Machine Learning Inspired Fault Detection of Dynamical Networks.

Dynamical networks are a framework commonly used to model large networks of interacting time-varying components such as power grids and epidemic disease networks. The connectivity structure of dynamical networks play a key role in enabling many interesting behaviours such as synchronisation and chimeras. However, dynamical networks can also be vulnerable to network attack, where the connectivity structure is externally altered. This can cause sudden failure and loss of stability in the network. The ability to detect these network attacks is useful in troubleshooting and preventing system failure. Recently, a back-propagation regression method inspired by RNN training algorithms was proposed to infer both local node dynamics and connectivity structure from measured node signals. This paper explores the application of backpropagation regression for fault detection in dynamical networks. We construct separate models for local dynamics and coupling structure to perform short-term freerun predictions. Due to the separation of models, abnormal increases in prediction error can be attributed to changes in the network structure. Automatic detection is achieved by comparing prediction error statistics across two windows that span a period before and after a network attack. This method is tested on a simulated dynamical network of chaotic Lorenz oscillators undergoing gradual edge corruption via three different processes: edge swapping, moving and deletion. We demonstrate that the correlation between increased prediction error and the occurrence of edge corruption can be used to reliably detect both the onset and approximate location of the attack within the network.