Neural Network Methods for Describing System Topologies
AssigneePassiveLogic, Inc.
InventorsTroy Aaron Harvey, Jeremy David Fillingim
Life of this patent
Estimating the term…
Abstract
A neural network in one embodiment is built by decomposing a structure into different building materials creating neurons that represent building materials and open spaces in a structure. Subsystems in the building have their neurons concatenated together to create same length neuron strings. In some embodiments, neurons in a short neuron string are split to make longer neuron strings. In some embodiments, neurons are added to some neuron strings to represent inside features, air features, and outside features.
Claims20
A neuron model creation system comprising:
- a memory;
- and
- at least one processor configured to:
- determine structure elements for at least a first subsystem and a second subsystem in a digital representation of a physical structure;
- build a neuron for each structure element in the first subsystem and the second subsystem, each neuron comprising an input, an output, and at least one parameter;
- modify at least one of the subsystems such that the first subsystem and the second subsystem are represented by an equal number of neurons;
- connect the neurons associated with the first subsystem to form a first neuron string and connect the neurons associated with the second subsystem to form a second neuron string, each neuron string being branchless and ordered;
- and
- execute the first neuron string and the second neuron string in parallel to model propagation of state through the physical structure.
A method of modeling physical behavior of a structure using a self-configuring neural model, the method comprising:
- deriving a digital representation of a physical structure comprising a plurality of subsystems;
- identifying physical structure elements within each subsystem;
- constructing a plurality of neuron strings, each neuron string corresponding to a subsystem and comprising an ordered, branchless sequence of neurons representing the physical structure elements of the subsystem;
- initializing parameters of the neurons based on physical properties of the corresponding physical structure elements;
- propagating one or more state values through the plurality of neuron strings to model transfer of state through the structure;
- and
- automatically modifying at least one of the plurality of neuron strings based on the digital representation of the structure,
- wherein the self-configuring neural model is generated and operates without training, without learned weights, and without use of historical training data.
A non-transient computer-readable storage medium storing instructions that, when executed by one or more processors, cause the one or more processors to implement a state-propagation engine configured to:
- receive a digital representation of a physical structure comprising multiple subsystems;
- generate, for each subsystem, a state-propagation path comprising an ordered sequence of computational nodes corresponding to physical structure elements;
- normalize the state-propagation paths to a common length across the multiple subsystems;
- propagate state values through the state-propagation paths to simulate physical transfer through the physical structure;
- and
- update parameters of the computational nodes during propagation based on the propagated state values,
- wherein the state-propagation engine operates deterministically based on the digital representation of the physical structure and does not perform a training phase or weight-learning operation.
Dependencies are inferred from the claim text.