Identification of Superheat Degree for Aluminum Electrolysis Based on Feature-Guide Deep Variational Autoencoder

Identification of Superheat Degree (SD) plays a critical role in decision-making of aluminum electrolysis process. With the development of industrial information technology, multimodal data can be collected from aluminum electrolysis cell, such as process data, flame hole images. To fully make use of process data and flame hole images, this paper proposes a SD identification method based on feature-guide deep variational autoencoder. According to flame hole images, the static features and dynamic features of flame hole video related to superheat are extracted. Feature-guide deep variational auto encoder is proposed to extract the deep features from flame hole images by feature alignment between the static-dynamic features and the deep features. Then, Fuzzy inference network is constructed to obtain the SD status according to process data, static-dynamic features, and deep features. The proposed method is validated on industrial aluminum electrolysis data with an accuracy of 93.2°/0, which is better than the existing algorithms.