Integration of Modeling Methods for Cyber-Physical Systems

Abstract

Cyber-physical systems (CPS) incorporate digital (cyber) and mechanical (physical) elements that interact in complex ways. Many safety-critical CPS, such as autonomous vehicles and drones, are becoming increasingly widespread and hence demand rigorous quality assurance. To this end, CPS engineering relies on modeling methods, which use models to represent the system and design-time analyses to interpret/change the models. Coming from diverse scientific and engineering fields, these modeling methods are difficult to combine, or integrate, due to implicit relationsand dependencies between them. CPS failures can lead to substantial damage or loss of life, and are often due to two key integration challenges: (i) inconsistencies between models — contradictions in models that do not add up to a cohesive design, and (ii) incorrect interactions of analyses — analyses performed out-of-order and in mismatched contexts, leading to erroneous analysis outputs.This thesis presents a novel approach to detect and prevent integration issues between CPS modeling methods during the design phase. To detect inconsistencies between models, the approach allows engineers to specify integration properties — quantified logical statements that relate elements of multiple models — in the Integration Property Language (IPL). IPL statements describe verifiable conditions that are equivalent to an absence of inconsistencies. To interface with the models,IPL relies on integration abstractions — simplified representations of models for integration purposes. This thesis proposes two abstractions: views (annotatedcomponent-and-connector models, inspired by software architecture) and behavioral properties (expressions in model-specific property languages, such as the linear temporal logic). Combining these abstractions lets engineers relate model structure and behavior in IPL statements. To ensure correct interactions of analyses, I introduce analysis contracts — a lightweight specification that captures inputs, outputs, assumptions, and guarantees for each analysis, in terms of the integration abstractions.Given these contracts, an analysis execution platform performs analyses in the order of their dependencies, and only in the contexts that guarantee correct outputs.My approach to integration was validated on four case studies of CPS modeling methods in different systems: energy-aware planning in a mobile robot, collisionavoidance in a mobile robot, thread/battery scheduling in a quadrotor, and reliable/secure sensing in an autonomous vehicle. This validation has shown that the approach can find safety-critical errors by specifying expressive integration properties and soundly checking them within practical constraints — all while being customizable to heterogeneous models, analyses, and domains.