GRU-based stacked sparse autoencoder with attention mechanism for process monitoring

To ensure the operation safety of industrial processes, process monitoring techniques have been receiving considerably increasing attention from both academia and industry. The advent of deep learning has revolutionized data-driven process monitoring by effectively dealing with the inherent nonlinearity of industrial processes. In this work, a gated recurrent unit-based stacked sparse autoencoder with attention mechanism (GRU-SSAE-AM) model is developed for process monitoring. This model leverages the power of recurrent neural networks in conjunction with the sparse stacked autoencoder to tackle the temporal and nonlinear features present in process data. To enhance the model's ability to distill critical information indicative of process status, an attention mechanism is naturally integrated within the encoder–decoder framework. Experiments on an industrial benchmark of Tennessee Eastman process (TEP) and a real blast furnace ironmaking process (BFIP) are carried out to verify the capability and effectiveness of the proposed GRU-SSAE-AM based process monitoring method by comparison with other related methods.