Fault Detection and Isolation in Industrial Networks using Graph Convolutional Neural Networks

Industrial networks represent large-scale systems that consist of several interacting components. In this paper, we present a solution for Fault Detection and Isolation (FDI) in industrial networks with same type of components (electrical, mechanical, etc.) such as power grids, and water supply networks. Traditional FDI algorithms are trained to detect and isolate faults on the level of a single component by considering features from this component and sometimes from nearby components. These algorithms are sub-optimal as they are independently applied to individual components without explicitly taking into consideration the dependency between the several components that co-exist in industrial network. The interaction between the components makes fault isolation challenging. An operation change or a fault in a component can affect the neighboring components. This negatively impact the diagnosis accuracy when we design an independent diagnoser for each component. On the other hand, designing a global diagnoser without considering the network structure can lead to overfitting which can degrade the diagnosis performance significantly specially when the training data is limited. Moreover, because of the large number of components in these systems, the single fault assumption may not stand. This increases the complexity of the problem. In order to solve this problem, we first model the industrial network as a weighted undirected graph structure. The graph structure represents the connected components. The weights quantify these connections. We then apply Graph Convolutional Neural Networks (GCNN) to detect and isolate faulty components in these systems. In the case study, we apply our solution to a simulated industrial network with 100 components. The case study shows that GCNN outperforms several baseline algorithms.