Periodic Reporting for period 1 - COMPLETE (Model Completion through Nonlinear System Identification)
Berichtszeitraum: 2023-05-01 bis 2025-10-31
Systems and control engineers aim to master increasingly complex dynamical systems while including stronger performance, operational and energy constraints. As model-based control design remains the dominant paradigm, this results in an increasing need for nonlinear modeling. However, model inter-pretability and generalization capabilities form important roadblocks for a wide adaptation and ap-plicability of nonlinear system identification methods.
Strong prior knowledge is given by existing models, provided by system designers and engineers, even though they do not capture all the nonlinear dynamics of the real-life system. These models are currently not accounted for during black-box system identification. COMPLETE aims to develop a comprehensive nonlinear system identification framework to obtain accurate and interpretable models of measured complex system dynamics by completing an approximate pre-existing model through black-box nonlinear system identification. New theory and algorithms are put in place to 1) provide model structures, algorithms and theory that flexibly interconnect the pre-existing model and the black-box completion 2) ensure that data-driven completion models are interpretable and preserve key system theoretic aspects 3) data-driven experiment design strategies to detect, quantify and localize model errors at low experimental cost.
These objectives are far beyond the actual abilities of system identification, lifting the model completion for dynamical systems from ad-hoc approaches to a systematic, flexible, theoretically supported framework. My leading expertise on structured nonlinear system identification, and recent proof-of-concept results ensure the feasibility of the project. The resulting system identification framework is applicable over a wide range of engineering disciplines (mechanical, electrical, biomedical) and pro-vides system engineers with the necessary insight to guide them towards better solutions for tomor-row's industry.
Strong prior knowledge is given by existing models, provided by system designers and engineers, even though they do not capture all the nonlinear dynamics of the real-life system. These models are currently not accounted for during black-box system identification. COMPLETE aims to develop a comprehensive nonlinear system identification framework to obtain accurate and interpretable models of measured complex system dynamics by completing an approximate pre-existing model through black-box nonlinear system identification. New theory and algorithms are put in place to 1) provide model structures, algorithms and theory that flexibly interconnect the pre-existing model and the black-box completion 2) ensure that data-driven completion models are interpretable and preserve key system theoretic aspects 3) data-driven experiment design strategies to detect, quantify and localize model errors at low experimental cost.
These objectives are far beyond the actual abilities of system identification, lifting the model completion for dynamical systems from ad-hoc approaches to a systematic, flexible, theoretically supported framework. My leading expertise on structured nonlinear system identification, and recent proof-of-concept results ensure the feasibility of the project. The resulting system identification framework is applicable over a wide range of engineering disciplines (mechanical, electrical, biomedical) and pro-vides system engineers with the necessary insight to guide them towards better solutions for tomor-row's industry.
Significant progress has been reported on 4 aspects
I. The introduction of the modular model structure, and the development of an associated identification algorithm for model completion.
This result introduces a model completion representation and identification algorithm that unifies previously available model structures, and allows for even more flexible representations. This model structure and the associated learning algorithm allows for the flexible model completion of given baseline models. The presence of the baseline model enable a faster and better model training, as well as an improved model interpretability.
II. The introduction of port-Hamiltonian models that embed the power flow of interconnected systems.
Complex dynamical systems can be viewed as an interconnection of subsystems. This result introduces a method for the identification of such interconnected nonlinear systems starting from measured input output data, while ensuring passivity of the identified models. This approach will open the door towards identifying the missing dynamics of one subsystem using data collected while it was operating as part of the complete system.
III. The development of a flexible space-filling experiment design algorithm for nonlinear dynamical systems.
The development of a space-filling experiment design algorithm for nonlinear dynamical systems that allows for a very flexible input signal parametrization, and the inclusions of constraints during the optimization. This result allows for generating high-quality data for model completion and model error detection.
IV. Performing nonlinearity detection and location in, and identification of interconnected nonlinear systems.
Locating dominant sources of nonlinearities allows for a targeted model completion and model structure selection strategy, resulting in more accurate, but simpler models. This result presents a frequency-domain approach for the quantification and location of nonlinearities for interconnected nonlinear systems.
I. The introduction of the modular model structure, and the development of an associated identification algorithm for model completion.
This result introduces a model completion representation and identification algorithm that unifies previously available model structures, and allows for even more flexible representations. This model structure and the associated learning algorithm allows for the flexible model completion of given baseline models. The presence of the baseline model enable a faster and better model training, as well as an improved model interpretability.
II. The introduction of port-Hamiltonian models that embed the power flow of interconnected systems.
Complex dynamical systems can be viewed as an interconnection of subsystems. This result introduces a method for the identification of such interconnected nonlinear systems starting from measured input output data, while ensuring passivity of the identified models. This approach will open the door towards identifying the missing dynamics of one subsystem using data collected while it was operating as part of the complete system.
III. The development of a flexible space-filling experiment design algorithm for nonlinear dynamical systems.
The development of a space-filling experiment design algorithm for nonlinear dynamical systems that allows for a very flexible input signal parametrization, and the inclusions of constraints during the optimization. This result allows for generating high-quality data for model completion and model error detection.
IV. Performing nonlinearity detection and location in, and identification of interconnected nonlinear systems.
Locating dominant sources of nonlinearities allows for a targeted model completion and model structure selection strategy, resulting in more accurate, but simpler models. This result presents a frequency-domain approach for the quantification and location of nonlinearities for interconnected nonlinear systems.
The currently obtained results have delivered the cornerstones for the development of a data-driven model completion setting. The key theory and algorithms are in place to realize a full model completion system identification pipleine from data generation, over model estimation, to model validation and interpretation. This allows the effective integration of prior knowledge, in the form of previously available approximate baseline models, in the data-driven modelling framework. Future research will consolidate these results, and further investigate the interpretability of the obtained models, as well as developping theory and methods to include other system knowledge - beyond the inclusion of a baseline model. Finally, the full toolchain will also be integrated in one toolbox (DeepSI). An initial version of this toolbox is already available online. With the help of this toolbox, we will validate the developped theory and algorithms on industry relevant problems, for example mechatronics and thermal deformation problems. The research has the potential, facilitated by the toolbox, to lead to off-shelf solutions for present and future technological problems in the high-tech, automotive and process domains.