By Hyun-Soo Lee, Head of Software, Sasung Power
Test data from structural cell endurance runs shows impedance drift 45 to 80 cycles before measurable capacity loss. The pattern repeats: a 0.3 to 0.5 milliohm rise in the high-frequency semicircle of the EIS spectrum aligns with early SEI instability and appears ahead of the degradation knee. A trained model can flag it. Its practical value depends on the response and whether the monitoring system resolves signals at that granularity.
Prediction and Detection Require Different Designs
Fault prediction and fault detection need different data and support different system architectures. Detection asks whether a fault has already occurred. Its answer is binary, and the data burden is modest: voltage, temperature, or impedance crosses a calibrated limit. The difficult part is setting that limit to catch real faults without reacting to measurement noise or normal state-of-charge variation.
Prediction asks whether a cell is following a trajectory that will produce a fault within N cycles. That requires a time-series model to follow how electrochemical parameters change over tens to hundreds of cycles, rather than reading only their current values. A 5 mV impedance shift may mark slow degradation that settles within 20 cycles, or acceleration preceding a thermal event within 50 cycles. The current value cannot separate those outcomes. The trajectory can.
Predictive modeling needs far more labeled data than a detection threshold. Training requires histories from cells that reached failure, including the precursor sequence for each failure mode. Such records are scarce for structural cells under combined mechanical and electrochemical loading.
Features That Carry the Useful Signal
Useful features for structural-cell fault prediction come from three domains: electrochemical, thermal, and mechanical state. Their value differs, and correlated signals can make combined features more predictive than either signal alone.
The strongest electrochemical features we have found are incremental-capacity peak shift, where dQ/dV follows lithiation transitions and their drift signals early degradation; low-frequency EIS arc diameter, which reflects separator aging and ionic diffusion; and high-frequency real impedance at 1 kHz, which tracks contact resistance and current collector corrosion.
Thermal information is more useful as the difference between each cell group and the pack mean, measured over a 20-cycle moving window, than as absolute temperature. A 2°C elevation that remains steady for 100 cycles is less concerning than a 2°C elevation that has increased by 0.8°C during the last 30 cycles.
For structural cells, a key mechanical feature is the relationship between applied strain and voltage-offset error against the linear strain-voltage model. Divergence from that cell's established linear relationship can indicate electrode delamination or binder degradation, both precursors to accelerated capacity fade. Conventional packs do not provide this feature.
Small Training Sets Set a Hard Limit
Structural battery cells for vehicle integration are costly and slow to make in test quantities. One endurance run of 1,000 cycles using 20 cells at a 0.5C rate takes approximately 8 months. Relevant failures, such as separator delamination under cyclic load, occur below 5% in well-designed cells at moderate loads. Reaching 50 labeled failure examples requires approximately 1,000 cells in endurance testing, an unreasonable program scale for an early-stage technology company.
This constraint limits model complexity. A deep neural network trained on 30 labeled failure examples from 18 cells will memorize its training set. It will generalize poorly to out-of-distribution load profiles or different cell geometries. Models that work in this setting are physics-constrained: simpler architectures whose structure encodes known electrochemical and mechanical relationships, reducing the freedom available to fit noise.
In practice, the most dependable limited-data approach is a physics-informed recurrent model, such as an LSTM or GRU whose architecture maps to the cell's equivalent circuit model. It learns aging-rate parameters instead of relying on black-box hidden state. That physical structure supplies a strong inductive bias and lowers the effective sample size needed for useful generalization.
False Positives Expose Limits of Accuracy Metrics
A fault prediction model with 95% accuracy on held-out test data may sound strong. In a 96-cell pack completing daily cycles, a 1% false positive rate per cell per charge cycle produces roughly one spurious fault alert per charge cycle. If every alert starts a maintenance inspection, false-positive costs will exceed the faults prevented within weeks.
For safety-critical alerts, false positive rate should be stated in operational terms. Alerts per cell per 100 cycles reveal more than aggregate test-set accuracy when the test distribution matches training data. The metrics diverge sharply when faults are rare, as they should be in a functioning pack.
We specify fewer than 0.5 false-positive events per cell per 1,000 cycles for predictive fault alerts, with a lead time of not less than 40 cycles before conventional detection methods can measure the predicted fault type. This is considerably stricter than "95% accuracy" and is the relevant specification here.
Where Physics Takes Over
Some structural-cell failures cannot be predicted from electrochemical telemetry alone, regardless of model sophistication. An overload, such as an off-road impact or emergency-braking load spike while the cell is charged, can cause separator delamination that appears as an instantaneous fault rather than gradual drift. There is no precursor sequence. The event itself is the fault.
For load-induced failures, the correct response is detection with mechanical context, not prediction: did the cell experience an out-of-specification load in the recent window? If so, initiate an EIS scan at the next available rest interval to evaluate post-event health. Because the load-failure physics is understood, a trained model is not required to identify an event beyond the cell design envelope.
ML models provide the most value for slow-drift failures, including SEI instability and lithium-plating precursors that develop over tens to hundreds of cycles. This is where physics-based thresholds struggle, because the signature is statistical rather than deterministic. ML should handle statistical drift detection while physics rules handle mechanical-event response. A single model covering both regimes would require impractical training data and perform neither task well.
Not Yet Ready for Production
No one in this field, including us, has fully validated an end-to-end structural-pack fault prediction system across a representative duty-cycle population for production vehicles. The per-cell EIS infrastructure required for high-quality features is only now appearing in prototype vehicle integrations, and labeled failure data under combined load remains limited.
Production-ready capability is detection: calibrated thresholds with strain-aware correction for mechanical noise, plus trajectory tracking for dQ/dV peak position and high-frequency impedance. Together, these layers detect faults substantially earlier than conventional pack-level monitoring without the deep datasets true prediction requires. Full prediction remains a research effort, so deploy detection and trajectory tracking first, then expand prediction as representative failure data accumulates.