Graph Neural Networks With Trainable Adjacency Matrices for Fault Diagnosis on Multivariate Sensor Data

Timely detection and accurate diagnosis of faults in technological processes can significantly reduce production costs in manufacturing. Modern industrial equipment, equipped with numerous sensors, generates vast amounts of data, providing opportunities for advanced fault detection and diagnosis. While convolutional and recurrent neural networks have achieved state-of-the-art performance, they often overlook the correlations and hidden relationships among sensor signals. To address this, we propose a graph neural network (GNN) architecture that constructs graphs of sensor relationships from data. We evaluated five methods for training different types of adjacency matrices allowing to set certain restrictions on the structure of the graph. The resulting graph structures were analyzed and potential for their use in transfer learning was evaluated. Additionally, we developed an architecture that uses multiple adjacency matrices, which reduces the number of trainable parameters while maintaining high prediction quality. Our models demonstrated state-of-the-art performance on the Tennessee Eastman Process dataset, showcasing their potential for fault diagnosis on multivariate sensor data.