A particle filter tracks battery SOC by maintaining many possible state estimates, then predicts RUL by propagating those estimates through a degradation model. Each particle represents a possible combination of SOC, voltage-related states, temperature effects, and aging parameters. As current, voltage, and temperature measurements arrive, the framework predicts each particle forward, weights it according to measurement agreement, resamples the population, and projects the surviving particles until an end-of-discharge or end-of-life condition is reached.
Core takeaway: Particle filtering is useful because it does not force a nonlinear, aging battery into a single Gaussian estimate. It produces a probability distribution for SOC and degradation, allowing RUL to be reported as a forecast with uncertainty rather than as one unsupported number.
Why Particle Filtering Fits Non-Linear Battery Testing
Battery behavior is not adequately represented by a fixed linear model
Battery capacity loss, self-discharge, internal resistance growth, temperature sensitivity, and rate effects vary with operating history, SOC, and age. Pulsed or highly variable loads make these effects more pronounced.
A simplified linear estimator can become inaccurate when the battery operates outside the conditions used for calibration. A particle filter instead uses a nonlinear state-transition model that can represent changing parameters and irregular operating profiles.
The state includes more than SOC
A battery state vector may include:
[ x_k = \begin{bmatrix} SOC_k & V_{1,k} & V_{2,k} & R_k & Q_k & \theta_k \end{bmatrix}^{T} ]
Here, (SOC_k) is state of charge, (V_1) and (V_2) can represent dynamic voltage states, (R_k) can represent internal resistance, (Q_k) usable capacity, and (\theta_k) other degradation or model parameters.
The exact state depends on the selected battery model. The important principle is that SOC and aging-related parameters can be estimated together, rather than treating battery health as a fixed input.
How the Filter Tracks SOC
1. Initialize a population of possible battery states
At the beginning of a test, the filter creates particles:
[ {x_0^i, w_0^i}_{i=1}^{N} ]
Each particle (x_0^i) is a plausible initial state, and (w_0^i) is its probability weight. The particles can reflect uncertainty in initial SOC, capacity, resistance, temperature effects, and other model parameters.
For example, if the initial SOC is uncertain, the particles span a range rather than assigning the battery one exact value.
2. Propagate every particle through the battery model
When the battery is cycled, each particle is advanced using a state-evolution model:
[ x_k^i = f_{k-1}(x_{k-1}^i,u_k,\theta_k^i)+w_k^i ]
The input (u_k) normally includes measured current and may also include temperature, load conditions, or other operating variables. The process-noise term accounts for model error and unmeasured effects.
This step predicts how each possible battery state should evolve during the next sampling interval.
3. Predict the measurable output
The filter converts each predicted state into an expected measurement:
[ \hat{z}_k^i = h_k(x_k^i,u_k)+v_k^i ]
For SOC estimation, the measured output (z_k) may be terminal voltage, current, temperature, or a combination of these signals. The observation model links internal states such as SOC and resistance to the voltage seen by the test system.
4. Update particle weights using measurements
Once the actual measurement arrives, each particle is scored according to how closely its prediction matches that measurement:
[ w_k^i \propto w_{k-1}^i,p(z_k\mid x_k^i) ]
Particles that predict the measured voltage and other signals well receive higher weights. Particles that disagree receive lower weights.
This measurement update corrects errors accumulated during model-based propagation. In practical terms, the filter combines coulomb counting and model prediction with the information contained in measured voltage and other sensors.
5. Calculate SOC from the weighted particle distribution
The estimated SOC can be obtained from the weighted mean:
[ \widehat{SOC}k = \sum{i=1}^{N} w_k^i,SOC_k^i ]
The particle spread provides additional information about confidence. A narrow distribution indicates that the measurements and model constrain SOC relatively well; a wide distribution indicates greater uncertainty.
This is a major distinction from a single-point SOC estimate: the particle filter can show both the best estimate and the range of plausible SOC values.
How the Filter Predicts RUL
RUL requires a degradation state or health indicator
SOC describes the current charge condition. RUL requires an estimate of how the battery’s ability is changing over time.
The filter can therefore include degradation variables such as usable capacity, internal resistance, or a parameter associated with SEI-layer growth. These variables evolve alongside SOC during cycling.
A health index derived from capacity, resistance, or another degradation signal can also be used, provided the signal is cleaned of measurement noise and treated consistently with the battery’s expected degradation behavior.
Model degradation as a stochastic process
A degradation state may be represented generically as:
[ d_k = g(d_{k-1},u_k,T_k,\text{history})+\eta_k ]
The model can account for operating conditions such as current profile, temperature, duration, and accumulated cycling history. The process-noise term allows individual cells to deviate from the nominal degradation path.
This matters because two cells with the same present SOC may have different capacity-loss rates and therefore different remaining lives.
Forecast each particle into the future
After assimilating the latest test data, the filter propagates every surviving particle forward under the expected future load and environmental conditions.
Each particle produces a possible future trajectory for capacity, resistance, voltage, and SOC. The trajectories are then checked against a defined endpoint, such as:
- End of discharge (EOD): the battery reaches its voltage limit under the specified load.
- End of life (EOL): the battery reaches a capacity, resistance, performance, or safety criterion.
- Operational failure: the battery can no longer meet the required application demand.
The number of future cycles or time steps until the endpoint is reached becomes that particle’s RUL prediction.
Convert endpoint times into a probability distribution
The collection of particle-level endpoint predictions forms an RUL distribution:
[ p(RUL\mid z_{1:k}) ]
From this distribution, the system can report an expected RUL, a median forecast, or prediction intervals. The interval is often more useful than a single value because it exposes uncertainty caused by noisy measurements, uncertain initial conditions, model error, and cell-to-cell variation.
How Resampling Maintains a Useful Estimate
Particle weights naturally become uneven
After repeated prediction and measurement updates, many particles may receive nearly zero weight while a small number carry most of the probability mass. This condition is called particle degeneracy.
If it is not addressed, the filter effectively loses the diversity needed to represent uncertainty and alternative degradation paths.
Resampling concentrates computation on plausible states
When degeneracy becomes significant, the algorithm removes low-weight particles and replicates high-weight particles. The new population is then propagated again with process noise so that it does not collapse into identical copies.
Resampling does not create new measurement information. It reallocates computational effort toward states that are consistent with the observed battery behavior.
The resampling trigger should be controlled
Resampling at every time step can reduce particle diversity, while never resampling allows degeneracy to accumulate. Implementations commonly use an effective sample-size criterion or a similar diagnostic to determine when resampling is needed.
The appropriate threshold depends on the model, particle count, sensor quality, and computational budget.
What the Testing Workflow Looks Like
Prepare and synchronize the test data
Battery testing systems must capture current and voltage synchronously. Temperature and operating-profile information are also important because dynamic load, thermal conditions, and pulse duration can strongly affect both SOC estimation and degradation prediction.
High-current pulses and rapidly changing loads make sampling quality especially important. Poor synchronization can cause the filter to attribute measurement timing errors to incorrect battery states.
Condition the degradation signals
Raw capacity, resistance, and voltage-derived health indicators may contain local noise. A degradation signal can therefore appear to recover temporarily even though irreversible aging has not reversed.
Smoothing methods such as moving averages or Savitzky–Golay filtering can reduce local noise. A monotonicity constraint may also be applied where the chosen health index is physically expected to degrade monotonically, preventing implausible recovery artifacts from entering the prognostic model.
Calibrate the model offline
Historical aging data from multiple cells can be used to estimate population-level degradation behavior and cell-to-cell variation. This offline stage helps establish plausible parameter ranges, process-noise levels, and failure-time behavior.
Mixed-effects or related population models can represent both common aging trends and individual-cell deviations.
Update the individual cell online
During a live test, the particle filter uses current measurements to adapt the population-level model to the specific cell being observed. Real-time resistance or capacity-related measurements can update the individual degradation trajectory.
This two-stage approach is often more robust than relying exclusively on either historical averages or a single cell’s noisy measurements.
Understanding the Trade-offs
Particle count affects accuracy and computation
More particles can represent complex probability distributions more accurately, especially when the state dimension is large or the posterior distribution is multimodal. The cost is greater memory use and computation time.
Too few particles may produce unstable estimates or miss plausible degradation paths. The particle count should therefore be selected through validation rather than assumed to guarantee accuracy.
Model quality still matters
A particle filter does not compensate for a fundamentally unsuitable battery model. If the state-transition or measurement model omits important temperature, hysteresis, pulse, or aging effects, the filter may assign high probability to physically incorrect states.
The algorithm is best understood as a nonlinear Bayesian estimator, not as a replacement for battery characterization.
RUL depends on future operating conditions
An RUL forecast is conditional on assumed future loads, temperatures, and operating policies. A cell may reach an endpoint earlier under sustained high-current demand than under a lighter profile.
The test or production system should therefore state the forecast scenario explicitly. Otherwise, a precise-looking RUL number can be misleading.
Fixed thresholds can be inappropriate
A single resistance or capacity threshold does not always represent failure accurately, particularly when the application has variable performance requirements or when hard failure is not sharply defined.
Where appropriate, degradation measurements can be linked to time-to-failure through probabilistic or hazard-based models. This allows the forecast to reflect observed failure behavior rather than relying on an arbitrary cutoff alone.
Uncertainty can grow rapidly during long forecasts
The farther the filter projects beyond the latest measurement, the more process uncertainty and uncertainty in future conditions influence the result. Long-horizon RUL predictions should therefore be updated as new cycling data becomes available.
A credible system reports changing uncertainty rather than presenting an unchanged forecast until failure occurs.
Making the Right Choice for Your Goal
Particle filtering is most effective when the estimation and prognostic workflow is designed around the test objective.
- If your primary focus is SOC accuracy: Use a nonlinear electrical model, synchronize current and voltage measurements, include relevant temperature effects, and report both the estimated SOC and its uncertainty.
- If your primary focus is RUL prediction: Include degradation variables such as capacity or resistance in the particle state, define the EOD or EOL criterion explicitly, and propagate particles under a stated future operating profile.
- If your primary focus is robust cell-to-cell prediction: Calibrate population behavior offline, represent manufacturing variation, and update each cell’s degradation path online with its own measurements.
- If your primary focus is real-time deployment: Control particle count and resampling frequency, preprocess noisy health indicators, and validate the computational latency against the test system’s sampling requirements.
A well-designed particle-filter framework turns nonlinear battery measurements into an evolving probability distribution for both present SOC and future battery life, giving engineers a basis for decisions rather than a false sense of precision.
Summary Table:
| Aspect | Description |
|---|---|
| Core Principle | Uses a population of particles to represent possible battery states, enabling non-linear and non-Gaussian estimation. |
| SOC Tracking | Propagates particles through a battery model, updates weights based on measurements, and computes SOC as a weighted average. |
| RUL Prediction | Projects particles into the future under assumed conditions, defining end-of-life criteria, and produces a probability distribution of RUL. |
| Key Advantage | Provides uncertainty bounds alongside estimates, improving decision-making under noisy or incomplete data. |
| Resampling | Prevents particle degeneracy by resampling, focusing computational effort on plausible states. |
| Trade-offs | Requires careful selection of particle count, model fidelity, and future operating scenarios to balance accuracy and computation. |
Ready to Enhance Your Battery Testing with Advanced Prognostics?
At KINTEK, we provide comprehensive laboratory equipment for battery R&D and advanced materials research. Our portfolio covers the entire cell fabrication workflow—from slurry mixing, coating, and precision pressing to cell assembly and testing systems. For accurate SOC and RUL estimation, our precision test equipment and robust data acquisition systems integrate seamlessly with frameworks like particle filtering. Whether you're in battery R&D or materials science, our solutions help you achieve reliable, insightful results.
Contact us today to discuss how KINTEK can support your battery testing needs and elevate your research.