A novel semi-supervised soft sensor modeling method based on deep dynamic and semantic information extraction for concentrate grade prediction in froth flotation

Abstract

With the rapid development of machine vision technology, this technology is being widely used in grade prediction for the froth flotation process. Several previously proposed froth visual features based methods to predict grade have yielded promising results, but some problems still exist, i.e., limited labeled data and complex data description. This can make it difficult for models to learn valid data representation and thus reduce prediction accuracy. To address the above-mentioned problem, a deep learning based semi-supervised soft sensor model, integrating dynamic and semantic information, is proposed to predict the concentrate grade. More specifically, two autoencoders, Stacked autoencoder (SAE) and Long Short-term Memory autoencoder (LSTM-AE), are pre-trained without supervision to respectively extract the deep semantic and dynamic temporal behavior hidden in all available input data (labeled and unlabeled data). Then, the representation vectors obtained from the two autoencoders transferred are concatenated to train a specific-task model on top using the limited training data. In addition, a self-evaluation module is introduced for tracking the predictive performance of our proposed model. Finally, a real-world antimony concentrate flotation process is used to validate our proposed method. Computational results show that our model eliminates the need to get large amounts of labeled data, and leverages unlabeled data, which is superior to traditional statistical models.