Unsupervised Fault Detection of Pharmaceutical Processes Using Long Short-Term Memory Autoencoders

Abstract

Unsupervised multilayer long short-term memory autoencoder (LSTM-AE) models are proposed for monitoring nonlinear batch processes. The methodology is demonstrated for a simulation-based study of an industrial-scale penicillin process and for an industrial vaccine manufacturing process, using production data. The LSTM-AE model was trained with two different loss functions: minimizing mean square error (MSE) between the input and reconstructed data and maximizing the average fault detection rate (FDR¯) in the training data set. Two algorithms are also proposed for obtaining contribution plots for the diagnosis of faults. For the industrial case study, where the faults are not known a priori, the contribution plots are found to be a valuable tool for identifying possible sources of faults. Furthermore, a semisupervised procedure has been proposed to select the normal process region for training the model. Two metrics are also presented to evaluate the performance of the proposed methodology: one for the simulator case study in which fault knowledge is available and one for the industrial case study in which fault knowledge is not available a priori. The proposed unsupervised algorithms exhibit a clear improvement in accuracy over linear methods or nonlinear techniques that do not explicitly account for dynamic behavior.