Towards Building a Digital Twin of Complex System Using Causal Modelling

Armen Aghasaryan

Abstract: Complex systems, such as communication networks, generate thousands of new data points about the system state every minute. Even if faults are rare events, they can easily propagate, which makes it challenging to distinguish root causes of errors from effects among the thousands of highly correlated alerts appearing simultaneously in high volumes of data. In this context, the need for automated Root Cause Analysis (RCA) tools emerges, along with the creation of a causal model of the real system, which can be regarded as a digital twin. The advantage of such model is twofold: (i) it assists in reasoning on the system state, given partial system observations; and (ii) it allows generating labelled synthetic data, in order to benchmark causal discovery techniques or create previously unseen faulty scenarios (counterfactual reasoning). The problem addressed in this paper is the creation of a causal model which can mimic the behavior of the real system by encoding the appearance, propagation and persistence of faults through time. The model extends Structural Causal Models (SCMs) with the use of logical noisy-OR gates to incorporate the time dimension and represent propagation behaviors. Finally, the soundness of the approach is experimentally verified by generating synthetic alert logs and discovering both the structure and parameters of the underlying causal model.