Mahalanobis Distance for Fleet Sensor Health Scoring

Fleet sensor health is inherently multivariate — camera yaw, radar range bias, and CAN timing anomalies rarely fail in isolation. Mahalanobis distance measures how far today's sensor agreement deviates from a baseline while accounting for covariance between modalities. Operators do not need linear algebra to act on the result: NADIR maps Mahalanobis-style envelopes to NOMINAL, CAUTION, and CRITICAL tiers in shadow mode. This guide explains the intuition, fleet use cases, integration paths, and audit evidence without exposing proprietary scoring weights.

By Dhruv Hegde · NADIR calibration intelligence

Why Euclidean thresholds fail mixed fleets

Simple distance metrics treat every residual dimension equally. Real ADAS telemetry correlates — camera pitch drift accompanies ride height change; radar elevation bias tracks bumper replacements. Applying one threshold per sensor produces false alerts on highway-heavy routes and misses coupled drift after glass campaigns.

Mahalanobis distance scales each dimension by baseline variance and cross-correlation learned per vehicle or cohort. A spike in one modality counts more when it is unexpected given the others. Fleet health scoring becomes statistically grounded rather than a pile of independent red lights.

NADIR learns baselines during shadow weeks — never assuming generic OEM bulletins replace vehicle-specific history. Mixed OEM rosters benefit because org-scoped models normalize ingest profiles on the NADIR platform.

Operator tiers without the matrix math

Console and fleet API consumers see three tiers, not eigenvalues. NOMINAL means residuals stay inside the learned envelope. CAUTION signals sustained trend or an unusual multivariate spike warranting inspection within your playbook window. CRITICAL prioritizes bay calibration and validation before high-risk dispatch.

Tiers are not DTCs. Vehicles may read CRITICAL in NADIR while dashboards stay clean — the gap continuous monitoring closes. Tier transitions generate evidence bundle candidates documented on the evidence page.

Safety boards should pre-define SLA language: CAUTION scheduling within fourteen days, CRITICAL holds within twenty-four hours — policies NADIR supports but does not dictate.

Fleet aggregation and cohort analytics

Fleet summaries rank VINs by tier, region, duty cycle, and repair history. Mahalanobis outliers cluster — a windshield supplier lot, a collision center with elevated repeat drift, a desert route with thermal swing. Geographic and site filters turn residuals into QA conversations backed by data.

Compare cohort health before and after policy changes: new dispatch holds, hub-and-spoke calibration routing, insurer-mandated monitoring. Export CSV histograms for renewal packets — increasingly common in 2026 fleet insurance reviews covered in our liability guide.

OEM validation teams segment by platform and climate without exposing competitor fleets — org isolation enforced at the API layer.

Integration and shadow validation

Telematics partners forward batch frames to NADIR ingest with idempotency keys for flaky cellular routes. The SDK quickstart replays synthetic fixtures before production keys. Shadow mode scores every eligible VIN while dispatch rules stay fixed — validate Mahalanobis-derived alert rates honestly before automation.

The HORIZON pilot tour shows executives how tier forecasts communicate risk before engineering completes API wiring. Week-two false alert review boards separate soiling from extrinsic shift — essential when Mahalanobis spikes correlate with weather.

Edge ingest patterns — offline queues, source tagging, SLIs — appear in the telematics edge ingest guide.

Relation to Kalman and cross-modal scoring

Kalman innovations emphasize temporal prediction error; Mahalanobis envelopes emphasize multivariate agreement at a snapshot or window. NADIR combines changepoint detection with covariance-aware scoring — see Kalman filtering for drift and cross-modal residual scoring for complementary views.

Engineers call authenticated residual endpoints for custom BI; executives read tier histograms. Separation keeps steering decks simple while data science explores modality contributions.

Audit and insurer workflows

Signed bundles include tier timelines, export hashes, and validation closure after shop work. Insurers prefer structured Mahalanobis-tier history over ad hoc screenshots — reduced discovery cost when ADAS performance is disputed. MSO partners attach the same bundles to franchise scorecards — see MSO chain of custody.

Cohort baselines and mixed OEM normalization

Mahalanobis envelopes learn from vehicle history — not generic thresholds copied from a single OEM bulletin. New acquisitions need two shadow weeks before tiers stabilize; used vehicles with unknown glass history may enter CAUTION until baselines form. Fleet analytics should tag duty cycle separately: last-mile stop-and-go versus linehaul cruise produces different covariance structures — mixing them inflates false alerts.

Supplier lot analytics attach build metadata when telematics forwards part numbers or RO fields include windshield batch codes. Procurement negotiates with repeat drift data, not warranty anecdotes alone. Repair site filters highlight franchise partners with elevated multivariate outliers — QA conversations backed by envelopes, not subjective audit checklists.

Insurance partners increasingly request tier histograms with renewal packets. Mahalanobis-derived CAUTION spikes after regional weather events should recover without bay visits when root cause is soiling — persistent elevation indicates extrinsic shift worth structural inspection. Document both patterns in pilot readouts so carriers trust monitoring maturity.

Engineering access versus operator simplicity

Data science teams may call authenticated endpoints for custom BI — eigenstructure stays server-side. Executives read three tiers. Separation keeps board decks honest while engineers explore modality contributions without exposing proprietary weights in slide screenshots that leak IP.

Webhook integrations push CRITICAL transitions into ServiceNow, Jira, or repair portals after shadow validates volume — configure signing and retry policies on the security page before production. Batch CSV exports feed quarterly OEM supplier reviews when field drift clusters by bracket vendor.

Compare Mahalanobis tier trends before and after dispatch policy changes — did CRITICAL holds reduce high-risk mileage exposure? Did CAUTION scheduling clear backlog without missing onset on highway-heavy units? Fleet health scoring becomes a recurring ops rhythm when metrics attach to executive dashboards alongside fuel and tire KPIs.

Weekly executive rhythm

Publish five numbers every Monday: tier mix percentage, median days to detect onset, false CAUTION after review, shop closure within SLA, export latency. Mahalanobis outliers without operational rhythm become shelfware — Console access alone does not change maintenance behavior. Pair metrics with named owners in safety, maintenance, and legal so CAUTION tickets close instead of aging silently in email.

Board slides should show one synthetic timeline — detection, shop action, validation — sourced from pilot exports. Avoid covariance slides; focus on MTTD, closure, audit completeness. Legal reviewers prefer chain-of-custody narratives to undocumented confidence scores without export hashes.

FAQ

Next steps

Review the NADIR platform, explore the HORIZON pilot walkthrough, and open the Calibration Lab before wiring fleet telemetry. Shadow pilots score every eligible VIN without changing dispatch — the default entry path on the NADIR homepage.

Request a four-week cohort via the footer pilot form or team@nadirai.net with fleet size, telematics partner, and target KPIs for mean time to detect and shop closure rates.