Learning constitutive equations of physical components with constraints discovery
The following relates generally to system modeling. Some embodiments described herein learn a representation of the parameter feasibility space that make model parameter tuning easier by constraining the search space, thus enabling physical interpretation of the learned model. They also enable model-based system analytics (controls, diagnosis, prognostics) by providing a system model.
1. An apparatus for modeling a physical component of a partially known system, wherein the physical component has a vector of unknown parameters, and the apparatus comprises:
at least one processor;
at least one memory including computer program code; and
the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to:
build a binary function that models feasibility of points based on the vector of unknown parameters;
at a first iteration, build a sequence of approximations of the binary function, wherein at the first iteration, the sequence of approximations is set to 0.5 for all the vector of unknown parameters;
at a second iteration, update the sequence of approximations of the binary function at least in part based on a minimum of a cost function of the vector of unknown parameters; and
generate model-based diagnostics of the partially known system based on the updated sequence of approximations of the binary function.
2. The apparatus of claim 1 , wherein the binary function is p(w)∈{0,1}, where:
w is the vector of unknown parameters; and
p(w)=1 if w is feasible and zero otherwise.
3. The apparatus of claim 1 , wherein:
the sequence of approximations is represented by q(w), where w is the vector of unknown parameters.
4. The apparatus of claim 1 , wherein:
the cost function is J(w), where w is the vector of unknown parameters;
the update of the sequence of approximations of the binary function happens by using an optimization problem of the J(w); and
the optimization problem is constrained by g i (w)≤0, where i=1, 2 and g i is a constraint function, where g i (w) are functions of approximations functions q(w).
5. The apparatus of claim 1 , wherein:
at the first iteration, the sequence of approximations of the binary function is built by minimizing a cross-entropy loss function.
6. The apparatus of claim 1 , wherein the at least one memory and the computer program code configured to, with the at least one processor, further cause the apparatus to:
subsequent to the second iteration, iterate by updating the sequence of approximations of the binary function at least in part based on the minimum of the cost function of the vector of unknown parameters; and
at a subsequent iteration, generate an audio or visual alarm based on an updated sequence of approximations of the binary function.
7. The apparatus of claim 1 , wherein the component is one of: a rail switch, a circuit, an electric motor, a wind turbine, an elevator, or a mechanical break.
8. The apparatus of claim 1 , wherein the at least one memory and the computer program code configured to, with the at least one processor, further cause the apparatus to:
enable model-based diagnostics of the component using the updated sequence of approximations.
9. An apparatus for modeling a physical component of a partially known system, wherein the physical component has a vector of unknown parameters, and the apparatus comprises:
at least one processor;
at least one memory including computer program code; and
the at least one memory and the computer program code configured to, with the at least one processor, cause the apparatus at least to:
build a binary function that models feasibility of points based on the vector of unknown parameters;
build a sequence of approximations of the binary function;
follow an iterative process with at least two iterations, each iteration including:
updating the sequence of approximations of the binary function by taking into account a minimum of a cost function of the vector of unknown parameters, wherein cost function is subject to a constraint that has an effect of prioritizing exploration of the feasibility of points with a first level of uncertainty over exploration of the feasibility of points with a second level of uncertainty, and the effect of prioritizing is a weight for finding the best feasibility of points for the sequence of approximations, wherein the first level is greater than the second level; and
generate model-based diagnostics of the partially known system based on the updated sequence of approximations of the binary function.
10. The apparatus of claim 9 , wherein, as the process iterates, the prioritizing effect of the constraint diminishes.
11. The apparatus of claim 10 , wherein the constraint satisfies a requirement that the prioritizing effect approaches zero as a number of iterations approaches infinity.
12. The apparatus of claim 9 , wherein the cost function is subject to a constraint that the prioritizing effect diminishes as the process iterates.
13. The apparatus of claim 9 , wherein the cost function is subject to the constraint that has the effect of prioritizing exploration of a point that has a first uncertainty over exploration of another point that has a second uncertainty, wherein the first uncertainty is greater than the second uncertainty.
14. A method for modeling a physical component of a partially known system, wherein the physical component has a vector of unknown parameters, and the method comprises:
building a binary function that models feasibility of points based on the vector of unknown parameters;
during a first iteration, building a sequence of approximations of the binary function, wherein at the first iteration, a sequence of approximations is set to 0.5 for all the vector of unknown parameters;
during a second iteration, updating the sequence of approximations of the binary function at least in part based on a minimum of a cost function of the vector of unknown parameters; and
generate model-based diagnostics of the partially known system based on the updated sequence of approximations of the binary function.