Study On End Point Prediction And Batching Optimization Of Converter

Converter steelmaking is the most common and most widely used in industry of steelmaking technology,furnace steelmaking process,complicated physical and chemical reaction is characterized by hot metal temperature itself,plus the exothermic reaction results in a high temperature in the furnace,quickly lead to short of molten steel smelting period,these features make it is very difficult to judge and respond to furnace accurately,in turn,affects the final quality of molten steel.Therefore,in order to improve the quality of molten steel,it is necessary to accurately control the reaction process of converter steelmaking and accurately predict the smelting end point.At present,the control methods of converter steelmaking mainly include the adjustment of initial batching and the determination of steel time by the operator through manual experience,the static adjustment of each batch according to historical data,and the dynamic adjustment according to real-time detection of molten steel quality.At present,the most widely used and relatively mature control methods in steel mills are the static adjustment based on material balance and heat balance and the dynamic adjustment based on the real time detection of sub-gun.Static adjustment is based on the target molten steel carbon content,temperature and historical data,using the material balance and heat balance,combined with expert experience to build a prediction or evaluation model,so as to compensate and optimize the initial ingredients such as scrap steel,molten iron,coolant,slag-making agent,so as to achieve the purpose of saving resources and improving the quality of molten steel.Because the static adjustment belongs to off-line control and lacks the real-time feedback of the intermediate process,the static adjustment has a certain influence on the prediction of the end point of the molten steel quality,and then affects the quality of the molten steel in the final tundish.At present,the initial batching of many steel converters is given through experience and simple batching formula,which leads to the waste of raw materials and the reduction of the first hit rate.At the same time,different steel plant operators have different experience,converter steelmaking process is complex,resulting in the application of static regulation is not very ideal.However,due to technical and financial reasons,many converters in China still adopt the simple static regulation combined with manual experience regulation.In view of the large number of converters in China and the fact that many of them adopt static regulation,it is of great research value and practical significance to develop a static control model of converters with high accuracy which is suitable for the field environment if the shortcomings of static regulation can be improved by combining with image processing and other technologies.The main research contents of this paper are as follows:(1)Through the acquisition of converter steelmaking ladle slag detection under the video,the video data of liquid steel frame by frame into images,because the movement of the liquid steel can influence due to environmental factors such as dust,temperature,noise plus frame by frame editing part will cause image distortion,and clip two images by frame interval too short,also need to remove some pictures,so you need to image preprocessing,eliminate distortion,closer to the adjacent images,after steel color image also need to make the best pictures covered with steel,avoid other factors cause the loss of recognition accuracy;According to the expert experience and the specific image data,the majority of normal molten steel is produced,but there are also various abnormal slaging.it is the existence of abnormal slaging.it is the existence of abnormal slaging.it directly reflects the decrease of the purity of molten steel and indirectly reflects the quality of the molten steel in this furnace.(2)ResNet101 convolution neural network is put forward based on the improved converter molten steel quality prediction evaluation model,analyzes the shortcomings of traditional ResNet structure,convolution was added between the network and the connection layer SPP-.net convolution group,you can enter any size of pictures,do not need to be tailored normalized scaling operations such as the size,increase the generalization ability of the model.After convolution kernels the ResNet101 convolution neural network model vector size,adjust and do comparison test for many times,established the steel is suitable for extracting image characteristics of molten steel quality prediction evaluation model of converter steel-making,continuous real-time detection of anomalies appear when under the molten steel slag and higher precision,and can be used to improve slag operation,do no slag tapping,can also be used to predict the evaluation of this furnace molten steel quality.(3)In this paper,a model of converter steelmaking initial ingredients compensation based on data fusion,analysis and simple formula ratio of converter steelmaking initial ingredients on the basis of experience,will be on a batch of target the quality of molten steel quality and the actual difference as an input,forecast a furnace on the basis of molten steel quality evaluation model of the output of the molten steel quality as another input,combined with expert experience to establish fuzzy rules,on the basis of two input to adjust the furnace time of the initial ingredients,and then improve the hit ratio,improve the quality of molten steel.In this paper,starting from the basic theory of bof steelmaking process,using the converter smelting process expert knowledge and data,image,first proposed the improved ResNet101 convolution converter molten steel quality evaluation model of neural network,then establish a model of converter steelmaking initial ingredients compensation based on data fusion,the finish quality of smelting process can be controlled by the computer automatically closed-loop optimization.