TRAINING OF MACHINE LEARNING MODELS WITH HARDWARE-IN-THE-LOOP SIMULATIONS

Number: 20240078465
Published: 2024-03-07
Filed: 2022-09-01
Assignee: DISNEY ENTERPRISES, INC.
Inventors: Milluzzi; Andrew Jesse et al.
CPC: G05B17/02; G06N20/00; G06F9/455
Verdict: High Notable software
Source: Google Patents · FreePatentsOnline

The keeper's note

Hardware-in-the-loop ML training technique (robotics/ride-adjacent).

Abstract

A method for training a control system model includes introducing a simulated fault into a software simulation of a physical system and generating emulated sensor data based on the simulated fault, where the emulated sensor data emulates output from one or more sensors of the physical system. The method further includes obtaining output data from a test control system provided with the emulated sensor data, where the test control system emulates a control system of the physical system and tagging the output data with the simulated fault to create training data. The method further includes utilizing the training data to train the control system model, where the control system model is a machine learning model for use with the control system of the physical system during operation of the physical system.

Background

BACKGROUND

Control systems are generally used to control complex mechanical and electromechanical systems. For example, theme park attractions, such as rollercoasters, rides, or the like, including control that monitor and actuate ride vehicles, and other aspects of the attraction may be controlled by control systems. Various control systems may also be used to assist in repair and/or maintenance of such systems, by identifying components within such systems in need of repair or replacement. However, control systems often evaluate the current state of a mechanical or electromechanical system and may not identify component malfunctions before they occur or before components need to be replaced for a system to remain operational. Further, such control systems may be unable to identify rare events occurring within a system, such as external physical conditions, system components that fail infrequently, and the like.

Control systems, including software controllers, are frequently used to control and monitor complex mechanical systems. For example, amusement park attractions may include complex ride systems controlled by control systems. Such attractions may further include show systems provided along a length of a ride system and/or about a travel path of ride vehicles, which show systems may be controlled and monitored by control systems to provide proper presentation of shows provided by the show systems. Other types of complex mechanical systems, such as automate

Claims

1. A method comprising: introducing a simulated fault into a software simulation of a physical system; generating emulated sensor data based on the simulated fault, the emulated sensor data emulating output from one or more sensors of the physical system; obtaining output data from a test control system provided with the emulated sensor data, the test control system emulating a control system of the physical system; tagging the output data with the simulated fault to create training data; and utilizing the training data to train a control system model, the control system model being a machine learning model for use with the control system of the physical system during operation of the physical system. || 9. A system comprising: a software simulation system including a software simulation of a physical system, the software simulation system being configured to generate emulated sensor data based on input to the software simulation system, the emulated sensor data emulating output from one or more sensors of the physical system; and a test control system emulating a control system of the physical system, the test control system configured to generate control system output based on the emulated sensor data. || 15. A method comprising: receiving, at hardware-in-loop (HIL) simulation of a physical system, programmatic data input including at least one injected fault, the at least one injected fault being representative of a physical condition in a physical system, the at least one injected fault being injected into the programmatic data input using Monte Carlo techniques; generating, by the HIL simulation, control system outputs, the control system outputs corresponding to outputs of a control system controlling the physical system; generating a dataset by tagging the control system outputs using the programmatic data input including the at least one fault; and training a control model using the dataset, the control model being trained to identify the physical condition in the physical system based on the outputs of the control system controlling the physical system.