A Unified Framework for Structural - Temporal Coherence in Graph Neural Networks for Data-Driven Fault Identification

In modern industrial systems, accurate fault identification is crucial for the early isolation of broken parts and for further system restoration. Furthermore, data-driven machine learning applications gain popularity because of the increased availability of sensor data and their effectiveness. Graph neural networks, which are neural processes that operate on graph-structured data, are ideal for representing non-Euclidean data captured from multiple sensors. In the current paper, we benefit from the representation ability of graph structures through the presentation of a novel neural graph model that utilizes residual connections and leverages the spatial structure of the graph and the local information of graph nodes. Adaptive structural information and temporal knowledge insight can be integrated by the suggested graph neural framework. The latter is accomplished by feeding the temporal encoding of the graph nodes into a spectral graph neural network for training. Simulation results on the widely used fault identification benchmark of the Tennessee Eastman industrial chemical process verify that the proposed method outperforms competitive machine learning methods and state-of-the-art graph neural models, strengthens graph neural network training, and can be used to accurately identify faults in real industrial scenarios.