	US 11,868,887 B2
	Making time-series predictions of a computer-controlled system
Melih Kandemir, Stuttgart (DE); Sebastian Gerwinn, Leonberg (DE); Andreas Look, Stuttgart (DE); and Barbara Rakitsch, Stuttgart (DE)
Assigned to ROBERT BOSCH GMBH, Stuttgart (DE)
Filed by Robert Bosch GmbH, Stuttgart (DE)
Filed on Jun. 7, 2021, as Appl. No. 17/340,978.
Claims priority of application No. 20180249 (EP), filed on Jun. 16, 2020.
Prior Publication US 2021/0397955 A1, Dec. 23, 2021
Int. Cl. G06N 3/08 (2023.01); G05B 13/02 (2006.01); G05B 13/04 (2006.01); G06N 3/047 (2023.01)

CPC G06N 3/08 (2013.01) [G05B 13/027 (2013.01); G05B 13/048 (2013.01); G06N 3/047 (2023.01)]

20 Claims

1. A computer-implemented method for generating and using a model of a computer-controlled system with which the computer-controlled system predicts states of the computer-controlled system for operation of the computer-controlled system, the prediction being performed by applying sensor measurements of the computer-controlled system to the trained model, the method comprising

accessing, by at least one processor, model data defining a model that uses a stochastic differential equation (SDE) that includes a set of variables having a drift component and a diffusion component, wherein the drift component includes a predefined part whose parameters are fixed and a trainable part whose parameters are adjustable by a training;

using, by the at least one processor, at least one sensor of a computer-controlled system to obtain a set of observations of the computer-controlled system, each observation of the set of observations including a time series of measurements of physical quantities of the computer-controlled system and/or an environment of the computer-controlled system obtained using the at least one sensor;

training, by the at least one processor, the model based on the set of observations, the training the model including iteratively performing the following until obtainments of a final version of the adjustable parameters of the trainable part of the drift component:

using the model to make a time-series prediction of a state of the computer-controlled system, at a current time point based on values of the set of variables of the SDE that represent the measurements of the physical quantities at a previous time point, wherein the making of the prediction includes:

evaluating the predefined part of the drift component to obtain a first drift;

evaluating the trainable part of the drift component to obtain a second drift;

combining the first drift and the second drift to obtain an overall drift; and

identifying, as the predicted state, a state corresponding to the overall drift with a stochasticity characterized by the diffusion component;

deriving a training signal by comparing the time-series prediction to an observation of the set of observations; and

adjusting at least the parameters of the trainable part of the drift component based on the training signal to optimize a contribution of the trainable part of the drift component to the overall drift; and

during the operation of the computer-controlled system, a processor of the computer-controlled system:

continually obtaining, by a processor of the computer-controlled system and from the at least one sensor of the computer-controlled system, the sensor measurements, the sensor measurements being of the physical quantities; and

predicting, by the processor of the computer-controlled system, a current state of the computer-controlled system by applying to the model, and from the continually obtained measurements, one or more of the obtained measurements of an earlier time.