An innovative hybrid model for ash content prediction in froth flotation based on contrastive learning

Accurate prediction of concentrate ash content is crucial for ensuring the stability and efficiency of the coal flotation process. Although various methods have been developed to improve the prediction performance, most of them did not consider the impact of multi-frequency features and heavily relied on the performance of single neural network. In this paper, a Novel Hybrid Frequency-Receiving Independent Coal Ash Content Neural Network (AfricaNet) was proposed to achieve an accurate multi-step prediction of ash content from coal flotation images. First, we alter the lighting parameters of the raw images to create a dataset with identical labels for contrastive learning, aiming to reduce data non-smoothness and uncertainty. After that, the wavelet transforms was also combined with an attention mechanism, forming a feature extraction module for coal froth images. The feature selection strategy was also optimized for ash content prediction by integrating the Coding RAte reduction TransformEr (CRATE) framework with contrastive learning loss function, achieving an R2 score of 0.9978, the lowest MAE of 0.0191. The experimental results demonstrate the proposed model’s superiority in terms of both accuracy and stability, highlighting its potential as a highly robust and interoperable solution for coal flotation.

Recommended citation: Zhou, M., Zhang, Z., Cheng, X., Yang, S., Wen, Z., & Zhou, C. (2025). An innovative hybrid model for ash content prediction in froth flotation based on contrastive learning. International Journal of Coal Preparation and Utilization, 1–25.
Download Paper | Download Slides | Download Bibtex