The Battery's Brain
The Battery's Brain: Designing a Battery Management System
A battery management system is three things at once: the eyes that measure every cell, the accountant that estimates how much charge is left, and the guard that disconnects the pack before a fault becomes a fire. This is how to design each one, the measurement accuracy that decides how much capacity you waste, the SOC estimation that has no sensor to read, the balancing choice, and the protection thresholds with real margin, with a worked demo on a 48 V pack.
A battery pack is a team of rowers in one boat. Thirteen cells in series, each pulling its own oar, each with slightly different strength, each ageing at a slightly different rate. The boat only goes as fast as the weakest rower, and if one rower collapses, the whole boat stops, or worse, capsizes.
The battery management system is the coxswain. It watches every rower, counts the energy the team has left, keeps the boat inside the safe part of the river, and stops the boat before anyone rows past the limit. It does three jobs at once, and the three jobs need different skills: the eyes that measure every cell, the accountant that estimates how much charge is left, and the guard that disconnects the pack before a fault becomes a fire.
This article designs all three, and the arithmetic that decides whether each one works.
The words, in plain English
BMS (battery management system). The electronics that keep a lithium pack safe, balanced, and honest. Without one, a lithium battery cannot be safely charged or discharged: it will overcharge, over-discharge, overheat, or drift out of balance, and any of those permanently damages the cells or causes fire.
AFE (analog front-end). The measuring chip. It reads each cell's voltage and temperature, and the pack's current, and hands the numbers to the microcontroller. The AFE is the BMS's eyes, and the accuracy of its eyes decides how much of the pack's capacity the design wastes.
Fuel gauge. The algorithm that estimates state of charge (SOC) and state of health (SOH). It is not a sensor. There is no sensor that reads percent. The fuel gauge infers the percent from proxies, the current integrated over time, the open-circuit voltage, and a model of the cell, the way Article #4's Kalman filter infers position from noisy sensors.
SOC (state of charge). The percentage full, the ratio of the remaining charge to the cell's present full capacity. The target accuracy is ±3% over the pack's lifetime.
SOH (state of health). How much of the cell's original capacity is left, and how much its internal resistance has grown. End of life is typically defined as SOH below 80% for EV applications, or below 70% for grid storage.
OCV (open-circuit voltage). The cell's terminal voltage in thermodynamic equilibrium, after it has rested long enough for the voltage to settle, typically one to four hours. Each SOC corresponds to a specific OCV, at a given temperature. The OCV curve is the lookup table the fuel gauge uses to re-anchor its estimate.
Coulomb counting. Integrating the current over time, counting the charge that has flowed in and out, the way an accountant counts money in and out. It is exact in the absence of measurement error, and it drifts without bound in the presence of any offset, because every error integrates forever.
Kalman filter / EKF (extended Kalman filter). The estimator from Article #4, applied to the battery. It treats the cell as a dynamic system, an OCV source in series with a resistance and one or two RC polarization networks, and estimates SOC as a hidden state, predicting with the coulomb-count equation and correcting with the voltage measurement. The foundational reference is Gregory Plett's three-part series on extended Kalman filtering for lithium-ion polymer packs, published in the Journal of Power Sources in 2004.
Cell balancing. Keeping the cells at the same state of charge, so the weakest cell does not limit the whole pack. Passive balancing bleeds excess charge from the highest-voltage cells through a resistor, as heat. Active balancing transfers the charge from the high cells to the low cells, recovering it.
OVP / UVP / OCP. Overvoltage protection, undervoltage protection, overcurrent protection, the three thresholds that disconnect the pack. OVP stops charging when a cell reaches its maximum safe voltage (4.20 V for NMC/NCA). UVP stops discharging when a cell drops below its minimum (typically 2.8–3.0 V). OCP limits the current to the cells' continuous rating, and short-circuit protection triggers in microseconds at much higher currents.
Contactor. A heavy-duty relay that connects and disconnects the high-voltage pack. The BMS commands it to open. A contactor can weld shut under a severe fault, which is why the protection has a layer after it.
Pyro fuse / pyrotechnic disconnect. A fuse that is fired by a small explosive charge, on command, to sever the high-voltage path in milliseconds. It is the layer that works when the contactor has welded and the software is gone.
Welded-contactor detection. The BMS's check that the contactor actually opened after it was commanded to open, by measuring the voltage across it. A contactor that does not open is a protection layer that has silently failed, and the fuse is the layer that catches it.
Isolation / insulation monitoring. The continuous measurement of the resistance between the high-voltage bus and ground. In a 400–800 V pack, insulation degradation is a shock and fire hazard, and the BMS raises an alarm, and in severe cases isolates the pack, when the resistance falls below a configured threshold.
C-rate. The charge or discharge current expressed as a multiple of the cell's capacity. A 3.5 Ah cell discharged at 3.5 A is at 1C; at 7 A it is at 2C.
Plating. Lithium metal depositing on the anode instead of intercalating into it, the damage caused by charging above 4.2 V, or by charging below 0 °C. It irreversibly reduces capacity and creates internal short-circuit risk.
SEI (solid electrolyte interphase). The thin passivation film that forms on the graphite anode during the first charges. It protects the electrolyte from the reactive anode. Its breakdown, at 80–120 °C, is the first exothermic step in thermal runaway, Article #14's first rung.
ASIL. Automotive Safety Integrity Level, from ISO 26262, the rating of how much risk a function carries, and how rigorous its design and testing must be. The BMS's protection functions are typically ASIL C or D, the highest levels.
The three blocks
A BMS has three functional blocks, and they fail in different ways.
One, the analog front-end. It measures each cell's voltage and temperature, and the pack current through a shunt resistor. The measurement has to be accurate enough that the protection thresholds mean something: a protection threshold is only as good as the measurement behind it, and a measurement error of 25 mV is a protection threshold that is wrong by 25 mV.
Two, the fuel gauge. It estimates SOC and SOH. It cannot measure either one directly, so it combines three methods that each fail in different ways, and the engineering challenge is combining them so the failures do not overlap.
Three, the protection circuitry. It disconnects the pack when voltage, current, or temperature exceeds safe limits. It is layered, because no single layer can be trusted: software limits first, then an independent hardware protection IC, then contactors, then fuses and pyrotechnic disconnects.
Block one: the eyes, and what their accuracy costs
AFE accuracy for automotive-grade parts spans roughly 1–10 mV, depending on the part. The two anchors, both from their datasheets:
- ADI LTC6811: a multicell battery stack monitor that measures up to 12 series cells with a 1.2 mV maximum total measurement error, better than 0.04% accuracy, all 12 cells in 290 µs, with a 16-bit delta-sigma ADC and a 1 Mbps isolated serial interface, fully specified from −40 °C to +125 °C, and engineered for ISO 26262-compliant systems.
- TI BQ76952: a 3- to 16-series battery monitor and protector for Li-ion, Li-polymer, and LiFePO4 packs, with high-accuracy cell voltage measurement of less than 10 mV (typical), a high-accuracy coulomb counter with an input offset error below 1 µV (typical), synchronized voltage and current sampling, and DEEPSLEEP at 9–10 µA with SHUTDOWN at 1 µA.
The pipeline's "TI BQ769x2 ±3 mV" could not be confirmed, the datasheet says less than 10 mV typical, not ±3 mV, and the article uses the datasheet figure. The point the pipeline made still holds, and it is the more important one: the fuel-gauge algorithm matters more for SOC accuracy than the AFE's last millivolt. A 1.2 mV AFE with a naive coulomb counter drifts past ±3% in a week; a 10 mV AFE with a good Kalman filter holds ±2–3%. The demo prices both.
And TI's own technical article gives the reason accuracy matters at the top of charge: if the pack must be disabled when a cell voltage exceeds 4.350 V, but the accuracy of the voltage measurement is only within ±25 mV, then the controller needs to disable the pack when the reported measurement exceeds 4.325 V, in case the error is −25 mV. The actual error may instead be +25 mV, meaning the actual cell voltage is only 4.300 V, so as much as 50 mV of usable battery capacity is unavailable due to the inaccuracy of the measurement.
The more accurate the voltage measurement, the less cell capacity is wasted. The demo turns that into a number.
Block two: the accountant that has no sensor to read
State of charge is not a physical quantity you can probe with a sensor. It is an internal thermodynamic state, accessible only through imperfect proxies: voltage, current, and temperature. Every production BMS stacks three methods, and the three fail in complementary ways.
Method one, coulomb counting. Bookkeeping. SOC(t) = SOC(0) − (1/Q) × ∫ I(τ) dτ, where Q is the cell's full capacity and η, the coulombic efficiency, is close to 1 for lithium-ion. It costs one multiply-add per sample and is exact in the absence of measurement error. Its weakness is that it is an open-loop integrator: any constant bias in the current reading becomes a ramp in the SOC estimate, and nothing in the method corrects it. A 1% current error at 1C over a full discharge is a 1% SOC error, but a 1 mA offset on a 10 A full-scale sensor over 24 hours is a fixed drift of 2.4% per day that eventually saturates the estimate. After a few hours of driving, the estimate can be off by 5–10%.
Method two, OCV lookup. Rest the cell until the terminal voltage settles, typically one to four hours, and read the SOC off the OCV curve. It is absolute: it does not drift, because it does not integrate. Its weakness is that it needs rest, and the OCV-SOC relationship is chemistry-dependent. For NMC, a rested OCV gives ±1–2% at a known temperature. For LFP, the curve is almost flat across the middle of the range, the resting voltage changes less than 50 mV across 70% of the SOC range, so voltage-based correction is nearly useless between 20% and 90% SOC, and the accuracy there is ±5–15%, fundamentally limited by the flat curve. An LFP BMS must anchor at the endpoints only.
Method three, the Kalman filter. The extended Kalman filter treats the cell as a dynamic system, a Thevenin equivalent, an OCV source in series with a resistance and one or two RC polarization networks, and estimates SOC as a hidden state. It predicts with the coulomb-count equation and corrects with the voltage innovation, using the OCV-SOC slope as the gain that tells the filter how much voltage error implies SOC error. The rule of thumb: an EKF buys ±2–3% accuracy where naive coulomb counting drifts to ±10% after a few days without anchoring. For NMC, the Kalman gain is high in the mid-SOC range because the OCV slope provides strong observability. For LFP in the plateau, the OCV slope is near zero, the gain collapses to near zero, and the filter essentially behaves as pure coulomb counting, which is why the LFP problem is not solved by a better filter, but by anchoring at the endpoints and by counting well.
The hybrid architecture is what production gauges use: coulomb counting between rest points, with the OCV measurement used to re-anchor the coulomb counter whenever the cell rests long enough, more than 30 minutes, per the guide from Article #6, and the Kalman filter blending the two under load. The accuracy table, from the checked guides: OCV lookup ±5–10% (flat chemistries worse); coulomb counting ±1% short-term, drifting with time and offset; EKF/model ±2–3%, self-correcting.
And the practical standard: ±3% is acceptable for consumer EVs; commercial vehicles may need ±1% for fleet optimisation.
SOH, the second number. State of health is the capacity fade and the internal-resistance growth. End of life is typically defined as SOH below 80% for EV applications, or below 70% for grid storage. A Grade-A LiFePO4 cell can deliver 4,000–6,000 charge-discharge cycles under ideal conditions, but only if the BMS keeps it inside the box for all of them. A poor BMS degrades good cells; a great BMS extends the life of average cells.
Block three: balancing: burn it, or move it
No two cells are perfectly identical, they have slightly different capacities, internal resistances, and self-discharge rates. Manufacturing cell-to-cell capacity variation in the same lot is ±1–3%. Without balancing, the weakest cell limits the whole pack: the BMS triggers undervoltage protection when the weakest cell reaches 2.8–3.0 V, even if the other cells still have usable energy, and it triggers overvoltage protection when the strongest cell reaches 4.2 V, even if the others are not yet full.
A 5% capacity spread costs the pack 5% of its capacity, and the charge terminates early, every cycle, for the life of the pack.
Passive balancing dissipates the excess charge from the high-SOC cells as heat through a resistor. It is simple, cheap, and reliable, and it is universally used in production EVs, because of its simplicity and reliability. Typical passive balancing current is 50–200 mA; integrated BMS chips limit it to 50–100 mA to avoid excessive PCB temperatures and localised hot spots, and the resistor dissipates 0.3–0.5 W per cell. A 100 mA current through a 37 Ω balancing resistor dissipates 370 mW per cell. The energy is wasted, efficiency 0%, by definition, and the balancing is slow: in one verified test on a 16S 280 Ah LFP pack, redistributing a 50 Ah SOC difference at 87 mA took 574 hours and generated 46.2 W of heat across the 16 cells.
Active balancing transfers the charge from the high-SOC cells to the low-SOC cells, using switched capacitors, inductors, or isolated DC/DC converters. The transferred energy is recovered into the pack rather than dissipated as heat. Typical currents are 0.5–5 A depending on topology, flying capacitors 100–500 mA, isolated DC/DC 500 mA–2 A, inductive 1–10 A, at 80–95% transfer efficiency (capacitive 80–90%, inductive 85–95%, DC/DC 85–95%). In the same 16S 280 Ah test, a 5.2 A inductive balancer corrected the 50 Ah difference in 9.6 hours with 2.3 W of heat, and a 3.1 A capacitive balancer in 16.1 hours with 0.8 W, roughly 60 times faster than passive, with twenty times less heat. The cost is the reason it is not universal: relative to passive at 1.0×, switched-capacitor active costs 2.5–3.5×, and inductive or flyback active costs 4.0–6.0×. The pipeline's "3–5× BOM cost" sits inside that range.
The choice, from the guides: passive for small packs, under 100 Ah per cell, short strings up to 8S, with tight initial cell matching, where simplicity and cost matter. Active for large packs, over 100 Ah per cell, long strings above 8S, where energy recovery and fast balancing matter. And the maintenance point the guides all make: passive balancing only works during charging or idle, when the pack is not delivering current, because the bleed resistor cannot keep up with a load.
A pack that is cycled hard and never rests cannot be balanced by a resistor, which is why the balancing decision is a duty-cycle decision, not a cost decision.
Protection thresholds need margin, and the margin is a tolerance stack
The protection thresholds are the most commonly misconfigured part of a BMS. The rule the guides agree on: set the protection thresholds just inside the cell's absolute maximum ratings, providing a safety margin without unnecessarily restricting the battery's usable capacity, and account for the tolerance of the detection circuit, because a nominal threshold is not the threshold the pack actually trips at.
Overvoltage. The charger terminates at the cell's maximum charge voltage, 4.20 V for NMC/NCA, 3.65 V for LiFePO4. The hard protection threshold sits above it. The sources split on exactly where, and the split is the teaching: one guide says set OVP at 4.20 V per cell, the standard charge termination voltage, and warns that setting it higher does not give the cells more charge, it exposes them to electrolyte oxidation and thermal stress, because cathode oxygen release accelerates above 4.25 V for NMC and thermal runaway risk rises sharply above 4.30 V. Another guide says a typical OVP setting for a 4.20 V cell might be 4.25 V ±0.025 V, to account for the detection tolerance and to protect against voltage overshoot and component drift over time. The honest design is the two-layer scheme: the charger stops at 4.20 V, and the hard protection trips at a nominal 4.25 V, with the tolerance stack analysed, because a ±25 mV detector at a 4.25 V nominal trips anywhere from 4.225 V to 4.275 V, and the top of that range is past the 4.25 V point where oxygen release accelerates. The demo shows the fix: set the nominal at 4.225 V, so the worst case is 4.25 V, or use a tighter detector.
The pipeline's "4.25–4.28 V" is at the high end, and the top of it is wrong: a 4.28 V nominal with a ±25 mV detector trips as high as 4.305 V, past the damage onset, and into the range where thermal runaway risk rises sharply. The article does not use it.
Undervoltage. The cell's minimum discharge voltage is typically 2.5 V for NMC/NCA, and the safe UVP setting is 2.7–2.8 V, above the absolute minimum, because the cell voltage drops steeply below 3.0 V and the remaining energy is minimal, and because deep discharge causes copper dissolution from the anode current collector, which can create internal short circuits on the next charge cycle. Most boards set UVP at 2.8–3.0 V. Setting the UVP too close to the absolute minimum causes nuisance trips; setting it too low costs the pack its life.
Overcurrent and short circuit. The BMS limits discharge current to the cells' continuous rating, and its short-circuit protection triggers in microseconds at much higher currents, 100+ A, to catch hard shorts rather than moderate overloads. The timing matters: voltage and temperature thresholds are usually evaluated on a 100 ms to 1 s cadence, which is adequate because those parameters change slowly; overcurrent and short-circuit protection must react in milliseconds, which is why current sensing for protection is often handled by a separate fast comparator path rather than the telemetry ADC loop.
The recovery gap. The BMS will not reconnect the charge path until the triggering cell's voltage drops below the OVP recovery threshold, typically 4.10 V for a ternary cell, 3.55–3.60 V for LiFePO4. The gap prevents rapid reconnect-trip cycling: protection activation, voltage drop, protection release, recharging, re-protection, a cycle that wears the pack and the switching devices for no benefit.
Temperature, and the 0 °C rule. The BMS halts charging below 0 °C, because charging below freezing causes lithium plating, permanent anode damage that cannot be reversed. Charging stops above 45 °C, discharge disconnects at 60 °C, and the rate-of-rise alarm from Article #14 fires on a temperature rise above 1 °C per minute (warning) and above 5 °C per minute (emergency disconnect). For LiFePO4, the safe range is 2.5–3.65 V per cell; going outside it, even briefly, causes permanent capacity loss.
The topology: where the boards live
Three topologies, and the choice is a pack-size decision.
Centralised. All measurement and control circuitry concentrated in a single control unit, directly connected to each battery cell. Compact, reduced cabling, smaller footprint, lower cost, and the high-voltage and low-voltage modules sit on the same board, so electrical isolation and safety clearances need special attention. Typical use: e-cycles, scooters, e-rickshaws, small energy storage, up to about 16–24 series cells. Beyond that, the long sense harnesses become heavy, noisy, and hard to assemble.
Distributed. Each cell or group of cells has a dedicated monitoring module with full BMS functionality, operating independently or with minimal coordination. The monitoring electronics sit on each module or cell group; the master handles estimation, contactors, and CAN. Typical use: high-voltage car and bus packs at 400–800 V.
Modular (master–slave). The combination: each module has a slave unit that manages a subset of cells and communicates with a central master controller, usually over an isolated daisy chain (isoSPI, TPL, or iso-UART, depending on the chip maker). It keeps the sense wires short, the modules test independently, and the design scales across pack sizes, which is why it dominates EV and stationary storage. Modular designs have better scalability and increased fault isolation than centralised ones, and are well suited to electric vehicles, renewable energy storage, and mission-critical applications. A fourth form is emerging: wireless BMS, where the module monitors talk to the master by radio instead of a wired daisy chain, cutting harness weight, on some newer car platforms.
Protection is layered: because every layer can fail
Lithium battery protection is not a contest to see which device opens first. The BMS, the contactors or MOSFETs, the pack fuse, and any cell- or module-level fusible links must each address a defined fault range. A design is successful when normal current and permitted pulses pass without interruption, recoverable faults are handled by active protection, and severe faults are cleared before conductors, cells, switching devices, or the enclosure exceed their withstand limits.
Layer one, software limits. The BMS controller samples cell voltages, pack voltage, string currents, module temperatures, and insulation resistance, and evaluates them against configured thresholds. When a value crosses a limit, the BMS executes a pre-defined response, typically opening the contactors that isolate the battery from the load and the charger. Software is the most flexible layer, and the least trustworthy: it depends on sensors, power, software, and an operable switching device.
Layer two, the independent hardware protection IC. A separate chip, with its own thresholds, that cannot be overridden by the software. The "Three Gates" framing from the vehicle-architecture literature: hardware protection in the chips cannot be overridden; the BMS software's protection requests can be graded; and the vehicle control plane decides how to honour those requests. The hardware layer is what catches the fault when the software has crashed, or when the software's own threshold is wrong.
Layer three, the contactors, with welded-contactor detection. The contactor is the commanded isolation: it opens when told to, and it has a finite breaking capacity. Under severe stress it may fail short, or weld. The BMS checks that the contactor actually opened after it was commanded to open, by measuring the voltage across it, and a contactor that does not open is a protection layer that has silently failed.
Layer four, the fuse and the pyrotechnic disconnect. The main pack fuse is independent of BMS software and control power. It protects against high-energy faults that exceed the switching device's interrupting capability, as well as faults that remain energised if a contactor welds or a MOSFET fails short. The fuse is therefore a final protective barrier, not a substitute for correct sensing, pre-charge, contactor selection, conductor sizing, or fault containment. The coordinated assignment: BMS logic handles monitored, recoverable conditions; contactors provide commanded isolation; the fuse clears faults beyond the safe switching capability; a pyro fuse, if used, provides triggered emergency disconnection. The fuse must clear before the fault exceeds the verified withstand of cables, busbars, contactors, MOSFETs, connectors, PCB conductors, insulation, and downstream equipment.
And the layer the BMS cannot provide, from Article #14: the BMS measures the casing and the busbar, not the inside of the cell, so there is a lag between the internal fault and the sensor's reading. The BMS disconnect does not save the failing cell, it buys time for the other layers. A safety case that ends at the BMS is a safety case that has not followed the failure past the disconnect.
The standards, and the market
BMS design is governed by ISO 26262 (automotive functional safety, the ASIL ratings), IEC 62619 (industrial lithium-ion safety), and UL 1973 (stationary storage). The LTC6811 is engineered for ISO 26262-compliant systems, which is why it is the anchor part in this article.
The market figures disagree by scope, and the article says so rather than picking one: the 2024 estimates cluster around $7–9.1 billion (ResearchAndMarkets $9.1B, Expert Market Research $8.65B, Newstrail's automotive-only figure $7.9B), and the forecasts to 2029–2034 range from $15 billion (Exactitude, 8.1% CAGR) to $54 billion (Expert Market Research, 20.1% CAGR). The pipeline's "$7.9B in 2024, projected $22.3B by 2032 (~13.8% CAGR)" is one vendor-guide projection, its arithmetic is consistent (13.9% CAGR over eight years), but it does not match any single source's figure in this round, and the automotive-only Newstrail report that gives the $7.9B starting point projects $29.85B by 2032 at 18.1% CAGR.
The honest point is the one every source agrees on: every lithium pack needs a BMS, and the market is growing fast because the margin between safe operation and thermal runaway gets thinner as energy densities pass 270 Wh/kg.
The BMS's own power: the guard that eats the envelope
One trap the guides all name and the budgets often miss: the BMS itself draws current, its monitoring, its balancing, its protection FETs. On a small pack, the BMS quiescent current is a line in the budget, next to the regulator and the floating pin. The BQ76952's DEEPSLEEP is 9–10 µA and its SHUTDOWN is 1 µA, on the demo pack, 10 µA is 0.24 mAh a day, 0.007% of the pack, negligible.
A discrete BMS drawing 1 mA is 24 mAh a day, 0.69% of the pack a day, the same order as the coulomb-counting offset drift in Step 2, and enough to matter on a pack that sits for months. A guard that eats the envelope is a guard that needs budgeting too.
Example Labs is a made-up company we use for worked examples.
Example Labs' 48 V bench pack, designed on paper
The pack is the one the bench arm runs on, from Articles #20 and #23: the shoulder drive needs 199 W mechanical, 221 W electrical, from a 48 V bus. The pack is 13S1P NMC, 3.5 Ah per cell, 48.1 V nominal, 54.6 V at full charge, 168 Wh. Every number below was run in Python before writing.
Step 1: the accuracy budget: what the AFE's millivolts cost
The NMC OCV curve is steep at the ends and flat in the middle. In the 20–80% SOC range it spans roughly 3.6–3.9 V, 300 mV over 60% of charge, or 5 mV per 1% SOC. At the top of charge, 90–100% SOC spans roughly 4.05–4.20 V, 150 mV over 10%, or 15 mV per 1% SOC.
- A 10 mV AFE error (the BQ76952's typical) maps to 2.0% SOC error through the mid-range slope, two-thirds of the ±3% budget, from the measurement alone
- A 1.2 mV AFE error (the LTC6811's maximum) maps to 0.24% SOC error, one-twelfth of the budget
- TI's own example, at the top of charge: a ±25 mV detector forces the disable at 4.325 V for a 4.350 V limit, wasting 50 mV of usable capacity, and at 15 mV per 1% SOC, that is 3.3% of the pack's capacity thrown away by the measurement, not by the cells
- The lesson is the pipeline's, with the numbers behind it: the fuel-gauge algorithm matters more than the last millivolt, but the last millivolt still costs capacity, and the cost is measurable
Step 2: coulomb counting drifts, and the anchor catches it
The pack is 3,500 mAh. A 1 mA offset in the current-sense chain, the size of the BQ76952's coulomb-counter input offset error across a 1 mΩ shunt, integrates to 24 mAh a day:
- 0.69% of the pack per day, from a 1 mA offset alone
- 4.8% after seven days without a rest or an anchor, past the ±3% target, from an offset the size of the datasheet's typical figure
- The fix is the hybrid architecture: coulomb counting between rest points, with the OCV measurement re-anchoring the counter whenever the cell rests for more than 30 minutes, and the Kalman filter blending the two under load. The EKF buys ±2–3% where naive coulomb counting drifts to ±10% after a few days without anchoring, and for LFP, where the mid-range OCV slope is near zero, the anchor only works at the endpoints, so the counting has to be good and the rest has to be real
Step 3: balancing: 5% apart, and the two ways to close it
A 5% capacity spread on the 3.5 Ah pack is 175 mAh to move, and without balancing, the pack loses 5% of its capacity, 8.4 Wh of 168 Wh, on every cycle, because the weakest cell triggers the protection first.
- Passive, at 100 mA: 175 mAh ÷ 100 mA = 1.75 hours of bleeding, at 370 mW of heat per cell through a 37 Ω resistor, and at 50 mA, 3.5 hours. It only works while the pack is charging or idle, because the bleed resistor cannot keep up with a load
- Active, at 2 A: 175 mAh ÷ 2 A = 5.25 minutes, recovering 85–95% of the energy instead of burning it
- The big-pack illustration, from the verified test: a 16S 280 Ah LFP pack with a 50 Ah SOC difference, passive at 87 mA took 574 hours and generated 46.2 W of heat; active inductive at 5.2 A took 9.6 hours with 2.3 W; active capacitive at 3.1 A took 16.1 hours with 0.8 W. Sixty times faster, twenty times less heat, at 2.5–6× the cost
- The choice, on the numbers: passive for this 3.5 Ah bench pack, where 1.75 hours of bleeding during an overnight charge is free; active for the 280 Ah pack, where 574 hours is not a balancing time but a calendar
Step 4: the protection thresholds, with the tolerance stack
The charger terminates at 4.20 V. The hard overvoltage protection is set above it, and the setting is a tolerance stack, not a number.
- Nominal 4.25 V, detector tolerance ±25 mV: the actual trip falls anywhere from 4.225 V to 4.275 V, and the top of that range is 25 mV past the 4.25 V point where cathode oxygen release accelerates for NMC
- The fix: set the nominal at 4.225 V, so the worst case is 4.25 V, or use a tighter detector. The margin is in the stack, not in the nominal
- The pipeline's 4.28 V nominal, with the same ±25 mV detector: the actual trip falls from 4.255 V to 4.305 V, 55 mV past the damage onset, into the range where thermal runaway risk rises sharply. That is why the article does not use it
- The recovery gap: the charge path does not reconnect until the cell falls to 4.10 V, the gap that prevents the trip-reconnect-trip cycle
- The undervoltage side, for completeness: the cell's minimum is 2.5 V, the UVP is set at 2.8–3.0 V, and the same tolerance logic applies, the worst case must stay above the copper-dissolution threshold, not the nominal
Step 5: the short circuit, and which layer catches it
A hard short across the pack terminals, through a loop of 10 mΩ of wiring and busbar, with each of the 13 cells contributing 25 mΩ of internal resistance (the article's stated assumption for a typical 18650):
- Loop resistance: 13 × 25 mΩ + 10 mΩ = 335 mΩ
- Prospective short-circuit current: 48.1 V ÷ 0.335 Ω = 144 A, and the power in the loop is 6.9 kW
- The software loop (100 ms to 1 s cadence) is too slow for a hard short, the current is already at 144 A before the software has taken its next sample
- The fast comparator path and the hardware protection IC must react in milliseconds, and the contactor must open, and if the contactor welds, the fuse must clear the fault before the conductors, the cells, or the enclosure exceed their withstand limits
- The BMS's own draw, for the budget: the BQ76952 in DEEPSLEEP at 10 µA is 0.24 mAh a day, 0.007% of the pack, negligible. A discrete BMS at 1 mA is 24 mAh a day, 0.69% of the pack a day, the same order as the coulomb-counting drift in Step 2, and the reason the BMS's quiescent current is a line in the budget
The decision test
Five questions, in order. If you cannot answer the first, the rest do not matter.
- What does your AFE actually measure, and what does its error cost? If the error maps to 2% of SOC through the OCV slope, two-thirds of the ±3% budget is spent before the algorithm starts. If nobody knows the AFE's error, the SOC number is a guess with a display.
- How does your fuel gauge re-anchor? Coulomb counting alone drifts without bound, 0.69% a day on the demo pack from a 1 mA offset. If there is no rest-based OCV anchor and no Kalman filter, the gauge is a clock that runs fast, and the only question is how fast.
- What is your balancing time, on your duty cycle? If the pack is cycled hard and never rests, passive balancing cannot keep up, and the weakest cell will limit the pack by 5% a cycle, every cycle, for the life of the pack.
- What is the worst case of your protection thresholds, not the nominal? If the OVP nominal is 4.25 V and the detector is ±25 mV, the worst case is 4.275 V, past the damage onset. The margin is in the stack, not in the nominal.
- What catches the fault when the software is gone, and when the contactor welds? If the answer is nothing, no independent hardware protection IC, no welded-contactor detection, no fuse, then the protection is a single point of failure, and the safety case ends at the software.
Objections
"But our cells are well matched, so we don't need balancing." Manufacturing cell-to-cell capacity variation in the same lot is ±1–3%, and the cells age at different rates, so the spread grows with every cycle. A pack that starts matched does not stay matched. The balancing decision is not about the first cycle. It is about the five-hundredth.
"But the BMS chip does SOC, so we don't need to think about it." The chip does the arithmetic. The algorithm is the design: the coulomb counter, the OCV anchor, and the Kalman filter, and the choice of which one to trust when they disagree. A chip with a naive coulomb counter drifts past ±3% in a week on the demo pack. The chip is not the fuel gauge. The fuel gauge is the chip plus the algorithm plus the rest points.
"But we set the OVP at 4.28 V and it has never tripped." It has never tripped because the charger terminates at 4.20 V, and the day it trips, it will trip at up to 4.305 V, past the damage onset. A protection threshold that has never been tested is not a protection threshold. It is a hope, with a number on it.
"But LFP is safe, so we don't need the layers." LFP is safer, and safer is not safe. LFP cells can enter thermal runaway: the SEI still breaks down at 80–120 °C, the separator still melts at 130–165 °C, and the electrolyte still vents violently. What LFP removes is the cathode oxygen, the leg of the fire triangle that makes the fire self-sustaining, so the same barrier buys more time and the same vent path has less to carry. The layers still matter, because the cell still gets extremely hot, still vents toxic gas, and still heats its neighbours. And LFP makes the fuel gauge harder, not easier, because the flat OCV curve removes the anchor the NMC gauge relies on.
Sources
- Analog Devices, LTC6811-1/LTC6811-2 datasheet (Rev. C)
What we used
Measures up to 12 series cells with a 1.2 mV maximum total measurement error (better than 0.04% accuracy); all 12 cells measured in 290 µs; 16-bit delta-sigma ADC with programmable noise filter; built-in isoSPI interface, 1 Mbps isolated serial communications; engineered for ISO 26262-compliant systems; fully specified from −40 °C to +125 °C
- Texas Instruments, BQ76952 datasheet (SLUSE13B, Rev. B, Nov 2021)
What we used
3- to 16-series high-accuracy battery monitor and protector for Li-ion, Li-polymer, and LiFePO4 packs; high-accuracy cell voltage measurement less than 10 mV (typical); high-accuracy coulomb counter with input offset error under 1 µV (typical); wide-range current applications (±200 mV across the sense resistor); support for simultaneous current and voltage sampling; DEEPSLEEP 9–10 µA, SHUTDOWN 1 µA; high-side charge-pump NFET drivers
- Texas Instruments, "Improving voltage measurement accuracy in battery monitoring systems" (SSZT315, technical article)
What we used
If the pack must be disabled when a cell voltage exceeds 4.350 V but the accuracy is only within ±25 mV, the controller needs to disable at 4.325 V, and as much as 50 mV of usable battery capacity is unavailable due to the inaccuracy; the more accurate the voltage measurement, the less cell capacity is wasted; the BQ76942 and BQ76952 support synchronized voltage and current measurements, with a 45 ms loop time for a 10S pack and 63 ms for a 16S pack
- InnovChip, "Battery SOC Estimation Guide: Coulomb Counting, OCV Lookup and EKF"
What we used
SOC cannot be measured directly; the OCV-SOC curve is measured by resting the cell until the terminal voltage settles (typically 1–4 hours); coulomb counting accuracy is set almost entirely by the current-sense chain (a 1% current error at 1C over a full discharge is a 1% SOC error; a 1 mA offset on a 10 A full-scale sensor over 24 hours is a fixed drift of 2.4% per day); the EKF propagates the state with the coulomb-count equation and corrects it with the voltage innovation, using the OCV-SOC slope as the gain; an EKF buys ±2–3% accuracy where naive coulomb counting drifts to ±10% after a few days without anchoring; OCV lookup ±5–10% (flat chemistries worse), coulomb counting ±1% short-term, EKF/model ±2–3%
- InnovChip, "Battery Fuel Gauge Design Guide"
What we used
The fuel gauge converts measurements into a trusted SOC, SOH, and remaining runtime; a good gauge holds 1–3% error across temperature and load; production gauges fuse coulomb counting between rest points with OCV re-anchoring, and the most capable parts add a full state-space/Kalman estimator; the stated accuracy (±1–3%) is only reachable when the cell is genuinely rested at an OCV opportunity and the current ADC offset is nulled; simple voltage table ±10–20%, coulomb counter only ±5–10% drifting, integrated fuel-gauge IC ±1–3%, full state-space/learning under 1%
- EVPULSE, "How a BMS Knows What It Cannot See"
What we used
SOC estimation uses three algorithms in combination: coulomb counting for dynamic tracking, OCV lookup for periodic anchoring, and a Kalman filter for optimal fusion; LFP's flat OCV curve makes voltage-based correction nearly useless in the 20–90% SOC range; LFP BMS must anchor at endpoints only; the EKF can recover from a wrong initial SOC estimate, unlike pure coulomb counting; coulomb counting ±2–3% with OCV anchoring (well-tuned, 50 cycles), OCV-based NMC ±1–2% (rested 2 hours), OCV-based LFP mid-plateau ±5–15%, EKF ±1–2% (well-parameterised, 25 °C); for LFP, even resting voltage changes less than 50 mV across 70% of the SOC range; for NMC, a 10 mV terminal voltage error at 50% SOC corresponds to a well-defined SOC correction; for LFP in the plateau, the Kalman gain collapses to near zero and the filter behaves as pure coulomb counting
- Skill-Lync, "SOC Estimation" (EV Fundamentals)
What we used
SOC must be estimated from measurable quantities; coulomb counting drifts and after a few hours of driving the estimate can be off by 5–10%; the Kalman filter is the gold standard, combining coulomb counting (predict) with OCV measurement (correct); accuracy comparison: coulomb counting ±5–10% with drift, OCV lookup ±2–5% at rest, Kalman filter ±1–3% with minimal drift; practical accuracy ±3% is acceptable for consumer EVs, commercial vehicles may need ±1%; accuracy of ±1–3% is achievable with modern algorithms
- IoT Digital Twin PLM, "State of Charge Estimation in a Battery Management System"
What we used
The foundational reference for model-based SOC work is Gregory Plett's three-part series on extended Kalman filtering for lithium-ion polymer battery packs, published in the Journal of Power Sources in 2004; current integration is accurate over seconds and wrong over days, voltage lookup is accurate at rest and wrong under load, and a filter that fuses both with a model gets the strengths of each
- SSLBattery, "Active vs Passive Battery Balancing: When Each Wins"
What we used
Passive balancing dissipates excess charge as heat (typically 50–200 mA balance current), simple and low-cost but slower and wastes energy; active balancing transfers charge at 0.4–10 A with 80–95% transfer efficiency; passive 50–200 mA, active (capacitor) 1–5 A at 80–90%, active (inductor) 5–10 A at 85–95%; balance time for a 100 Ah pack at 5% imbalance: passive 10–25 hours, capacitive 1–3 hours, inductive 0.5–1 hour
- HoloBattery, "Active vs Passive BMS Balancing for HV Battery Packs"
What we used
Passive balancing dissipates excess charge through resistors at 50–100 mA; integrated BMS chips limit passive current to 50–100 mA to avoid excessive PCB temperatures and localised hot spots, and the resistor dissipates 0.3–0.5 W per cell; measured on a 16S 280 Ah LFP pack at 25 °C: passive 87 mA, 50 Ah SOC difference corrected in 574 hours with 46.2 W of heat; active inductive 5.2 A in 9.6 hours with 2.3 W; active capacitive 3.1 A in 16.1 hours with 0.8 W; efficiency 0% passive, 89% inductive, 96% capacitive (measured); energy recovery 85–92% inductive, 95–98% capacitive
- EVPULSE, "Cell Balancing in EV Battery Packs"
What we used
Manufacturing cell-to-cell capacity variation (same lot) ±1–3%; typical passive balancing current 50–200 mA; typical active balancing current 0.5–2 A; passive efficiency 0% (all energy dissipated as heat); active efficiency (per-cell isolated DC/DC) 85–95%; 100 mA through a 37 Ω balancing resistor dissipates 370 mW per cell; passive is universally used in production EVs because of its simplicity and reliability, active is reserved for high-value applications due to component cost and complexity
- Electrical Equipment Factory, "Lithium Ion Cell Balancing: BMS Engineering Design Guide"
What we used
Passive balancing restricted to bypass currents below 300 mA due to enclosure thermal dissipation limits; active architectures achieve currents up to 5 A at efficiencies exceeding 85%; relative cost factor 1.0× passive (baseline), 2.5–3.5× switched capacitor, 4.0–6.0× inductive/flyback; worked example: a 5.6 Ah deficit with an 88.5 mA passive shunt takes 63.3 hours
- Maxwell Energy, "Cell Balancing in BMS"
What we used
Passive balancing burns off excess charge as heat (efficiency 0%); active balancing moves it to weaker cells at 80–95% efficiency; passive 50–200 mA typical, active 0.5–5 A+ depending on topology; 50–200 mA is typically sufficient for maintaining balance in a well-matched pack but slow for correcting large imbalances
- Eneronix, "BMS Protection Explained: Overvoltage, Undervoltage, Overcurrent & Short Circuit"
What we used
OVP is the most fundamental BMS protection function and the one most commonly misconfigured; a cell pushed significantly past 4.25 V is a thermal runaway candidate; NMC damage onset 4.25–4.35 V, cathode oxygen release accelerates above 4.25 V, thermal runaway risk rises sharply above 4.30 V; NCA most sensitive, oxygen release begins at lower overvoltage, significant risk above 4.25 V; LiFePO4 OVP 3.65 V, damage onset 3.70–3.75 V; for NMC and NCA, set OVP at 4.20 V per cell, the standard charge termination voltages, setting OVP higher does not give the cells more charge, it exposes them to electrolyte oxidation and thermal stress; the hard protection response occurs within 50–100 ms; the BMS will not reconnect the charge path until the triggering cell's voltage drops below the OVP recovery threshold, typically 3.55–3.60 V for LiFePO4, which prevents rapid reconnect-trip cycling
- Link-B2B, "Guide to Lithium Battery Protection Board Cutoff Settings"
What we used
The OVP cutoff should be set at or slightly below the cell's specified maximum charge voltage; for a standard cell rated at 4.20 V, a typical OVP setting might be 4.25 V ±0.025 V, accounting for the tolerance of the protection board's detection circuit; setting OVP exactly at the cell's maximum voltage might not provide enough buffer, while setting it too low (4.15 V) prematurely stops charging; the UVP cutoff must be set above the cell's minimum discharge voltage, if a datasheet specifies 2.5 V minimum, a safe UVP setting is 2.7 V or 2.8 V
- The Day Time Report, "Lithium-Ion Battery BMS: When Does Overcharge Protection Trigger & How to Recover?"
What we used
For ternary lithium batteries the full charge voltage is 4.2 V per cell, so the BMS single-cell overcharge protection is set to 4.25 V, and the total voltage protection condition is 4.2 V times the number of cells; the overcharge recovery voltage is 4.1 V × N for ternary cells; the voltage gap between the trigger and the recovery threshold prevents the protection activation → voltage drop → protection release → recharging → re-protection cycle
- Electronic Saviors, "How Much Can Voltage Vary in a BMS?"
What we used
NMC/NCA maximum OVP 4.20–4.25 V (high-voltage variants up to 4.35 V); UVP 2.80–3.00 V (critical shutdown 2.50 V); cell voltage variations above 150 mV trigger a fault/safety cutoff; a 1–3× quota is the healthy pipeline coverage range for the demo's coverage analogy
- Large Battery, "Battery Pack Fuse Selection"
What we used
Lithium battery protection is not a contest to see which device opens first; the BMS, contactors or MOSFETs, pack fuse, and cell- or module-level fusible links must each address a defined fault range; the main fuse is independent of BMS software and control power, and protects against high-energy faults that exceed the switching device's interrupting capability, as well as faults that remain energised if a contactor welds or a MOSFET fails short; the fuse must clear before the fault exceeds the verified withstand of cables, busbars, contactors, MOSFETs, connectors, PCB conductors, insulation, and downstream equipment
- AserIndustry, "Fuse Selection for Battery Management Systems"
What we used
The BMS may open contactors when it detects overcurrent, but a contactor is not a substitute for a fuse; detection can fail, control power can disappear, and contactors can weld; a coordinated system assigns roles: BMS logic handles monitored, recoverable conditions; contactors provide commanded isolation; the fuse clears faults beyond the safe switching capability; a pyro fuse, if used, provides triggered emergency disconnection
- Battery Storage BMS, "BMS Protection Functions"
What we used
Voltage and temperature thresholds are usually evaluated on a 100 ms to 1 s cadence, which is adequate because those parameters change slowly; overcurrent and short-circuit protection must react in milliseconds, which is why current sensing for protection is often handled by a separate fast comparator path rather than the telemetry ADC loop; a 750 V lithium battery bank can deliver fault currents measured in kiloamperes within milliseconds of an internal short
- Archi-Intelligence, "Opening Authority: The Battery Management System and Vehicle E/E Architecture"
What we used
The "Three Gates": hardware protection in the chips cannot be overridden; the BMS software's protection requests can be graded; and the vehicle control plane decides how to honour those requests; in 2022 and again in 2024, US regulatory filings recorded recalls of the same kind: the high-voltage contactor opened while driving, and the vehicle lost propulsion
- E3S Web of Conferences, "Comparative study between two BMS architecture in a PV-battery off-grid: Centralized and modular topologies"
What we used
The top three BMS topologies are centralised (all measurement and control in a single unit, directly connected to each cell), distributed (each cell or group of cells has a dedicated monitoring module with full BMS functionality), and modular master-slave (each module manages a subset of cells and communicates with a central master controller); modular BMSs have better scalability and increased fault isolation, and are well suited to electric vehicles, renewable energy storage, and mission-critical applications
- PMC (National Library of Medicine), "Scalable, Decentralized Battery Management System Based on Self-organizing Nodes"
What we used
The classification divides BMS topologies into centralised, modularised, distributed, and decentralised; centralised BMSs with a single control unit are increasingly replaced by decentralised management, whereby sensor, control, and computing resources are distributed
- DIYguru, "BMS Deep Dive"
What we used
Centralised: one board, monitoring ICs, MCU, and switches, with sense wires running to every cell, for e-cycles, scooters, e-rickshaws, small ESS (up to ~16–24 series cells); modular master-slave: one monitoring board per module, linked to a master controller by an isolated daisy chain, for 3W/4W packs, buses, larger ESS; distributed: monitoring electronics on each module or cell group, the master handling estimation, contactors, and CAN, for high-voltage car and bus packs (400–800 V); wireless BMS on some newer car platforms; as packs grow beyond about 16–24 cells in series, long sense harnesses become heavy, noisy, and hard to assemble, so designs move to modular boards linked by isolated daisy chains
- ABK Battery, "PACK Manufacturing Process Series: Core Functions, System Architecture, and Hardware Structure of Battery Pack BMS"
What we used
The core function of BMS is to monitor battery voltage, temperature, and current in real time, optimise performance through SOC estimation and balancing control, and provide fault protection for overvoltage, overcurrent, and overtemperature; distributed BMS delegates acquisition to individual battery modules, with multiple slave control units achieving distributed sampling and preliminary data processing while the master control unit handles system-level management
- Article #6, "Power Design for Battery Products" (Senova, 2026-11-19)
What we used
The BMS thresholds, balancing, and SOC/SOH as the first layer; OVP 4.20–4.25 V per cell, UVP 2.8–3.0 V, no charging below 0 °C, LiFePO4 2.5–3.65 V, SOC ±3% via coulomb counting + Kalman correction, SOH end of life below 80% (EV) or below 70% (grid storage); standards ISO 26262, IEC 62619, UL 1973; BMS market $7.9B (2024) → $22.3B (2032), labelled as a vendor-guide figure
- Article #14, "Thermal Runaway" (Senova, 2026-11-19)
What we used
The temperature ladder (SEI breakdown 80–120 °C, separator melt 130–165 °C, cathode oxygen release 150–220 °C NMC / 250 °C+ LFP, full runaway above 300 °C, internal temperatures over 800 °C); the rate-of-rise alarm (1 °C/min warning, 5 °C/min emergency disconnect); charging stops above 45 °C, discharge disconnects at 60 °C; the BMS cannot see inside the cell, and the disconnect buys time for the other layers rather than curing an internal short
- Article #4, "The Kalman Filter: Estimating What You Cannot Measure" (Senova, 2026-11-19)
What we used
The estimator the BMS uses to fuse its sensors, and the same mathematics the fuel gauge applies to state of charge
- ResearchAndMarkets / Expert Market Research / Exactitude / Newstrail market reports (checked 2026-11-19)
What we used
The 2024 BMS market estimates cluster at $7–9.1 billion (ResearchAndMarkets $9.1B, Expert Market Research $8.65B, Newstrail automotive-only $7.9B); forecasts to 2029–2034 range from $15 billion (Exactitude, 8.1% CAGR) to $54 billion (Expert Market Research, 20.1% CAGR); Newstrail's automotive-only report projects $7.9B in 2024 to $29.85B by 2032 at 18.1% CAGR