Physical-Knowledge Embedded Convolutional Neural Network for Aluminum Electrolysis Superheat Degree Identification

The superheat degree (SD) is a crucial production indicator in the aluminum electrolysis process. However, due to the harsh environment characterized by high temperature, strong magnetic fields, and strong corrosion in the aluminum electrolysis industry site, only a limited amount of local flame hole video data of aluminum electrolysis can be collected for SD identification. Furthermore, as the identification of SD becomes more refined, the number of categories to be identified increases. The increase in categories results in similarities between different categories, which presents challenges in the identification process. To address these problems, we investigate how to integrate extensive professional knowledge with deep neural network architectures and developed a physical-knowledge embedded convolutional neural network (PKE-CNN) method. First, the physical features of the flame hole video are extracted by incorporating the prior knowledge. Subsequently, the PKE-CNN architecture is proposed, which incorporates physical knowledge into CNN, and an enhanced cross-entropy loss function is proposed to distinguish samples with high similarity. Finally, the parallel particle swarm optimization (PPSO)-based flame hole knowledge selection and optimization method is proposed to deal with limited data samples by achieving efficient knowledge embedding. In order to verify the effectiveness of the proposed method, an industrial aluminum electrolysis flame hole video experiment is conducted, extensive results demonstrate that the proposed method can overcome the interference of similar samples and learn features that are effective for identifying tasks. Compared with some state-of-the-art methods, our method has achieved significant performance improvements.