Performance of Convolutional Neural Networks for Feature Extraction in Froth Flotation Sensing

Image-based soft sensors are of interest in process industries due to their cost-effective and non-intrusive properties. Unlike most multivariate inputs, images are highly dimensional, requiring the use of feature extractors to produce lower dimension representations. These extractors have a large impact on final sensor performance. Traditional texture feature extraction methods consider limited feature types, requiring expert knowledge to select and may be sensitive to changing imaging conditions. Deep learning methods are an alternative which does not suffer these drawbacks. A specific deep learning method, Convolutional Neural Networks (CNNs), mitigates the curse of dimensionality inherent in fully connected networks but must be trained, unlike other feature extractors. This allows both textural and spectral features to be discovered and utilised. A case study consisting of platinum flotation froth images at four distinct platinum-grades was used. Extracted feature sets were used to train linear and nonlinear soft sensor models. The quality of CNN features was compared to those from traditional texture feature extraction methods. Performance of CNNs as feature extractors was found to be competitive, showing similar performance to the other texture feature extractors. However, the dataset also exhibits strong spectral features, complicating comparison between texture feature extractors. The results gathered do not provide sufficient information to distinguish between the types of features detected by the CNN and further investigation is required.