative correlation created by how the sample was selected.
Distortions baked into how evidence was collected — they bend conclusions even with perfect math.
Selection bias — your sample isn’t representative because of how it was chosen (e.g. only warranty claims).
Survivorship bias — you only see the units that “survived” to be observed; the worst cases went silent and are missing.
Reporting bias — only some events get reported (vivid, severe, or engaged-owner cases), skewing rates.
Lead-time / immortal-time bias — time-related traps in survival data that make a group look better or worse than it is.
Confirmation bias — favouring evidence that fits the hypothesis you already hold.
Anchoring — over-weighting the first number or idea you encountered (e.g. the DTC’s text).
Premature closure — stopping at the first plausible cause before testing it.
Double-counting — treating correlated copies of one fact as independent evidence.
The heart of prognostics: the conditions that drive damage, and the damage itself.
Stressor — an operating or environmental condition that drives degradation (temperature, current, voltage, vibration, humidity, cycling, load).
Dominant stressor — the one stressor that contributes most to a given failure mode; the lever a prognostic model leans on.
Load / loading — the demand placed on a component (mechanical, electrical, thermal); high load usually means faster wear.
Duty cycle — the pattern of on/off, high/low operation over time.
Usage / mission profile — the representative mix of conditions a unit actually experiences in service.
Operating envelope — the full map of conditions a system runs in; a fault may live in only one corner of it.
Condition coverage — whether your data captured the fault under the condition that triggers it — “rich enough” isn’t about count.
Degradation — the gradual, monotonic worsening of a component toward failure.
Degradation path / trajectory — the curve a health measure traces over time toward the failure threshold.
Damage accumulation / Miner’s rule — the idea that damage piles up with use; Miner’s rule sums fractional fatigue from each load cycle.
Stress–strength interference — failure happens when the applied stress overlaps the part’s strength distribution.
Fatigue / wear / corrosion — failure by repeated loading / surface loss by rubbing / chemical attack of a material.
Calendar aging vs cycle aging — degradation that accrues with time vs with use (e.g. battery shelf-aging vs charge cycles).
Capacity fade / internal-resistance rise — a battery holding less charge / impeding current more, the two classic aging signatures.
Lithium plating / SEI growth / dendrite — battery degradation mechanisms: metallic lithium deposits / a growing surface film / needle-like growths that can short a cell.
Health indicator / condition indicator / feature — a quantity computed from raw signals that tracks how degraded a unit is.
Health index — a single 0–100%-style score combining several indicators (e.g. State of Health).
Baseline / nominal behaviour — what a healthy unit looks like, against which anomalies and degradation are measured.
Failure threshold / end-of-life (EOL) — the health level at which the unit is declared failed (e.g. 70% State of Health).
The math of how long things last and how long this one has left.
Prognostics — predicting future failure (versus diagnostics, which explains the present).
PHM (Prognostics and Health Management) — the discipline of monitoring health, predicting failure, and acting on it.
RUL (Remaining Useful Life) — how much working life a unit has left before it crosses the failure threshold.
Prognostic horizon — how far ahead a prediction stays trustworthy.
Model-based / data-driven / hybrid prognostics — RUL from a physics model / from learned patterns in data / from a blend of both.
Physics-of-failure (first principles) — predicting life from the actual degradation physics, not just statistics.
Surrogate model — a fast approximation that stands in for an expensive physics simulation.
Digital twin — a live model of a specific unit, updated with its data, used to simulate and forecast it.
Failure rate (λ) / bathtub curve — failures per unit time / the classic shape: high infant-mortality, flat middle, rising wear-out.
Hazard rate — the chance a still-healthy unit fails in the next instant.
Reliability / survival function R(t) — the probability a unit is still working at time t.
Weibull distribution (shape β, scale η) — a flexible life model; β<1 = infant mortality, β≈1 = random, β>1 = wear-out.
B10 / B50 life — the age by which 10% / 50% of units have failed (B50 = median life).
Censoring — units that haven’t failed yet are partial information (“survived at least this long”), not zeros.
Kaplan–Meier / Cox proportional hazards — a non-parametric survival curve / a model relating hazard to covariates.
MTTF / MTBF / MTTR / availability — mean time to failure / between failures / to repair / fraction of time a system is up.
Kalman / particle filter (for RUL) — recursive estimators that track a degradation state and project it forward with uncertainty.
Prediction interval / confidence bounds — the uncertainty band around an RUL estimate (not just a single number).
Drift — slow change over time in a sensor (sensor drift) or in the data distribution a model sees (concept drift).
The shop-floor and design-office toolkit.
Root-cause analysis (RCA) — the disciplined search for the deepest fixable cause of a problem.
Five Whys — asking “why” repeatedly to walk a symptom back to its root.
Fishbone / Ishikawa diagram — a cause-and-effect diagram organising candidate causes by category.
Fault tree analysis (FTA) / event tree — a top-down logic tree of how a failure can occur / a forward tree of how an initiating event unfolds.
FMEA / FMECA (DFMEA, PFMEA) — a structured catalogue of failure modes and their effects (with Criticality); Design and Process variants.
RPN (severity × occurrence × detection) — a Risk Priority Number ranking which failure modes to fix first.
Reliability block diagram — a diagram of how components’ reliability combines (series/parallel) into the whole.
Redundancy / derating — extra spare capacity for fault tolerance / running parts below their limits to extend life.
Burn-in / HALT / HASS — early stressing to weed out infant-mortality units / Highly Accelerated Life (and Stress Screen) testing.
NFF / CND — “No Fault Found” / “Could Not Duplicate” — a returned part that tests fine (often an intermittent or a phantom).
Containment → correction → prevention — stop the bleeding / fix this batch / make it impossible — three distinct actions.
CAPA / 8D / PDCA — Corrective-and-Preventive-Action; structured problem-solving (8 Disciplines; Plan-Do-Check-Act).
Poka-yoke (mistake-proofing) — a design that makes the error physically impossible.
Control plan / SPC / Cp, Cpk / control limits — the plan that holds a process in spec / Statistical Process Control / capability indices / the chart’s action lines.
Escape point — the test or check that should have caught the fault but let it through.
Pareto (80/20) — the rule that a few causes account for most of the problems; focus there first.
Warranty / field return / recall / campaign / TSB — post-sale failure data / a returned part / a fix program; a remediation campaign; a Technical Service Bulletin to dealers.
Borrowed largely from epidemiology — reasoning about a distribution, not a unit.
Cohort — a defined group of vehicles sharing attributes (plant, firmware, lot, climate).
Cohort boundary — the build attribute that cleanly separates affected from healthy cars — it usually names the cause.
Case-control — comparing affected units (cases) against matched healthy ones (controls).
Prevalence / incidence — how common a fault is now / how many new cases appear per period.
Exposure / population at risk / denominator — the suspected stressor a group experienced / the cars that could fail / what you divide by to get a rate.
Stratification / Mantel–Haenszel — comparing within subgroups / combining subgroup rates into one number without the Simpson trap.
Mixture / de-mixing — one code hiding several causes / splitting the population into homogeneous sub-groups before diagnosing.
Clustering / BIC / soft assignment / noise tail — grouping cars by signature / a criterion that picks how many clusters / letting a car belong partly to two / the handful that belong to neither.
Surveillance — always-on monitoring of fleet rates to catch a problem while it is small.
Onset curve / change-point — how the rate rises over time / the moment it changed, often tied to an OTA or a build change.
Two-level aggregation / hierarchical model — summarise per car first, then reason across cars (one car = one vote).
Remediation scope — the decision of who gets fixed and how: OTA, targeted recall, or monitor-only.
Prevented loss vs campaign cost / mandatory-recall threshold — the economic trade-off behind a campaign / the safety line above which a recall is required regardless of cost.
Verify-the-fix (prevalence bend) — confirming a campaign worked by watching the treated cohort’s rate bend down.
False-recall rate — how often the system urges a campaign against a fault that isn’t real — the costliest error.
Turning “the curve bent after we shipped” into real proof.
DAG (directed acyclic graph) — a diagram of what causes what, used to decide what to adjust for.
do-operator / do-calculus — the math of an intervention (“if we force X”) as opposed to mere observation.
Back-door path — a confounding route between cause and effect that you must block to get a clean estimate.
Randomized experiment (A/B test) — randomly assigning treatment vs control so the difference is a clean causal effect.
Treatment / control group — the units that get the change / the comparison units that don’t.
Staged rollout / canary — releasing a fix to a small slice first; if the order is randomized, it’s a real experiment.
Difference-in-differences (DiD) / parallel-trends — comparing treated vs untreated before/after to cancel shared trends / the assumption it relies on.
Instrumental variable / natural experiment — using an as-good-as-random nudge to recover causation when you can’t randomize.
Intent-to-treat — analysing units by the group they were assigned to, even if some didn’t comply.
Guardrail metric / ramp / stopping rule — a safety metric watched during a rollout / the gradual % increase / the pre-set condition to halt or proceed.
The embedded and automotive vocabulary.
ECU — Electronic Control Unit; the computer running and watching a vehicle subsystem.
DTC (Diagnostic Trouble Code) — the standardised fault code an ECU stores (e.g. P0AE0); a symptom pointer.
Freeze-frame / snapshot — conditions captured at the moment a code set / a richer recorded data window around an event.
MIL — Malfunction Indicator Lamp; the dashboard warning light.
Limp mode / derate — the ECU restricting performance to protect the system after a fault.
UDS / OBD (ISO 14229 / SAE J1979) — the diagnostic protocols a scan tool uses to read codes and data.
CAN / CAN-FD / J1939 / DBC — the in-vehicle network / a faster version / the heavy-duty higher-layer protocol / the file that decodes its messages.
Fault maturation / debounce / healing counter — requiring a fault to persist over several cycles before a code is stored, and clearing it after enough clean cycles.
False-alarm budget — the strict limit on crying wolf that forces on-board detection to be conservative.
ISO 26262 / ASIL / functional safety / safety goal — the automotive safety standard / its risk-integrity levels / designing against hazards / a top-level safety requirement.
OTA / firmware / calibration — a remote update / the embedded software / the tunable parameters in it.
VIN / build record / supplier lot / production date — the unique vehicle ID / its as-built attributes / the batch a part came from / when it was made.
BMS / SOC / SOH — Battery Management System / State of Charge (how full) / State of Health (how degraded).
The plumbing and the multi-agent product that run the method at scale.
Telemetry / time series — streamed sensor data from vehicles / measurements ordered in time.
VIN spine / entity resolution / record linkage (Fellegi–Sunter) — joining all records on the vehicle ID / matching records that disagree on identity / the classic fuzzy-matching framework.
Lakehouse (bronze → silver → gold) — the data platform; raw → cleaned → analysis-ready (“one row per car”) layers.
Feature store / vector store / knowledge graph / ontology — stores for model features / document embeddings (semantic search) / a graph of parts-faults-fixes / its formal vocabulary.
Semantic / metrics layer — one shared definition of terms like “prevalence” so humans and tools compute them identically.
Lineage — the recorded trail of where each data value came from, so a result can be re-run.
Missingness (MCAR / MAR / MNAR) / imputation — why data is absent (at random, or not) / filling gaps — risky when missingness is biased.
Pipeline / ETL–ELT / batch vs stream — the automated data flow / extract-transform-load steps / processing in chunks vs continuously.
RAG (retrieval-augmented generation) / embeddings — grounding a model’s answer in retrieved documents / numeric representations that make text searchable by meaning.
Agent / orchestrator / worker / blackboard — an AI that acts / a planner that routes work / a single-job agent / the shared notebook (the case) they read and write.
MCP (Model Context Protocol) — the open standard for giving an AI model tools and data.
Human-in-the-loop (HITL) / gate / guardrail / prompt injection — a person approves irreversible steps / the approval point / safety rules / an attack where untrusted text smuggles in instructions.
Eval harness / synthetic fleet / ground truth / IoU — the rig that grades the pipeline / fake data with a known injected cause / the known answer / “Intersection over Union,” how well a found set matches the true one.
Precision / recall — of the flags raised, how many were right / of the real cases, how many were caught.
Hallucination / groundedness — a model asserting something unsupported / whether outputs are backed by real evidence.
Observability / tracing / SLO / drift monitoring — seeing what the system did / step-by-step logs / a service-level target / watching for the data or model going stale.
Durable execution / budget governor / model router — a workflow that can pause and resume / a cap on cost per case / sending easy steps to a cheap model and hard ones to a frontier model.
Memory: working / episodic / semantic / procedural / long-term — scratchpad / past-case library / knowledge graph + doctrine / playbooks / confirmed facts.
Write-gate (memory governance) — only human-confirmed, verified facts graduate into durable memory, so the system can’t poison itself.
Multi-tenancy / RBAC — isolating each customer’s data / Role-Based Access Control on who can do what.