A lightweight convolutional neural network model for quantitative analysis of phosphate ore slurry based on laser-induced breakdown spectroscopy

Abstract

The phosphorus content is an important control parameter in the flotation process of phosphate ore slurry. The real-time and on-stream monitoring of the phosphorus content can improve the control stability and flotation performance. Laser-induced breakdown spectroscopy (LIBS) is very suitable for online monitoring of the phosphorus content in the flotation process due to the advantages of no sample preparation and online detection. However, on account of the matrix effect, self-absorption effect and limited sample size with more dimensions, the accuracy of quantitative analysis is not satisfactory. To solve the above problems, we proposed a lightweight convolutional neural network model, referred to as the L-CNN spectra model. The model extracts spectral low-level features by the first three convolution layers. Unlike the traditional CNN model, we thinned the CNN by removing the activation function and pooling layers. Subsequently, the cross-channel high-level features on various scales are integrated by the inception module. The predicted concentration of phosphorus pentoxide is the output of the fully connected layer. The experimental results demonstrated that the L-CNN spectra model can improve the quantitative analysis accuracy of the flow slurry. We also discussed the impact of activation function and pooling operation after the convolutional layers in the feature extraction process. It is proved that the proposed L-CNN spectra model outperforms three competing CNN models as well as the partial least squares regression (PLSR) model and the support vector machine regression (SVR) model.