Fleet-Scale Root-Cause Analysis: One Code, a Population of Evidence

This is the fifth and final post in the diagnosis set, after Root-Cause Analysis Under Uncertain, Heterogeneous Evidence, When Is the Evidence Enough?, and Harder Cases in Diagnosis. Those treated a single vehicle. This one scales up: the same trouble code appears across thousands of cars, and the rest of the evidence — telemetry summaries, warranty claims, service notes, build records — arrives as one big batch.

That sounds like more of the same, but it is a different problem. With one car you have a single observation and you ask “what broke?” With a fleet you have a distribution, and a valid explanation has to account for three things at once: how common the failure is, who it happens to, and when it started spreading. Those extra constraints make impostors much easier to kill — but they hide a new trap, the one that catches every fleet investigation: the same code is rarely the same cause. A common DTC across a fleet is usually a mixture of several mechanisms wearing one mask, and if you diagnose the batch as a single failure you will blend two real causes into one fictitious one.

So the fleet method has a different first move — de-mix the population — and a different last move: turn the diagnosis into a remediation scope (who gets an over-the-air fix, who gets a recall, who gets nothing). In between, the cohort contrast that was a side-check on one car becomes the main engine.

What it covers

Thirteen sections, about twenty-six minutes, in plain language.

§ 1 — From a car to a population. Why a fleet hypothesis must explain prevalence, stratification, and onset — and why that makes diagnosis easier, not harder.

Part A · One code, many causes

§ 2 — The mixture trap. Why one DTC across the fleet is usually several root causes, and what averaging them costs you.

§ 3 — De-mix first. Clustering the batch into homogeneous sub-populations and running the method per cluster, with an algorithm and a block diagram.

Part B · The cohort contrast is the engine

§ 4 — Let the boundary name the cause. Stratifying affected vs. spared by build plant, firmware, lot, and climate — and reading the split.

§ 5 — Confounding & Simpson’s paradox. Why a fleet-level correlation can reverse inside every subgroup, and why stratification is now mandatory.

Part C · Aggregating batched evidence

§ 6 — Two levels, not one. Summarizing per vehicle first, then reasoning across vehicles — and not letting volume manufacture false confidence.

§ 7 — Messy by default. Heterogeneous trust, missing data, and the selection and survivorship biases that warranty batches carry.

Part D · Time and the decision

§ 8 — Read the curve. Prevalence and onset over time, and change-points that pin the cause to a production change or an OTA.

§ 9 — From cause to campaign. Forecasting how many more will fail, and choosing the remediation scope — OTA, recall, or targeted service.

§ 10 — The fleet procedure. The whole method as one flowchart, with the master algorithm.

§ 11 — A worked scenario. A 6,000-vehicle fleet, one charging code, two hidden causes — start to finish.

§ 12 — Practical notes. Field rules for population-scale diagnosis.

§ 13 — References & further reading. Epidemiology, reliability, mixtures, and causal inference, with pointers.

Read it

Open the monograph →

The monograph lives at its own URL in the warm-paper layout of the series, with a mixture diagram, a stratification tree, a Simpson’s-paradox panel, a two-level aggregation block diagram, prevalence and onset curves, a remediation-decision figure, the fleet RCA flowchart, three algorithms, and a worked 6,000-vehicle scenario.

Where it sits

The diagnosis series, in order:

I · The core method

II · At fleet scale

Fleet-Scale Root-Cause Analysis — (this post)
Fleet RCA, Made Rigorous

III · Proof, foundations & prevention

Full map: The Diagnosis Series index.

← Back to Autonomy