(362aa) Process Monitoring and Online Fault Detection and Diagnosis Using Deep Recurrent Neural Networks on Plant Data

We proposed a deep recurrent neural network for online data analytics and fault detection for smart manufacturing (SM) applications. The model consists of a novel setup of two recurrent neural networks (RNN) with gated recurrent units. The novel setup of the RNNs models the uncertainty in the data and allows for process monitoring in the deterministic and probabilistic spaces. The RNN layers also allows the model to learn dynamic relationships and retain long-term dependencies in the plant data.

Our experiments are designed to evaluate the false positives and missed detection rates of our proposed model on two case studies: (i) the simulated Tennessee Eastman Process (TEP), (ii) real plant data from an industrial Air Separation Unit (ASU). We show results from adaptively updating model parameters online when there is a change in the normal operating condition of the process.

We also compare the performances of our proposed model with state-of-the-art methods in statistical and other deep learning-based methods of fault detection and diagnosis. The false positives and missed detection comparison show that our model outperforms current state-of-the-art.