FrothNet-LightGBM: A two-stage multimodal model for concentrate ash prediction in coal flotation

Concentrate ash content is a key indicator of product quality and process adjustment in coal flotation. In industrial plants, ash content is usually measured by offline laboratory analysis, which limits continuous quality monitoring. This study proposes a FrothNet-LightGBM two-stage multimodal framework for industrial ash content prediction under delayed assay labels. In Stage 1, froth images and four process variables are used to train a lightweight FrothNet image encoder for ash-related visual representation learning. In Stage 2, the trained encoder is frozen to extract 256-dimensional deep froth features. These features are fused with process variables and modelled by LightGBM. Fold-inside Top-K compression is then applied to reduce redundancy in the visual feature space. On 948 synchronized industrial samples, FrothNet achieved MAE = 0.3102, RMSE = 0.4015, and R 2 = 0.8400 with only 0.35 M parameters. In Stage 2, a compact 9-dimensional input, consisting of Top-5 visual features and four process variables, achieved MAE = 0.2026 ± 0.0066, RMSE = 0.2867 ± 0.0075, and R 2 = 0.9216 ± 0.0041. With the expressive LightGBM configuration, the performance further improved to MAE = 0.1922 ± 0.0081, RMSE = 0.2843 ± 0.0215, and R 2 = 0.9229 ± 0.0132. These results indicate that froth images and process variables provide complementary information, and that moderate compression of deep visual features improves the stability of industrial ash-content prediction.