Kalman Filtering

The algorithm operates in a perpetual predict-correct loop. The prediction step projects the current state estimate and its uncertainty forward in time using a dynamic process model. The correction step then fuses a new noisy measurement with the prediction, weighting each by their respective uncertainties via the Kalman Gain. This recursion means the filter never needs to store the entire history of past measurements; only the previous state estimate and its covariance matrix are required to process the next observation.

If the system dynamics are linear and both the process noise and measurement noise are additive, white, and Gaussian, the Kalman filter is the minimum mean squared error (MMSE) estimator. No other algorithm can produce a state estimate with a smaller error variance under these conditions. The filter maintains an explicit representation of uncertainty through the state covariance matrix P, which quantifies the confidence in the estimate and is essential for sensor fusion.

The Kalman Gain (K) is the critical weighting factor that determines how much the filter trusts the new measurement versus the prediction. It is computed from the predicted error covariance and the measurement noise covariance R.

• High measurement noise (large R): Kalman Gain is low; the filter trusts the prediction more.
• Low measurement noise (small R): Kalman Gain is high; the filter rapidly corrects toward the measurement. This dynamic weighting allows the filter to optimally handle sensors with time-varying reliability.

A core strength of the Kalman filter is its natural ability to combine data from multiple disparate sensors. In the correction step, measurements from LiDAR, encoders, IMUs, or cameras can be sequentially ingested. Each sensor's measurement is incorporated using its own noise covariance matrix, allowing the filter to seamlessly integrate a fast-but-drifty sensor with a slow-but-absolute sensor to produce an estimate that is both smooth and accurate. This is fundamental to autonomous vehicle localization.

The standard Kalman filter assumes linear dynamics. For real-world non-linear systems, two primary extensions exist:

• Extended Kalman Filter (EKF): Linearizes the system around the current estimate using a first-order Taylor series expansion (Jacobian matrices). It is widely used but can diverge for highly non-linear systems.
• Unscented Kalman Filter (UKF): Uses a deterministic sampling technique (sigma points) to capture the true mean and covariance after a non-linear transformation without linearization, providing superior accuracy for many applications.

Kalman filtering is ubiquitous in systems requiring real-time state estimation from noisy data:

• Manufacturing: Tracking the position of a robotic arm end-effector by fusing encoder and vision data.
• Aerospace: The Apollo navigation computer used a Kalman filter to guide the lunar module.
• Finance: Estimating unobservable parameters like stochastic volatility in time-series models.
• Battery Management: Estimating the State of Charge (SoC) of a lithium-ion cell by combining voltage, current, and temperature readings with an electrochemical model.

An optimization-based state estimation technique that solves a constrained optimization problem over a sliding window of past measurements. Unlike the Kalman filter, MHE explicitly handles physical constraints like valve saturation limits and non-negative concentrations.

• Uses a receding horizon of N past measurements
• Computationally heavier than Kalman filtering but handles non-linear constraints
• Preferred in chemical process control where state limits are critical

The field of building mathematical models of dynamic systems from observed input-output data. Before a Kalman filter can be deployed, system identification must provide the state transition matrix and process noise covariance.

• Estimates parameters of a candidate model structure
• Produces the A, B, and Q matrices required by the Kalman filter
• Poor system identification leads to filter divergence

The continuous, bidirectional data flow that ensures a virtual representation mirrors its physical counterpart in real-time. Kalman filters provide the state estimation backbone that drives this synchronization by correcting the digital model with live sensor data.

• Kalman innovation signals trigger model updates
• Enables predictive what-if simulations on the synchronized twin
• Critical for closed-loop manufacturing optimization

A non-parametric Bayesian method that models a distribution over possible functions, providing both predictions and calibrated uncertainty estimates. While Kalman filters assume linear dynamics, Gaussian processes can model complex non-linear relationships without a predefined structure.

• Complements Kalman filtering for non-parametric state estimation
• Provides variance estimates that can feed into the Kalman measurement noise matrix R
• Used in virtual metrology for quality prediction

An automated diagnostic layer that continuously evaluates the statistical performance of control loops. CPM systems analyze the innovation sequence from a Kalman filter to detect filter divergence, sensor drift, or model mismatch.

• Checks if innovations are zero-mean white noise
• A non-white innovation sequence signals filter degradation
• Triggers re-tuning or maintenance alerts in adaptive process control