Flotation froth point-cloud operating-condition recognition based on DGCNN with integrated optimization

To address pronounced variations in the scale of local froth structures across flotation operating conditions, difficulties in adaptively matching a fixed neighborhood scale, and reduced recognition accuracy caused by strong mechanical vibrations at industrial sites, this study proposed a method for recognizing operating conditions from three-dimensional flotation froth point clouds based on a Dynamic Graph Convolutional Neural Network with Integrated Optimization (DGCNN-IO). Three-dimensional froth point clouds from multiple flotation cells were collected using an automatic slide-rail RGB-D inspection and acquisition platform. Original coordinates, normal vectors, surface curvature, local roughness, relative height, and local density were combined to construct 10-dimensional enhanced geometric features. An online neighborhood-scale optimization mechanism based on a Multi-Armed Bandit (MAB) was introduced into a Dynamic Graph Convolutional Neural Network (DGCNN). Before each training epoch, the model selected the graph convolution neighborhood scale for that epoch from a set of candidate scales, and the corresponding value estimate was updated based on validation-set rewards to alleviate scale mismatch caused by a fixed neighborhood parameter. A joint optimization training strategy was further introduced to jointly optimize model parameters through knowledge distillation and consistency constraints, improving robustness to complex industrial disturbances. Experimental results showed that DGCNN-IO achieved an overall classification accuracy of 97.0% and an F1 score of 96.9% on a dataset collected under typical operating conditions, including excessive and insufficient frother dosages. Under Gaussian noise simulating mechanical vibrations, the noise-induced degradation rate was only 1.1%. These results outperform those of the comparison models, including PointNet, PointNet++, DGCNN, PointMLP, and PointNeXt, demonstrating that DGCNN-IO has good operating-condition discrimination capability and robustness to disturbances.