Prediction of Froth Flotation Performance Using Convolutional Neural Networks

Abstract

Deep learning is a subset of machine learning that uses artificial neural networks for extracting high-level features from image data. In the present study, a soft sensor is proposed for the prediction of the flotation performance through froth features generated by the use of pre-trained convolutional neural networks. Several state-of-the-art convolutional neural networks (AlexNet, GoogLeNet, VGGNet, ResNet, and SqueezeNet) pre-trained on the ImageNet database are used to predict the metallurgical performance of two flotation systems. The first case study is a batch copper flotation system video-captured over a wide range of process conditions. The second case study is an industrial coal flotation column equipped with a continuous video recording system. The pre-trained networks are used to extract features from the froth images, and these features are subsequently used to predict the flotation conditions and performance. The prediction results by the pre-trained algorithms were compared with the traditional image processing algorithms. This demonstrates the ability of the pre-trained structures to generalize to images outside the ImageNet database. GoogLeNet outperforms other network architectures and provides more accurate predictions of the flotation process behavior and performance.