Ash determination of coal flotation concentrate by analyzing froth image using a novel hybrid model based on deep learning algorithms and attention mechanism

Flotation is an important separation method for coal preparation, where ash content is critical to coal product quality. However, the absence of fast and accurate ash determination of coal flotation concentrate restricts the automation of flotation. Therefore, this paper presents a novel hybrid model, named as convolution-attention parallel network (CAPNet), for rapid and accurate determination of the ash content of coal flotation concentrate by analyzing froth images. First, we construct the CAPNet model by combining the classic CNN model (ResNet) and attention mechanism. Two parts are run in parallel so that they can learn from each other without mutual interference. Second, the hyperparameters of CAPNet are optimized using the orthogonal experimental design (OED) method. Finally, the proposed CAPNet is extensively compared with baseline models. Results show that CAPNet outweighs other methods in terms of accuracy and stability. It can achieve a R2 of 0.926, which is about 5%–10% greater than those of baseline CNN models, and over 30% higher than those of machine learning (ML) methods. As for other metrics, such as MAE, MAPE, RMSE, TIC, MPD, MGD and Var, the proposed CAPNet achieves 10%–50% of improvement compared to CNN models, and 50%–80% of improvement compared to ML methods. Extensive cross-comparison of performance between models clearly indicates that the CAPNet is superior to its competitors for the ash determination of coal flotation concentrate using froth images. Furthermore, CAPNet can also reduce the ash determination time from hours needed by existing standard method to 6 ms, which is ideal for engineering applications. We believe that the application of CAPNet in real production will significantly improve the automation and intelligence level of coal flotation, which can also increase economic benefits.