Augmenting Extended Kalman Filters with Dynamic Temperature Derivative State Models

Silicon

Micromachined capacitive proof masses shift resonance and baseline capacitance under changing thermal loads. Heat conducts through ceramic packaging into the die substrate, creating transient spatial gradients across sensing anchors. When temperatures shift faster than 2.0 degrees Celsius per minute, internal thermomechanical stress warps the frame asymmetrically.

Integrated temperature-sensing diodes measure local junction conditions, but they lag the proof mass temperature by several seconds.

Standard factory calibration records offset values at static equilibrium steps, holding temperatures at fixed plateaus like -40, +25, and +85 degrees Celsius for thirty minutes before taking readings. This static mapping assumes that temperature readings accurately reflect internal strain on the sensing element ~ an assumption that fails during rapid environmental transitions.

When thermal flux enters the package through the top surface, the temperature diode on the ASIC layer registers the rise before heat reaches the MEMS anchor. That delay creates a dynamic hysteresis loop where offset shifts depend on the rate of temperature change, not just absolute temperature.

Static calibration tables understate zero-rate output drift by 410 percent when thermal transition rates exceed 3.5 degrees Celsius per minute.

Modeling bias as a static third-order polynomial captures less than twenty percent of total drift during active ramping. Unmodeled rate-dependent deformation skews filter biases, driving linear position drift past ten meters over a five-minute thermal transient.

Augmentation

Expanding the filter state vector means appending temperature and its temporal derivative to the kinematic parameters. Traditional inertial estimation models track position, velocity, orientation, acceleration bias, and angular rate bias. But under HVAC venting, rapid altitude climbs, or engine heat soak, zero-rate biases fluctuate with both thermal amplitude and rate of change.

Incorporating the dynamic derivative requires extending the state vector to include temperature and rate-of-change states alongside core kinematic terms. In continuous form, angular velocity bias connects to these environmental states through partial derivative sensitivity matrices.

The state transition model uses a first-order differential equation for package heat transfer, based on a single lumped thermal capacity model where temperature rate decays toward ambient over a characteristic time constant. The continuous dynamic matrix incorporates cross-coupling terms that link angular rate bias updates directly to the dynamic thermal rate state.

• State Vector Definition Append absolute temperature and first-order temperature time derivative to the estimation array to capture thermal lag dynamics.
• Continuous Transition Matrix Construction Insert spatial thermal time constants into off-diagonal blocks connecting bias drift states to temperature derivative states.
• Polynomial Sensitivity Expansion Evaluate partial derivatives of zero-rate bias with respect to absolute temperature and dynamic thermal rate at the current state estimate.
• Discrete Transition Discretization Convert continuous system matrices to discrete form using matrix exponential approximations calculated over the filter step interval.

Calculating the state transition matrix requires evaluating the continuous matrix exponential over discrete sampling intervals. The dynamic calibration equation expresses zero-rate bias as a linear combination of static polynomial temperature terms, dynamic rate linear terms, and cross-coupling products. Evaluating the Jacobian requires analytical partial derivatives with respect to state variables at each filter step.

Calculating these partial derivatives yields explicit matrix entries. The derivative of bias with respect to absolute temperature includes quadratic terms, whereas the derivative with respect to thermal rate isolates dynamic coupling coefficients. The dynamic temperature derivative state itself evolves as an autoregressive random walk driven by zero-mean Gaussian white process noise.

By placing package thermal conduction constants directly in off-diagonal terms, the continuous state transition matrix predicts bias shifts before heat reaches the mechanical anchors.

Covariance

Process noise density values assigned to temperature derivative states dictate filter responsiveness during rapid environmental shifts. Setting high process noise lets the filter track steep thermal ramps, but injects measurement noise into bias states under isothermal conditions. Setting low process noise smooths static estimates, but causes lag and estimator divergence during sharp temperature steps.

Is Thermal Rate State Augmentation Observable under Isothermal Operations?

Constant operating temperatures drive the rate of change to zero, zeroing out dynamic derivative entries in the state measurement matrix. Observability matrices maintain full rank as long as kinematic excitation couples with temperature measurement updates. System states remain estimable during prolonged thermal plateaus because absolute temperature measurements constrain baseline polynomial terms while derivative states collapse to zero variance.

Tuning process noise entries requires matching real-world environmental rate limits and accounting for the power spectral density of micro-climate variations. If an enclosure experiences maximum thermal swings of 10.0 degrees Celsius per minute, the process noise variance must accommodate that spectral bandwidth without over-filtering.

Tuning process noise covariance above ambient thermal transition limits introduces numerical jitter into bias estimate states.

Measurement noise variance for internal temperature sensors must reflect analog-to-digital converter quantization levels and thermal noise floors. When sensors exhibit 0.0625 degree Celsius quantization steps, discrete derivative calculations generate impulse spikes in the innovation sequence. Filter gains must down-weight single-step temperature changes to avoid feeding quantization artifacts into kinematic states.

Whether adaptive process covariance scaling can separate high-frequency thermal rate noise from true mechanical acceleration spikes under random vibration remains an open question in embedded sensor estimation.

Chamber

Environmental characterization profiles must impose controlled thermal acceleration to isolate dynamic coefficients from static polynomial terms. Standard soak tests hold fixed temperature steps until package stress reaches mechanical equilibrium before logging sensor data, whereas dynamic characterization runs continuous linear ramps at varying slopes across the full operating range.

Characterizing a sensor involves running tests inside a thermal chamber across ramp speeds of 1.0, 3.0, 5.0, and 10.0 degrees Celsius per minute. Gyroscope and accelerometer outputs recorded during positive ramps diverge systematically from outputs recorded during negative ramps of identical magnitude, forming a characteristic hysteresis loop.

Isolating dynamic coefficients involves subtracting static offset curves from the dynamic ramp datasets. The remaining residual offset correlates directly with the calculated temperature rate of change, and fitting linear and cross-term models to this residual data yields the scaling factors needed for state matrix augmentation.

• Uncalibrated Thermal Shock Exposure Rapid ambient shifts generate localized substrate gradients that exceed model linearization boundaries, causing temporary filter divergence.
• Inadequate Thermal Chamber Airflow Low-velocity air circulation produces uneven package heating, corrupting extracted dynamic lag time constants.
• Diode Quantization Overshoot Coarse temperature sensor resolution introduces discrete step functions into derivative calculations, inducing artificial bias state corrections.
• Mechanical Mounting Stress Coupling Fastening printed circuit boards directly to aluminum chamber plates introduces thermal expansion strains that corrupt pure thermal drift metrics.

Sensing element thermal mass delays internal response. As thermal ramping accelerates, package thermal isolation shifts the apparent offset curve along the temperature axis. This shift corresponds directly to the spatial time delay constant between the package exterior and the silicon substrate anchor.

Compliance with ISO 16750 thermal shock specifications demands continuous dynamic bias tracking throughout a fifty degree per minute transient profile.

Calibrating these dynamic values eliminates the hysteresis loop, collapsing positive and negative ramp bias tracks into a single unified state estimate. Static calibration tables are often treated as sufficient, with transient drift spikes misattributed to printed circuit board layout rather than packaging thermal dynamics.

Execution

Embedded processor clock cycles scale steeply as state matrix dimensions expand to accommodate environmental derivatives. Adding temperature and rate-of-change states enlarges the matrix inverted during Kalman gain calculations, increasing real-time floating-point operations with each added variable.

Processing an augmented state filter on an ARM Cortex-M4 floating-point unit requires careful memory alignment and execution timing optimization. Partitioning state vector update steps allows high-rate kinematic updates to run independently of lower-rate environmental calculations.

1. Sample primary inertial sensors at 200 Hz to update high-rate kinematic integration loops.
2. Trigger internal temperature sensor conversion at a sub-sampled frequency of 10 Hz.
3. Calculate raw thermal rate of change using a finite impulse response differentiator filter.
4. Execute extended filter gain updates for augmented thermal states at the 10 Hz rate.
5. Propagate dynamic bias updates into the 200 Hz kinematic integration loop.

A multi-rate filter architecture limits CPU load while preserving thermal rate tracking. Kinematic covariance propagation runs at full sensor data rates, while off-diagonal thermal state covariances execute only when new temperature samples arrive. Running covariance updates faster than the physical sensor’s analog front-end response time injects quantization noise into derivative state estimates.

Procurement

Sourcing high-grade inertial sensors for high-transient environments requires evaluating internal temperature sensor architecture alongside baseline noise figures. Integrated temperature sensors vary widely in resolution, sampling rate, physical location, and thermal coupling to the primary substrate. Datasheets quoting low static noise floors often conceal low-bandwidth, highly quantized temperature channels unsuitable for dynamic rate derivation.

Sourcing secondary vendors for identical packaging footprints introduces hidden risk if the vendor uses a different substrate layout or mold compound. Changes in encapsulation materials alter thermal conductivity, shifting the spatial time constant and invalidating original dynamic filter calibration parameters.

Unannounced substrate layout alterations between silicon revisions change internal thermal response time constants without shifting static datasheet figures.

Engineers must verify that component qualifications adhere to AEC-Q100 Grade 1 or AEC-Q103 standards, checking specifically for dynamic thermal transient test protocols. Second-source qualification requires re-running continuous thermal ramp testing across all proposed vendor samples to derive part-specific dynamic coefficients. Inserting a mandatory engineering change notice clause specifying written confirmation for any die metalization or substrate revision protects downstream filtering algorithms from unverified thermal lag shifts.