Time-Scale-Aware Recursive Estimation Under a Real-Time Deadline

Embedded estimators that must finish each update before the next sensor sample often simplify their physics. One common simplification is to treat very fast internal dynamics, such as the pressure in a small engine volume or the fast voltage relaxation in a battery, as settling instantly. These fast parts are then written as algebraic equations rather than differential equations. This lets the solver take steps roughly ten times larger without blowing up. When the same estimator is also asked to learn the system's physical parameters online, the simplification has a second effect that is usually ignored. Collapsing a fast mode merges its parameters with neighboring ones. The parameters that remain identifiable are then combinations, not physical quantities, and that is not a fixed fact about the system. Whether a mode counts as "fast" depends on the ratio of its time constant to the solver step, and the time constant itself depends on parameters that drift with temperature, aging, and operating point. An estimator that ignores this either reports biased physical parameters or keeps updating directions it cannot observe. In the second case its uncertainty estimate swells without bound, which is the familiar "covariance windup" failure.

We propose an estimator that reads its own current parameter estimate to decide, at every step, which modes are fast relative to its fixed real-time step. It then changes three things at once: the model structure (full differential form or reduced algebraic form), the set of parameter directions it updates, and the way it carries its uncertainty across a structure change. The central claim is that model reduction for speed and parameter identifiability should be decided jointly and online, by the same estimated quantity. Doing so should remove the parameter bias and uncertainty blow-up that static reductions cause under drift, while still meeting a hard per-step deadline.

Core Claim and What Is New

The closest work is Real-time performance of DAE and ODE based estimators evaluated on a diesel engine. It rewrote the stiff intercooler pressure state of a diesel air path as an algebraic constraint, giving a differential-algebraic model (a DAE). It absorbed the intercooler volume into the intake manifold. On test-bench data, the resulting extended Kalman filter (EKF, a Kalman filter that linearizes a nonlinear model at each step) stayed stable at 125 ms steps, while the ordinary-differential-equation (ODE) filter failed beyond about 15 ms. That paper motivates our work and supplies the reduction technique. It differs from us in three ways: the reduction was chosen once by an engineer, only states were estimated, and the volumes and discharge coefficients were fixed constants.

The recursive-estimation literature works from the other side. Co-Estimation of State-of-Charge and State-of-Health for Lithium-Ion Batteries Considering Temperature and Ageing and Online Broadband Impedance Identification for Lithium-Ion Batteries Based on a Nonlinear Equivalent Circuit Model track drifting battery parameters with forgetting-factor recursive least squares. Forgetting-factor RLS is a running least-squares fit that discounts old data. These papers constrain our design, because they show that thermal shocks move parameters substantially. They also leave open what happens when that movement pushes a modeled time constant across the sampling interval.

The gap is the coupling between the two. Under drift, the stiffness ratio that justified a reduction changes. Under reduction, the identifiable parameter set changes too.

The deciding quantity is the stiffness ratio of each candidate fast mode $i$:

$$\rho_i(\hat\theta_t) = \frac{\Delta t}{\tau_i(\hat\theta_t)}.$$

Here $\Delta t$ is the fixed real-time step, and $\tau_i(\hat\theta_t)$ is the mode's time constant computed from the current parameter estimate $\hat\theta_t$. It comes from the eigenvalues of the small linearized fast block, so it is cheap to compute. With hysteresis thresholds $\rho_{lo} < \rho_{hi}$, the estimator reduces mode $i$ when $\rho_i > \rho_{hi}$ and restores it when $\rho_i < \rho_{lo}$. The hysteresis prevents chattering near a single threshold.

The new ingredient is what happens to the parameter estimate when the structure changes. Write $s(\theta)$ for the lumped parameters that the reduced model actually depends on. Examples are the effective manifold volume, or an apparent ohmic resistance that absorbs a fast RC pair. Let $S = \partial s / \partial \theta$ be their sensitivity. In reduced mode, the Kalman gain is projected onto the row space of $S$. The random-walk noise that normally keeps parameter uncertainty alive is switched off in the null space of $S$. As a result, directions that cannot be observed at this step size are frozen rather than inflated.

When a mode is restored, the fast state is placed on its quasi-steady manifold, $x_f = h(x_s, \hat\theta)$, where $x_s$ are the slow states. Its covariance and cross-covariance are set through the Jacobian of $h$. This keeps the filter statistically consistent through the switch instead of starting the fast state from scratch.

A deadline guard runs alongside. Each structure has a measured worst-case execution time on the target processor. If a structure's worst case would exceed the deadline, it is never allowed at the current $\Delta t$.

Research Program

We will build the estimator as an augmented-state EKF, with parameters appended to the state as slow random walks, in two domains with known physics and controllable ground truth.

The first domain is a battery equivalent-circuit model with an ohmic resistor and two RC pairs, sampled at a fixed step of 1 s. Temperature drives the fast RC time constant from well below the sample period to near or above it. This creates a clean identifiability trap. When the fast pair is lumped, a naive estimator reports an "ohmic resistance" equal to $R_0 + R_1$. Since growth in $R_0$ is a standard state-of-health indicator, this bias contaminates aging estimates exactly when temperature changes.

The second domain is a mean-value diesel air-path model patterned on the DAE-EKF paper's structure, driven by the World Harmonized Transient Cycle (WHTC) operating trajectory. Here, pressure time constants in the small volumes depend on mass flow. Low-load phases therefore move modes back out of the "fast" regime, and the discharge coefficients and effective volumes become the drifting parameters.

The analysis has two parts. First, we will derive the bias that appears when parameters are estimated through a reduced model with $\rho$ finite rather than infinite. This should give the threshold at which projecting onto $S$ beats keeping the full parameter set, as a function of $\rho$ and noise level. Second, we will measure worst-case execution time for each structure on an embedded ARM-class CPU, so the deadline guard rests on measurements rather than estimates.

Baselines

The do-nothing reference is a full-ODE augmented EKF with explicit integration at the largest step it can sustain, run at that same step in every scenario. It shows what happens when the deadline is ignored or the step is forced. We expect divergence at the real-time step, mirroring the 15 ms stability limit reported in the DAE-EKF paper.

The strongest existing method is a fixed reduction plus parameter augmentation. This is the DAE-EKF design from Real-time performance of DAE and ODE based estimators evaluated on a diesel engine, with the reduction chosen for the nominal operating point and parameters appended as random-walk states. Its reported result is qualitative parity. The DAE-EKF at 125 ms had RMSE similar to the ODE-EKF at 10 ms on WHTC bench data, and the file gives no absolute RMSE value. For the battery domain, the comparable baseline is forgetting-factor RLS as in the battery co-estimation work cited above. No reported result was found on Lacuna for parameter-tracking error of forgetting-factor RLS under temperature-driven time-constant crossover.

Both methods get the same simulated data, the same step size, and the same noise model. They also get the same tuning budget, a fixed grid over process noise and forgetting factor, chosen on a held-out drift profile. Beating the strongest baseline means four things together:

  1. Lower physical-parameter RMSE.
  2. Lower bias in the ohmic resistance (battery) or effective volume (diesel) during and after crossover.
  3. Normalized estimation error squared (NEES, the squared error divided by the filter's own predicted variance) staying inside its chi-square consistency band.
  4. Zero deadline misses.

Three ablations isolate the ingredients. The first keeps online structure switching but uses naive covariance handling at switches and no null-space freezing. The second keeps null-space freezing but uses an always-reduced structure. The third keeps everything but sets thresholds by hand rather than computing them from $\hat\theta$.

Why not just…?

Why not just use an implicit, stiffly stable integrator on the full ODE at the large step? This is the real threat. An L-stable implicit step, meaning one that strongly damps very fast modes, drives a stiff mode toward its quasi-steady value automatically. That is very close to the algebraic reduction. We will run implicit Euler and a Rosenbrock method on the full augmented model as a separate arm. Our prediction is that this arm is numerically stable but carries the same identifiability problem in a hidden form. The fast-mode parameters stay in the state, their sensitivities vanish at large $\rho$, and their covariance winds up. If the implicit arm matches our method on parameter bias and NEES within its deadline, the contribution shrinks to the covariance-projection rule. That rule can be bolted onto any integrator, and we would say so.

Why not just reduce once for the worst case and never switch back? A permanent reduction is cheap and always meets the deadline. However, it permanently loses the fast-mode parameters in regimes where they are identifiable, such as warm batteries and low-load engine phases. The always-reduced ablation measures exactly what is lost there.

Why not just use directional forgetting or regularized RLS to stop windup? Directional forgetting discounts only directions that recent data has excited, and it does curb windup under weak excitation. But it cannot distinguish a direction that is unexcited because of the input from one that is unobservable because of the step size. It will therefore keep partially updating merged parameters. We include it as a tuned arm on the temperature-crossover scenario. If it removes the $R_0$ bias, our structural argument fails.

Proposed Experiments

The decisive check is the battery cold-start profile. Temperature ramps from cold to warm under a dynamic current profile, so the fast RC time constant crosses the 1 s step during the run. We will plot $\hat R_0$ against its true value for every arm. Our method should track the true $R_0$ through the crossover. The fixed-reduction baseline should report $R_0 + R_1$ while cold, and the full-model arms should show NEES leaving the consistency band.

The second check is the WHTC-driven diesel model with discharge coefficients aged mid-cycle. Here we test whether switching occurs at the predicted low-load phases without deadline misses, and whether the restore step keeps NEES consistent.

Across both domains, we will sweep the hysteresis width and the noise level. This tests whether the predicted bias-versus-$\rho$ curve from the analysis matches the empirical crossover point.

A negative result on either decisive check would be informative. If the implicit-integrator arm or directional forgetting performs as well, that would show the identifiability issue can be absorbed by existing tools.