Answers are grounded in this paper and cite their sources.
Industrial processes are frequently characterized by dynamic temporal correlations and multimode operating conditions, which considerably increase the difficulty of developing accurate soft sensor models. Although variational autoencoder (VAE)-based approaches have demonstrated promising representation capability, their unimodal latent prior assumption limits the characterization of complex multimode distributions and consequently degrades prediction performance under varying operating regimes. To address this issue, a mode-aware dynamic variational autoencoder regression (MADVAR) model is proposed for online soft sensing of dynamic multimode industrial processes. The proposed framework employs a temporal encoder-decoder architecture to extract latent representations from sequential process data, while a mode-aware classifier is incorporated into the latent space to explicitly capture multimode characteristics. Meanwhile, a regression network is established in the latent space to directly infer quality-related variables from learned latent features. Moreover, a Wasserstein-distance-based regularization strategy is introduced to reduce the geometric discrepancy between conditional posterior components and their corresponding Gaussian prior components. The proposed method is validated on a numerical example and a chemical process, where superior prediction performance is achieved.
No takes yet. Share an insight, caveat, or question.
Dai et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: