Self-Supervised Graph Structure Learning for Cyber-Physical Systems

Cyber-physical systems (CPS) often exhibit complex and dynamic behaviors that depend on their inherent structure. Knowing these connections is crucial for understanding a system's normal behavior or detecting changes in its state. We introduce Self-Supervised Graph Structure Learning for Cyber-Physical Systems using denoising autoencoders, or CP4SL, a method consisting of a graph generator module and a denoising autoencoder module in the form of a Graph Neural Network (GNN). Our approach addresses the challenge of inferring the structure of a CPS in terms of weighted connections between signals of multivariate time series. To account for temporal dependencies characteristic of CPS, we extend existing GSL methods by providing custom graph convolutional layers that use dilated causal convolutions. In contrast to prior works that have explored GNNs for CPS focusing on task-specific metrics, we measure the quality of learned connections explicitly. We evaluate our model on a system of synchronizing coupled oscillators and highlight remaining real-world challenges. Our code is available at https://github.com/alemamm/cp4s1-lightning Copyright (c) 2024 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)