Soft Sensor Enhancement for Multimodal Industrial Process Data: Meta Regression Gaussian Mixture Variational Autoencoder
Lei Chen, Yuan Xu, Qunxiong Zhu, Yan‐Lin He
Abstract
Traditional industrial soft sensors often treat industrial process data as uniformly distributed or unimodal. However, in reality, due to variations in operating conditions, industrial process data frequently exhibit multimodal characteristics. Overlooking the multimodal nature of data can lead to soft sensor models failing to accurately capture changing patterns and relationships under different operating conditions, thereby compromising the accuracy of prediction results. To address this issue, a method called meta regression Gaussian mixture variational autoencoder (MR-GMVAE) is proposed. First, a regression Gaussian mixture variational autoencoder (R-GMVAE) is introduced to extract multimodal distributions and dynamically model the process variables. An adaptive module is incorporated to allocate importance to different modes. Second, the concept of meta-learning is introduced for training on multimodal data, enabling the R-GMVAE to adaptively customize its training based on the data from different modes. Finally, the effectiveness and adaptability of the proposed MR-GMVAE are validated using two sets of industrial data.