Enhanced variational autoencoder with continual learning capability for multimode process monitoring

Abstract

Existing mode segmentation methods in multimode process monitoring generally fail to simultaneously determine the number of modes and segmentation points. In addition, traditional monitoring methods typically suffer from “catastrophic forgetting”, resulting in poor monitoring performance. This paper proposes a novel mode segmentation method called latent variable mapping greedy Gaussian segmentation (LMGGS) and enhances the variational autoencoder (VAE) with continual learning (CL-VAE) capability to address the problem of catastrophic forgetting. First, the LMGGS is used for mode segmentation, which reformulates the mode segmentation problem as a covariance-regularized maximum likelihood estimation problem. Second, weights in VAE deemed unimportant were set to zero, and the remaining ones were updated by training the model with important samples in a direction orthogonal to the gradient space of the previous modes. Finally, statistics and thresholds based on the reconstruction error were established to determine the system states. The LMGGS can simultaneously determine the segmentation point and the number of modes, while the CL-VAE can effectively address catastrophic forgetting and reduce data storage requirements. The superiority of the proposed methods was validated through experiments on two simulated datasets and an actual penicillin fermentation dataset.