Method and device for determining control parameters for regulating a technical system
A method using pre-trained variational autoencoders and Neural Posterior Estimation models automates control parameterization in technical systems, overcoming manual intervention and data limitations, efficiently determining optimal parameter sets for desired behaviors.
Patent Information
- Application Number
- DE102024208128
- Authority / Receiving Office
- DE · DE
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2026-03-05
AI Technical Summary
Existing methods for parameterizing control functions in technical systems, such as braking systems and steering systems, require manual intervention by experienced engineers and are limited by the need for explicit performance functions and consistent data dimensionality, restricting the expressiveness of learned embeddings and simulation data usage.
A computer-implemented method using a pre-trained variational autoencoder and Neural Posterior Estimation model to iteratively determine optimized parameter sets by simulating system behavior, mapping time series data to latent vectors, and refining probability distributions to achieve desired system behavior without explicit performance functions.
Automates the parameterization process, reduces the need for extensive simulation data, and allows handling of time series of varying lengths, effectively identifying parameter sets that achieve the desired system behavior.
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
Technical field
[0001] The invention relates to controlled technical systems, and in particular to the design of controls for technical systems by suitable parameterization of the control. Technical background
[0002] Technical systems, such as braking systems, and in particular functions like ABS and ESP, hydraulic drives, and steering and lane assist systems, require control functions to ensure specific functionalities. These control functions are typically defined by specifying a functional instruction and a multitude of parameters, and thus adapted to the characteristics of the technical system being controlled.
[0003] The application, i.e., the parameterization of the controls, is often done manually by experienced application engineers. These engineers often have years of experience with the technical system and the functions to be controlled. During the application process, the control parameters are set and the system behavior is then tested.
[0004] Methods are already known that use machine learning models, especially variational autoencoders, in combination with Bayesian optimization, as described, for example, in Antoine Grosnit et al., "High-Dimensional Bayesian Optimisation with Variational Autoencoders and Deep Metric Learning", https: / / arxiv.org / pdf / 2106.03609.pdf, to perform parameterization. However, this approach requires the explicit specification of a performance function for the Bayesian optimization as well as consistent data dimensionality for the variational autoencoder.
[0005] Furthermore, the Bayesian optimization approach limits the possible amount of simulation data that can be considered, or severely restricts the expressiveness of the learned embeddings.
[0006] Alternative strategies without unsupervised pretraining, which only use simulation-based inference (SBI) approaches, rely on so much simulated data that the necessary simulation time is rarely achievable in practice.
[0007] The object of the present invention is to provide an improved method for parameterizing control functions, which combines and, in particular, automates the application process. Disclosure of the invention
[0008] This problem is solved by the method for parameterizing a control function for a control of a technical system according to claim 1 and a corresponding device according to the dependent claim.
[0009] Further details are specified in the dependent claims.
[0010] According to a first aspect, a computer-implemented method for determining parameter values of parameters of an optimized parameter set for operating a specific technical system is provided, wherein the parameter set corresponds in particular to control parameters, wherein the behavior of the parameterized technical system can be simulated by operating variable profiles that specify time profiles of at least one input variable, at least one output variable and at least one state variable, with the following steps: - Providing a data-based representation model that is trained to map temporal operational size profiles of one or more technical systems to a latent representation vector that characterizes the behavior of the technical system; - Providing a data-based distribution model that is trained to map latent representation vectors resulting from simulated operational characteristic curves of the specific technical system to a probability distribution of parameter values of the parameter sets; the following steps are performed iteratively: - Providing parameter values of an initial parameter set or selecting parameter values of a parameter set from a probability distribution of parameters by random selection; - Simulating or measuring the technical system parameterized with the parameter values of the parameter set in order to obtain temporal operating parameter profiles; - Evaluating the obtained operational size profile with the data-based representation model to obtain a latent representation vector; - Further development or retraining of the data-based distribution model with a training data set from the obtained latent representation vector and the parameter set, so that an updated probability distribution of parameters results from the data-based distribution model.
[0011] Furthermore, the optimized parameter set can be determined based on the posterior probability distribution resulting from the further developed or retrained distribution model, whereby the parameter values of the optimized parameter set are determined as those that most likely produce a given reference behavior.
[0012] Furthermore, the representation model can be provided in the form of an encoder part of a variational autoencoder, configured as a recurrent or convolutional neural network or as a transformer network, and trained / will be trained unsupervised for a variety of temporal operational parameter profiles of different technical systems.
[0013] To enable the evaluation of time series of measurement data of varying lengths, representing the system behavior of a technical system with parameters from a specific parameter set, a variational autoencoder can be trained to generate a latent representation vector with predefined dimensionality for the given parameter set. The time series data can be obtained through measurement or simulation. In other words, a physical model of the technical system, a control function parameterized according to the parameter set under consideration, and time series data are generated. This time series data is then used for the unsupervised training of a variational autoencoder.
[0014] The latent representation vector can now be viewed as a low-dimensional representation of the parameters used. This vector serves to evaluate the subsequently simulated system behavior in relation to the reference behavior, which is provided in the form of simulated time-dependent operational parameters. For this purpose, a model is first created using a variational autoencoder, which maps time series data for already parameterized technical systems onto the latent representation vector. In principle, any variational autoencoder with recurrent or convolutional neural networks (including LSTM or GRU networks) or networks with transformer architectures is suitable as a representation model.
[0015] Alternatively, the representation model can be formed in the form of an MVTS transformer model with an imputation task, as revealed, for example, in George Zerveas et al., “A Transformer-based Framework for Multivariate Time Series Representation Learning”, https: / / arxiv.org / abs / 2010.02803.
[0016] Furthermore, the distribution model, especially in the form of a Flow Matching Posterior model or a Posterior Approximation model, can be trained with training datasets of latent representation vectors of the operational variable profiles and associated parameter values.
[0017] Pre-trained representation models make it possible to use data in the form of operational characteristic curves of a parameterized technical system collected from various sources and to map them in a latent representation vector. For a subsequent application process, correspondingly less data needs to be collected or simulated, since the fundamental characteristic time series data has already been learned.
[0018] The latent representation vector can be used as a comparison measure between the desired and observed system behavior of the technical system.
[0019] In particular, a distance measure to the latent representation vector can be used to evaluate the evaluated parameterization.
[0020] Subsequently, parameters for a parameter set are randomly selected from a prior distribution (or later, a posterior distribution).
[0021] Time series of operational parameter profiles are determined for randomly selected parameters through simulation or measurement.
[0022] The encoder part of the previously trained variational autoencoder is used as a parameter model to determine a latent representation vector for the resulting time series of operational variable profiles. This allows the time series data to be represented in a vector of fixed size.
[0023] The distribution model can be designed as a Neural Posterior Estimation model and, for latent representation vectors resulting from simulations of the specific technical system, specify probability distributions of parameter values for the parameters.
[0024] Using the distribution model, a posterior probability distribution can be learned over possible parameter values that most closely approximate the reference behavior. The reference behavior corresponds to predefined operating characteristic curves that represent the desired behavior of a suitable parameterized technical system. The evaluation is performed based on the latently embedded operating characteristic curves of the already evaluated parameter sets. From the learned posterior distribution, the parameter values for the next iteration are then randomly selected (drawn) and simulated again, thus executing the process iteratively. The representation model is used to generate a latent representation vector of fixed length for the simulated time series of operating characteristic curves.
[0025] The distribution model, in the form of the Neural Posterior Estimation model, is iteratively trained, thereby improving the posterior probability distribution for possible parameter values. This process is executed iteratively until a predefined termination condition is reached (e.g., by reaching a predetermined maximum number of iterations). Subsequently, the parameters from a parameter set that are most likely to produce the reference behavior can be applied to the target system and evaluated there.
[0026] Pretraining the representation model allows past application processes of one or more technical systems to be used for parameterizing a new technical system. This automatically trains many relationships known to experts. The representation model thus uses the general relationships between the behavior of a technical system parameterized with a specific parameter set and the corresponding parameter set as a reference to find the appropriate parameter set for a control system to be applied.
[0027] The method supports the creation of posterior probability functions for selecting parameter values for the parameter set with respect to the specific technical system to be applied. This allows the appropriate parameter set to be found by iteratively approximating the optimized parameter values. For this purpose, time series data in the form of operational variable profiles are simulated for the technical system in order to identify parameter values for the desired parameter set based on the increasingly precise posterior probability distribution.
[0028] In contrast to state-of-the-art methods based on Bayesian optimization, no explicit specification of a performance function is necessary, and time series of different lengths can also be processed.
[0029] By appropriately choosing the size of the latent representation vector, even complex time series data can be represented without requiring too many simulation runs, as is known, for example, in simulation-based inference methods. Brief description of the drawings
[0030] The embodiments are explained in more detail below with reference to the accompanying drawings. These show: Fig. 1 a schematic representation of a technical system with a control structure of a control loop; Fig. 2. A flowchart illustrating a procedure for optimizing parameter values for a specific technical system. Description of embodiments
[0031] The following describes a technical system 1 whose function is determined by one or more parameters of a parameter set.
[0032] Examples of such systems include a braking system with corresponding ABS and ESP functions, hydraulic drives, steering and lane-keeping functions, each of which has complex control structures adapted to the technical system. The control structure comprises a control function 2 with a parameterization consisting of a plurality of parameters from a parameter set P, and a control loop 3, which corresponds to a technical device to be controlled.
[0033] To implement a control system for a novel function, a control function is typically defined first, which must then be parameterized in an application process using control parameters. Parameter selection is currently done manually or based on Bayesian optimization methods, which, however, require the definition of a performance function, the choice of which significantly influences the performance of the parameter set.
[0034] A procedure is now used to parameterize the control function, as shown in the flowchart of the Fig. 2 will be explained in more detail.
[0035] In step S1, a pre-trained representation model is first provided. This pre-trained representation model serves to provide the system behavior of a large number of technical systems with various control functions, which have already been parameterized in an arbitrary or optimized way using other methods.
[0036] Thus, various dynamic technical systems are considered, each with an arbitrary or optimal parameterization, although they implement different control functions. The system behavior is determined using time series data of operating variables that represent the dynamic system behavior. These time-based operating variable profiles can include input variables (control variables), output variables (manipulated variables), and measured or modeled state variables (such as temperature, pressure, rotational speed, and the like) of the technical system.
[0037] To determine the representation model, it is pre-trained unsupervised with the time series data using a variational autoencoder (as the encoder component) or a suitable generative model, such as an MVTS transformer model with an imputation task, as described in George Zerveas et al., "A Transformer-based Framework for Multivariate Time Series Representation Learning", https: / / arxiv.org / abs / 2010.02803_bekannt. This allows relevant characteristics of a wide variety of technical systems to be mapped into a latent representation. These are provided as a low-dimensional latent representation vector.
[0038] The latent representation vector serves to evaluate the subsequently simulated system behavior in relation to the reference behavior, which is provided in the form of simulated operating parameter profiles.
[0039] Subsequently, in step S2, initial parameter values of a parameter set (prior parameter set) are provided, or parameter values are randomly selected from a previously learned posterior probability distribution.
[0040] In step S3, the system behavior is simulated based on the provided or selected parameter set in order to obtain simulated operating parameter profiles as time series data.
[0041] In step S4, the representation model is used to generate a latent representation vector for the operational variable profiles simulated from the simulation.
[0042] In step S5, a distribution model is trained or further developed, which can, for example, be a Neural Posterior Estimation (NPE) model. This model learns a posterior probability distribution over possible parameter values that most closely approximate the reference behavior. The NPE corresponds to a mapping of latent vectors to parameter sets. The NPE must be invertible, or conditionable to latent representations, in order to predict a probability distribution over the possible values of a parameter set.
[0043] The distribution model maps the latent representation vectors to the parameter values of the parameter set. The model is trained in a known manner. To predict the latent representation from the parameters of the trained model and thus arrive at a prediction of possible parameter values for a given reference behavior, Bayes' theorem is used to invert the model so that the parameter values for a reference can be determined.
[0044] The process is iterated and the posterior model is further developed until a termination condition is met, for example, given by a maximum number of iterations, a maximum optimization time, and the like.
[0045] In step S6, the termination condition is checked. If the obtained parameter values are sufficiently optimized (alternative: yes), the procedure continues with step S7. Otherwise (alternative: no), the process returns to step S2.
[0046] In step S2, the trained distribution model can now be used to determine parameter values (posterior) for the next iteration.
[0047] Subsequently, in step S6, the parameter values that are most likely to produce the reference behavior, based on the posterior probability distribution derived from the distribution model, can be used to parameterize the control function. This can serve as the starting point for further adjustment of the parameter set using expert knowledge. QUOTES INCLUDED IN THE DESCRIPTION
[0000] This list of documents cited by the applicant was automatically generated and is included solely for the reader's convenience. The list is not part of the German patent or utility model application. The DPMA accepts no liability for any errors or omissions. Cited non-patent literature
[0000] Antoine Grosnit et al., “High-Dimensional Bayesian Optimization with Variational Autoencoders and Deep Metric Learning,” https: / / arxiv.org / pdf / 2106.03609.pdf
[0004] George Zerveas et al., “A Transformer-based Framework for Multivariate Time Series Representation Learning,” https: / / arxiv.org / abs / 2010.02803
[0015] George Zerveas et al., “A Transformer-based Framework for Multivariate Time Series Representation Learning,” https: / / arxiv.org / abs / 2010.02803_
[0037]
Claims
[1] Computer-implemented method for determining parameter values of parameters of an optimized parameter set for operating a specific technical system (1), wherein the behavior of the parameterized technical system (1) can be simulated by operating variable profiles that specify time profiles of at least one input variable, at least one output variable and at least one state variable, comprising the following steps: - Providing (S1) a data-based representation model that is trained to map temporal operational size profiles of one or more technical systems to a latent representation vector that characterizes the behavior of the technical system (1); - Providing a data-based distribution model trained to map latent representation vectors resulting from simulated operational characteristic curves of the specified technical system to a probability distribution of parameter values of the parameter sets, iteratively performing the following steps: - Providing (S2) parameter values of an initial parameter set or selecting parameter values of a parameter set from a probability distribution of parameters by random selection; - Simulating (S3) or measuring the technical system parameterized with the parameter values of the parameter set in order to obtain temporal operational variable profiles; - Evaluate (S4) the obtained operational size profile with the data-based representation model to obtain a latent representation vector; - Further training or retraining (S5) of the data-based distribution model with training data set from the obtained latent representation vector and the parameter set, so that an updated probability distribution of parameters results from the data-based distribution model. [2] Method according to claim 1, with the further step of: - Determining (S7) the optimized parameter set based on the posterior probability distribution resulting from the further developed or retrained distribution model, wherein the parameter values of the optimized parameter set are determined as those that most likely produce a given reference behavior. [3] Method according to claim 1 or 2, wherein the representation model is provided in the form of an encoder part of a variational autoencoder configured as a recurrent neural network or as a convolutional neural network or as a transformer network, and which is / is trained unsupervised for a variety of operating parameter profiles of different technical systems. [4] Method according to claim 1 or 2, wherein the representation model is / is formed in the form of an MVTS transformer model with an imputation task. [5] Method according to any one of claims 1 to 4, wherein the distribution model, in particular in the form of a flow matching posterior model or a posterior approximation model, is trained with training data sets of latent representation vectors of the operational variable profiles and associated parameter values. [6] Method according to any one of claims 1 to 5, wherein the distribution model is designed as a Neural Posterior Estimation model, and specify probability distributions of parameter values for the parameters for latent representation vectors resulting from simulations of the specified technical system. [7] Device for carrying out one of the methods according to any one of claims 1 to 6. [8] Computer program product comprising instructions which, when the program is executed by at least one data processing device, cause it to perform the steps of the method according to any one of claims 1 to 6. [9] Machine-readable storage medium comprising instructions which, when executed by at least one data processing device, cause it to perform the steps of the method according to any one of claims 1 to 6.