Anomaly detection tells you something is wrong. A classifier tells you which fault it is. In vehicle health monitoring, the model that finally puts a name on the fault is almost always a tree ensemble — a Random Forest or a gradient-boosted model like XGBoost. This post builds them up from the single decision tree, in plain words, and runs one worked example all the way through.

What it covers

Fourteen sections, about twenty-five minutes, in simple language with diagrams, code, and the arithmetic shown.

Part A · From one tree to a crowd

§ 2 — One decision tree (CART). The flowchart the computer writes for itself: splits, purity (Gini/entropy), why a single tree is readable but overfits — the interpretable baseline.

§ 3 — Why a crowd. The one idea behind ensembles — errors that disagree cancel out — and the two ways to build one: bagging (kills variance) and boosting (kills bias).

Part B · The models

§ 4 — Random Forest. Bootstrap bags plus random feature subsets; the vote; out-of-bag validation for free; the safe default.

§ 5 — Boosting. AdaBoost and gradient boosting — trees that fix each other’s mistakes — and how XGBoost, LightGBM, and CatBoost differ, with a when-to-use table.

§ 6 — Extra Trees. The even-more-random forest: faster to train, often just as accurate.

§ 7 — Choosing & tuning. A “which model do I pick?” decision flowchart and a hyperparameter cheat-sheet with VHM defaults.

§ 8 — Feature importance. Gini vs. permutation vs. SHAP — turning a label into a lead, and the confounder caveat.

Part C · In practice

§ 9 — Worked example. Classifying the P0AE0 charging fault into healthy / firmware bug / connector lot, with runnable Python, a confusion matrix, and the importance ranking.

§ 10 — From score to decision. Calibration, and choosing the threshold by the dollar cost of a miss vs. a false flag.

§ 11 — Beyond classification. The same trees in regression mode for Remaining Useful Life, with prediction-interval bands.

§ 12 — The limits of trees. Where they fail — extrapolation, raw signals, missing data — and what to reach for instead.

§ 13 — Practical notes. Rare faults, label leakage, group-and-time splits, calibration, tuning, edge vs. cloud, and drift.

§ 14 — References & further reading.

Read it

Open the monograph →

The monograph lives at its own URL in the warm-paper layout, with a worked decision tree, bagging-vs-boosting and Random Forest block diagrams, a gradient-boosting flowchart, a feature-importance plot, two algorithm boxes, runnable Python, a confusion matrix, practical notes, and references.

It sits alongside Isolation Forest (the unsupervised cousin), the boosting family in depth, and Fault Detection and Isolation. For the jargon, see the Diagnosis & Prognosis Glossary.

← Back to Autonomy