An Integrated Multitask Perceptual Framework for Classification and Grade Prediction of Flotation Froth Conditions through Self-supervised Pre-training

Froth flotation represents a critical technique in mineral sorting, with the characteristics of flotation froth being key to adjusting production process parameters. Accurately interpreting the essential information contained within the froth is vital for optimizing flotation efficiency. Traditional methods, reliant on either manual feature extraction or deep learning from limited labeled data, are hindered by insufficient labeled data and a failure to fully exploit the latent information in froth images, resulting in models with low accuracy and poor robustness. Additionally, these methods necessitate multiple distinct networks for different detection tasks, leading to cumbersome processes and reduced computational efficiency. In response, this research introduces a self-supervised, pre-trained multi-task perception framework to classify flotation froth and predict mineral grades simultaneously. Leveraging a large-scale unlabeled image repository for pre-training, this framework employs vector quantization and knowledge distillation techniques to extract the essence of froth features, establishing an efficient self-supervised learning loop. The multi-task learning architecture not only facilitates cross-task knowledge transfer but also minimizes redundant computations, enhancing overall computational efficiency within a unified network architecture that accomplishes both condition classification and grade prediction. Experimental results indicate that the proposed method achieves efficient and accurate judgments under limited annotated data, offering a novel approach for transparent detection in the flotation process, predicting key performance indicators, and improving product quality.