AUTHOR=Hu Feng , Zhou Mengran , Dai Rongying , Liu Yu TITLE=Recognition method of coal and gangue based on multispectral spectral characteristics combined with one-dimensional convolutional neural network JOURNAL=Frontiers in Earth Science VOLUME=10 YEAR=2022 URL=https://www.frontiersin.org/journals/earth-science/articles/10.3389/feart.2022.893485 DOI=10.3389/feart.2022.893485 ISSN=2296-6463 ABSTRACT=

Accurate identification of coal and gangue is very important for realizing efficient separation of coal and gangue and clean utilization of coal. Therefore, a method for identifying coal and gangue by using multispectral spectral information and a convolutional neural network (CNN) model is proposed. First, 200 pieces of coal and 200 pieces of gangue in the Huainan mining area were collected as the experimental materials. The multispectral information of coal and gangue was collected, and the average value of each wavelength position was calculated to obtain the spectral information of the whole band. Then, based on the one-dimensional CNN (1D-CNN), namely, 1D-CNN-A and 1D-CNN-B, and with the help of stochastic gradient descent (SGD), Adam, Adamax, and Nadam optimizers, the rectified linear unit (ReLU) function and its improved function were used as the activation function to compare the identification ability of the identification models with different network structures and parameters. The best 1D-CNN model for identification of coal and gangue based on multispectral spectral information is obtained as follows: a network model containing three one-dimensional convolution units B, PReLU is used as the activation function, and Nadam is used as an optimizer to achieve the best identification effect. At this time, 40 coal samples in the test set can be accurately identified, and only one gangue sample in 40 gangue samples is wrongly predicted as coal. Finally, compared with the traditional recognition strategy (different combinations of principal component analysis and support vector machine), the excellent performance of this method is further proven. The results show that the combination of multispectral imaging and 1D-CNN can achieve accurate identification of coal and gangue without considering how to select appropriate preprocessing and feature extraction methods, which is of great significance in promoting the development of separation technology for coal and gangue.