Semi-supervised clustering for operating condition recognition in sulfide ore flotation froth

Flotation is a core technology for the upgrading and purification of sulfide ores, and it has been widely applied in industrial production. In industrial flotation processes, operators heavily rely on visual inspection of froth conditions to adjust operating parameters, which leads to strong subjectivity, poor stability, and delayed feedback. In recent years, machine vision technology has been widely adopted for flotation froth image classification. However, most existing methods rely on traditional clustering or other classification models, which struggle to handle practical challenges such as label scarcity and large variations in operating conditions. Moreover, the mapping between image features and flotation conditions remains unclear. To address these challenges, this study proposed a semi-supervised clustering-based method for fine classification of flotation froth images, aiming to establish a quantitative relationship between image features and flotation conditions. First, based on a self-constructed flotation experiment platform under multiple operating conditions, froth images were collected under varying impeller speeds and collector dosages. Subsequently, multidimensional features including color, texture, and size were extracted. A mutual information-based method was then employed for dimensionality reduction and feature selection, resulting in compact feature vectors. On this basis, an “Eff-Collector” kernel function was constructed by incorporating operating conditions. Unsupervised clustering pseudo-labels were used to guide Gaussian kernel-based clustering, thereby building a semi-supervised clustering model constrained by operating conditions. This method effectively improves froth image classification accuracy under limited labeling conditions and enhances the ability of clustering results to represent actual operating conditions. The average accuracy of operating condition identification from froth images exceeds 94.89 %, with the standard deviation of major clustering categories across multiple trials remaining below 3 %. These results provide a solid theoretical foundation and a practical technical pathway for reagent regulation and intelligent monitoring in flotation processes.