Deep learning-based ash content prediction of coal flotation concentrate using convolutional neural network

•A new prediction method of ash content of coal flotation concentrate is proposed.•The proposed CNN combines deep features and local sub-images.•Matches state-of-the-art performance on the industrial flotation dataset.•The maximum frequency of the sub-image is used to as a predictor.•The accuracy of 97.1 % obtained for an industrial prediction time of 0.019 s. Convolutional neural networks, as the current state-of-the-art in image classification, are regarded as a promising way for flotation soft sensors based on froth images. This paper proposes a flotation soft sensor solution that predicts the concentrate ash content of coal flotation by using froth images and convolutional neural network architecture. A froth image dataset from the coal flotation site was divided into seven interval classes based on concentrate ash content, and data augmentation was used for this dataset. Then, several state-of-the-art convolutional neural networks (AlexNet, VGG_16, VGG_19, ResNet_18, ResNet_34, ResNet_50, ResNet_101) are trained to classifying the froth images with different concentrate ash content interval (integer ± 0.5). The classification performance of each model, the relationship between model performance and hyperparameters, and the abstract pixel feature of the best model are examined and visualized. The highest classification accuracy (97.1%) is obtained by the ResNet_101 network after fine-tuning. CNN extracts the features into a multi-channel set of abstract pixel features (7 × 7) before entered into the fully-connected layer. The results show that the abstract pixel feature significantly outperformed the manually engineered features. Finally, it takes only 0.31 s for the model to classify a froth image, which indicates that the convolutional neural network has a good industrial performance.