A Tracking-Based Burst Bubble Recognition in Flotation Using Deep Graph Neural Network

Bubble burst and coalescence are key indicators of froth stability, and play a vital role in flotation process monitoring. This article presents a tracking-based framework to improve recall in burst bubble recognition. The framework conceptualizes bubble trajectories over the entire froth video as a directed acyclic graph, where the minimum-cost flow is optimized through a deep graph neural network. Initial node embeddings are extracted using a temporal-aware pyramid network, which leverages hierarchical spatial-attention features to preserve small bubble representations and mitigate information loss in downsampling layers. In addition, an edge-regularized jumping graph aggregates sparsity-constrained embeddings across multiple depths, effectively addressing the oversmoothing problem while capturing long-term trajectory information. Finally, burst and coalescence bubbles are identified by analyzing the relative motion of neighboring bubbles through nearest supportive neighbors. Experiments in industrial coal and galena flotation process demonstrate that our framework improves recall by 18.45% and F1-score by 11.80% compared to existing methods.