Granted Patent B2

US 12,530,595 · App. 17/165,714 · Granted Jan 20, 2026

Inventors: Ryan Alan Goldhahn (Livermore, CA); Priyadip Ray (Livermore, CA); Braden C. Soper (Livermore, CA); Hao Chen (Boise, ID); Deepak Rajan (Livermore, CA)

Assignee: LAWRENCE LIVERMORE NATIONAL SECURITY, LLC

G06N3/126G06N3/086G06N7/01

Abstract

Claims (43)

1 . A method performed by a first node for identifying a source location of a potentially hazardous physical event, the method comprising:

collecting, via a sensor in the first node, a measurement of a physical quantity and is associated with the potentially hazardous physical event, wherein the first node is one of a plurality of nodes that are spatially distributed across a geographic area, each node of the plurality of nodes including a sensor and a computing device for identifying the source location of the potentially hazardous physical event, the computing device in each of the plurality of nodes including a processing unit and a memory;

repeatedly re-calculating parameters representing a probability distribution of the source location of the potentially hazardous physical event, by the computing device in the first node, until a termination criterion is satisfied, to iteratively increase an accuracy of an approximation of the source location of the potentially hazardous physical event, including,

receiving, from each other node of the plurality of nodes, a probability distribution description and a gradient, each said probability distribution description representing a probability distribution of the source location as determined by a corresponding other node, wherein each received probability distribution description is based on a model of measurements of the potentially hazardous physical event over time and space, the model including either of: i) a Black Box Variational Inference (BBVI) algorithm and an Evidence Lower Bound (ELBO), or ii) a Distributed Markov chain Monte Carlo (“MCMC”) algorithm,

wherein, for a message routed through intermediate nodes, probability distribution descriptions and gradients generated by said intermediate nodes are integrated into the message to increase an amount of received data contributing to the approximation of the source location without increasing network communication overhead between the first node and each other node;

executing, by the first node, either of: i) a BBVI algorithm and an ELBO, or ii) a MCMC algorithm with an output of an additional probability distribution description as the approximation of the source location of the potentially hazardous physical event, based on the probability distribution descriptions received from the other nodes and the gradients received from the other nodes;

sampling the additional probability distribution description of the first node to produce a plurality of sampled probability distribution values of the probability distribution description generated by the first node;

calculating, by the first node, a gradient from the sampled probability distribution values of the probability distribution description generated by the first node, based on an ELBO relating to the probability distribution values;

sending, from the first node to a subset of the other nodes, the probability distribution description generated by the first node and the gradient calculated by the first node;

reducing propagation of unhelpful data to conserve network bandwidth in response to determining, based on a signal-to-noise ratio of the measurement or a magnitude of the calculated gradient, that the generated probability distribution description and calculated gradient are not informative;

filtering defective data in response to determining that the generated probability distribution description, the calculated gradient, or both are inconsistent with those received from other nodes; and

updating the parameters based on the probability distribution descriptions and gradients generated by said intermediate nodes,

wherein when the termination criterion is satisfied, the probability distribution generated by the first node is indicative of the approximation of the source location of the potentially hazardous physical event; and

outputting, by the first node, an indication of the approximation of the source location of the potentially hazardous physical event to an external entity.

2 . The method of claim 1 wherein the physical quantity is a gas that is released at the source location.

3 . The method of claim 1 wherein the physical quantity is radiation that is released at the source location.

4 . The method of claim 1 wherein the probability distribution is a normal distribution and the parameters are mean and variance of the normal distribution.

5 . The method of claim 1 wherein the termination criterion is based on the calculated gradient.

6 . The method of claim 1 wherein the termination criterion is based on a number of iterations of re-calculating the parameters.

7 . A node for identifying a source location of a potentially hazardous physical event based on a measurement relating to the potentially hazardous physical event and based on probability distribution descriptions received from other nodes, each probability distribution description indicating a probability distribution of the source location, the node comprising:

a sensor to collect a measurement of a physical quantity and is relating to the potentially hazardous physical event, wherein the sensor comprises at least one of a gas sensor or a radiation sensor and the potentially hazardous physical event comprises at least one of a gas release or a radiation release;

a communication component using which the node can communicate with other nodes; and

a computing device that includes:

one or more processors; and

one or more non-transitory computer-readable storage mediums storing computer-executable instructions, execution of which by the one or more processors causes the computing device to perform a process of iteratively increasing an accuracy of an approximation of the source location of the potentially hazardous physical event, wherein the process includes to:

receive via the communication component from each of a plurality of other nodes a probability distribution description and a gradient derived from samplings of the probability distributions by the other nodes, each node of the plurality of other nodes including a sensor and a computing device to identify the source location of the potentially hazardous physical event, the computing device in each of the plurality of other nodes including a processor and a memory, wherein each received probability distribution description is based on a model of measurements of the potentially hazardous physical event over time and space, the model including either of: i) a Black Box Variational Inference (BBVI) algorithm and an Evidence Lower Bound (ELBO), or ii) a Distributed Markov chain Monte Carlo (“MCMC”) algorithm,

wherein, for a message routed through intermediate nodes, probability distribution descriptions and gradients generated by said intermediate nodes are integrated into the message to increase an amount of received data contributing to the approximation of the source location without increasing network communication overhead between the node and each other node;

execute either of: i) a BBVI algorithm and an ELBO, or ii) a MCMC algorithm with an output of an additional another probability distribution description as the approximation of the source location of the potentially hazardous physical event, based on the received probability distribution descriptions and gradients, wherein the probability distribution description generated by the node is based on a model of measurements of the potentially hazardous physical event over time and space

sample the additional probability distribution description of the node to produce a plurality of sampled probability distribution values of the probability distribution description generated by the node;

calculate a gradient derived from the sampled probability distribution values of the probability distribution description generated by the node, based on an ELBO relating to the probability distribution values;

send via the communication component to a subset of the other nodes, the generated probability distribution description and the calculated gradient;

reduce propagation of unhelpful data to conserve network bandwidth in response to determining, based on a signal-to-noise ratio of the measurement or a magnitude of the calculated gradient, that the generated probability distribution description and calculated gradient are not informative;

filter defective data in response to determining that the generated probability distribution description, the calculated gradient, or both are inconsistent with those received from other nodes;

update the parameters based on the probability distribution descriptions and gradients generated by said intermediate nodes,

wherein when a termination criterion is satisfied, the generated probability distribution represents the approximation of the source location of the potentially hazardous physical event; and

output an indication of the approximation of the source location of the potentially hazardous physical event to an external entity.

8 . The node of claim 7 wherein the other nodes are a subset of nodes that are selected randomly.

9 . The node of claim 7 wherein the computer-executable instructions further include instructions to determine whether the generated probability distribution description and calculated gradient are informative and when not informative, suppress the sending of the generated probability distribution description and the calculated gradient.

10 . The node of claim 7 wherein the computer-executable instructions further include instructions to receive from a sensor an indication to forward the calculated gradient to another node and forward the generated probability distribution description and calculated gradient to the other node.

11 . The node of claim 7 wherein the computer-executable instructions further include instructions to, prior to sending the generated probability distribution description and calculated gradient to other nodes, quantize the generated probability distribution description, the calculated gradient, or both to reduce bandwidth needed for the sending.

12 . The node of claim 7 wherein the computer-executable instructions to determine whether the generated probability distribution description, the calculated gradient, or both are inconsistent with those received from other nodes, and when inconsistent, suppress the sending of the generated probability distribution description and the calculated gradient.

13 . The method of claim 1 , further comprising randomly selecting the subset of the other nodes, to which to send the generated parameters and the calculated gradient.

14 . The node of claim 7 wherein the node is at least one of an unmanned aerial vehicle (“UAV”), an unmanned ground vehicle (“UGV”), an unmanned underwater vehicle (“UUV”) or an unmanned space vehicle (“USV”).

Assignments (2)

Continuity (2)

References Cited (15)