A meta-learning-based online adaptive fine-tuning VAE for industrial process data streams

Data-driven deep learning soft sensor has been recognized as an efficient industrial modeling technique for addressing high-dimensional nonlinear problems. However, when confronted with industrial process data streams that simultaneously exhibit high nonlineality and strong time-varying characteristics, offline deep learning models with poor adaptability struggle to achieve accurate quality parameter predictions. To address these challenges, this paper proposes a meta-learning-based online adaptive VAE method for data streams. Specifically, this method introduces a variational autoencoder regression (VAER) model, which incorporates Gaussian distribution constraints into latent variable learning, thereby alleviating the adverse effects of complex variable distributions on predictions. Additionally, a meta-learning algorithm is incorporated into model pre-training to obtain the meta-parameters, thereby enhancing generalization capabilities. Finally, a just-in-time learning fine-tuning update process based on meta-parameters of VAER is completed in online test phase. The effectiveness and practicality of the proposed method are validated through a case study on substrate concentration prediction in the penicillin fermentation process.