Computational Augmentation of Model Based System Engineering: Supporting Mechatronic System Model Development with AI Technologies

Efforts in applying computational support for automatic design synthesis and configuration generation as well as efforts to support descriptive and computational model development for system design and verification has been approached with semantic formalisation of modelling languages and of generic structural and functional concepts using meta-models. Modelling the system using descriptive models helps the designer to explicitly document dependencies between properties and parameters of system and external entities. The descriptive models thus produced often do not consider physics based justification for presence and/or absence of relations. It is often the case, the simulation results obtained at later stages requires changing requirements as well as modifying logical (modelling relations between high level functions parameters/properties and parameters/properties of high level entities) and physical architectures (modelling relations between component’s parameters and properties) to accommodate those requirements. The current MBSE (Model Based System Engineering) tools have capabilities to verify construction of models according to predefined model formats i.e. meta-models. However, these tools and current research in augmenting capabilities of these tools lacks the focus on evaluating content inside the models i.e. whether the system modelled by models represents a system that can be physically realized. This work has tried to avail the potential of available AI (Artificial Intelligence) technologies for assisting modelling activities performed for requirement definition and analysis, architecture design and verification phase of system development process by directing designer to tools that can formalise outputs of model development activities. The proposed problem formulation is based on the insight that a system modelled at both conceptual and detailed design level can be represented by logical and mathematical relations between the properties and parameters of internal and external components or functions of the system and domain. Therefore formulation defines concepts used in requirement, logical architecture and physical architecture models using relation between parameters and properties in those models. Concepts, such as operational requirements (or non-functional requirements for particular use case scenario), are defined through the usage of sets and linking value domains of those sets to particular system application domain for which system model is being developed. These relations enables systematic elaboration of requirements into logical and physical architecture models as well as storage and retrieval of existing model knowledge using existing AI tools. A novel framework has been developed to retrieve existing descriptive structure and function models using logical reasoning as well as to retrieve existing simulation models stored in embedding space of auto-encoder neural network. Beside adopting the concepts of semantic formalisation and meta-model based descriptive knowledge retrieval it utilises novel application of unsupervised representation learning capability of neural network auto-encoders to store known physically and technologically feasible designs in low dimensional representation that cluster similar designs therefore inducing similarity or distance metric that can be used to retrieve the known design with similar behaviour as new required behaviour. Framework also enable application of generic and domain specific logical constraints (as other works has done before) and introduces new concept of system application domain to ensures that at every stage of the model development leading to conceptual physical design architecture stays inside the physical constraints as per system usage domain. The instantiated meta-model elements which are classified to a system application domain (SAD) are implicitly constraint by system usage context constraints (e.g. parameter value restriction), similarly known simulation models can also be categorised to different SADs. The proposed framework extends the conventional approach of automated design synthesis which is only based only on decomposition of high level function (summarizing input to output mapping) into basic functions and selecting components to realize those basic functions. "A system is designed with the aim that it can execute its function(s) as per performance requirements of that function(s) in required operational conditions"- By concentrating on this statement it can be seen that conventional approach of functional decomposition and function allocation to known structural components cannot guarantee to yield a working system in required scenarios by ignoring the dependencies between environment or operating conditions and operating modes of prospective designs satisfying high level function. The results obtained from the implementation of domain specific knowledge representation and retrieval (involving mixture of numerical and logical constraints) as well as the results obtained from implementation of neural network auto-encoder for representation and retrieval of domain specific simulation model demonstrates the viability of these technologies to support the proposed framework.