Fault Prognosis and Detection for Carbonate Chemical Process under Multimodal Conditions Based on Transformer and Self-Adaptive Deep Learning

Fault prognosis and detection are of vital importance for safe and efficient operation of complex chemical processes, which have become a research hotspot by employing mainstream deep learning methods. However, nonsignificant variability of characteristic variables under some fault conditions and multioperation mode makes the corresponding task difficult to achieve. In the present work, transformer-based multivariate multistep prediction is utilized to establish the fault prognosis model, which intends to realize early fault prediction. In addition, by using samples under multiple normal states, local adaptive standardization is employed to preprocess the local moving window data. Then, variational autoencoder combined with a series of bidirectional long short-term memory layers is developed for fault detection, which intends to make fault detection model adaptive for multiple operating modes. By using a practical chemical process for high-purity carbonate ester production as a benchmark, the efficiency of fault prognosis and detection model is validated, and the corresponding process monitoring performance is clearly clarified.