# US 2025/0165679 A1 · State Estimation using Physics-Constrained Machine Learning

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 2025/0165679 A1
- Document ID: US20250165679A1
- Status: Published application
- Kind: A1
- Application number: 18/512,319
- Filed: 2023-11-17
- Published: 2025-05-22
- Earliest claimed: 2023-11-17
- Assignees: Mitsubishi Electric Research Laboratories, Inc.
- Applicants: Mitsubishi Electric Research Laboratories, Inc.
- Inventors: Vedang Deshpande; Ankush Chakrabarty; Christopher Laughman; Abraham Vinod
- Web page: https://patentgrep.com/patent/US20250165679A1

## CPC

- G05B 13/0265: {the criterion being a learning criterion}
- G06F 30/27: using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
- G05B 17/02: electric
- G05B: CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS (systems for controlling or regulating non-electric variables G05D; systems for regulating electric or magnetic variables G05F; control devices or systems insofar as characterised by mechanical features only G05G)
- G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N)

## Abstract

To perform a task based on an internal state of a digital twin simulating an operation of a mechanical system, where at least some state variables of the internal states of the digital twin are subject to constraints derived from the physics of a structure of the mechanical system, a processor executes a neural network including an autoencoder trained to process a current internal state of the digital twin and a current control input to the mechanical system to produce a current output of the mechanical system caused by the current control input and an unconstrained next internal state of the digital twin transitioned from the current internal state based on the current control input, and a neural operator trained to modify the unconstrained next internal state of the digital twin to produce a constrained next internal state of the digital twin satisfying the constraints.

## Claims

### Claim 1

A method for performing a task based on an internal state of a digital twin simulating an operation of a mechanical system, wherein at least some state variables of the internal states of the digital twin are subject to constraints derived from physics of a structure of the mechanical system, wherein the method uses a processor coupled with stored instructions implementing the method, wherein the instructions, when executed by the processor carry out steps of the method, comprising:
executing a neural network including (1) an autoencoder trained to process a current internal state of the digital twin and a current control input to the mechanical system to produce a current output of the mechanical system caused by the current control input and an unconstrained next internal state of the digital twin transitioned from the current internal state based on the current control input, and (2) a neural operator trained to modify the unconstrained next internal state of the digital twin to produce a constrained next internal state of the digital twin satisfying the constraints; and
performing the task based on the constrained next internal state of the digital twin.

### Claim 2 (depends on 1)

The method of claim 1, wherein the constraints define a linear inequality in terms of the internal states, and wherein the neural operator encodes the constraints using an equivalent rays-based description of a polytope defined by the linear inequality, such that the unconstrained next internal state is passed through a fully connected neural layer which produces a nonnegative vector of a dimension compatible with a matrix of rays describing the polytope, and a product of the nonnegative vector with the matrix of rays results in the constrained next internal state satisfying the constraints.

### Claim 3 (depends on 1)

The method of claim 1, wherein the mechanical system is a vapor compression system (VCS), and wherein the constraints mandate the internal states of the digital twin of the VCS to satisfy decreasing pressure in a direction of refrigerant flow in the VCS.

### Claim 4 (depends on 1)

The method of claim 1, wherein the constraints define box constraints for one or a combination upper and lower bound on the internal states of the digital twin.

### Claim 5 (depends on 1)

The method of claim 1, wherein the mechanical system is a vapor compression system (VCS), and wherein the task includes one or a combination of an estimation of heating/cooling capacities delivered by the VCS to a space, a computation of refrigerant mass in the VCS, control of the VCS, a performance monitoring of the operation of the VCS, anomaly detection in the operation of the VCS, a model predictive control of the VCS.

### Claim 6 (depends on 1)

The method of claim 1, further comprising:
collecting operational data including a sequence of control inputs for controlling the mechanical system and a corresponding sequence of outputs of the mechanical system controlled by the sequence of control inputs; and
iteratively executing the neural network to estimate a sequence of internal states of the digital twin corresponding to the sequence of control inputs to the mechanical system and the corresponding sequence of outputs of the mechanical system.

### Claim 7 (depends on 6)

The method of claim 6, further comprising:
executing a probabilistic filter estimating the internal states of the digital twin to produce the sequence of internal states using a prediction model subject to process noise and a measurement model subject to measurement noise, wherein at least one of the prediction model and the measurement model include the neural network.

### Claim 8 (depends on 7)

The method of claim 7, further comprising:
passing the internal states estimated in each iteration of the probabilistic filter through an encoder and a decoder of the autoencoder to produce decodings of the estimations of the probabilistic filter; and
processing the decodings of the estimations of the probabilistic filter with the neural operator to enforce the constraints.

### Claim 9 (depends on 6)

The method of claim 6, wherein the probabilistic filter is an extended Kalman filter with a model linearization obtained by differentiation of the neural network.

### Claim 10 (depends on 1)

The method of claim 1, wherein the autoencoder architecture comprises:
a first encoder subnetwork configured to encode the current internal state into latent space to produce an encoding of the current internal state in the latent space;
a second encoder subnetwork configured to encode the current control input into the latent space to produce an encoding of the current control input in the latent space;
a latent subnetwork configured to propagate forward in time one or a combination of the encoding of the current internal state and the encoding of the current control input to produce an encoding of the next internal state in the latent space;
a first decoder subnetwork configured to decode the encoding of the next internal state from the latent space to produce the unconstrained next internal state;
a second decoder subnetwork configured to decode the encoding of the current internal state to produce the current output corresponding to the current internal state and the current internal input.

### Claim 11 (depends on 10)

The method of claim 10, wherein the latent subnetwork is configured to propagate the encoding of the current internal state forward in time and combine the propagated encoding of the current internal state with the encoding of the current control input to produce the encoding of the next internal state in the latent space.

### Claim 12 (depends on 10)

The method of claim 10, wherein the latent subnetwork is configured to propagate a mix of the encoding of the current internal state and the encoding of the current control input forward in time to produce the encoding of the next internal state in the latent space.

### Claim 13 (depends on 10)

The method of claim 10, wherein the autoencoder is trained to minimize a loss function of a difference of a sequence of the internal states of the digital twin and a sequence of outputs of the VCS measured over a prediction horizon.

### Claim 14 (depends on 13)

The method of claim 13, wherein the loss function includes
a first lost term penalizing a reconstruction loss between the current internal state and a decoding of the current internal state;
a second lost term penalizing a reconstruction loss between the next internal state and a decoding of the next internal state from the encoding of the current internal state; and
a third lost term penalizing a reconstruction loss between the current output of the VCS and the decoding of the current output of the VCS.

### Claim 15 (depends on 10)

The method of claim 10, wherein the autoencoder is trained jointly with the neural operator.

### Claim 16 (depends on 1)

The method of claim 1, further comprising:
controlling the mechanical system based on the constrained next internal states of the digital twin to perform the task.

### Claim 17

A device for performing a task based on an internal state of a digital twin simulating an operation of a mechanical system comprising:
one or more processors configured to:
execute a neural network including (1) an autoencoder trained to process a current internal state of the digital twin and a current control input to the mechanical system to produce a current output of the mechanical system caused by the current control input and an unconstrained next internal state of the digital twin transitioned from the current internal state based on the current control input, and (2) a neural operator trained to modify the unconstrained next internal state of the digital twin to produce a constrained next internal state of the digital twin satisfying the constraints; and
perform the task based on the constrained next internal state of the digital twin.

### Claim 18 (depends on 17)

The device of claim 17, wherein the constraints define a linear inequality in terms of the internal states, and the neural operator encodes the constraints using an equivalent rays-based description of a polytope defined by the linear inequality, such that the unconstrained next internal state is passed through a fully connected neural layer which produces a nonnegative vector of a dimension compatible with a matrix of rays describing the polytope, and a product of the nonnegative vector with the matrix of rays results in the constrained next internal state satisfying the constraints.

### Claim 19 (depends on 17)

The device of claim 17, wherein the mechanical system is a vapor compression system (VCS), and the constraints mandate the internal states of the digital twin of the VCS to satisfy decreasing pressure in a direction of refrigerant flow in the VCS.

### Claim 20 (depends on 17)

The device of claim 17, wherein the one or more processors are further configured to:
collect operational data including a sequence of control inputs for controlling the mechanical system and a corresponding sequence of outputs of the mechanical system controlled by the sequence of control inputs; and
iteratively execute the neural network to estimate a sequence of internal states of the digital twin corresponding to the sequence of control inputs to the mechanical system and the corresponding sequence of outputs of the mechanical system.