Superheat Degree Category Refinement of Aluminum Electrolysis Cell Using Flame Hole Video Apparent Spatio-Temporal Feature Clustering Network

Flame hole video (FHV) is an important data source that can effectively evaluate the energy balance (superheat degree category, SDC) of modern aluminum electrolysis cell. This data source has time-varying apparent spatio-temporal features, and exists the issues of high acquisition cost, small amount of data, manual labeling, and few label categories. To address the above issues, a flame hole video apparent spatio-temporal feature clustering network (ASTFCN) is proposed for superheat degree category refinement of aluminum electrolysis cell (SDCRAEC). Firstly, the method of dividing FHV into multiple spatial region nodes is proposed to better capture detailed apparent spatial features. Secondly, the Gaussian filter and the DoG filter are respectively used to obtain the apparent low-frequency features and the apparent high-frequency features of each frame in the spatial-regional node, and the feature dimension of the spatial-regional node is reduced to 1-dimension by 3D convolution and pooling operations. Then, the apparent features obtained by matrix splicing of low-frequency features and high-frequency features are used as the input of the structural deep clustering network (SDCN) to perform the FHV clustering task and achieve the superheat degree category refinement. Finally, experiments are conducted on an actual industrial aluminum electrolysis FHV dataset, and the clustering results verify the effectiveness of the proposed method.