Model-Driven Software Engineering of Machine Learning-Enabled Cyber-Physical Systems & IoT Services

The Model-Driven Software Engineering (MDSE) paradigm has proven beneficial in highly complex domains, such as those related to Cyber-Physical Systems (CPS) and the Internet of Things (IoT). However, CPS applications and IoT services are becoming increasingly data-driven and smart, using Artificial Intelligence (AI), in particular Machine Learning (ML). In this paper, we propose a novel approach to enable domain-specific MDSE for smart CPS applications and smart IoT services. We enable the infrastructure for edge analytics and federated ML. Federated ML is vital for the IoT, especially due to its privacy-preserving nature. Our focus is primarily on unsupervised and semi-supervised ML. This is required when the data instances that are used for training the ML models may be unlabeled or only partially labeled. We argue that this is a pragmatic approach, as the majority of data in many use case scenarios are typically unlabeled. We implement the proposed approach, and validate it using a case study, as well as through empirical evaluation with a number of experts. The provided model-to-code transformations (i.e., code generators) produce the full source code of the target IoT services out of the software model instances in an automated manner