Search

CN-121996937-A - Method for predicting the time course of a physical target variable by means of a machine learning model

CN121996937ACN 121996937 ACN121996937 ACN 121996937ACN-121996937-A

Abstract

A method has the steps of providing, for each of a plurality of physical variables, multivariable sensor data having sensor data representing a time course of the physical variable, each physical variable being assigned a text description describing the physical variable and its measurement environment, dividing the sensor data into a plurality of sensor data segments for each physical variable, determining, for each sensor data segment, a sensor data segment representation representing the sensor data segment having a predefined dimension, determining input elements using the sensor data segment representation, time-dependent position information representing the position of the sensor data segment within a time period, and the text description of the physical variable, and predicting the time course of the physical target variable by means of a machine learning model in response to input of all input elements and at least one target parameter query representing the position of the time course to be predicted within the time period and the text description of the physical target variable into the machine learning model.

Inventors

  • S. Golvin
  • M. SiGe
  • M. Moshkowitz

Assignees

  • 罗伯特·博世有限公司

Dates

Publication Date
20260508
Application Date
20251103
Priority Date
20241104

Claims (10)

  1. 1. Method (100) for predicting a time course (214) of a physical target variable by means of a machine learning model (212), the method (100) having: -providing (102) multivariable sensor data, which is assigned to a time period and which has, for each of a plurality of physical quantities, a respective sensor data (210), which represents a time course of the physical quantity over the time period, wherein each physical quantity is assigned a respective text description, which describes the physical quantity; For each physical parameter (104) of the plurality of physical parameters: Dividing the respective sensor data into a respective plurality of sensor data segments; For the plurality of sensor data each of the sensor data segments: o determining a respective sensor data segment representation, the respective sensor data segment representation representing the sensor data segment, and the respective sensor data segment representation (irrespective of the number of data points of the sensor data segment) having a predefined dimension, O determining a respective input element using the respective sensor data segment representation, time-dependent position information representing the position of the sensor data segment within the time period, and the respective textual description of the physical parameter; -predicting (106) the time course (214) of the physical object parameter by means of the machine learning model (212) in response to inputting all input elements and at least one object parameter query into the machine learning model (212), the object parameter query representing a position of the time course (214) to be predicted within the time period and a textual description of the physical object parameter.
  2. 2. The method (100) of claim 1, Wherein the respective plurality of sensor data segments of at least one physical parameter has at least two sensor data segments with mutually different numbers of data points.
  3. 3. The method (100) according to claim 1 or 2, Wherein the time-related position information represents a start time point and an end time point within the time period.
  4. 4. The method (100) according to claim 1 to 3, Wherein the machine learning model (212) has a transformer model, and the encoder (212-1) and/or decoder (212-2) of the transformer model has an attention layer to which all input elements are fed.
  5. 5. The method (100) according to any one of claims 1 to 4, Wherein a corresponding sensor data segment representation for a sensor data segment is determined by means of an attention unit having a learned sensor data segment-specific parameter vector as a query and having the sensor data segment as a key and as a value, and/or Wherein the respective input element is determined using the respective sensor data segment representation, the respective position representation and the respective text representation of the physical variable, wherein the position representation is determined by means of an attention unit having the learned, position-specific parameter vector as a query and the time-dependent position information as a key and as a value.
  6. 6. The method (100) according to any one of claims 1 to 5, Wherein the machine learning model (212) has a transformer model, one or more attention layers of the transformer model in the encoder (212-1) and/or decoder (212-2) having an attention unit to which the target parametric query is delivered.
  7. 7. A system (200), the system having: -a device (202) designed for carrying out the technical process; one or more sensors (208) for detecting the multivariate sensor data (210), and Control device (204) which is designed to predict a time-dependent course (214) of the physical target variable according to any one of claims 1 to 6 and to control the technical process taking into account the prediction.
  8. 8. A data processing unit set up for implementing the method according to any one of claims 1 to 6.
  9. 9. A computer program having instructions which, when implemented by a processor, cause the processor to perform the method (100) according to any one of claims 1 to 6.
  10. 10. A computer readable medium storing instructions which, when executed by a processor, cause the processor to perform the method (100) according to any one of claims 1 to 6.

Description

Method for predicting the time course of a physical target variable by means of a machine learning model Prior Art It may be desirable for various technical (e.g. physical or chemical) processes to predict the time-varying course of a physical parameter from multivariate time-series data of other physical parameters and/or to predict anomalies from multivariate time-series data of a plurality of physical parameters. It may be desirable, for example, to predict the state of health (state-of-health) or hydrogen loading of the fuel cell from the time course of the current intensity and voltage, or in the case of a drilling machine, which material is drilled from the time course of the current intensity and voltage, or to predict anomalies from the time course of the current intensity and voltage, etc. For this purpose, machine learning models can generally be trained for exactly one use case (for example for predicting the health of a fuel cell). Disclosure of Invention The present disclosure relates to a method for predicting a time course of a physical target parameter by means of a machine learning model from multivariate sensor data, wherein the multivariate sensor data may be irregularly sampled sensor data. If sensor data of different sensors are detected, these sensor data may have mutually different sampling rates. Data points may also be missing in some sensor data (e.g., due to measurement errors or because the data points were removed due to excessive uncertainty, etc.). The time segments in which the sensor data is present may also have mutually different durations. Intuitively, it may occur that not every data point in the first sensor data can be correlated with a data point in the second sensor data that is different from the first sensor data. The method described herein enables prediction of the time course of a physical target variable also in the case of such irregular sensor data. This is achieved, for example, by dividing the sensor data into sensor data segments and then determining for each sensor data segment a respective sensor data segment representation having the same predefined dimension for all sensor data segments. Thus, the dimension represented by the sensor data segment is independent of the regularity (e.g., sampling rate, presence of data points, etc.) of the data points of the sensor data segment. The machine learning model described herein may also be trained on predictions of corresponding physical target parameters for a plurality of different tasks having at least partially different physical parameters. This can be, for example, that the laws of physics that apply across different tasks have been learned efficiently. Such training is achieved because the methods described herein are capable of handling irregular multi-variable sensor data. The various aspects relate to a method for predicting a time course of a physical target variable by means of a machine learning model, comprising providing multivariable sensor data which are assigned to a time segment and which have, for each of a plurality of physical variables, respective sensor data which represent the time course of the physical variable in the time segment, wherein each physical variable is assigned a respective text description which describes the physical variable (and optionally also describes the measuring environment in which the respective sensor data is detected), dividing, for each physical variable, the respective sensor data into a respective plurality (for example, disjoint) of sensor data segments, determining, for each of the plurality of sensor data segments, a respective sensor data segment representation which represents the sensor data segment and which represents the number of data points of the sensor data segment, in dependence on the position of the respective sensor data segment, of the sensor data segment in a respective time-dependent manner of the sensor data segment, and a time-dependent element of the input model being predicted by means of the machine learning model, and the position of the respective sensor data of the physical variable in the time segment being input-dependent on at least one of the time-dependent model, and the position of the respective sensor data of the physical variable being input-dependent on the time-dependent model, the target parameter query represents a (e.g., temporal) location of the time course to be predicted over the period of time and a textual description of the physical target parameter. The following describes various embodiments. Example 1 is a method for predicting a time course of a physical object parameter by means of a machine learning model as described previously. Example 2 is set up according to example 1, wherein a corresponding plurality of sensor data segments of the at least one physical parameter have at least two sensor data segments with mutually different numbers of data points. Since each sensor data segment is mapped onto a c