Intelligent optimization of reagent dosage in coal slime flotation using multi-source feature fusion and a Transformer-LSTM model

Existing intelligent reagent dosage methods for coal slime flotation often rely on narrow feature sets. They also suffer from temporal misalignment among multi-source signals. These limitations constrain optimization accuracy. To overcome these issues, this paper proposes an intelligent dosing approach integrating multi-source feature fusion with a Transformer-LSTM model. First, we establish a plant-wide process characterization system. It includes tailings ash content, residual reagent concentration in recirculating water, and coarse particle loss. This system improves the characterization of complex working conditions and captures cross-process correlations. Next, we introduce a mutual information-based temporal alignment method. It quantifies dependencies between features and reagent dosage, thereby synchronizing multi-source signals. The aligned features are then fed into a hybrid model. It combines the Transformer’s global dependency capture with the LSTM’s temporal dynamics modeling, enabling high-accuracy dosage decisions. Finally, a B/S architecture-based flotation dosing system is developed, integrating data acquisition, temporal alignment, intelligent decision-making, and visualization to support real-time monitoring and online optimization. Industrial tests show that the proposed method raises the qualified rate of froth concentrate ash content from 75.7% under manual operation to 90.0%, significantly reducing labor intensity while ensuring product quality and downstream process stability. This study offers a new direction for multi-source feature fusion and intelligent control in complex industrial contexts.