BMS Design Under Uncertainty — The Problems That Don't Have Clean Answers

Electrode slippage is the parameter estimation problem that conventional EKF cannot solve — it requires ICA or DVA to distinguish capacity loss from lithium inventory loss, and most production BMS designs cannot do either.

This article addresses the problems at the architectural level — the ones that persist after the basic EKF is correctly implemented, the model is well-parameterised, and the sensors are calibrated. These are the problems that appear at 500+ cycles in demanding Indian commercial fleet conditions: electrode slippage, parameter covariance identifiability, multi-cell observability limits, and the fundamental information-theoretic boundaries on what a voltage measurement can tell you about a battery's state.

The Parameter Identifiability Problem

A 2RC equivalent circuit model has 7 parameters per temperature point: R₀, R₁, C₁, R₂, C₂, Q_rated, and η. These parameters change with aging. The EKF can estimate states (SOC, RC voltages) well. Can it also estimate parameters online?

In principle, yes — this is the dual estimation / AEKF approach. In practice, there is an identifiability constraint: you can only estimate parameters that are observable from the available measurements. For a battery, the available measurements are terminal voltage and current. From these, you can observe:

• R₀: Observable from instantaneous voltage response to current steps
• R₁, τ₁: Observable from fast transient voltage response (sub-10-second dynamics)
• R₂, τ₂: Observable from slow relaxation dynamics (10–600 seconds)
• Q_rated: Observable only from a full discharge — the total charge extracted divided by the SOC change. Not observable from partial cycles.
• η: Near-1 for Li-ion; not practically observable from measurement in normal operation

The implication: Q_rated cannot be estimated online from partial cycles. A battery that is always opportunity-charged (never fully depleted) never provides the observation needed to update Q_rated. The BMS must either:

1. Force periodic full calibration cycles (full charge → full discharge)
2. Accept that Q_rated is unknown and use a conservative lower bound
3. Use ICA/DVA to extract capacity information from slow partial charges

import numpy as np

def monte_carlo_soc_uncertainty(n_samples: int = 10000):
    """
    Propagate parameter uncertainties through SOC estimator.
    Returns distribution of SOC error at 50% SOC.
    """
    rng = np.random.default_rng(42)

    # Uncertain parameters (mean, std)
    Q_Ah   = rng.normal(60.0, 3.0,  n_samples)   # capacity +/-5% (aging)
    I_bias = rng.normal(0.0,  0.15, n_samples)   # current sensor +/-0.15 A

    # Nominal: 1C discharge for 1800 s -> SOC drops from 1.0 to 0.5
    I_true   = 60.0   # 1C
    soc_true = 0.5

    # Measured current with bias
    I_meas  = I_true + I_bias
    soc_est = 1.0 - (I_meas * 1800.0) / (Q_Ah * 3600.0)

    error = soc_est - soc_true
    print(f"SOC error -- mean: {error.mean()*100:.2f}%  "
          f"std: {error.std()*100:.2f}%  "
          f"95th pctile: {np.percentile(np.abs(error), 95)*100:.2f}%")
    return error

monte_carlo_soc_uncertainty()

Electrode Slippage: The Hidden Degradation Mode

Conventional SOH estimation tracks total capacity (Q_measured/Q_rated). Electrode slippage is a degradation mode that changes the internal electrode balance without necessarily reducing total capacity in the short term — making it invisible to standard capacity tracking.

Mechanism: In a new LFP cell, the anode capacity is ~10% larger than the cathode capacity (anode-limited design). As the cell ages:

• Anode capacity decreases (lithium plating at low temperatures, SEI growth consuming Li inventory)
• Cathode capacity decreases (iron dissolution, structural changes)

If anode capacity degrades faster than cathode (common in hot cycling conditions), the relative electrode balance changes. The OCV-SOC curve shifts because the cathode phase transitions — which determine the LFP plateau features — now occur at different absolute SOC points.

Why this matters for BMS accuracy: The BMS's OCV-SOC table was characterised for a fresh cell. After electrode slippage, the OCV-SOC relationship has changed. The table is wrong. OCV corrections and EKF observations based on the wrong table produce systematic SOC estimation errors that grow with cycle count.

import numpy as np

def recursive_least_squares(V_history, I_history, dt_s, lam=0.99):
    """
    Online identification of R0 and R1 via Recursive Least Squares.
    Simple 1-RC model: V = OCV - R0*I - V_RC
    Returns R0, R1 estimates.
    """
    n     = len(V_history)
    theta = np.array([0.005, 0.003])   # [R0, R1] initial guess
    P     = np.eye(2) * 100.0          # initial covariance

    for k in range(1, n):
        dV_RC = np.exp(-dt_s / (theta[1] * 3000)) * I_history[k-1]
        phi   = np.array([-I_history[k], -dV_RC])   # regressor vector
        y_k   = V_history[k]                          # measurement
        e_k   = y_k - phi @ theta                    # prediction error
        P     = (P - (P @ np.outer(phi, phi) @ P) /
                 (lam + phi @ P @ phi)) / lam
        theta = theta + P @ phi * e_k

    return theta   # [R0_est, R1_est]

Incremental Capacity Analysis: The Diagnostic Tool

ICA computes the incremental capacity dQ/dV — how much charge enters the battery per millivolt of voltage increase during a slow charge. For a fresh LFP cell, this curve shows distinct peaks corresponding to the phase transitions in the LFP olivine structure. As the battery ages, these peaks:

• Shift left or right (electrode slippage)
• Decrease in height (active material loss)
• Broaden (increased internal resistance and polarisation)

Running ICA requires:

• Charge rate of C/20 to C/10 (very slow, to minimise overpotential)
• Voltage resolution of ±0.5 mV or better
• Complete charge from 0% to 100% SOC

For a fleet depot with overnight charging capability, ICA can be run on 1–3% of the fleet each night without impacting operational readiness. A standard BMS with 12-bit ADC (1 mV resolution) is adequate for ICA at C/10 rates. The analysis requires post-processing (the dQ/dV computation is noise-sensitive and requires smoothing) that can be done off-board in fleet management software.

The Adaptive Extended Kalman Filter

The standard EKF uses fixed Q (process noise covariance) and R (measurement noise covariance). The AEKF adapts these matrices online using the innovation sequence — the sequence of differences between measured and predicted terminal voltage.

Innovation-based Q adaptation: If the innovation is consistently larger than expected (measurement disagreements persist), the battery is behaving differently from the model. This could mean temperature changed rapidly, the cell is degrading faster than modelled, or a sensor is drifting. The AEKF increases Q to acknowledge higher model uncertainty, allowing faster state corrections.

Innovation-based R adaptation: If the innovation is smaller and more predictable than expected, the measurements are cleaner than the R matrix assumed. The AEKF decreases R, giving more weight to measurements. Conversely, if noise suddenly increases (vibration event, alternator load), R increases and the filter smooths rather than chasing noise.

The AEKF's advantage in Indian commercial fleet conditions:

• Temperature swings (road surface, traffic, time of day) cause rapid parameter changes
• Sensor degradation (contact resistance, connector corrosion) is progressive and hard to track
• Cell aging patterns vary across the fleet — a single fixed model does not describe all cells after 200+ cycles

Multi-Cell Observability: When Pack-Level BMS Is Not Enough

For packs with cell-to-cell variance (realistic after 200+ cycles), pack-level state estimation has a fundamental limitation: a single pack voltage measurement cannot be deconvolved into individual cell states.

Example: A 96S pack with two cells at 30% SOC and 94 cells at 60% SOC will show a pack OCV that is indistinguishable from a pack with all 96 cells at approximately 59.4% SOC. The pack-level BMS sees the average; the individual cell states are invisible.

This is why individual cell voltage monitoring — which all production BMS designs implement — is not equivalent to individual cell SOC estimation. Even with per-cell voltage measurement, the EKF must be run per-cell with per-cell parameters to achieve accurate per-cell SOC.

Running 96 parallel EKFs on a resource-constrained MCU (typical BMS: Cortex-M4 at 168 MHz) is feasible but requires careful implementation:

• Sparse matrix operations (most P matrices are nearly diagonal for independent cells)
• Fixed-point computation with overflow protection
• Prioritised computation (cells near protection thresholds get first computation slots)
• Temperature-indexed parameter lookup from shared table (saves MCU memory)

Fundamental Limits on SOC Observability

The Kalman filter Kalman gain:

K = P × Hᵀ × (H × P × Hᵀ + R)⁻¹

For LFP in the plateau, H = ∂OCV/∂SOC ≈ 0. The gain K → 0 regardless of sensor quality, model accuracy, or algorithm sophistication.

This is not a limitation of the EKF specifically — it is a fundamental information-theoretic result: when the measurement is insensitive to the state being estimated, no algorithm can estimate that state from the measurement. For LFP in mid-plateau, the battery's terminal voltage carries essentially no information about SOC. All estimation must come from current integration.

The practical implication: for LFP commercial EVs with opportunity charging (no regular full charges), the SOC estimate will drift unbounded and no BMS algorithm can prevent this. The only solutions are:

1. Periodic full charges (provides OCV endpoint anchor)
2. Impedance spectroscopy (EIS) for SOC estimation independent of OCV (expensive, not standard in automotive BMS)
3. Accepting higher SOC error and widening the safety margins accordingly

Key Takeaways