Method and device for constructing virtual sensor system
By combining physical models and machine learning systems, training the overall model and determining the approximation using symbolic regression to generate virtual sensor systems, the problem that virtual sensor systems in the prior art is difficult to achieve high modeling quality and interpretability in safety-critical applications. The effect of effectively determining the status of the technical system is achieved.
Patent Information
- Application Number
- CN202411049615.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-08-02
- Filing Date
- 2024-08-01
- Publication Date
- 2025-02-14
AI Technical Summary
In the prior art, when machine learning-based virtual sensor systems are used in safety-critical applications, it is difficult to achieve high modeling quality and good interpretability. At the same time, in the absence of precise physical models, it is difficult to effectively determine the status of the technical system.
By providing physical models and measurement pairs, adapting parameters of the physical model to minimize the deviation between the estimated value and the state variable, and in combination with machine learning systems such as Gaussian processes or neural networks, the overall model is trained to determine the overall estimate of the state variables. . Use symbolic regression to determine the approximation of machine learning systems, and perform logical operations on the physical model with the approximation to generate a virtual sensor system.
The high modeling quality and good interpretability of machine learning-based virtual sensor systems in safety-critical applications can be realized, and the status of the technical system can be effectively determined, making up for the shortcomings in the absence of precise physical models.
Smart Images
Figure CN119442826A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a method for automatically generating a system for determining state variables that characterize a state of a technical system, a virtual sensor system, a computer program and a machine-readable storage medium. Background Art
[0002] Physical models can be applied to correlate physical actions with virtual sensors However, physical models with sufficient accuracy are not known for all application cases of virtual sensors. Machine learning methods can fill this gap.
[0003] Thus, a method for generating a virtual sensor using an autoencoder is known from DE 10 2017 218 922 A1.
[0004] However, virtual sensor systems based on such machine learning systems are relatively new and therefore less tried and tested. Gaining insights into such machine learning systems is also often difficult. This is a hurdle when using them in safety-critical applications.
[0005] A method for creating physical equations using machine learning methods is known from DE 10 2020215138 A1.
[0006] Advantages of the present invention
[0007] In contrast, the present invention has the following advantage: the generated virtual sensor system combines the high modeling quality of a virtual sensor system based on machine learning methods with good interpretability and thus with the possibility of protecting a virtual sensor system based on a physical model.
[0008] Further aspects of the invention are the subject matter of the independent, parallel claims. Advantageous developments are the subject matter of the dependent claims. Summary of the invention
[0009] In a first aspect, the invention relates to a method for automatically generating a system for determining a state variable characterizing a state of a technical system from a measured variable, i.e. a measured variable characterizing a second state of the technical system, (in other words, a method for automatically generating a virtual sensor system), wherein the method comprises the following steps:
[0010] o providing a first model, in particular a physical model, wherein the first model is designed to determine an estimated value of the state variable as a function of the measured variable;
[0011] oProvide measurement pairs consisting of a measured variable and a respectively assigned state variable
[0012] o parameters characterizing the behavior of the first model are adapted or adapted as a function of the measurement pair, in particular so adapted that a deviation between an estimate determined from the measured variables using the first model and the state variable is minimized;
[0013] o providing a machine learning system, in particular a Gaussian process or a neural network, wherein the machine learning system is operated on the first model to generate an overall model, which is designed to determine an overall estimated value of the state variable as a function of the measured variables;
[0014] o The machine learning system is trained to minimize an error function or is trained to minimize an error function, where the error function comprises the deviation between a state variable and a population estimate of the state variable;
[0015] o determining an approximation of the machine learning system from the machine learning system using symbolic regression, in particular determining such that the approximation of the machine learning system approximates the machine learning system as best as possible;
[0016] o and wherein a logical operation performed by the first model and the symbolic regression is provided as the generated system.
[0017] The above-mentioned advantages are thereby achieved.
[0018] In one embodiment, it can be provided that, for symbolic regression using a machine learning system, a system is generated consisting of input data of the machine learning system and associated The approximation is determined by regressing these input data and output data.
[0019] In other words, input data is provided and the machine learning system is used to provide associated output data. Therefore, such pairs consisting of input data and associated output data can be easily constructed in large quantities, which makes symbolic regression particularly simple.
[0020] Alternatively or additionally, provision can be made that in symbolic regression, regression candidates and the respectively associated quality of fit are in particular suggested to the user, and the selected regression candidate is in particular accepted by the user and used as an approximation of the machine learning system. Existing expert knowledge can thus be incorporated in a particularly simple manner into the generated virtual sensor system. In particular, it is possible to suggest only regression candidates whose associated quality of fit exceeds a predeterminable minimum quality of fit.
[0021] In these methods, it can be provided that the technical system is an electric machine and / or that a voltage sensor, a temperature sensor or a rotational speed sensor is used to determine the measured variable.
[0022] Alternatively, it can be provided that the technical system is an energy storage device, in particular a battery or a fuel cell system, and / or that a voltage sensor or a temperature sensor is used to determine the measured variable.
[0023] Alternatively, it may be provided that the technical system is a braking and / or steering system of the motor vehicle and / or that a voltage sensor, a temperature sensor, a rotational speed sensor or a steering angle sensor is used to determine the measured variable.
[0024] In one embodiment, it can be provided that a state variable which characterizes the state of the technical system is determined by the generating system in the technical system using the measurement data.
[0025] On the other hand, the present invention relates to a virtual sensor system for determining a state variable characterizing a state of a technical system based on a measured variable characterizing a second state of the technical system, comprising a system generated using a method according to one of the above-mentioned methods, for determining the state variable based on the measured variable.
[0026] On the other hand, the present invention relates to a computer program which is configured to perform one of the above methods as shown above (i.e., the computer program contains instructions which, when a computer executes the computer program, cause the computer to perform one of these methods) and to a machine-readable storage medium having a computer program stored thereon. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Embodiments of the present invention are explained in more detail below with reference to the accompanying drawings. In these drawings:
[0028] Figure 1 Schematically illustrates an exemplary implementation for generating and applying a virtual sensor system;
[0029] Figure 2 An exemplary process of the method is shown in a flow chart. DETAILED DESCRIPTION
[0030] Figure 1An exemplary embodiment for implementing the present invention is schematically shown. A technical system (100), in some embodiments a fuel cell system, is shown. A sensor (101), in some embodiments a temperature sensor, is arranged on the technical system (100) for determining the gas temperature in the fuel cell system. The measured variable measured by the sensor (101) is transmitted to a computer (200) on which a virtual sensor system (210) is implemented, and a system generated by a second computer (300) determines a further state variable describing the state of the technical system, such as the gas pressure in the fuel cell system, based on the measured variable. The technical system (100) also includes an actuator (120), such as a valve or a pump, which is controlled based on the state variable determined by the generated system and characterizing the state of the technical system (100).
[0031] exist Figure 2 The method for generating the generated system exemplarily described in the embodiment is implemented as a computer program stored on a machine-readable storage medium (301) of the second computer (300).
[0032] Figure 2 A process for constructing an exemplary embodiment of the generated system is shown in a flowchart.
[0033] First, a physical model (1000) is provided which describes the interaction between the state variables to be characterized and the measured variables of the technical system sufficiently well. In other words, the physical model receives the measured variables at its input and provides at its output variables which represent the state variables of the technical system with a good approximation. In other words, an estimated value of the variable is provided at the output of the physical model.
[0034] In some embodiments, the physical system can be parameterized by parameters. Pairs of measured variables determined by sensors (101) and associated state variables are then provided (1100), wherein the associated state variables can be measured, for example, using additional sensor means in the technical system (100) during the development phase. If the physical system can be parameterized by parameters, these parameters are adapted (1200) in such a way that the estimated values of the state variables determined by the physical model from the measured variables correspond as closely as possible to the respective associated state variables contained in the pairs.
[0035] A machine learning system is then provided, in some embodiments a neural network, which is logically operated on, such as added or multiplied, with the physical model, and the machine learning system is trained (1300) so that the model thus logically operated determines the respectively associated state variables as well as possible from the measured variables. In other words, if the measured variables are provided at the input to the logically operated model, an overall estimate of the associated state variables is provided at the output. This training can be done in the usual way, for example, by minimizing a cost function that evaluates the deviation between the overall estimate of the state variables and the provided state variables.
[0036] The input data (in an embodiment, the measurement data described above) is now fed to the machine learning system, and the associated corresponding output data is determined at its output (1400). Based on these input data and output data, symbolic regression is used to determine an approximation of the machine learning system (1500).
[0037] In some embodiments, the symbolic regression includes providing regression candidates from a predeterminable function space, in particular, regression candidates having a complexity smaller than a predeterminable maximum complexity (1501).
[0038] These regression candidates are fitted to the input data and the associated output data (1502), i.e., the corresponding parameters of the regression candidates are adapted so that the function curve of the regression candidate corresponds to these pairs as well as possible. As a measure of how well the curve corresponds to these pairs, the corresponding fitting quality is determined. In an embodiment, the better the curve corresponds to these pairs (e.g., given the x-values), the better the fit quality. 2 The smaller the value is, the better the fit quality is.
[0039] In some embodiments, those regression candidates whose own fit quality is worse than a predeterminable minimum fit quality are removed ( 1503 ).
[0040] The regression candidates are now adopted as approximations of the machine learning system (1504). In some embodiments, this is done by selecting the regression candidate with the best quality of fit. In other preferred embodiments, the regression candidates are provided to the user along with their respective quality of fit (1505), and a selection is received from the user and provided as an approximation (1506).
[0041] The logical operation between the physical model and the approximation is provided as a virtual sensor, ie as a system for determining state variables characterizing the state of the technical system from the measured variables ( 1600 ).
[0042] Then, if Figure 1The technical system (100) in can be operated (1700) using the virtual sensor (210) obtained in this way.
[0043] This ends the method.
Claims
1. A method for automatically generating a system for determining a state variable characterizing a state of a technical system (100) from a measured variable, wherein the measured variable characterizes a second state of the technical system (100), wherein the method comprises the following steps: o providing a first model, wherein the first model is designed to determine an estimated value of the state variable from the measured variable; oprovide measurement pairs consisting of a measured variable and a respectively assigned state variable; o parameters characterizing the behavior of the first model are adapted based on the measurement pairs; oProviding a machine learning system, wherein: The machine learning system performs a logical operation with the first model to generate an overall model, wherein the overall model is configured to determine an overall estimated value of the state variable based on the measured variable; o The machine learning system is trained to minimize an error function, where the error function contains the deviation between the state variable and the population estimate of the state variable; o Determine an approximation of a machine learning system based on the machine learning system using symbolic regression; o and wherein the logical operations performed by the first model and the symbolic regression are provided as the generated system.
2. A method according to claim 1, wherein for symbolic regression using a machine learning system, pairs consisting of input data and associated output data of the machine learning system are generated, and the approximation is determined by regression of the input data and the output data.
3. A method according to any one of the preceding claims, wherein In symbolic regression, regression candidates and respectively associated qualities of fit are suggested, and the selected regression candidate is in particular received by a user and adopted as an approximation of the machine learning system.
4. A method according to any one of the preceding claims, wherein The technical system (100) is an electric machine and / or uses a voltage sensor, a temperature sensor or a rotational speed sensor to determine the measured variable.
5. A method according to any one of claims 1 to 3, wherein: The technical system (100) is an energy storage device, in particular a battery or a fuel cell system; and / or a voltage sensor or a temperature sensor is used to determine the measured variable.
6. A method according to any one of claims 1 to 3, wherein The technical system (100) is a braking and / or steering system of a motor vehicle and / or uses a voltage sensor, a temperature sensor, a rotational speed sensor or a steering angle sensor to determine a measured variable.
7. A method according to any one of the preceding claims, wherein A state variable characterizing a state of the technical system is determined by a generated system in the technical system (100) based on the measurement data.
8. A virtual sensor system for determining a state variable characterizing a state of a technical system (100) based on a measured variable characterizing a second state of the technical system (100), the virtual sensor system comprising a system generated using a method according to any one of claims 1 to 6, which is used to determine the state variable based on the measured variable.
9. A computer program, which is designed to carry out the method according to any one of claims 1 to 6.
10. A machine-readable storage medium having stored thereon a computer program according to claim 9.
Citation Information
Patent Citations
Method and device for data collection
DE102017218922A1
Method and apparatus for operating a technical system
DE102020215138A1