Analysis Of Clink Particles Based On DCNN

Cement is an important resource and fundamental product that affects the develop-ment of the national economy.As a basic material,cement is widely used in all aspects of people's lives such as civil construction,bridge highway construction,water conser-vancy construction,and industrial architecture.The quality monitoring and evaluation of the clinker in the intermediate product is an important link to ensure the cement quality.However,the current domestic cement plant stays in the way of manual inter-vention,sampling and analysis,which is very time-consuming and labor-consuming.Therefore,this paper has designed a clinker particle size detection algorithm based on machine vision.In this paper,based on the combination of particle size detection algorithm in the traditional mineral processing field,the deep convolution network UNet algorithm is introduced for the first time to segment the clinker particle image,and on this basis,improvements are made to increase the accuracy of particle segmentation.Finally,the particle size distribution of the segmented image is estimated using some typical particle size characteristics.The main research contents and achievements of this article are as follows:1.A deep convolutional network UNet algorithm was introduced to segment the clinker particle image.The traditional particle image segmentation algorithm needs complicated and professional process design and fine parameter adjust-ment,and the robustness is not good enough.Different styles of particle image segmentation require different algorithm flow.For the deep convolutional net-work algorithm UNet,the algorithm flow for segmentation of different kinds of particle images is the same,and fewer parameters need to be manually adjusted.2.UNet algorithm framework building,training and testing.Using the Keras tool based on Tensorflowto build the UNet framework,the training set data is simply pre-processed and combined with the hand-painted label image to train the UNet network.After the training,the test set is used to verify the particle segmentation effect of UNet algorithm and quantified.Finally,using the two downloaded par-ticle images with different styles of clinker particles to test the knowledge transfer characteristics of UNet algorithm.3.UNet algorithm improvements.Based on deepth the network and the loss func-tion,the UNet algorithm is improved:the pre-trained VGG16 network is used as the encoder in the UNet framework,and the positive sample points,ie,the particle gap points,is weighted in the loss function.The improved VGG_UNet algorithm is compared with the previous UNet algorithm.4.Clinker particle size estimation.Using the improved segmentation algorithm,the clinker grain image is segmented,and the grain size of the segmented grain image is estimated using some typical grain size feature parameters,and compared with the actual grain size distribution.