A hybrid feature selection method for production condition recognition in froth flotation with noisy labels

Production condition recognition plays a major role in the froth flotation process control, and its accuracy has great influence on the performance of froth flotation. In order to improve the performance, lots of features are extracted from froth images; however, not all of these features are useful. There are some redundant and irrelevant features among them, which may increase the computational cost and the difficulty of training classifier, or even reduce the recognition accuracy. In order to resolve this issue, a feature selection method named mRMR-BSTA is proposed in this paper to find the best feature subset with most useful and less redundant features. This proposed method is comprised of two phases: the filter phase based on minimal-redundancy-maximal-relevance (mRMR) criterion and the wrapper phase based on binary state transition algorithm (BSTA), and it is applied to the production condition recognition in gold-antimony froth flotation. Least squares support vector machine (LSSVM) is applied to recognition and used as a black box to evaluate the quality of selected features in wrapper phase. Especially, for removing the noisy data with wrong labels caused by the subjectivity of workers, a preprocessing approach based on fuzzy C-means (FCM) is proposed in this paper. Finally, the most efficient feature subset in froth flotation is selected, including hue, mean of blue, relative red component, coarseness, and high frequency energy. The production condition is classified into eight classes with high accuracy successfully, and the effectiveness of the proposed method for production condition recognition in froth flotation is validated.