Position, Navigation, and Timing Technologies in the 21st Century. Группа авторов

non‐Gaussian pdfs.

      36.3.2 Gaussian Sum Filters

      One approach for modeling systems with non‐Gaussian pdfs is the use of composite random variables expressed as a sum of Gaussian random variables. The generalized Gaussian sum can be expressed as

      (36.31)

where w^[j] is a scalar weighting factor, y_j is a Gaussian random variable with mean , and covariance . These individually weighted Gaussian random variables can represent the overall distribution of the state vector. An example of a density function created using a sum of Gaussian random variables is shown in Figure 36.1.

      36.3.2.1 Multiple Model Adaptive Estimation

      One implementation of the Gaussian sum filtering approach is known as multiple model adaptive estimation (MMAE). The MMAE filter uses a weighted Gaussian sum to address the situation where unknown or uncertain parameters exist within the system model. Some examples of these types of situations include modeling discrete failure modes, unknown structural parameters, or processes with multiple discrete modes of operation (e.g. “jump” processes).

Figure 36.1 Gaussian sum illustration. The random variable x_sum is represented by a weighted sum of three individual Gaussian densities. In this example, x_sum = 0.25x₁ + 0.5x₂ + 0.25x₃.

      Consider our standard linear Gaussian process and observation models, repeated from Eqs. 36.17 and 36.18 for clarity:

      (36.32)

      (36.33)

      In the previous development, it was assumed that the system model parameters (i.e. ) were known. Let us now consider the situation where some of the system model parameters are unknown.

      To address this situation, we can define a vector of the unknown system parameters, a, and jointly estimate these parameters along with the state vector. In other words, we must now solve for the following density:

      (36.34)

      which, after applying Bayes’ rule, can be expressed as

(36.35)

      It is important to note that this expression is the product of the “known‐system model” pdf, p(x_k| a, ℤ_k), and a new density function, p(a| ℤ_k), which is the pdf of the unknown system parameters, conditioned on the observation set. Assuming a ∈ ℝⁿ, the parameter density can be written as

      (36.36)

Applying Bayes’ rule yields

      (36.37)

      Marginalizing the denominator about the parameter vector results in a more familiar form:

(36.38)

      where p(z_k| a, ℤ_k−1) is the measurement prediction density, which, given our linear observation model, is expressed as the following normal distribution:

      (36.39)

      Unfortunately, the integral in the denominator is intractable in general, which requires an additional constraint. If the system parameters can be chosen from a finite set (e.g. a ∈ {a^[1], a^[2], ⋯, a^[j]}), the parameter density can be expressed as the sum of the individual probabilities of the finite set. This results in a system parameter pdf defined as

(36.40)

where is the probability of the j‐th parameter vector at time k‐1, and δ(·) is the delta function. It can be observed that the sum of the weights must be unity in order to represent a probability density. Substituting Eq. 36.40 into Eq. 36.38:

      (36.41)

      Moving the position of the summation operators and parameter weight vector:

      (36.42)

      The properties of the delta function can be exploited to rewrite the numerator and eliminate the integral from the denominator:

(36.43)

      At this point, we have established the posterior pdf of the parameter vector as a finite weighted set. Revisiting our system parameter pdf, now defined at time k

(36.44)

and substituting into Eq. 36.43 yields the parameter density update relationship

(36.45)

      In the above equation, the predicted measurement pdf, p(z_k| a^[j], ℤ_{k − 1}), is evaluated at the measurement realization at time k, which yields the likelihood of realizing