Meaning
Mathematical frameworks define the relationship between the hidden state of a system and the noisy measurements collected from sensors or logs. An observation model maps the internal variables of a business, such as true customer sentiment, to the external data points actually recorded, like survey scores. This mapping includes a description of the noise or error inherent in the data collection process.
It fails to apply if the sensors provide no information about the underlying state.
Measurement Error
Random noise and systematic bias distort the signals received by a tracker. The observation model quantifies these distortions as a probability density function. High noise levels require more smoothing to extract the truth.
Calibration is the process of identifying the parameters of this error.
State Projection
Transformations link the multidimensional reality of a process to a reduced measurable output. In an observation model, this projection might be a linear function or a complex nonlinear equation. It determines what part of the state is actually visible.
Parts of the system that are not projected into the observations are called unobservable.
Signal Integrity
Reliability of the data stream determines the trust placed in the model output. If the observation model is inaccurate, the state estimate will diverge from reality even if the internal dynamics are well understood. Verifying the model involves comparing predicted measurements against actual results.
This check ensures the sensors are providing a consistent window into the system.
