Enhancing Monitoring Performance of Pharmaceutical Processes Using Dual-Attention Latent Dynamic Conditional State-Space Model

The batch fermentation of penicillin is an important process in the pharmaceutical industry. As the product quality is highly dependent on the operating condition during the production, online process monitoring is necessary to make sure that the product meets the market demand. However, the 3-dimensional structure and uneven-length batch data, nonlinearity, process dynamics are inherited in the data. In such a circumstance, it is hard to build the process monitoring model. This paper proposed a supervised learning scheme with the concept of sequence prediction to increase the robustness of the model in process monitoring. The proposed method is called the dual-attention latent dynamic conditional state-space model (DA-LDCSSM). DA-LDCSSM embeds the spatial and temporal attention to extract the important features among variables and among sequence time points. Also, the process dynamics are described in the noise-free latent space by LDCSSM to reduce the effect of noise disturbance on the model. The data simulated by the industrial-scale penicillin fermentation simulation are used to test DA-LDCSSM for monitoring performance.