Fault detection Neural Differential Auto-encoders

In this study, we propose a Graph neural Differential Auto-encoder (GNDAE) model for fault detection and process monitoring. The GNDAE framework is capable of dealing with graph data inputs which are inherently non-Euclidean objects, and is an extension of the Neural Ordinary Differential Equations (NODE), which provides continuous depth to the discrete set of hidden layers of neural network. After GNDAE model is trained on the normal operating data, the reconstructed output and the hidden features of the input graphs are extracted, and a normal operable range (NOR) with 99% confidence interval is set as the cut-off tolerance limit. T 2 and SPE statistics are then computed for the proposed model within this defined range. The performance of the proposed framework was superior to various baselines when validated on a couple of datasets, namely the benchmark Tennessee Eastman Process (TEP) dataset and the Aved & oslash;re wastewater treatment plant dataset.