부산대학교 도서관

Concentrate grade and recovery are key production indexes (KPIs) for industrial flotation process. To establish the soft sensor model of the concentrate grade and recovery, a lot of froth image features are extracted as the input variables. However, these image features contain some redundant and irrelevant features. To improve the efficiency without degrading the performance of the soft sensor model, a sensitive feature selection (FS) method is proposed in this article. Sensitivity coefficient is defined to weigh the attribute significance of features to label, which is calculated by gray correlation analysis. Then, the criterion of sensitive FS based on minimal-redundancy-maximal-relevance (mRMR) is proposed. To solve the FS problem, a multiswarm particle swarm optimization (PSO) with collaborative search PSO (CS-PSO) is developed. Information exchange mechanism among three particle swarms in CS is proposed to improve the search effect and search accuracy. Self-adjusting structure RBFNN (SA-RBFNN) is employed to establish the soft sensor model to predict the concentrate grade based on the selected froth image features. The effectiveness of the proposed method is validated by the industrial flotation process data by comparing with other methods.