摘要：AbstractHydrocyclones are widely used in mineral processing to separate coarse and fine particles. Despite the effect that their performance can have on downstream operations, they are not generally monitored or controlled, since online estimation of particle size is challenging. In this paper, it is shown that a pretrained convolutional neural network can be used to provide useful estimates of the particle sizes in the underflow of hydrocyclones. More specifically, GoogLeNet is used in a transfer learning mode to extract features from images of the underflow of a hydrocyclone, which can then be used as predictors of particle size in a regression model. The model was able to explain approximately 85% of the variance in the particle sizes associated with the underflow images. However, with additional training of the last feature layers of the model, it was able to explain approximately 91% of the variance of the particle sizes.
