Detecting coal content in gangue via machine vision and genetic algorithm-backpropagation neural network

Abstract

Coal content in gangue is one of the important technical and economic indicators of coal preparation plants and cannot be detected online at present. A novel approach based on binocular machine vision and genetic algorithm-backpropagation neural network (GA-BPNN) was proposed. First, the sample image was segmented, and each region was judged to be coal or gangue. Four size and eleven density features were extracted from each region. For the size features, the values of each feature were summed up by class, whereas the average operation was performed for the density features. The values of gangue features were divided by those of coal features to obtain an initial feature set based on a two-dimensional (2D) image. By contrast, binocular images were collected to estimate a three-dimensional (3D) feature, that is, the height features of coal and gangue particles. Subsequently, the projected area and the height feature were multiplied to obtain another 3D feature, that is, the volume feature. The Spearman correlation coefficient was adopted as the feature selection method, and eight optimal features were extracted from the final feature set, including both 2D and 3D features. Then, the coal content in gangue was modeled using a BPNN optimized by GA. Finally, three models were built based on 17 complete features, 6 optimal 2D features, and 8 optimal 3D features; another model without GA optimization was also built. The gangue sample experiment of Hongliu Coal Preparation Plant showed that the average relative error of our method was 7.98%, which is much better than that of the other three models.