Method and device for determining control parameters for controlling a technical system
The method automates control parameter determination in technical systems by using pre-trained variational autoencoders and neural models to iteratively refine parameter sets, addressing manual adjustment limitations and data constraints, enhancing efficiency and adaptability.
Patent Information
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2025-08-26
- Publication Date
- 2026-03-11
AI Technical Summary
Existing methods for parameterizing control functions in technical systems, such as braking systems and steering assistance, often require manual adjustment and rely on limited data or extensive simulation, lacking an integrated and automated approach.
A computer-implemented method using pre-trained variational autoencoders and neural posterior estimation models to iteratively determine optimal control parameters by simulating system behavior and updating probability distributions based on latent representation vectors, eliminating the need for explicit quality functions and handling varying data lengths.
Enables automated and efficient parameterization of control functions, reducing reliance on expert knowledge and simulation time, while effectively adapting to different system behaviors.
Smart Images
Figure 2026042757000001_ABST
Abstract
Description
[Technical Field]
[0001] Technical Field The present invention relates to controlled technical systems, and in particular to designing controls for technical systems by appropriately parameterizing the controls. [Background technology]
[0002] Background technology Technical systems, such as braking systems, especially ABS and ESP, hydraulic drives, and functions such as steering assistance and lane assistance, require control functions to ensure a certain functionality. These control functions are usually defined by setting function rules and a number of parameters, and are thus adapted to the characteristics of the technical system to be controlled.
[0003] In many cases, the application of the control, i.e., parameterization, is performed manually and by skilled applicators who often have many years of experience with the technical systems and the functions to be controlled. For application, the control parameters are adjusted and then the system behavior is tested.
[0004] Methods for parameterizing machine learning models, especially variational autoencoders, in combination with Bayesian optimization are already known, e.g., Antoine Grosnit et al., "High-Dimensional Bayesian Optimization with Variational Autoencoders and Deep Metric Learning," https: / / arxiv.org / pdf / 2106.03609.pdf. However, these methods assume an explicit specification of a quality function for Bayesian optimization and consistent data dimensionality for the variational autoencoder.
[0005] Furthermore, approaches that use Bayesian optimization either limit the possible amount of simulation data that can be considered, or severely limit the representation strength of the learned embeddings.
[0006] Alternative strategies that do not involve unsupervised pre-training, relying solely on Simulation Based Inference (SBI) approaches, rely on large amounts of simulated data, so the simulation time required is rarely feasible in practice. [Prior art documents] [Non-patent literature]
[0007] [Non-Patent Document 1] Antoine Grosnit et al., “High-Dimensional Bayesian Optimization with Variational Autoencoders and Deep Metric Learning”, https: / / arxiv.org / pdf / 2106.03609.pdf Summary of the Invention [Problem to be solved by the invention]
[0008] The object of the present invention is to provide an improved method for parameterizing control functions, which is implemented in an integrated and, in particular, automated manner in application processes. [Means for solving the problem]
[0009] Disclosure of the Invention The above problem is solved by a method for parameterizing a control function for controlling a technical system as set forth in claim 1 and a corresponding device as set forth in the parallel claim.
[0010] Further embodiments are set forth in the dependent claims.
[0011] According to a first aspect, there is provided a computer-implemented method for determining parameter values for parameters of an optimized parameter set for operating a given technical system, comprising: the parameter set corresponds in particular to control parameters, and the behavior of the parameterized technical system can be simulated by operating variable transitions which show the time course of at least one input variable, the time course of at least one output variable and the time course of at least one state variable, The computer-implemented method comprises the following steps: - providing a representation model based on the data, the representation model being trained to map the time evolution of operational quantities of one or more technical systems to a respective latent representation vector characterizing the behavior of the technical system; providing a data-based distribution model, the distribution model being trained to map latent representation vectors obtained from simulated operational evolution of a given technical system to probability distributions of parameter values for the parameters of the parameter set; Equipped with The following steps: providing parameter values of an initial parameter set or selecting parameter values of the parameter set from a parameter probability distribution by random selection; - simulating or measuring a technical system parameterized by the parameter values of the parameter set in order to obtain the time course of the operating quantity; evaluating the obtained motion vectors using a data-based representation model to obtain latent representation vectors; retraining or re-educating the data-based distribution model using a training dataset consisting of the obtained latent representation vectors and parameter set, thereby obtaining an updated probability distribution of the parameters from the data-based distribution model; is carried out repeatedly, A method is envisaged.
[0012] Furthermore, an optimized parameter set can be determined based on the posterior probability distribution obtained from the retrained or re-educated distribution model, and the parameter values of the optimized parameter set are determined as the parameter values that most likely produce a given reference behavior.
[0013] Furthermore, the representation model can be provided in the form of the encoder part of a variational autoencoder, configured as a recurrent neural network, or as a convolutional neural network, or as a transformer network, and the variational autoencoder has / is trained in an unsupervised manner for a plurality of operational transitions of a plurality of different technical systems.
[0014] To enable evaluation of time series of measurement data of different lengths that represent the system behavior of a technical system with parameters of a predetermined parameter set, a variational autoencoder can be trained to obtain latent representation vectors with predefined dimensions for a predetermined parameter set. In this case, the time series of data can be determined by measurement or simulation. In other words, a physical model of the technical system, a control function parameterized according to the parameter set to be observed, and time series data are determined. These time series data are used for unsupervised training of the variational autoencoder.
[0015] Here, the latent representation vector can be considered as a low-dimensional representation of the parameters used. The latent representation vector is used to evaluate the subsequent simulated system behavior in relation to a reference behavior provided in the form of a simulated time course of the operating quantities, respectively. For this purpose, a model that maps time series data to latent representation vectors for an already parameterized technical system is first created using a variational autoencoder. Essentially, any variational autoencoder with a recurrent neural network or a convolutional neural network (including LSTM or GRU networks) or with a network with a transformer architecture is suitable as a representation model.
[0016] Alternatively, the representation model may be / can be formed in the form of an MVTS Transformer model with an imputation task, as disclosed, for example, in "A Transformer-based Framework for Multivariate Time Series Representation Learning" by George Zerveas et al., https: / / arxiv.org / abs / 2010.02803.
[0017] Furthermore, a distribution model, particularly in the form of a flow-matching a posteriori model or a posteriori approximation model, can be trained using a training data set consisting of latent representation vectors of motion quantity transitions and their corresponding parameter values.
[0018] Pre-trained representation models allow for the use of data in the form of transitions of parameterized operational quantities of technical systems collected from multiple different sources and mapped to latent representation vectors, so that correspondingly less data needs to be collected or simulated for the subsequent application process, since the basic characteristic time series data have already been learned.
[0019] The latent representation vector can be used as a comparison measure between the desired and observed system behavior of a technical system.
[0020] In particular, a distance measure for the latent representation vectors can be used to perform the evaluation of the evaluated parameterizations.
[0021] Then, parameters for the parameter set are randomly selected from the prior distribution (or later the posterior distribution).
[0022] A time series of the transition of the operating quantity is determined for the randomly selected parameters by simulation or measurement.
[0023] Using the encoder part of a previously trained variational autoencoder as a parameter model, a latent representation vector is identified for the time series of motion data obtained in this way, allowing the time series data to be represented by a vector of fixed size.
[0024] The distribution model may be configured as a neural posterior estimation model, which can show a probability distribution of parameter values for parameters given a latent representation vector obtained from a simulation of a given technical system.
[0025] Using the distribution model, a posterior probability distribution for parameter values that are most likely to approximate a reference behavior can be learned. The reference behavior corresponds to a given operational transition that corresponds to the desired behavior of a suitable parameterized technical system. The evaluation is performed based on the latent embedded operational transitions of the already evaluated parameter set. Then, parameter values for the next iteration are randomly selected (extracted) from the learned posterior distribution and re-simulated, thereby performing the method iteratively. Using the representation model, a new latent representation vector of a fixed length is generated for the simulated time series of operational transitions.
[0026] A distribution model in the form of a neural posterior estimation model is thus iteratively further trained, thereby improving the posterior probability distribution over possible parameter values. The method is performed iteratively until a given stopping condition is reached (e.g., by reaching a predetermined maximum number of iterations). The parameters of the parameter set that most likely produce the reference behavior can then be applied to the target system and evaluated there.
[0027] By pre-training the representation model, the past application processes of one or more technical systems can be used for parameterizing a new technical system, whereby many relationships known to experts are automatically trained. That is, the representation model uses the general relationship between the behavior of a technical system parameterized with a given parameter set and the corresponding parameter set as a reference to find a suitable parameter set for the control to be applied.
[0028] In this case, the method supports the creation of a posterior probability function for selecting parameter values for a parameter set for a specific technical system to be applied, thereby making it possible to find a suitable parameter set by iteratively approximating to optimized parameter values. For this purpose, time series data in the form of operational volume transitions are simulated for the technical system, and parameter values for the parameter set to be found are found based on the increasingly accurate posterior probability distribution.
[0029] In contrast to prior art methods based on Bayesian optimization, no explicit specification of the quality function is required and it is also possible to handle time series of different lengths.
[0030] By appropriately selecting the size of the latent representation vector, complex time series data can be represented without requiring excessive simulation runs, as is known for example in the case of simulation-based inference methods.
[0031] In the following, embodiments will be described in more detail based on the accompanying drawings. [Brief explanation of the drawings]
[0032] [Figure 1] 1 is a schematic diagram of a technical system with a control structure of a controlled object; [Figure 2] 1 is a flow chart illustrating a method for optimizing parameter values of parameters for a given technical system. DETAILED DESCRIPTION OF THE INVENTION
[0033] Description of the embodiment In the following, a technical system 1 is described, the functionality of which is determined by one or more parameters of a parameter set.
[0034] Examples of such systems may be braking systems with corresponding ABS and ESP functions, hydraulic drives, steering functions and lane keeping functions, each with a complex control structure adapted to the technical system. The control structure comprises a parameterized control function 2, which consists of a number of parameters from a parameter set P, and a control object 3, which corresponds to the technical device to be controlled.
[0035] To implement a control for a new type of function, a control function must first be provided, which must then be parameterized by control parameters in the application process. Parameter discovery has traditionally been performed manually or based on Bayesian optimization methods, which requires defining a quality function, and the quality of the parameter set is highly dependent on the selection of this quality function.
[0036] Here, a method is used to parameterize the control functions, as will be explained in detail with reference to the flow chart of FIG.
[0037] For this purpose, in step S1, firstly a pre-trained representation model is provided, which is used to provide the system behavior of a plurality of technical systems with various control functions that have already been parameterized in an arbitrary or optimized manner using other methods.
[0038] That is, the pre-trained representation model is based on a plurality of different dynamic technical systems, each of which is arbitrarily or optimally parameterized, despite implementing different control functions. The system behavior is determined based on time series data of operating quantities that map the dynamic system behavior. These operating quantity variations over time may include input quantities (driving quantities), output quantities (manipulated quantities), and measured or modeled state quantities (e.g., temperature, pressure, rotation speed, etc.) of the technical system.
[0039] To determine the representation model, the representation model is pre-trained in an unsupervised manner on time series data using a variational autoencoder (as the encoder part) or a suitable generative model, such as the MVTS transformer model with a complementary task, as known from George Zerveas et al., "A Transformer-based Framework for Multivariate Time Series Representation Learning," https: / / arxiv.org / abs / 2010.02803. This allows relevant characteristics of multiple technical systems to be mapped to latent representations. These relevant characteristics are provided as low-dimensional latent representation vectors.
[0040] The latent representation vectors are used to evaluate subsequent simulated system behaviors relative to reference behaviors provided in the form of respective simulated operating variable transitions.
[0041] Subsequently, in step S2, initial parameter values of the parameter set (prior parameter set) are provided, or parameter values are randomly selected from the already learned posterior probability distributions.
[0042] In step S3, the system behavior is simulated based on the provided or selected parameter set to obtain simulated operational quantity transitions as time series data.
[0043] In step S4, the representation model is used to generate latent representation vectors for the simulated motion quantity transitions from the simulation.
[0044] This leads to training or retraining, in step S5, a distribution model, which may be configured as, for example, a neural posterior estimation model. This distribution model learns a posterior probability distribution over the likely parameter values that most closely resemble the reference behavior. The neural posterior estimation model corresponds to a mapping of latent vectors to parameter sets. To predict a probability distribution over possible values of the parameter set, the neural posterior estimation model must be inverted or conditioned to the latent representation.
[0045] In this case, the distribution model maps latent expression vectors to parameter values of a parameter set. Training of the model is performed in a manner known per se. From a model trained to predict latent expressions from parameters, the model can be inverted using Bayes' theorem to arrive at predictions about likely parameter values for a given reference behavior, thereby determining parameter values relative to the reference.
[0046] The method is repeated and the posterior model is retrained until a stopping condition is met, defined, for example, by a maximum number of iterations, a maximum period of optimization, etc.
[0047] In step S6, the interrupt conditions are checked accordingly: if the obtained parameter values are sufficiently optimized (option: yes), the method proceeds to step S7; otherwise (option: no), the method returns to step S2.
[0048] Now, in step S2, the trained distribution model can be used to determine parameter values (posteriori) for the next iteration.
[0049] Subsequently, in step S6, the parameter values that most likely produce the reference behavior based on the posterior probability distribution obtained from the distribution model can be used to parameterize the control function, which can serve as a starting point for further adaptation of the parameter set by expert knowledge.
Claims
1. 1. A computer-implemented method for determining parameter values for parameters of an optimized parameter set for operating a given technical system (1), comprising: The behavior of the parameterized technical system (1) can be simulated by operational variable transitions that represent the time course of at least one input variable, the time course of at least one output variable, and the time course of at least one state variable, The computer-implemented method comprises the following steps: - a step (S1) of providing a data-based representation model, said representation model being trained to map the temporal evolution of operational quantities of one or more technical systems to a respective latent representation vector characterizing the behavior of said technical system (1); - providing a data-based distribution model, said distribution model being trained to map latent representation vectors obtained from simulated operational evolution of said given technical system to probability distributions of parameter values for the parameters of said parameter set; Equipped with The following steps: - providing (S2) parameter values of an initial parameter set or selecting parameter values of the parameter set from a parameter probability distribution by random selection; a step (S3) of simulating or measuring said technical system parameterized by the parameter values of said parameter set in order to obtain the behavioral variables over time; - evaluating the obtained motion vectors using a representation model based on the data (S4) to obtain a latent representation vector; a step (S5) of retraining or re-educating the data-based distribution model using a training dataset consisting of the acquired latent representation vectors and the parameter set, thereby obtaining an updated probability distribution of parameters from the data-based distribution model; is carried out iteratively.
2. The method comprises the following further steps: A step (S7) of determining an optimized parameter set based on the posterior probability distribution obtained from the retrained distribution model, wherein the parameter values of the optimized parameter set are determined as the parameter values that most likely produce a given reference behavior. Equipped with The method of claim 1.
3. the representation model is provided in the form of an encoder portion of a variational autoencoder; the variational autoencoder is configured as a recurrent neural network, as a convolutional neural network, or as a transformer network; The variational autoencoder has been / is trained in an unsupervised manner for a plurality of operational transitions of a plurality of different technical systems; 3. The method according to claim 1 or 2.
4. said representation model is / is formed in the form of an MVTS transformer model with a completion task, 3. The method according to claim 1 or 2.
5. said distribution model, in particular in the form of a flow matching posterior model or posterior approximation model, being trained using a training data set consisting of latent representation vectors of the motion quantities transitions and their corresponding parameter values; 5. The method according to any one of claims 1 to 4.
6. the distribution model is configured as a neural posterior estimation model and indicates a probability distribution of parameter values for the parameters for a latent expression vector obtained from a simulation of the predetermined technical system; 6. The method according to any one of claims 1 to 5.
7. Apparatus for carrying out the method according to any one of claims 1 to 6.
8. A computer program product comprising instructions for causing at least one data processing device to carry out the steps of the method according to any one of claims 1 to 6 when the program is executed by said at least one data processing device.
9. A machine-readable storage medium comprising instructions which, when executed by at least one data processing device, cause the at least one data processing device to perform the steps of the method of any one of claims 1 to 6.