How to Keep LFP Batteries Honest When Their Voltage Sensors Lie

Under a +3% voltage gain fault, conventional filters misestimate state-of-charge by over 22 percentage points. The new RBC-DEKF algorithm reduces that error to just 0.356—nearly eliminating the impact of sensor drift in one of the world’s most widely used battery chemistries.

That number isn’t theoretical. It comes from rigorous testing on real lithium iron phosphate (LFP) cells subjected to simulated sensor faults across dozens of drive cycles and temperatures. And while the math behind it runs deep, the implication is simple: we now have a way to keep LFP batteries honest even when their voltage sensors lie.

This matters because LFP batteries are winning the global energy transition. Safer, longer-lasting, and cheaper than nickel-rich alternatives, they power everything from Tesla Model 3s to utility-scale storage farms. But they come with a quirk: their voltage barely changes as they charge and discharge. Over 80% of their capacity, the terminal voltage stays so flat that a millivolt-level sensor error can translate into a massive state-of-charge (SOC) miscalculation. Get it wrong, and your EV might think it has 50% charge when it’s actually at 20%. Or worse—the battery management system (BMS) could push the cell into dangerous territory, triggering premature shutdowns or, in rare cases, thermal events.

Until now, most BMS designs assumed clean, well-calibrated sensors. But in reality, voltage sensors degrade. They drift. They suffer from aging, temperature swings, electromagnetic interference. And when that happens in an LFP battery, all bets are off.

The breakthrough here isn’t just detecting these faults—it’s surviving them without losing track of the true state.

The Science

At the heart of this work is a recognition: you can’t fix a problem if your diagnostic tool is part of the problem.

Traditional state estimators—like the ubiquitous extended Kalman filter (EKF)—rely heavily on voltage measurements to correct their internal state predictions. When the measured voltage doesn’t match the predicted voltage, the filter assumes the state is wrong and adjusts accordingly. That works fine for batteries with steep voltage curves, like NMC or LCO chemistries. But for LFP? Not so much.

On the flat plateau, a small, persistent voltage offset—say, due to a drifting sensor—looks identical to a large shift in SOC. The filter dutifully “corrects” the state, only to find the mismatch persists. So it keeps correcting. And correcting. Until the estimated SOC diverges completely from reality.

To break this cycle, the researchers built a dual-path estimator: the Residual-Bias Compensation Dual Extended Kalman Filter (RBC-DEKF). Instead of treating voltage residuals as pure state errors, it splits them into two components:

1. Electrochemical state evolution — governed by a reduced-order physics-based model called the Control-Oriented Parameter-Grouped Single-Particle Model with Thermal effects (CPG-SPMT)
2. Voltage-domain discrepancy — a separate, adaptive state that captures systematic sensor biases or gain errors

The CPG-SPMT is a simplified version of the full pseudo-two-dimensional (P2D) model, designed for real-time use. It tracks lithium concentrations in both the graphite anode and LFP cathode, modeling diffusion, reaction kinetics, and ohmic losses. Crucially, it preserves the paired-electrode dynamics that give rise to the flat OCV plateau.

Meanwhile, the voltage-discrepancy state acts like a continuously updated calibration offset. If the model consistently underpredicts voltage by 10 mV, the filter learns that and compensates—without touching the actual SOC estimate.

But there’s a catch: if you let this residual state adapt too early—especially when the initial SOC is unknown—you risk letting it absorb the real state error instead of the sensor fault. In other words, the fix becomes the problem.

So the team introduced a commissioning phase.

When the system starts up with uncertain SOC (a common scenario in real-world applications), the residual adaptation is temporarily suspended. Instead, the filter uses a short window of verified-healthy data—buffered current, temperature, and voltage—to align the electrochemical model with reality through trajectory matching. Only once the physical state is stabilized does it reactivate the residual channel, allowing it to begin tracking any persistent sensor discrepancies.

This staged approach—recover state first, then accommodate fault—is what makes the method robust.

What They Found

The results are striking. The researchers tested their RBC-DEKF against a conventional Single-EKF across 24 different combinations of temperature and drive cycle (DST, FUDS, US06), injecting both additive bias (constant voltage offset) and multiplicative gain (scaling error) faults.

Under additive bias—say, a sensor stuck reading 30 mV too high—the Single-EKF averaged a mean SOC RMSE of 7.646 percentage points. That means, on average, its state estimate was off by nearly 8% of total capacity. For context, most BMS targets aim for sub-3% error under normal conditions.

The RBC-DEKF? Just 0.170 percentage points—a 45-fold improvement.

For multiplicative gain faults—where the sensor scales all readings by, say, 1.03x—the gap was even wider. The Single-EKF’s error ballooned to 22.646 percentage points, effectively rendering the SOC estimate useless. At that level, the car might report 70% charge when it’s actually at 47%, or worse.

The RBC-DEKF held steady at 0.356 percentage points—less than half a percent off, even under this severe fault.

Mean SOC RMSE Under Additive Bias and Multiplicative Gain Faults

Label	Value
Additive Bias	7.646 percentage points
Multiplicative Gain	22.646 percentage points

Source: Guo et al., 2026. Lower values indicate more accurate SOC estimation.

They also tested how the system handles incorrect initial SOC—a realistic scenario when a battery is powered on mid-cycle or after a communication loss. Starting with deliberate errors, they ran 18 test segments anchored at 50% SOC. The commissioned RBC-DEKF achieved a full-record mean SOC RMSE of 1.747 percentage points, including the startup period. By contrast, both the always-on RBC-DEKF and a conventional joint EKF failed to recover properly, landing at 11.192 and 11.400 points respectively.

In other words: without the startup safeguard, even a smart filter can go astray.

Further analysis showed the method’s robustness across temperatures and drive cycles.

Temperature-Dependent SOC RMSE Under Severe Faults

Label	Value
-10°C	0.65 SOC RMSE (%)
0°C	0.48 SOC RMSE (%)
10°C	0.39 SOC RMSE (%)
25°C	0.31 SOC RMSE (%)
40°C	0.35 SOC RMSE (%)

Source: Guo et al., 2026. Performance remains stable across operating temperatures.

shows how SOC RMSE varies from -10°C to 40°C under severe +30 mV additive and +3% gain faults. Error increases slightly at low temperatures—expected, since kinetic limitations amplify model uncertainty—but never exceeds 0.85 percentage points. Across all conditions, the median SOC RMSE remained below 0.4 points.

And

Performance Across Temperature and Drive Cycles Under Severe Faults

Label	Value
DST	0.31 SOC RMSE (%)
FUDS	0.29 SOC RMSE (%)
US06	0.33 SOC RMSE (%)

Source: Guo et al., 2026. Radar plot shows consistency of performance across dynamic load profiles.

, a radar plot of performance across the three drive cycles under extreme faults, reveals something else: consistency. Whether it’s the aggressive acceleration of US06 or the stop-and-go rhythm of FUDS, the filter performs uniformly well. No single cycle exposes a fatal flaw.

Why This Changes Things

We don’t just need better batteries. We need batteries we can trust—especially as they become larger, more integrated, and harder to replace.

Today’s BMS designs are often brittle. They assume ideal conditions: perfect sensors, known initial states, predictable aging. But real-world operation is messy. Sensors degrade. Cars get repaired. Batteries are repurposed. And when those assumptions break, so does confidence in the SOC.

This paper offers a path beyond that fragility.

By decoupling physical state from sensor artifact, the RBC-DEKF treats voltage not as gospel, but as a signal to be interpreted—one that can be noisy, biased, or even deceptive.

That shift in perspective has ripple effects.

First, it enables longer service life. If a sensor develops a slow drift, today’s systems might flag a fault and require replacement. With this method, the BMS could silently compensate, extending hardware lifetime and reducing maintenance costs.

Second, it improves safety margins. Large SOC errors force conservative operation—keeping the battery away from true limits to avoid risk. With tighter error bounds, operators can safely utilize more of the available capacity, increasing effective range or storage duration.

Third, it opens doors for second-life applications. Retired EV batteries, often repurposed for grid storage, face uncertain histories. Initial SOC may be unknown; sensor calibration may be lost. A fault-tolerant estimator like this one could onboard such batteries reliably, without requiring full recalibration.

And fourth, it strengthens the case for LFP dominance. As automakers like Tesla, BYD, and Ford double down on LFP for standard-range models, concerns about state estimation have lingered. This work directly addresses that weakness, potentially accelerating adoption.

It’s worth noting that this isn’t just a software patch. It’s a rethinking of how estimation should work in systems where observability is fundamentally limited. The core insight—that you must sequence recovery and accommodation—could apply to other domains: fuel cells with sluggish oxygen sensors, supercapacitors with leakage currents, or even biological systems where indirect measurements dominate.

What’s Next

No solution is perfect. The RBC-DEKF still relies on a well-characterized electrochemical model. If the battery degrades in ways not captured by parameter drift—say, particle cracking or electrolyte dry-out—the model fidelity will erode, and with it, the filter’s ability to distinguish state from sensor error.

Moreover, the current implementation assumes static faults: constant bias or fixed gain. Real sensors can exhibit nonlinearities, hysteresis, or intermittent dropouts—challenges not yet addressed.

Future work could integrate this framework with online degradation tracking, creating a co-estimation system that adapts to both sensor faults and cell aging. Another direction: combining it with redundant sensing or distributed architectures, enabling fault isolation at the pack level.

There’s also the question of implementation cost. While the CPG-SPMT is far lighter than full P2D models, it’s still more complex than simple equivalent circuit models. Embedding this in low-cost microcontrollers will require careful optimization—though the authors note that modern BMS chips are increasingly capable.

Finally, field validation is essential. Lab tests are controlled. Real roads are not. Testing this algorithm on actual vehicle fleets, under real degradation patterns and environmental stress, will be the ultimate proof.

But already, the message is clear: we no longer have to choose between LFP’s safety and its estimation reliability. With smarter filtering, we can have both.

As the world electrifies, our machines will spend less time in labs and more time in the wild—where sensors fail, conditions vary, and trust must be earned, not assumed. This work is a step toward building systems that don’t just function under ideal conditions, but endure when things go wrong.