By Hyun-Soo Lee, Head of Software, Sasung Power
At cycle 340 of a 500-cycle endurance run on a 48-cell structural module, pack voltage followed a normal degradation slope. Aggregate capacity was down 1.8% from its starting point, within the expected 0.3 to 0.4% per 100 cycles for LFP chemistry. The pack monitor reported no anomaly. Yet Cell C4 had a resting impedance 6.2 mV above the module mean over the previous 60 cycles, a pattern associated with early lithium plating on the graphite anode and preceding measurable capacity loss by 40 to 80 additional cycles.
What Pack Monitoring Can Actually See
Pack monitoring combines measurements. It tracks the sum of series-string voltages, enclosure temperature averaged across sensor points, and current integrated at the pack terminal. That supports broad state-of-charge tracking and over-current protection. It does not provide enough resolution for early cell fault detection.
Statistical dilution is the core problem. In a 96-cell series string, one cell's impedance anomaly contributes roughly 1% of total string voltage. A 6 mV cell-level anomaly therefore becomes 0.06 mV in the string reading, below the noise floor of most battery management controllers, whose pack-terminal ADC resolution ranges from 0.1 to 1 mV depending on design. The fault remains in the data, but cannot be resolved without measuring each cell.
Pack monitoring also obscures where heat is forming. A hot spot in one cell group at an enclosure corner appears as only a 0.5 to 1.5°C rise at the pack-average temperature sensor, which is within normal variation. At cell resolution, the same location may sit 8 to 12°C above neighboring cells, crossing the threshold for accelerated electrolyte decomposition in many chemistries.
The 3 to 8 Millivolt Impedance Shift
The 3 to 8 mV range is not one threshold. It contains precursor signatures linked to different degradation mechanisms. Each part of the range needs a practical interpretation so detection limits can reduce both false positives and missed faults.
A 3 mV difference in resting cell voltage from neighboring pack positions is consistent with early active-material loss, usually lithium inventory loss caused by SEI, or solid electrolyte interphase, growth. At 3 mV, this is not an actionable fault. It is an early-cycle degradation signal that may stabilize and does not call for immediate intervention.
A 5 to 6 mV shift, together with rising impedance in the 1 to 10 Hz EIS frequency range, is consistent with graphite-anode lithium plating, especially during low-temperature charging above a 1C rate. This signature comes before internal short circuit by tens of cycles. At this level, the BMS should lower the affected cell group's charge rate and mark the cell for inspection at the next service interval.
An 8 mV or greater shift, combined with cell temperature at least 3°C above the pack mean, is a late precursor. In our test data, it correlates with capacity dive, defined as more than 5% capacity loss per 20 cycles. Early BMS protection can still prevent a thermal event, but the cell must be replaced within the current service period.
None of these signatures appears in pack-level aggregate data.
Sampling and Infrastructure for Precursor Detection
Detecting a 5 to 6 mV impedance precursor requires continuous per-cell voltage sampling at not less than 1 sample per second during discharge and rest, plus periodic electrochemical impedance spectroscopy, or EIS, in the 1 to 10 Hz range. Cell-level EIS needs dedicated hardware or software PRBS, pseudorandom binary sequence, injection through BMS balancing circuits. PRBS lowers hardware cost but sacrifices some frequency resolution. It is the method used on our current platform.
Telemetry from 96 cells sampled at 1 Hz produces only about 1.5 KB per second of voltage and temperature data. Bandwidth is not the difficult part. Sensor placement and wiring are. A conventional pack typically uses temperature sensors at 4 to 6 points. A structural pack with cells integrated into frame members may contain 20 to 30 separate thermal zones that need independent monitoring for useful spatial fault detection.
Our platform uses a distributed measurement network for cell telemetry. Multiple sub-BMS nodes, each handling 12 to 16 cells, communicate with the main BMS controller over a CAN bus backbone. Compared with a centralized topology, this reduces harness complexity and supports fault isolation at each node without bringing every cell sense line to one controller location.
Why Structural Cells Raise the Telemetry Noise Floor
Conventional cell telemetry is calibrated with cells held in rigid, zero-load fixtures. Structural cells occupy vehicle load paths and flex, compress, and twist as the vehicle moves. Those mechanical inputs add voltage noise that conventional packs do not experience.
In the carbon fiber composite LFP cells we observe, strain creates a voltage offset of about 1.5 to 3 mV per 1% strain along the in-plane fiber direction. During typical vehicle cornering at 0.3 to 0.5 g lateral, cells in the outer floor panel experience 0.1 to 0.3% strain, producing a 0.15 to 0.45 mV voltage noise floor. That is small beside the 5 to 6 mV lithium-plating signature, but material beside the 3 mV early-degradation signature at the detection limit.
Separating mechanical voltage noise from electrochemical state takes one of two approaches. A physical strain sensor can measure the mechanical input and subtract its expected voltage contribution. Alternatively, a statistical filter can distinguish mechanical load noise, which follows vehicle dynamics and is bandlimited by vehicle resonance at 2 to 15 Hz, from electrochemical impedance drift, which changes slowly and tracks cycle count rather than instantaneous load. Our software defaults to the statistical filter. Strain-sensor correction is available as a hardware add-on when the application requires the 3 mV detection limit.
When Cell-Level Telemetry Is Not Needed
Cell-level telemetry is not right for every battery system. A stationary energy storage installation with monthly service intervals, replaceable modules, stable thermal conditions, and little mechanical load may not need per-cell impedance monitoring. Pack monitoring is sufficient when degradation is spatially uniform, cells age similarly, and a missed early fault means module replacement rather than a thermal event in a safety-critical application. The point is to match monitoring depth to failure consequences.
At current component prices, per-cell telemetry hardware, including added voltage channels, temperature sensors, and processing capacity, increases BMS hardware cost by 8 to 15%. That expense is justified for structural cells under mechanical load, high-cycle applications where early fault isolation lowers total ownership cost, and platforms where thermal event risk must be minimized. It is not automatically justified as a default specification for every battery pack. Overstating the case would produce hardware over-specification where pack monitoring is adequate.
Set monitoring granularity during system specification, not firmware implementation. Deciding early avoids later cost and rework.