Unsupervised heat balance indicator construction based on variational autoencoder and its application to aluminum electrolysis process monitoring

Heat balance plays a significant role in reflecting the health state of aluminum electrolysis process (AEP). However, current methods hardly take into consideration the quantitative Heat Balance Indicator (HBI) construction by using the unlabeled data. In addition, it is limited to construct HBI by learning the complex relationship between degraded features and large-scale HBI labels in a supervised manner, because the labeled data are scarce and annotations are expensive in practice. To quantitatively construct HBI by using the unlabeled data, this paper proposes an unsupervised HBI construction method based on variational autoencoder (VAE). Firstly, we propose fuzzy evaluation strategy to estimate the tendency of cell temperature to highlight the trend of heat balance. Rather than simply using the latent features, we extract the feature representation of the heat balance state considering not only the latent features but also the reconstruction error. Finally, HBI is constructed by calculating the distance between the features representation of normal heat balance and degraded state. The applications of heat balance monitoring in a real-world aluminum electrolysis plant are performed to verify its effectiveness. The experimental results demonstrate that our proposed HBI construction method can better represent heat balance state of AEP, the average fault detection rate can achieve 80% for the monitoring electrolytic cells, increasing by more than 3% compared with these traditional monitoring statistics.