[en] Real-time systems are systems that have specific timing requirements. They are critical systems that play an important role in modern societies, be it for instance control systems in factories or automotives. In recent years, Ethernet has been increasingly adopted as layer 2 protocol in real-time systems. Indeed, the adoption of Ethernet provides many benefits, including COTS and cost-effective components, high data rates and flexible topology. The main drawback of Ethernet is that it does not offer "out-of-the-box" mechanisms to guarantee timing and reliability constraints. This is the reason why time-sensitive networking (TSN) mechanisms have been introduced to provide Quality-of-Service (QoS) on top of Ethernet and satisfy the requirements of real-time communication in critical systems. The promise of Ethernet TSN is the possibility to use a single network for different criticality levels, e.g, critical control traffic and infotainment traffic sharing the same network resources.
This thesis is about the design of Ethernet TSN networks, and specifically about techniques that help quantify the extent to which a network can support current and future communication needs. The context of this work is the increasing use of design-space exploration (DSE) in the industry to master the complexity of designing (e.g. in terms of architectural and technological choices) and configuring a TSN network. One of the main steps in DSE is performing schedulability analysis to conclude about the feasibility of a network configuration, i.e., whether all traffic streams satisfy their timing constraints. This step can take weeks of computations for a large set of candidate solutions with the simplest TSN mechanisms, while more complicated TSN mechanisms will require even longer time.
This thesis explores the use of Artificial Intelligence (AI) techniques to assist in the design of TSN networks by speeding up the DSE. Specifically, the thesis proposes the use of machine learning (ML) as an alternative approach to schedulability analysis. The application of ML involves two steps. In the first step, ML algorithms are trained with a large set of TSN configurations labeled as feasible or non-feasible. Due to its pattern recognition ability, ML algorithms can predict the feasibility of unseen configurations with a good accuracy. Importantly, the execution time of an ML model is only a fraction of conventional schedulability analysis and remains constant whatever the complexity of the network configurations.
Several contributions make up the body of the thesis. In the first contribution, we observe that the topology and the traffic of a TSN network can be used to derive simple features that are relevant to the network feasibility. Therefore, standard and simple machine learning (ML) algorithms such as k-Nearest Neighbors are used to take these features as inputs and predict the feasibility of TSN networks. This study suggests that ML algorithms can provide a viable alternative to conventional schedulability analysis due to fast execution time and high prediction accuracy. A hybrid approach combining ML and schedulability analyses is also introduced to control the prediction uncertainty.
In the next studies, we aim at further automating the feasibility prediction of TSN networks with the Graph Neural Network (GNN) model. GNN takes as inputs the raw data from the TSN configurations and encodes them as graphs. Synthetic features are generated by GNN, thus the manual feature selection step is eliminated. More importantly, the GNN model can generalize to a wide range of topologies and traffic patterns, in contrast to the standard ML algorithms tested before that can only work with a fixed topology. An ensemble of individual GNN models shows high prediction accuracies on many test cases containing realistic automotive topologies. We also explore possibilities to improve the performance of GNN with more advanced deep learning techniques. In particular, semi-supervised learning and self-supervised learning are experimented. Although these learning paradigms provide modest improvements, we consider them promising techniques due to the ability to leverage the massive amount of unlabeled training data.
While this thesis focuses on the feasibility prediction of TSN configurations, AI techniques have huge potentials to automate other tasks in real-time systems. A natural follow-up work of this thesis is to apply GNN to multiple TSN mechanisms and predict which mechanism can provide the best scheduling solution for a given configuration. Although we need distinct ML models for each TSN mechanism, this research direction is promising as TSN mechanisms may share similar feasibility features and thus transfer learning techniques can be applied to facilitate the training process. Furthermore, GNN can be used as a core block in deep reinforcement learning to find the feasible priority assignment of TSN configurations. This thesis aims to make a contribution towards DSE of TSN networks with AI.