Abstract
Centralized monitoring methods experience reduced fault detection sensitivity in large-scale industrial systems due to the masking effect arising from the aggregation of many interconnected variables. Decentralized monitoring, where variables are grouped into subsystems, has been shown to effectively address these limitations. However, the performance of this class of methods critically depends on how the network is partitioned, and the role of its structural factors on fault detection remains poorly understood. This work studies how network topology and causal structure affect decentralized monitoring in high-dimensional systems. Using SimCaNet, a DAG-based data simulator, where large-scale systems with 100-1000 variables were generated, we rigorously compared the performance of centralized and decentralized causal log-likelihood monitoring methods under process perturbations and sensor bias faults. Network partitioning is performed using the Leiden community detection algorithm and characterized at the network, community, and node levels. Fault Detection Sensitivity (AUCNORM) is quantified per variable, using the normalized area under the curve, which is computed from the true positive rate profiles across fault magnitudes. Results show that AUCNORM is primarily driven by the number and the size of communities across fault types. Smaller communities consistently achieve higher AUCNORM than larger communities, while excessive fragmentation or high modularity degrade performance. Community density and inter-community coupling further enhance fault detection sensitivity. The strength of node-level causal relationships explains most of the variability observed at root nodes for sensor bias faults and their robustness to information loss. Furthermore, ablation studies confirm the robustness of distributed systems to major disruptions in the system: more than 32% of community information can be removed without significantly impacting the fault detection sensitivity. These topology-driven partitioning findings constitute novel and valuable contributions to the theoretical foundations for designing better distributed monitoring systems in complex industrial plants.