Research On Image Recognition Algorithm Of Coal-fired Flame In Kiln Based On Deep Convolution Network

Rotary kiln is a large key thermal equipment applied in industrial fields such as building materials,metallurgy and chemical industry.Due to complex sintering process in kiln,the recognition and judgment of working conditions using coal-fired flame image in kiln is vital to follow-up stabilization control.In-depth learning is now common in speech recognition,image processing,Internet applications and other fields.This paper studies the recognition algorithm of coal-fired flame image in kiln through deep convolutional neural network,and uses immune genetic algorithm to perfect parameter design of convolutional network for more reliable image recognition performance,which is important to the application of industrial soft sensing in in-depth learning.The study is mainly divided into the following aspects:(1)Studied the architecture and basic theory of convolutional neural network(CNN).Analyzed the basic operation modules such as convolution computation,activation function,pooling,full connection,dropout layer and the core computations such as back propagation and gradient descent in CNN whose basic structure and calculation process were made based on this.(2)Based on the characteristics of coal-fired flame image in kiln,such as more dust and being sensitive to disturbance,the flame image was preprocessed by median filter after image graying.Based on the three working conditions of "over-burning","upward-burning" and " undersintering",the working condition recognition image data set of the sintering zone of rotary kiln was formed by manual marking.(3)Analyzed the four convolutional neural network models,namely,AlexNet,VGGNet,GoogleNet and ResNet.Based on Tensorflow simulation platform,the experiment of coal-fired flame image recognition in kiln was conducted by using the flame image set of the upper section of kiln.The experimental results showed that AlexNet failed to converge,VGGNet demonstrated over-fitting,and GoogleNet converged slowly,thereby selecting ResNet depth network as the image recognition model of coal-fired flame in kiln.(4)Finally,optimized the initial weights of convolutional layer and full connection layer of ResNet network through immune genetic algorithm and selected the immune genetic control parameters through the comparison of genetic optimization solutions and the comprehensive consideration of iteration time.The experimental comparison showed that the recognition rate of ResNet network optimized by immune genetic algorithm rose by 4%,and the convergence speed of loss function was well up.