Intermittent HV Isolation Faults

A warning that comes and goes with the weather

Healthy isolation sits in the tens of megohms. A transient tail is dropping below the half-megohm threshold.

The IMD continuously measures isolation resistance between the HV bus and chassis. When it falls below the threshold (here, MΩ), the vehicle logs an isolation DTC and may inhibit charging. The defining feature is that the fault is intermittent and self-clearing: it appears on cool, humid mornings and disappears once the vehicle warms — behavior no permanent insulation defect would show. Reports skew heavily coastal and toward the pre-dawn hours.

Because the symptom appears only under specific weather and time-of-day conditions, several environmental and component variables move together. Full telematics — weather, surface temperature, location, connector genealogy — make a disciplined attribution possible.

The instrumented events

500 morning measurement events were logged across the fleet, each pairing the IMD isolation reading with the local weather, the connector surface temperature, the coastal/inland location, salt exposure, and the connector supplier lot.

The dataset is synthetic, generated from a condensation-physics model with a known ground truth so the method can be graded. Dew-point margin and salt are the true drivers; the connector lot is built as a dealer-allocation proxy with no leakage pathway of its own.

The field of suspects

H1 and H2 are the environmental chain; H3/H4 are the tempting non-environmental explanations; H5/H6 are the "is it even real?" checks. The data must distinguish a genuine, weather-driven leakage path from a component or sensor artifact.

Where a hasty investigation ends

A Pareto ranks association. To reach causation we need the continuous environmental variables and the discipline of holding them constant.

What actually tracks the isolation drop

Dew-point margin is the standout. But lot, coastal, and humidity all correlate, and only the four-gate test — association, significance, materiality, mechanism — separates cause from proxy.

A dealer-allocation artifact

Lot Lc's correlation has a logistics explanation: Lc was allocated mostly to coastal dealerships. It labels "this vehicle lives by the sea," nothing more. Two tests confirm it.

Second, the mechanism gate: the connector's bulk resistivity and contact resistance for lot Lc are within material spec — there is no electrical pathway by which the lot itself could shed megohms of isolation. Strong correlation, no mechanism: a confounder. H3 is rejected, and the supplier hold is released.

Not "it's humid" — "the surface is below the dew point"

The regression makes the structure explicit. Lot explains little; humidity helps a lot; adding the dew-point margin and salt dominates and drives the lot coefficient to zero:

This is the abductive refinement. "It's humid" is incomplete — many humid mornings show no fault. The pattern that fits is sharper: faults appear when the connector surface temperature falls below the dew point, so moisture actually condenses on the creepage path, and worsen where salt makes that film conductive. Humidity is necessary background; the dew-point margin is the trigger. That refined, testable claim — Riso collapses when margin < 0, and recovers when the surface warms — is what §8 confirms.

Where the data meets the dew point

8.1 — The diurnal signature

If condensation is the cause, faults must cluster when surfaces are coldest — the pre-dawn hours — and clear as the sun warms the connector. The fault rate by hour shows exactly this:

8.2 — Isolation resistance vs dew-point margin

Plotting Riso against the dew-point margin reveals a sharp knee at zero — the onset of condensation — and a clear coastal/inland separation driven by salt:

8.3 — The fault map

The triage, distilled

The recurring per-event decision compresses to three checks, evaluable from the IMD reading plus weather telematics. Each leaf names cause and action; thresholds come from the data (fault below MΩ; condensation at dew-point margin < 0).

Hardening against the weather

The mechanism chain and its confounder, side by side:

The fix attacks the film and the nuisance trip without touching the (sound) connector:

Forecasting susceptibility and field exposure

Because the cause is environmental and mechanistic, both the per-event susceptibility and the seasonal field exposure can be forecast. The susceptibility curve maps the trigger directly:

Across a full year, the field exposure for a worst-case coastal vehicle — and the effect of the coating — is stark:

Field & design action

Immediate containment: deploy the IMD firmware debounce and active dry-out cycle fleet-wide so transient dew events no longer raise hard faults or block charging. Corrective action: add a hydrophobic conformal coating and improved sealing/venting at the HV connector in the next build, and qualify a small dry-out heater for high-exposure climates. Release the supplier hold on connector lot Lc — it was never causal. Verify by confirming the dawn fault peak disappears and Riso stays above threshold through condensation events in coastal validation vehicles, and track nuisance-fault rate by climate zone rather than by lot.

Kept or dropped

Reproducibility

The dataset and full analysis ship alongside as isolation_fault_rca.py (condensation-physics generator + diagnostic pipeline) and isolation_events.csv (500-event mockup). The core of the generator:

# condensation film bridges the HV creepage path
dewpoint_margin = surface_C - dewpoint_C        # < 0 => condensation
film = 1/(1+exp(dewpoint_margin/1.5)) + 0.15*max((RH-70)/30,0)
Riso = RCLEAN / (1 + 60*film*(1 + 1.5*salt_index))   # Mohm
fault = Riso < 0.5                                # isolation threshold

# confounder: lot Lc allocated mostly to coastal dealers
p_Lc = where(coastal==1, 0.75, 0.18)             # proxy, resistivity in spec

The pipeline reproduces every figure: the Pareto, correlation screen, partial-correlation confounder test (lot → r ≈ ), staged regression (R² ), the diurnal signature, the Riso–margin knee, and the prognosis.