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

Unsupervised multilayer long short-term memory autoencoder(LSTM-AE)models are proposed for monitoring nonlinear batch processes. Themethodology is demonstrated for a simulation-based study of an industrial-scalepenicillin process and for an industrial vaccine manufacturing process,using production data. The LSTM-AE model was trained with two differentloss functions: minimizing mean square error (MSE) between the inputand reconstructed data and maximizing the average fault detectionrate ( FDR ) in the training data set. Two algorithmsare also proposed for obtaining contribution plots for the diagnosisof faults. For the industrial case study, where the faults are notknown a priori, the contribution plots are found to be a valuabletool for identifying possible sources of faults. Furthermore, a semisupervisedprocedure has been proposed to select the normal process region fortraining the model. Two metrics are also presented to evaluate theperformance of the proposed methodology: one for the simulator casestudy in which fault knowledge is available and one for the industrialcase study in which fault knowledge is not available a priori. Theproposed unsupervised algorithms exhibit a clear improvement in accuracyover linear methods or nonlinear techniques that do not explicitlyaccount for dynamic behavior.