Real-time Prediction Of Final Carbon Temperature Of Converter Steelmaking Based On Flame Image Recognition Of Convolutional Neural Network

The determination of end point of converter steelmaking is an important link in converter production,the accuracy of determination directly affects the quality of the molten steel produced,the main decision is based on whether the carbon content and temperature of the molten steel in the furnace meet the requirements of steel production,therefore,the detection of carbon content and temperature in molten steel is the main factor in the determination process.The difficulty in measuring carbon content and temperature is continuous real-time accurate prediction,existing methods for achieving carbon temperature detection,such as the sub-gun detection method,are accurate but costly and difficult to popularize and cannot be continuously detected.The traditional converter end point determination method can realize the determination of the end point of the converter by using the extracted artificial flame characteristics,however,it is not possible to continuously predict the carbon content and temperature in the molten pool at the end point in real time,and there are problems such as being susceptible to environmental interference and losing the characteristic information of the flame image,which affects the accuracy.In response to this problem,the main research contents of this paper are described as follows:(1)Collecting the flame change video of the furnace mouth of the converter steelmaking process through an industrial camera,and transferring the flame video data frame by frame to the HSI color space more in line with human observation,analysis of the change trend of the Saturation and hue value of the flame image in the HSI color space as the blowing process progresses,the characteristics associated with the change of carbon temperature in the converter blowing process were found out,the BP neural network was used to fit and analyze the correlation between the flame image characteristics and the carbon content and temperature at the end of the converter steelmaking.(2)Propose a carbon content prediction method for converter steelmaking based on improved Lenet-5 convolutional neural network.Analyze the insufficiency of the traditional Lenet-5 structure,add local response normalization to the traditional structure,and increase the generalization ability of the model,and change the original Sigmoid activation function to Swish activation function to avoid the problem of gradient explosion or disappearing due to parameter saturation.After many adjustments to internal parameters(such as convolution kernel size,number of fully connected hidden layer neurons,learning rate,etc.)and comparative experiments,the prediction model of carbon content in converter end point based on Lenet-5 convolutional neural network suitable for extracting flame image features is established,which can continuously predict the carbon content at the end point in real time with higher precision.(3)A multi-scale fusion deep convolutional neural network is proposed to predict the carbon content of the converter steelmaking end point.Analyze the insufficiency of traditional convolutional neural network structure,and use residual structure to improve the depth of traditional convolutional neural network prediction model structure,combining multi-scale fusion convolutional layers,fully utilizing features of 1×1 convolution kernel dimension reduction and multi-channel information correlation,extracting rich and abstract flame image features,The extracted flame features are mapped to the marking space of the flame sample through the neural network,and the flame color image data set of the converter end point is made,and the model capable of continuously predicting the carbon content at the end point in real time is trained.(4)Propose a method for predicting carbon content in converter steelmaking by convolutional neural network with quaternion representation.Analyze the insufficiency of traditional convolutional layer sub-channels to extract the characteristics of flame images,and establish a quaternion convolutional quaternion fully connected neural network structure model,the quaternion convolutional layer can preserve the correlation between the channels and reduce the loss of associated information caused by the flame image due to the sub-channel processing.The influence of the traditional fully connected layer on the quaternion flame characteristics is analyzed.In order to protect the quaternion convolutional layer,the multi-channel correlation in the quaternion flame feature is not destroyed by the common fully connected layer.On the basis of the quaternion convolutional layer,a fully connected layer represented by a quaternion is added,and the quaternion fully connected layer is a 1×1 quaternion convolution operation satisfying the full connection rule,enabling the neural network to retain multichannel correlation features.And its effectiveness is verified by experiments.