TelemetryLab

← TelemetryLab Research · Vehicle Safety Telemetry · 2026

The Unread Safety Signal

A modern vehicle continuously broadcasts the raw signals from its brakes, steering, airbags, battery and every driver-assist feature onto an internal data bus. The data is already there for the taking. Processed into intelligence, it yields 39 safety insights almost no one extracts today.

OEM-neutral · grounded in public sourcesSignal names from the open opendbc CAN database39 safety insights, extractable from the raw data

distinct safety insights you can extract by processing data the fleet already broadcasts.

safety domains they span, from ADAS usage to battery-fire risk.

200+

CAN signals a single passive OBD tap already sees.

40,990

U.S. road deaths in 2023 (NHTSA estimate). Safety performance plays out in the field.

The premise. The instrument cluster shows the driver a handful of warning lights. The same vehicle pours dozens of raw safety-relevant signals onto its internal bus the whole time. The gap is not the data, the signals are right there. The gap is the processing: almost no one turns them into safety intelligence. This document maps what that processing can extract; it is not a product specification, and it does not detail the methods.

What the dashboard shows the driver

A few lights, after the fact

Check engineABSAirbagLow tireBatteryLane assist off

What processing that data can reveal

Dozens of insights, once it is processed

Which features the driver turns off, and whyFalse-alarm rate per featurePhantom-braking events, mappedBrake wear and fluid-pressure driftPer-tire pressure, not just “low”Battery thermal-runaway precursorsSilent airbag/SRS faultsSteering vs yaw consistencySensor availability and calibrationNear-misses per 1,000 miles

The map below itemizes the right column: all 39 insights, grouped into 13 safety domains.

The unprocessed layer

39 insights · 13 domains

The raw signals for 39 distinct safety insights are already on the bus. Extracting them is the work almost no one does.

Every tile below is a safety insight that has to be extracted by processing the raw signals a passive OBD tap already sees, computed from them, not read off them. The data spans every driver-assist feature and safety system the car runs. The signals are available; the analysis is the value, and it is the part that is missing.

Safety insights extracted from data the fleet already broadcasts1 passive OBD tap

39waiting to be extracted from the same data

Left: safety insights standard connected-car telematics extracts from this data today. Right: the insights you can extract from the same raw signals, processed right.

01 Features kept on 4

Feature opt-out ranking
The every-drive shut-off ritual
Time to first disable
Driving with the suite off

02 Alert quality 4

Alert precision per feature
Phantom braking, counted and mapped
Traffic-sign read accuracy
Alert load per hour

03 Usage and settings 4

Real engagement per feature
Following-distance choice and drift
Hands-off time and supervision drift
Speed-feature adoption and override

04 Did the feature do its job 5

Every automatic brake, tagged
Lane-keep correction quality
Blind-spot heed rate
Driver override = false-positive rate
Why cruise dropped out

05 Attention vs the regulators 3

Drowsiness and distraction warning rates
Attention escalation behavior
Hands-on engagement quality

06 Brakes 3

Brake wear / life trend
Brake-fluid leak / low pressure
Brake-blend / transition quality

07 Tires 2

Per-tire under-inflation, early
Slow leak vs a bad sensor

08 EV battery 3

Cell-voltage imbalance trend
Thermal-runaway precursors
Charge / thermal margin

09 Airbags and restraints 2

Silent SRS faults
Occupant-detection faults

10 Steering 2

EPS health trend
Steering vs yaw consistency

11 Availability and calibration 3

Feature availability %
Blocked-sensor cascade
Camera / radar degradation

12 Driving style and risk 2

Harsh-event rate, system-brake excluded
Time over the posted limit

13 Crash precursor and near-miss 2

Near-miss rate per 1,000 mi
Software version as the experiment

Usage is it accepted Quality does it perform Safety does it protect Sensor health is it available Driver behavior context

Four of the thirty-nine, in full

what the insights look like

Features kept on · Usage

A safety feature a third of the fleet switches off protects no one

Process the on/off state of every driver-assist feature, captured at each start, and the true acceptance ranking falls out, including the features drivers defeat within minutes of every single drive. The raw states are on the bus; the ranking is what you compute from them.

~1 in 3owners disable lane-keeping in field studies (directional; measure the fleet’s real number). IIHS opt-out research.

EV battery · Safety

An EV battery signals a fire before it starts

The raw signals are on the bus: cell voltage, pack temperature, venting gas. Processed against each pack’s own baseline, they surface a precursor in a fixed order, gas first, then an abnormal voltage signature, then a temperature rise, anywhere from tens of seconds to over an hour before the event.

Fig. 1 · EV battery

The thermal-runaway warning hierarchy, by signal

Documented advance warning before a thermal event. Vented gas appears first (lead time method-dependent, no single published figure), then voltage, then temperature. A separate fast-charging model resolves the final window to tens of seconds.

Sources: Nature Comms Eng 2025 (voltage ~1.47 h, temperature ~0.57 h); Nature Scientific Reports 2025 (deep-learning charging model, 9.95 to 27.95 s). Log scale.

Airbags and restraints · Safety

A restraint system can fail with the dashboard dark

The airbag controller stores fault codes in module memory, many of which set before, or without, a warning lamp the driver would notice. Pull and analyze those codes across the fleet and a degraded restraint system the owner never saw shows up, vehicle by vehicle.

20+SRS circuit parameters the module watches and logs (FMVSS 208), independent of the dashboard lamp.

Crash precursor · Safety

Count the precursors, not the accidents

The raw events are on the bus: hard brakes, forward-collision warnings, swerves, stability-control catches. Counted and trended, they become the aviation flight-data-monitoring signal applied to the road, and answer “is the fleet getting safer” years before crash data can. Near-misses are a validated leading indicator of crash risk.

Fig. 2 · Near-miss

The safety pyramid the bus can count

Aviation reads the routine precursors, not just the wreckage. A vehicle generates the same stream.

Illustrative ratio of precursor events to crashes (aviation flight-data-monitoring logic). Real ratios are fleet-specific and readable from the bus.

What this runs on, and why the signals are checkable

No factory connection, no data agreement, no access to anything the car does not already broadcast.

The device

A single passive, read-only OBD-II dongle in the standard diagnostic port. It reads what the vehicle already broadcasts and transmits nothing back, so it cannot touch any control system.

The signals

The signal names in this catalog are real, drawn from several automakers’ public CAN databases. Confirming the exact frames on a given vehicle is the first step of any deployment, an on-vehicle decode.

Where the names come from

The open, community-maintained opendbc database covers roughly 399 vehicle models. Across the ADAS, cruise, braking, steering, blind-spot, body and powertrain modules, the dongle decodes hundreds of signals.

Honesty on each read

Most insights are computable from signals already documented in public databases. Some battery, airbag and EPS insights are extracted from standard diagnostic codes. None of it assumes access to in-cabin video or OEM-internal data.

The signals are real, and not one brand’s

A sample of the named signals these insights are extracted from, verbatim from five automakers’ published CAN databases. The catalog above mixes them: it does not matter which insight comes from which badge.

Automaker	Steering	Wheel speed / brake	Cruise / assist	Source (opendbc)
Honda / Acura	STEER_MOTOR_TORQUE.MOTOR_TORQUE	WHEEL_SPEEDS.WHEEL_SPEED_FL; VSA_STATUS.USER_BRAKE	ACC_HUD.ACC_ON	_honda_common.dbc
Toyota / Lexus	STEER_ANGLE_SENSOR.STEER_ANGLE	BRAKE.BRAKE_AMOUNT	PCM_CRUISE_2	_toyota_2017.dbc
Subaru	ES_LKAS.LKAS_Request	ES_Brake.Brake_Pressure	Cruise_Status.Cruise_On	_subaru_global.dbc
Hyundai / Kia / Genesis	SAS11.SAS_Angle; LKAS11.CF_Lkas_LdwsSysState	WHL_SPD11.WHL_SPD_FL	SCC11.MainMode_ACC; .VSetDis	hyundai_can.dbc
GM	PSCMSteeringAngle.SteeringWheelAngle	EBCMBrakePedalPosition.BrakePedalPosition	ASCMActiveCruiseControlStatus	gm_global_a_*.dbc

Read verbatim from the public comma.ai/opendbc CAN databases (github.com/commaai/opendbc). Community reverse-engineered; on-vehicle decode confirms the exact frames per model.

Reading the evidence honestly

Settled · quote freely

Regulation and claims data

FMVSS / EU GSR / Euro NCAP thresholds, the battery precursor order, L2 showing no crash reduction beyond AEB, AEB-with-pedestrian cutting real claims, near-misses as validated crash surrogates, the steering-to-yaw relationship as textbook physics.

Directional · measure the real number

Single-fleet statistics

Opt-out percentages, hands-free mileage shares, nuisance-alarm ratios, the exact thermal-runaway lead time. Direction-true and number-soft, which is the point: no one has the real number for a given fleet, and it is readable.

Color only · do not build on it

Limits and caveats

Telemetry supports attribution, it does not prove fault alone; baselines are per-vehicle; the onboard recorder is short; camera-based attention carries demographic bias; driver-level use must be consent-based. On-vehicle confirmation is always step one.

Scope, and what is adjacent

This catalog covers one layer: what processing this telemetry reveals about safety-feature usage and functional-safety health. The same underlying data also supports recall scoping, homologation and software-update-history evidence (UN R156), and recurring regulatory reporting. Those are adjacent work, not detailed here.

Sources

public, authoritative, tiered

What is ours, and what is sourced. The groupings, the cross-signal analysis, and the “count the precursors” framing are TelemetryLab’s analysis. The underlying facts are not: signal definitions, regulatory thresholds, feature behaviors and study findings come from the public sources below. Signal names are from the community-maintained opendbc database. Vendor-stated or single-study figures are tagged, and read as directional.