Elevated Module Weld Resistance

A resistance the spec won't allow

Good joints sit near 35 µΩ. A tail of modules is landing well above 80.

End-of-line testing flags modules whose tab-to-busbar interconnect resistance exceeds the 80 µΩ upper spec limit. The concern is not cosmetic: a high-resistance joint dissipates more power, runs hotter under fast charge, and drifts further upward in service — raising the risk of thermal flags and warranty returns. The escape rate at EOL was running near , far above target.

Because every module carries full production traceability — station, electrode state, supplier lot, line speed, shift, plant conditions — the data supports a disciplined attribution rather than a guess. The challenge, as always, is that several of those variables move together.

The instrumented build

600 modules were logged at EOL with their full genealogy. Each record carries the weld station, the electrode wear (welds since the last tip dressing), the supplier foil lot, process parameters, plant conditions, and the measured interconnect resistance.

The dataset is synthetic, generated from a process model with a known ground truth so the method can be graded. Electrode wear is the only true driver of resistance; the supplier lot is built as a routing-linked proxy with no metallurgical path to the defect.

The manufacturing suspects

Seven hypotheses were carried, spanning the manufacturing failure-mode families:

As before, several are not rivals but links in a chain: a station's dressing interval (H3) governs electrode wear (H2), which sets the joint resistance. Recognizing that prevents mistaking the station for the fix.

Where a hasty investigation ends

The first instinct in a quality escape is a Pareto of defects by attribute. Sorted by supplier lot, it points an unambiguous finger:

A Pareto ranks association, not causation. To go further we need the continuous process variables and the discipline of holding them against one another.

What actually tracks the resistance

Electrode wear is far and away the strongest signal, but station and lot are non-trivial, and a Pareto would have stopped at the lot. The four-gate test — association, significance, materiality, mechanism — is what separates them.

A routing artifact, not a material defect

Lot L2's correlation has a mundane explanation: L2 was the foil lot fed predominantly to Station B. It is a label for "this module was built on Station B," nothing more. Two tests confirm it.

First, holding electrode wear constant collapses both the lot and the station associations to noise, while wear itself survives intact:

Second, the mechanism gate: foil resistivity for lot L2 is within material spec, so there is no metallurgical pathway by which L2 could add tens of µΩ to a joint. Strong correlation, no mechanism — the signature of a confounder. H1 is rejected; the supplier hold is released.

Following the chain to the dressing interval

The regression makes the mediation explicit. Lot alone explains little; adding station helps; adding the wear term dominates and drives the lot and station coefficients to near zero:

This is the abductive step. The best explanation for "Station B is loud, but only its high-wear modules, and lot L2 just happens to ride along" is that the actionable cause is neither the station nor the lot but the electrode condition — specifically, welds since the last dressing. Station B is implicated only because its lax dressing interval lets wear run deep. The refined, testable claim: resistance is set by cycles-since-dressing, and resets when the tip is dressed. That prediction is the subject of §8.1, and it is decisive.

Where the data meets the weld

8.1 — The control chart: a tooling-wear sawtooth

If wear-since-dressing is the cause, resistance on Station B should climb across each dressing cycle and snap back when the tip is dressed — a sawtooth, not a random scatter. It does exactly that:

8.2 — Weld strength falls as resistance rises

The metallurgical companion: as the tip wears, the weld nugget shrinks, so pull strength drops exactly as resistance climbs — two faces of the same loss of joint quality.

8.3 — Process capability

Capping the dressing interval so the tip never reaches deep wear transforms the resistance distribution relative to the spec limit:

The screen, distilled

The recurring per-module decision compresses to three checks, evaluable at EOL or on the line — each leaf names cause and action. Thresholds come from the data (spec 80 µΩ; the wear knee near 1200 cycles, deep wear past ~3000).

Closing the escape

The fix is a maintenance and screening change inside the existing line — no new hardware. The process and its SPC feedback loop:

The causal chain and the confounder, side by side:

Setting the interval, forecasting the field

Because the cause is mechanistic, both the line fix and the field risk can be forecast. First, the dressing interval is an optimization: escape rate rises sharply with cycles-since-dress, so the interval is chosen to hold escape below target.

Second, the field consequence of not screening: above-spec joints drift upward in service and eventually trip thermal flags.

Plant action

Immediate containment: institute a 100% EOL interconnect-resistance screen across all stations, and quarantine modules built on Station B since its last electrode dressing. Corrective action: set every station's dressing interval from the escape-vs-interval curve at the 0.5% target, add an SPC control chart on EOL resistance with an action limit that auto-triggers dressing, and pilot adaptive weld-energy compensation to flatten the wear sensitivity. Release the supplier hold on lot L2 — it was never causal. Verify by confirming the sawtooth no longer breaches spec and Cpk holds above target on each line, and track field thermal-flag rate by build-since-dressing cohort.

Kept or dropped

* Station B is the locus, but the actionable cause is its dressing interval, not station hardware.

Reproducibility

The dataset and full analysis ship alongside as weld_resistance_rca.py (process-model generator + diagnostic pipeline) and weld_modules.csv (600-module mockup). The core of the generator:

# resistance climbs super-linearly past the electrode-wear knee
wear_term = max((electrode_cycles - KNEE)/1000, 0) ** 1.6
R = 35 + 17*wear_term + 0.05*(weld_energy-100) + lot_resistivity + noise
defect = R > 80                              # microohm spec limit

# confounder: lot L2 routed mostly to Station B (the lax line)
p_L2 = where(station=="B", 0.80, 0.12)       # proxy, foil in spec
dress_interval = {"A":1500, "B":4000, "C":1700}  # B is lax

The pipeline reproduces every figure: the Pareto, correlation screen, partial-correlation confounder test (lot → r ≈ ), staged regression (R² ), the SPC sawtooth, capability (Cpk ), and the prognosis.