Patent Application
20250045496
The present invention relates to a method for automated generation of a system for ascertaining a state variable characterizing a state of a technical system, to a virtual sensor system, to a computer program and to a machine-readable storage medium.
Physical models can be used to model physical interactions for virtual sensors. However, physical models with sufficient accuracy are not available for all applications of virtual sensors. Machine learning methods can close this gap.
German Patent Application No. DE 10 2017 218 922 A1 describes a method for generating virtual sensors by means of autoencoders.
However, virtual sensor systems based on such machine learning systems are relatively new and therefore less tried and tested. Insights into such machine learning systems are also often difficult. This is an obstacle when used in safety-critical applications.
German Patent Application No. DE 10 2020 215 138 A1 describes a method for creating physical equations by means of machine learning methods.
The present invention has an advantage that the generated virtual sensor system combines a high modeling quality of the virtual sensor system based on machine learning methods with good interpretability and thus the possibility of validating virtual sensor systems based on physical models.
Further aspects of the present invention are disclosed herein. Advantageous further developments and example embodiment of the present invention are disclosed herein.
In a first aspect, the present invention relates to a method for automated generation of a system for ascertaining a state variable characterizing a state of a technical system depending on a measured variable, i.e. a measured variable characterizing a second state of the technical system (in other words, a method for automated generation of a virtual sensor system). According to an example embodiment of the present invention, the method comprises the following steps:
As a result, the advantages mentioned above may achieved.
In a further development of the present invention, pairs of input data and associated output data of the machine learning system can be generated for symbolic regression by means of the machine learning system and the approximation is ascertained by means of regression of this input data and output data.
In other words, input data is provided and associated output data is provided by means of the machine learning system. Thus, such pairs of input data and associated output data can be easily generated in large numbers, which makes symbolic regression particularly simple.
According to an example embodiment of the present invention, alternatively or additionally, in the symbolic regression, regression candidates and in each case associated fit qualities can be proposed in particular to a user and a selected regression candidate can be received in particular by the user and adopted as an approximation of the machine learning system. This makes it particularly easy to incorporate existing expert knowledge into the generated virtual sensor system. In particular, it is possible that only regression candidates are proposed, of which the associated fit qualities exceed a predeterminable minimum fit quality.
With these methods, the technical system can be an electrical machine, and/or the measured variable can be ascertained by means of a voltage sensor, a temperature sensor or a speed sensor.
Alternatively, according to an example embodiment of the present invention, the technical system can be an energy storage device, in particular a battery or a fuel cell system, and/or the measured variable can be ascertained by means of a voltage sensor or a temperature sensor.
Alternatively, according to an example embodiment of the present invention, the technical system can be a braking and/or steering system of a motor vehicle and/or the measured variable can be ascertained by means of a voltage sensor, a temperature sensor, a speed sensor or a steering angle sensor.
In a further development of the present invention, by means of the generated system in the technical system, the state variable characterizing the state of the technical system can be ascertained depending on measurement data.
In a further aspect, the present invention relates to a virtual sensor system for ascertaining a state variable characterizing a state of a technical system depending on a measured variable characterizing a second state of the technical system, comprising the virtual sensor system generated by the method according to one of the aforementioned methods of the present invention for ascertaining the state variable depending on the measured variable.
In further aspects, the present invention relates to a computer program which is configured to perform any of the aforementioned methods of the present invention as set forth above (i.e., that the computer program includes instructions to cause a computer to perform any of these methods when the computer program is executed by the computer) and a machine-readable storage medium on which the computer program is stored.
In the following, example embodiments of the present invention are explained in more detail with reference to the figures.
In embodiments, the method for generating the generated system described in
Initially, a physical model is provided (1000), which describes the correlation between the measured variable and the state variable of the technical system to be characterized sufficiently well. In other words, the physical model receives the measured variable at its input and provides a variable at its output that is a good approximation of the state variable of the technical system. In other words, an estimated value of this variable is provided at the output of the physical model.
In some embodiments, this physical system can be parameterized using parameters. Pairs of measured variables ascertained by means of the sensor (101) and associated state variables, which can be measured in the technical system (100) by means of additional sensor technology in a development phase, for example, are then provided (1100). If the physical system can be parameterized by parameters, these parameters are adjusted (1200) in such a way that estimated values of the state variable, which the physical model ascertains depending on the measured variables, correspond as well as possible to the particular associated state variables contained in the pairs.
A machine learning system, in some embodiments a neural network, is then provided, which is linked to the physical model, for example added or multiplied, and the machine learning system is trained (1300) in such a way that the model, linked in this way, ascertains the in each case associated state variable as well as possible from the measured variables. In other words, if the measured variable is provided to the linked model at the input, an overall estimated value of the associated state variable is provided at the output. This training can be done in the usual way, for example, by minimizing a cost function which evaluates a deviation between the overall estimated value of the state variable and the provided state variable.
Now, input data, in embodiments the above measurement data, is fed to the machine learning system and associated particular output data is ascertained at its output (1400).
Based on this input data and output data, an approximation of the machine learning system is ascertained by means of symbolic regression (1500).
In some embodiments, the symbolic regression comprises providing regression candidates from a predeterminable function space, in particular those with a complexity less than a predeterminable maximum complexity (1501).
These regression candidates are fitted to the input data and associated output data (1502), i.e. the particular parameters of the regression candidates are adjusted so that the functional course of the regression candidate matches the pairs as closely as possible. As a measure of how well the course corresponds to these pairs, a particular fit quality is ascertained, which, in embodiments, is greater the better the course corresponds to the pairs (for example, an inverse of the χ2 function).
In some embodiments, the regression candidates of which the fit quality is worse than a predeterminable minimum fit quality (1503) are removed.
A regression candidate is now adopted as an approximation of the machine learning system (1504). In some embodiments, this is done by selecting the regression candidate with the best fit. In other, preferred embodiments, the regression candidates are provided to a user together with their particular fit quality (1505) and a selection is received from the user and provided as an approximation (1506).
The link of physical model and approximation is provided as a virtual sensor, i.e. as a system for ascertaining the state variable characterizing a state of the technical system depending on the measured variable (1600).
The technical system (100) as shown in
This ends the method.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10 2023 207 393.5 | Aug 2023 | DE | national |
