Development of data-based modeling procedures that are able to combine the statistical reliability of current identification procedures with the efficiency and flexibility of machine learning-based methods. A key goal is to automate the entire modeling process (toolchain) from data collection to verification and to preserve both physical interpretability in modeling and to map unknown, physically difficult-to-describe, correlations and dynamic features.
In automotive, mechatronic and chemical applications, increasing performance requirements and explosive hardware complexity require that the current, mostly ideal, generic physical modeling be replaced by the development of procedures that target the unique specific dynamics of a given device and adapt to its changes. The resulting methods are capable of providing a digital copy (digital twin) for a given device, which goes far beyond the technologies currently used under this name for data aggregation and rudimentary estimation and simulation procedures. They are able to modify the physical base model (mode augmentation). Mapping unknown, physically difficult-to-describe relationships and dynamic features with the detail required by the modeling goal (e.g., prediction and prediction, predictive maintenance, modeling for control).
The research aims at automating the entire modeling chain:
data collection and experiment design,
model structure selection,
selection of modeling criteria and objective functions,
model estimation,
model verification based on modeling objectives.
In order to achieve this degree of automation and to effectively map the dynamic description of the system that is unknown but required for modeling purposes, it is necessary to combine the statistical reliability of current identification procedures with the efficiency and flexibility of machine learning-based methods.
In the first phase of the research, the efficient application of structures used in machine learning, such as Gaussian processes, neural networks in dynamic estimation and their combination with physical models needs to be realized by solving the necessary theoretical representation and statistical efficiency problems.
It is also important to develop learning-based modeling in batch-wise (offline) and continuous-learning recursive form (online), where in the latter the model learns continuously during operation, adapts to changing operating conditions through the learning component. An important issue is the analysis of the model from the point of view of identification, the examination of identifiable issues, the development of automatic experimental design methods that are able to cooperate with the system independently. It is also essential to estimate the reliability of the models and to recursively reduce the uncertainty ranges with further experiments with automatic design to automate the entire modeling chain, thus supporting the industrial application of autonomous systems.
Expected results:
Development of self-learning, physically interpretable and adaptive modeling algorithms and related experiment design and verification procedures, their implementation on real systems, demonstration of industrial applicability